From eb9c7d4ed8984bdff6585e38d04e7d17bf14155e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 3 Jan 2024 22:04:04 -0500 Subject: [PATCH 001/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 138 ++++++++++++++++++++++------------------- vendor/llama.cpp | 2 +- 2 files changed, 74 insertions(+), 66 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 75800c0d8..4aada532e 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -93,6 +93,9 @@ def _load_shared_library(lib_base_name: str): # llama.h bindings +_lib.llama_max_devices.argtypes = [] +_lib.llama_max_devices.restype = ctypes.c_int32 + LLAMA_MAX_DEVICES = _lib.llama_max_devices() # define LLAMA_DEFAULT_SEED 0xFFFFFFFF @@ -481,7 +484,7 @@ class llama_context_params(Structure): # // model quantization parameters # typedef struct llama_model_quantize_params { -# int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() +# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() # enum llama_ftype ftype; // quantize to this llama_ftype # bool allow_requantize; // allow quantizing non-f32/f16 tensors # bool quantize_output_tensor; // quantize output.weight @@ -499,7 +502,7 @@ class llama_model_quantize_params(Structure): only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored pure (bool): disable k-quant mixtures and quantize all tensors to the same type""" _fields_ = [ - ("nthread", c_int), + ("nthread", c_int32), ("ftype", c_int), ("allow_requantize", c_bool), ("quantize_output_tensor", c_bool), @@ -698,13 +701,13 @@ def llama_time_us() -> int: _lib.llama_time_us.restype = ctypes.c_int64 -# LLAMA_API int llama_max_devices (void); +# LLAMA_API int32_t llama_max_devices(void); def llama_max_devices() -> int: return _lib.llama_max_devices() _lib.llama_max_devices.argtypes = [] -_lib.llama_max_devices.restype = c_int +_lib.llama_max_devices.restype = ctypes.c_int32 # LLAMA_API bool llama_mmap_supported (void); @@ -734,7 +737,7 @@ def llama_get_model(ctx: llama_context_p) -> llama_model_p: _lib.llama_get_model.restype = llama_model_p -# LLAMA_API int llama_n_ctx (const struct llama_context * ctx); +# LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); def llama_n_ctx(ctx: llama_context_p) -> int: return _lib.llama_n_ctx(ctx) @@ -758,31 +761,31 @@ def llama_vocab_type(model: llama_model_p) -> int: _lib.llama_vocab_type.restype = c_int -# LLAMA_API int llama_n_vocab (const struct llama_model * model); +# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); def llama_n_vocab(model: llama_model_p) -> int: return _lib.llama_n_vocab(model) _lib.llama_n_vocab.argtypes = [llama_model_p] -_lib.llama_n_vocab.restype = c_int +_lib.llama_n_vocab.restype = c_int32 -# LLAMA_API int llama_n_ctx_train(const struct llama_model * model); +# LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); def llama_n_ctx_train(model: llama_model_p) -> int: return _lib.llama_n_ctx_train(model) _lib.llama_n_ctx_train.argtypes = [llama_model_p] -_lib.llama_n_ctx_train.restype = c_int +_lib.llama_n_ctx_train.restype = c_int32 -# LLAMA_API int llama_n_embd (const struct llama_model * model); +# LLAMA_API int32_t llama_n_embd (const struct llama_model * model); def llama_n_embd(model: llama_model_p) -> int: return _lib.llama_n_embd(model) _lib.llama_n_embd.argtypes = [llama_model_p] -_lib.llama_n_embd.restype = c_int +_lib.llama_n_embd.restype = c_int32 # // Get the model's RoPE frequency scaling factor @@ -802,7 +805,7 @@ def llama_rope_freq_scale_train(model: llama_model_p) -> float: # // Get metadata value as a string by key name -# LLAMA_API int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); def llama_model_meta_val_str( model: llama_model_p, key: Union[c_char_p, bytes], buf: bytes, buf_size: int ) -> int: @@ -811,22 +814,22 @@ def llama_model_meta_val_str( _lib.llama_model_meta_val_str.argtypes = [llama_model_p, c_char_p, c_char_p, c_size_t] -_lib.llama_model_meta_val_str.restype = c_int +_lib.llama_model_meta_val_str.restype = c_int32 # // Get the number of metadata key/value pairs -# LLAMA_API int llama_model_meta_count(const struct llama_model * model); +# LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model); def llama_model_meta_count(model: llama_model_p) -> int: """Get the number of metadata key/value pairs""" return _lib.llama_model_meta_count(model) _lib.llama_model_meta_count.argtypes = [llama_model_p] -_lib.llama_model_meta_count.restype = c_int +_lib.llama_model_meta_count.restype = c_int32 # // Get metadata key name by index -# LLAMA_API int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); def llama_model_meta_key_by_index( model: llama_model_p, i: Union[c_int, int], buf: bytes, buf_size: int ) -> int: @@ -834,12 +837,17 @@ def llama_model_meta_key_by_index( return _lib.llama_model_meta_key_by_index(model, i, buf, buf_size) -_lib.llama_model_meta_key_by_index.argtypes = [llama_model_p, c_int, c_char_p, c_size_t] -_lib.llama_model_meta_key_by_index.restype = c_int +_lib.llama_model_meta_key_by_index.argtypes = [ + llama_model_p, + c_int32, + c_char_p, + c_size_t, +] +_lib.llama_model_meta_key_by_index.restype = c_int32 # // Get metadata value as a string by index -# LLAMA_API int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); def llama_model_meta_val_str_by_index( model: llama_model_p, i: Union[c_int, int], buf: bytes, buf_size: int ) -> int: @@ -849,15 +857,15 @@ def llama_model_meta_val_str_by_index( _lib.llama_model_meta_val_str_by_index.argtypes = [ llama_model_p, - c_int, + c_int32, c_char_p, c_size_t, ] -_lib.llama_model_meta_val_str_by_index.restype = c_int +_lib.llama_model_meta_val_str_by_index.restype = c_int32 # // Get a string describing the model type -# LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); def llama_model_desc( model: llama_model_p, buf: bytes, buf_size: Union[c_size_t, int] ) -> int: @@ -866,7 +874,7 @@ def llama_model_desc( _lib.llama_model_desc.argtypes = [llama_model_p, c_char_p, c_size_t] -_lib.llama_model_desc.restype = c_int +_lib.llama_model_desc.restype = c_int32 # // Returns the total size of all the tensors in the model in bytes @@ -905,7 +913,7 @@ def llama_get_model_tensor( # // Returns 0 on success -# LLAMA_API int llama_model_quantize( +# LLAMA_API uint32_t llama_model_quantize( # const char * fname_inp, # const char * fname_out, # const llama_model_quantize_params * params); @@ -923,7 +931,7 @@ def llama_model_quantize( c_char_p, POINTER(llama_model_quantize_params), ] -_lib.llama_model_quantize.restype = c_int +_lib.llama_model_quantize.restype = c_uint32 # // Apply a LoRA adapter to a loaded model @@ -932,12 +940,12 @@ def llama_model_quantize( # // The model needs to be reloaded before applying a new adapter, otherwise the adapter # // will be applied on top of the previous one # // Returns 0 on success -# LLAMA_API DEPRECATED(int llama_apply_lora_from_file( +# LLAMA_API DEPRECATED(int32_t llama_apply_lora_from_file( # struct llama_context * ctx, # const char * path_lora, # float scale, # const char * path_base_model, -# int n_threads), +# int32_t n_threads), # "use llama_model_apply_lora_from_file instead"); def llama_apply_lora_from_file( ctx: llama_context_p, @@ -962,17 +970,17 @@ def llama_apply_lora_from_file( c_char_p, c_float, c_char_p, - c_int, + c_int32, ] -_lib.llama_apply_lora_from_file.restype = c_int +_lib.llama_apply_lora_from_file.restype = c_int32 -# LLAMA_API int llama_model_apply_lora_from_file( +# LLAMA_API int32_t llama_model_apply_lora_from_file( # const struct llama_model * model, # const char * path_lora, # float scale, # const char * path_base_model, -# int n_threads); +# int32_t n_threads); def llama_model_apply_lora_from_file( model: llama_model_p, path_lora: Union[c_char_p, bytes], @@ -990,9 +998,9 @@ def llama_model_apply_lora_from_file( c_char_p, c_float, c_char_p, - c_int, + c_int32, ] -_lib.llama_model_apply_lora_from_file.restype = c_int +_lib.llama_model_apply_lora_from_file.restype = c_int32 # // # // KV cache @@ -1094,7 +1102,7 @@ def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times -# LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx); +# LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); def llama_get_kv_cache_token_count(ctx: llama_context_p) -> int: """Returns the number of tokens in the KV cache (slow, use only for debug) If a KV cell has multiple sequences assigned to it, it will be counted multiple times @@ -1103,18 +1111,18 @@ def llama_get_kv_cache_token_count(ctx: llama_context_p) -> int: _lib.llama_get_kv_cache_token_count.argtypes = [llama_context_p] -_lib.llama_get_kv_cache_token_count.restype = c_int +_lib.llama_get_kv_cache_token_count.restype = c_int32 # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) -# LLAMA_API int llama_get_kv_cache_used_cells(const struct llama_context * ctx); +# LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); def llama_get_kv_cache_used_cells(ctx: llama_context_p) -> int: """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" return _lib.llama_get_kv_cache_used_cells(ctx) _lib.llama_get_kv_cache_used_cells.argtypes = [llama_context_p] -_lib.llama_get_kv_cache_used_cells.restype = c_int +_lib.llama_get_kv_cache_used_cells.restype = c_int32 # // Clear the KV cache @@ -1361,7 +1369,7 @@ def llama_save_session_file( # struct llama_context * ctx, # llama_token * tokens, # int32_t n_tokens, -# int n_past), +# int32_t n_past), # "use llama_decode() instead"); def llama_eval( ctx: llama_context_p, @@ -1377,7 +1385,7 @@ def llama_eval( return _lib.llama_eval(ctx, tokens, n_tokens, n_past) -_lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int, c_int] +_lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int32, c_int32] _lib.llama_eval.restype = c_int @@ -1387,7 +1395,7 @@ def llama_eval( # struct llama_context * ctx, # float * embd, # int32_t n_tokens, -# int n_past), +# int32_t n_past), # "use llama_decode() instead"); def llama_eval_embd( ctx: llama_context_p, @@ -1400,7 +1408,7 @@ def llama_eval_embd( return _lib.llama_eval_embd(ctx, embd, n_tokens, n_past) -_lib.llama_eval_embd.argtypes = [llama_context_p, c_float_p, c_int, c_int] +_lib.llama_eval_embd.argtypes = [llama_context_p, c_float_p, c_int32, c_int32] _lib.llama_eval_embd.restype = c_int @@ -1480,7 +1488,7 @@ def llama_batch_free(batch: llama_batch): # // 0 - success # // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) # // < 0 - error -# LLAMA_API int llama_decode( +# LLAMA_API int32_t llama_decode( # struct llama_context * ctx, # struct llama_batch batch); def llama_decode(ctx: llama_context_p, batch: llama_batch) -> int: @@ -1492,7 +1500,7 @@ def llama_decode(ctx: llama_context_p, batch: llama_batch) -> int: _lib.llama_decode.argtypes = [llama_context_p, llama_batch] -_lib.llama_decode.restype = c_int +_lib.llama_decode.restype = c_int32 # // Set the number of threads used for decoding @@ -1634,25 +1642,25 @@ def llama_token_nl(model: llama_model_p) -> int: # // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int llama_add_bos_token(const struct llama_model * model); +# LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); def llama_add_bos_token(model: llama_model_p) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" return _lib.llama_add_bos_token(model) _lib.llama_add_bos_token.argtypes = [llama_model_p] -_lib.llama_add_bos_token.restype = c_int +_lib.llama_add_bos_token.restype = c_int32 # // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int llama_add_eos_token(const struct llama_model * model); +# LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); def llama_add_eos_token(model: llama_model_p) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" return _lib.llama_add_eos_token(model) _lib.llama_add_eos_token.argtypes = [llama_model_p] -_lib.llama_add_eos_token.restype = c_int +_lib.llama_add_eos_token.restype = c_int32 # // codellama infill tokens @@ -1704,12 +1712,12 @@ def llama_token_eot(model: llama_model_p) -> int: # /// @return Returns a negative number on failure - the number of tokens that would have been returned # /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. # /// Does not insert a leading space. -# LLAMA_API int llama_tokenize( +# LLAMA_API int32_t llama_tokenize( # const struct llama_model * model, # const char * text, -# int text_len, +# int32_t text_len, # llama_token * tokens, -# int n_max_tokens, +# int32_t n_max_tokens, # bool add_bos, # bool special); def llama_tokenize( @@ -1730,24 +1738,24 @@ def llama_tokenize( _lib.llama_tokenize.argtypes = [ llama_model_p, c_char_p, - c_int, + c_int32, llama_token_p, - c_int, + c_int32, c_bool, c_bool, ] -_lib.llama_tokenize.restype = c_int +_lib.llama_tokenize.restype = c_int32 # // Token Id -> Piece. # // Uses the vocabulary in the provided context. # // Does not write null terminator to the buffer. # // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. -# LLAMA_API int llama_token_to_piece( +# LLAMA_API int32_t llama_token_to_piece( # const struct llama_model * model, # llama_token token, # char * buf, -# int length); +# int32_t length); def llama_token_to_piece( model: llama_model_p, token: Union[llama_token, int], @@ -1762,8 +1770,8 @@ def llama_token_to_piece( return _lib.llama_token_to_piece(model, token, buf, length) -_lib.llama_token_to_piece.argtypes = [llama_model_p, llama_token, c_char_p, c_int] -_lib.llama_token_to_piece.restype = c_int +_lib.llama_token_to_piece.argtypes = [llama_model_p, llama_token, c_char_p, c_int32] +_lib.llama_token_to_piece.restype = c_int32 # // @@ -1924,7 +1932,7 @@ def llama_sample_softmax( # LLAMA_API void llama_sample_top_k( # struct llama_context * ctx, # llama_token_data_array * candidates, -# int k, +# int32_t k, # size_t min_keep); def llama_sample_top_k( ctx: llama_context_p, @@ -1939,7 +1947,7 @@ def llama_sample_top_k( _lib.llama_sample_top_k.argtypes = [ llama_context_p, llama_token_data_array_p, - c_int, + c_int32, c_size_t, ] _lib.llama_sample_top_k.restype = None @@ -2129,7 +2137,7 @@ def llama_sample_grammar( # llama_token_data_array * candidates, # float tau, # float eta, -# int m, +# int32_t m, # float * mu); def llama_sample_token_mirostat( ctx: llama_context_p, @@ -2155,7 +2163,7 @@ def llama_sample_token_mirostat( llama_token_data_array_p, c_float, c_float, - c_int, + c_int32, c_float_p, ] _lib.llama_sample_token_mirostat.restype = llama_token @@ -2320,8 +2328,8 @@ class llama_beams_state(ctypes.Structure): # llama_beam_search_callback_fn_t callback, # void * callback_data, # size_t n_beams, -# int n_past, -# int n_predict); +# int32_t n_past, +# int32_t n_predict); def llama_beam_search( ctx: llama_context_p, callback: "ctypes._CFuncPtr[None, c_void_p, llama_beams_state]", # type: ignore @@ -2340,8 +2348,8 @@ def llama_beam_search( llama_beam_search_callback_fn_t, c_void_p, c_size_t, - c_int, - c_int, + c_int32, + c_int32, ] _lib.llama_beam_search.restype = None diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f6793491b..cb1e2818e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f6793491b5af6da75edad34d6f503ef86d31b09f +Subproject commit cb1e2818e0e12ec99f7236ec5d4f3ffd8bcc2f4a From cf743ec5d32cc84e68295da8442ccf3a64e635f1 Mon Sep 17 00:00:00 2001 From: xaviviro Date: Fri, 5 Jan 2024 00:12:02 +0100 Subject: [PATCH 002/624] Added ChatGLM chat format (#1059) Co-authored-by: Xavier Vinaixa Rosello --- llama_cpp/llama_chat_format.py | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 037f96a2d..6f402e08b 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -172,6 +172,20 @@ def _format_chatml( ret += role + "\n" return ret +def _format_chatglm3( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the chatglm3 style.""" + ret = "" + if system_message: + ret += system_message + for role, message in messages: + if message: + ret += role + "\n" + " " + message + else: + ret += role + return ret + @dataclasses.dataclass class ChatFormatterResponse: @@ -685,6 +699,22 @@ def format_chatml( _prompt = _format_chatml(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt, stop=_sep) +@register_chat_format("chatglm3") +def format_chatglm3( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + system_template = """<|system|> +{system_message}""" + system_message = _get_system_message(messages) + system_message = system_template.format(system_message=system_message) + _roles = dict(user="<|user|>", assistant="<|assistant|>") + _sep = "" + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_chatglm3(system_message, _messages, _sep) + return ChatFormatterResponse(prompt=_prompt, stop=_sep) + @register_chat_format("openchat") def format_openchat( From f766b70c9a63801f6f27dc92b4ab822f92055bc9 Mon Sep 17 00:00:00 2001 From: Caleb Hoff Date: Thu, 4 Jan 2024 17:12:32 -0600 Subject: [PATCH 003/624] Fix: Correct typo in README.md (#1058) In Llama.create_chat_completion, the `tool_choice` property does not have an s on the end. --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 97fe6e180..b2e879ec4 100644 --- a/README.md +++ b/README.md @@ -238,7 +238,7 @@ The gguf-converted files for this model can be found here: [functionary-7b-v1](h } } }], - tool_choices=[{ + tool_choice=[{ "type": "function", "function": { "name": "UserDetail" From 907b9e9d4281336072519fbf11e885768ad0ff0b Mon Sep 17 00:00:00 2001 From: Fedor Moiseev Date: Fri, 5 Jan 2024 06:12:58 +0700 Subject: [PATCH 004/624] Add Saiga chat format. (#1050) --- llama_cpp/llama_chat_format.py | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 6f402e08b..0ef7bd4a8 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -734,6 +734,28 @@ def format_openchat( return ChatFormatterResponse(prompt=_prompt, stop=_sep) +# Chat format for Saiga models, see more details and available models: +# https://huggingface.co/collections/IlyaGusev/saiga2-saigamistral-6505d4ccc3d1e53166b636cd +@register_chat_format("saiga") +def format_saiga( + messages: list[llama_types.ChatCompletionRequestMessage], + **kwargs, +) -> ChatFormatterResponse: + _message_template = "{role}\n{content}" + _roles = dict(user="user", bot="bot", system="system") + _messages = _map_roles(messages, _roles) + + _prompt = "" + for role, content in _messages: + if content: + _prompt += _message_template.format(role=role, content=content) + else: + _prompt += f"{role}\n" + # Response template + _prompt += "bot" + return ChatFormatterResponse(prompt=_prompt.strip()) + + @register_chat_completion_handler("functionary") def functionary_chat_handler( llama: llama.Llama, From fffcd0181c2b58a084daebc6df659520d0c73337 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 4 Jan 2024 18:26:00 -0500 Subject: [PATCH 005/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index cb1e2818e..b3a7c20b5 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit cb1e2818e0e12ec99f7236ec5d4f3ffd8bcc2f4a +Subproject commit b3a7c20b5c035250257d2b62851c379b159c899a From 75d0527fd782a792af8612e55b0a3f2dad469ae9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 4 Jan 2024 18:30:12 -0500 Subject: [PATCH 006/624] Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 228e3b99b..bcf166554 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.27] + +- feat: Update llama.cpp to ggerganov/llama.cpp@b3a7c20b5c035250257d2b62851c379b159c899a +- feat: Add `saiga` chat format by @femoiseev in #1050 +- feat: Added `chatglm3` chat format by @xaviviro in #1059 +- fix: Correct typo in README.md by @qeleb in (#1058) + ## [0.2.26] - feat: Update llama.cpp to ggerganov/llama.cpp@f6793491b5af6da75edad34d6f503ef86d31b09f diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 5db466434..d3fe66b3a 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.26" \ No newline at end of file +__version__ = "0.2.27" \ No newline at end of file From 142a9e1bc3227c511a204f941fc4a058c2db4577 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Jan 2024 16:20:50 -0500 Subject: [PATCH 007/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index b3a7c20b5..eec22a1c6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit b3a7c20b5c035250257d2b62851c379b159c899a +Subproject commit eec22a1c6378d9a013943cbddb4330c0da621442 From 1ae05c102b79ce036fff195347404c1afde2bbc8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 8 Jan 2024 14:51:29 -0500 Subject: [PATCH 008/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 94 +++++++++++++++++++++++++++++++++--------- vendor/llama.cpp | 2 +- 2 files changed, 76 insertions(+), 20 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 4aada532e..abd7b4cde 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -104,7 +104,7 @@ def _load_shared_library(lib_base_name: str): # define LLAMA_MAX_RNG_STATE (64*1024) LLAMA_MAX_RNG_STATE = 64 * 1024 -#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla' +# define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla' LLAMA_FILE_MAGIC_GGLA = 0x67676C61 # define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn' @@ -179,6 +179,7 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_Q5_K_S = 16, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -222,11 +223,12 @@ def _load_shared_library(lib_base_name: str): # } llama_token_data; class llama_token_data(Structure): """Used to store token data - + Attributes: id (llama_token): token id logit (float): log-odds of the token p (float): probability of the token""" + _fields_ = [ ("id", llama_token), ("logit", c_float), @@ -244,11 +246,12 @@ class llama_token_data(Structure): # } llama_token_data_array; class llama_token_data_array(Structure): """Used to sample tokens given logits - + Attributes: data (ctypes.Array[llama_token_data]): token data size (int): size of the array sorted (bool): whether the array is sorted""" + _fields_ = [ ("data", llama_token_data_p), ("size", c_size_t), @@ -303,7 +306,8 @@ class llama_batch(Structure): token (ctypes.Array[llama_token]): the token ids of the input (used when embd is NULL) embd (ctypes.Array[ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL) pos (ctypes.Array[ctypes.Array[llama_pos]]): the positions of the respective token in the sequence - seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs""" + seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs + """ _fields_ = [ ("n_tokens", c_int32), @@ -318,6 +322,7 @@ class llama_batch(Structure): ("all_seq_id", llama_seq_id), ] + # enum llama_model_kv_override_type { # LLAMA_KV_OVERRIDE_INT, # LLAMA_KV_OVERRIDE_FLOAT, @@ -327,6 +332,7 @@ class llama_batch(Structure): LLAMA_KV_OVERRIDE_FLOAT = 1 LLAMA_KV_OVERRIDE_BOOL = 2 + # struct llama_model_kv_override { # char key[128]; # enum llama_model_kv_override_type tag; @@ -343,6 +349,7 @@ class llama_model_kv_override_value(CtypesUnion): ("bool_value", c_bool), ] + class llama_model_kv_override(Structure): _fields_ = [ ("key", ctypes.c_char * 128), @@ -350,6 +357,7 @@ class llama_model_kv_override(Structure): ("value", llama_model_kv_override_value), ] + # struct llama_model_params { # int32_t n_gpu_layers; // number of layers to store in VRAM # int32_t main_gpu; // the GPU that is used for scratch and small tensors @@ -365,6 +373,7 @@ class llama_model_kv_override(Structure): # // override key-value pairs of the model meta data # const struct llama_model_kv_override * kv_overrides; + # // Keep the booleans together to avoid misalignment during copy-by-value. # bool vocab_only; // only load the vocabulary, no weights # bool use_mmap; // use mmap if possible @@ -372,7 +381,7 @@ class llama_model_kv_override(Structure): # }; class llama_model_params(Structure): """Parameters for llama_model - + Attributes: n_gpu_layers (int): number of layers to store in VRAM main_gpu (int): the GPU that is used for scratch and small tensors @@ -383,6 +392,7 @@ class llama_model_params(Structure): vocab_only (bool): only load the vocabulary, no weights use_mmap (bool): use mmap if possible use_mlock (bool): force system to keep model in RAM""" + _fields_ = [ ("n_gpu_layers", c_int32), ("main_gpu", c_int32), @@ -416,6 +426,7 @@ class llama_model_params(Structure): # enum ggml_type type_k; // data type for K cache # enum ggml_type type_v; // data type for V cache + # // Keep the booleans together to avoid misalignment during copy-by-value. # bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) # bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) @@ -424,7 +435,7 @@ class llama_model_params(Structure): # }; class llama_context_params(Structure): """Parameters for llama_context - + Attributes: seed (int): RNG seed, -1 for random n_ctx (int): text context, 0 = from model @@ -444,7 +455,9 @@ class llama_context_params(Structure): mul_mat_q (bool): if true, use experimental mul_mat_q kernels (DEPRECATED - always true) logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) embedding (bool): embedding mode only - offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU""" + offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU + """ + _fields_ = [ ("seed", c_uint32), ("n_ctx", c_uint32), @@ -493,14 +506,16 @@ class llama_context_params(Structure): # } llama_model_quantize_params; class llama_model_quantize_params(Structure): """Parameters for llama_model_quantize - + Attributes: nthread (int): number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() ftype (int): quantize to this llama_ftype allow_requantize (bool): allow quantizing non-f32/f16 tensors quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored - pure (bool): disable k-quant mixtures and quantize all tensors to the same type""" + pure (bool): disable k-quant mixtures and quantize all tensors to the same type + """ + _fields_ = [ ("nthread", c_int32), ("ftype", c_int), @@ -745,13 +760,16 @@ def llama_n_ctx(ctx: llama_context_p) -> int: _lib.llama_n_ctx.argtypes = [llama_context_p] _lib.llama_n_ctx.restype = c_uint32 + # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); def llama_n_batch(ctx: llama_context_p) -> int: return _lib.llama_n_batch(ctx) + _lib.llama_n_batch.argtypes = [llama_context_p] _lib.llama_n_batch.restype = c_uint32 + # LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); def llama_vocab_type(model: llama_model_p) -> int: return _lib.llama_vocab_type(model) @@ -1080,7 +1098,7 @@ def llama_kv_cache_view_init( # // Free a KV cache view. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); -def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore +def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore """Free a KV cache view. (use only for debugging purposes)""" return _lib.llama_kv_cache_view_free(view) @@ -1091,7 +1109,7 @@ def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]"): # typ # // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); -def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore +def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore """Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)""" return _lib.llama_kv_cache_view_update(ctx, view) @@ -1251,6 +1269,40 @@ def llama_kv_cache_seq_shift( ] _lib.llama_kv_cache_seq_shift.restype = None + +# // Integer division of the positions by factor of `d > 1` +# // If the KV cache is RoPEd, the KV data is updated accordingly +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_div( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d); +def llama_kv_cache_seq_div( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + d: Union[c_int, int], +): + """Integer division of the positions by factor of `d > 1` + If the KV cache is RoPEd, the KV data is updated accordingly + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + return _lib.llama_kv_cache_seq_div(ctx, seq_id, p0, p1, d) + + +_lib.llama_kv_cache_seq_div.argtypes = [ + llama_context_p, + llama_seq_id, + llama_pos, + llama_pos, + c_int, +] +_lib.llama_kv_cache_seq_div.restype = None + # // # // State / sessions # // @@ -2063,10 +2115,11 @@ def llama_sample_temp( temp: Union[c_float, float], ): """Temperature sampling described in academic paper "Generating Long Sequences with Sparse Transformers" https://arxiv.org/abs/1904.10509 - + Parameters: candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - temp: The temperature value to use for the sampling. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.""" + temp: The temperature value to use for the sampling. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. + """ return _lib.llama_sample_temp(ctx, candidates, temp) @@ -2111,10 +2164,11 @@ def llama_sample_grammar( grammar, # type: llama_grammar_p ): """Apply constraints from grammar - + Parameters: candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - grammar: A grammar object containing the rules and constraints to apply to the generated text.""" + grammar: A grammar object containing the rules and constraints to apply to the generated text. + """ return _lib.llama_sample_grammar(ctx, candidates, grammar) @@ -2148,13 +2202,14 @@ def llama_sample_token_mirostat( mu, # type: _Pointer[c_float] ) -> int: """Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - + Parameters: candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. m: The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. - mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.""" + mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. + """ return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu) @@ -2188,12 +2243,13 @@ def llama_sample_token_mirostat_v2( mu, # type: _Pointer[c_float] ) -> int: """Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - + Parameters: candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.""" + mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. + """ return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index eec22a1c6..1fc2f265f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit eec22a1c6378d9a013943cbddb4330c0da621442 +Subproject commit 1fc2f265ff9377a37fd2c61eae9cd813a3491bea From 431cb3ec81f3b9de95f41d519d465effad6cb412 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jan 2024 15:32:39 -0500 Subject: [PATCH 009/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1fc2f265f..6efb8eb30 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1fc2f265ff9377a37fd2c61eae9cd813a3491bea +Subproject commit 6efb8eb30e7025b168f3fda3ff83b9b386428ad6 From 2ddce7294ec51fb719fba126fc1229450f18c976 Mon Sep 17 00:00:00 2001 From: Joseph Turian Date: Wed, 10 Jan 2024 02:46:03 -0500 Subject: [PATCH 010/624] print_grammar to stderr (#1052) --- llama_cpp/llama_grammar.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index aec023cd6..0c3b2e0ef 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -72,7 +72,7 @@ def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": ) if verbose: print(f"{cls.from_string.__name__} grammar:", file=sys.stderr) - print_grammar(sys.stdout, parsed_grammar) + print_grammar(sys.stderr, parsed_grammar) print(file=sys.stderr) return cls(parsed_grammar) From df3be58d6c5a96fafa53f65a3bdb2d0fe21960e9 Mon Sep 17 00:00:00 2001 From: Stephen Hankinson Date: Wed, 10 Jan 2024 03:46:27 -0400 Subject: [PATCH 011/624] Add ability to pass in penalize_nl param (#1068) --- llama_cpp/llama.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 9178a2225..7c819b0d6 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1201,6 +1201,7 @@ def generate( mirostat_mode: int = 0, mirostat_tau: float = 5.0, mirostat_eta: float = 0.1, + penalize_nl: bool = True, logits_processor: Optional[LogitsProcessorList] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, grammar: Optional[LlamaGrammar] = None, @@ -1261,6 +1262,7 @@ def generate( mirostat_eta=mirostat_eta, logits_processor=logits_processor, grammar=grammar, + penalize_nl=penalize_nl, ) if stopping_criteria is not None and stopping_criteria( self._input_ids, self._scores[-1, :] From f0159663d922a6a693b1f9aee335d8667ae7f39e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 10 Jan 2024 02:51:17 -0500 Subject: [PATCH 012/624] Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index bcf166554..5061247a1 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.28] + +- feat: Update llama.cpp to ggerganov/llama.cpp@6efb8eb30e7025b168f3fda3ff83b9b386428ad6 +- feat: Add ability to pass in penalize_nl param by @shankinson in #1068 +- fix: print_grammar to stderr by @turian in #1052 + ## [0.2.27] - feat: Update llama.cpp to ggerganov/llama.cpp@b3a7c20b5c035250257d2b62851c379b159c899a diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index d3fe66b3a..33234fb5f 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.27" \ No newline at end of file +__version__ = "0.2.28" \ No newline at end of file From bb610b94282982d184129b8e8fc30e8fa04c4448 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 11 Jan 2024 22:51:12 -0500 Subject: [PATCH 013/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 5 +++++ vendor/llama.cpp | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index abd7b4cde..92ec29956 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -180,6 +180,8 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -200,6 +202,9 @@ def _load_shared_library(lib_base_name: str): LLAMA_FTYPE_MOSTLY_Q5_K_S = 16 LLAMA_FTYPE_MOSTLY_Q5_K_M = 17 LLAMA_FTYPE_MOSTLY_Q6_K = 18 +LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19 +LLAMA_FTYPE_MOSTLY_IQ2_XS = 20 +LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6efb8eb30..1d118386f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6efb8eb30e7025b168f3fda3ff83b9b386428ad6 +Subproject commit 1d118386fea031f01550f8cd47a5c86296e5333f From 7c898d5684a147ed9fdcbbd48831d982bf3a4523 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 13 Jan 2024 22:37:49 -0500 Subject: [PATCH 014/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 32 +++++++++++++++++++++++++------- vendor/llama.cpp | 2 +- 2 files changed, 26 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 92ec29956..989b67a1e 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -112,8 +112,8 @@ def _load_shared_library(lib_base_name: str): # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 3 -LLAMA_SESSION_VERSION = 3 +# define LLAMA_SESSION_VERSION 4 +LLAMA_SESSION_VERSION = 4 # struct llama_model; @@ -220,6 +220,14 @@ def _load_shared_library(lib_base_name: str): LLAMA_ROPE_SCALING_YARN = 2 LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN +# enum llama_split_mode { +# LLAMA_SPLIT_NONE = 0, // single GPU +# LLAMA_SPLIT_LAYER = 1, // split layers and KV across GPUs +# LLAMA_SPLIT_ROW = 2, // split rows across GPUs +# }; +LLAMA_SPLIT_NONE = 0 +LLAMA_SPLIT_LAYER = 1 +LLAMA_SPLIT_ROW = 2 # typedef struct llama_token_data { # llama_token id; // token id @@ -365,20 +373,28 @@ class llama_model_kv_override(Structure): # struct llama_model_params { # int32_t n_gpu_layers; // number of layers to store in VRAM -# int32_t main_gpu; // the GPU that is used for scratch and small tensors -# const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) +# enum llama_split_mode split_mode; // how to split the model across multiple GPUs + +# // main_gpu interpretation depends on split_mode: +# // LLAMA_SPLIT_NONE: the GPU that is used for the entire model +# // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results +# // LLAMA_SPLIT_LAYER: ignored +# int32_t main_gpu; + +# // proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES +# const float * tensor_split; # // Called with a progress value between 0.0 and 1.0. Pass NULL to disable. # // If the provided progress_callback returns true, model loading continues. # // If it returns false, model loading is immediately aborted. # llama_progress_callback progress_callback; + # // context pointer passed to the progress callback # void * progress_callback_user_data; # // override key-value pairs of the model meta data # const struct llama_model_kv_override * kv_overrides; - # // Keep the booleans together to avoid misalignment during copy-by-value. # bool vocab_only; // only load the vocabulary, no weights # bool use_mmap; // use mmap if possible @@ -389,8 +405,9 @@ class llama_model_params(Structure): Attributes: n_gpu_layers (int): number of layers to store in VRAM - main_gpu (int): the GPU that is used for scratch and small tensors - tensor_split (ctypes.Array[ctypes.c_float]): how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) + split_mode (int): how to split the model across multiple GPUs + main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored + tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. progress_callback_user_data (ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data @@ -400,6 +417,7 @@ class llama_model_params(Structure): _fields_ = [ ("n_gpu_layers", c_int32), + ("split_mode", c_int), ("main_gpu", c_int32), ("tensor_split", c_float_p), ("progress_callback", llama_progress_callback), diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1d118386f..76484fbfd 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1d118386fea031f01550f8cd47a5c86296e5333f +Subproject commit 76484fbfd355df388f71d6edaa98e1692a74de7e From 359ae736432cae5cdba50b011839d277a4b1ec8d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 14 Jan 2024 08:17:22 -0500 Subject: [PATCH 015/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++++ vendor/llama.cpp | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 989b67a1e..3b261cd23 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -526,6 +526,7 @@ class llama_context_params(Structure): # bool quantize_output_tensor; // quantize output.weight # bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored # bool pure; // disable k-quant mixtures and quantize all tensors to the same type +# void * imatrix; // pointer to importance matrix data # } llama_model_quantize_params; class llama_model_quantize_params(Structure): """Parameters for llama_model_quantize @@ -537,6 +538,7 @@ class llama_model_quantize_params(Structure): quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored pure (bool): disable k-quant mixtures and quantize all tensors to the same type + imatrix (ctypes.c_void_p): pointer to importance matrix data """ _fields_ = [ @@ -545,6 +547,8 @@ class llama_model_quantize_params(Structure): ("allow_requantize", c_bool), ("quantize_output_tensor", c_bool), ("only_copy", c_bool), + ("pure", c_bool), + ("imatrix", c_void_p), ] diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 76484fbfd..bb0c13924 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 76484fbfd355df388f71d6edaa98e1692a74de7e +Subproject commit bb0c1392479398f9aba86d9ec98db0b95ede6e6d From 5502ac8876047c231a86fdaa1ee5f3dae6a9b3a5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 15 Jan 2024 10:12:10 -0500 Subject: [PATCH 016/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 43 ++++++++++++++++++++++++++++++++++-------- vendor/llama.cpp | 2 +- 2 files changed, 36 insertions(+), 9 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 3b261cd23..9e8e3cec7 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -229,6 +229,7 @@ def _load_shared_library(lib_base_name: str): LLAMA_SPLIT_LAYER = 1 LLAMA_SPLIT_ROW = 2 + # typedef struct llama_token_data { # llama_token id; // token id # float logit; // log-odds of the token @@ -395,6 +396,7 @@ class llama_model_kv_override(Structure): # // override key-value pairs of the model meta data # const struct llama_model_kv_override * kv_overrides; + # // Keep the booleans together to avoid misalignment during copy-by-value. # bool vocab_only; // only load the vocabulary, no weights # bool use_mmap; // use mmap if possible @@ -407,7 +409,7 @@ class llama_model_params(Structure): n_gpu_layers (int): number of layers to store in VRAM split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored - tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES + tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. progress_callback_user_data (ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data @@ -1960,14 +1962,39 @@ def llama_sample_repetition_penalties( # /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 -# /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted. -# /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. -# /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance. -# LLAMA_API void llama_sample_classifier_free_guidance( -# struct llama_context * ctx, +# /// @param logits Logits extracted from the original generation context. +# /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. +# /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance. +# LLAMA_API void llama_sample_apply_guidance( +# struct llama_context * ctx, +# float * logits, +# float * logits_guidance, +# float scale); +def llama_sample_apply_guidance( + ctx: llama_context_p, + logits, # type: _Pointer[c_float] + logits_guidance, # type: _Pointer[c_float] + scale: Union[c_float, float], +): + """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" + return _lib.llama_sample_apply_guidance(ctx, logits, logits_guidance, scale) + + +_lib.llama_sample_apply_guidance.argtypes = [ + llama_context_p, + c_float_p, + c_float_p, + c_float, +] +_lib.llama_sample_apply_guidance.restype = None + + +# LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance( +# struct llama_context * ctx, # llama_token_data_array * candidates, -# struct llama_context * guidance_ctx, -# float scale); +# struct llama_context * guidance_ctx, +# float scale), +# "use llama_sample_apply_guidance() instead"); def llama_sample_classifier_free_guidance( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] diff --git a/vendor/llama.cpp b/vendor/llama.cpp index bb0c13924..448339675 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit bb0c1392479398f9aba86d9ec98db0b95ede6e6d +Subproject commit 4483396751c79dea540808b9cb9238245d06da2b From c689ccc7286b6b77289a8587b6534ec948d55611 Mon Sep 17 00:00:00 2001 From: Mark Neumann Date: Mon, 15 Jan 2024 07:45:57 -0800 Subject: [PATCH 017/624] Fix Pydantic model parsing (#1087) --- llama_cpp/llama_grammar.py | 1 - tests/test_grammar.py | 39 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 39 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 0c3b2e0ef..c02e65642 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1433,7 +1433,6 @@ def _add_rule(self, name: str, rule: str): def visit(self, schema: Dict[str, Any], name: str) -> str: schema_type: Optional[str] = schema.get("type") # type: ignore - assert isinstance(schema_type, str), f"Unrecognized schema: {schema}" rule_name = name or "root" if "$defs" in schema: diff --git a/tests/test_grammar.py b/tests/test_grammar.py index 2e24903c2..ef9392b7a 100644 --- a/tests/test_grammar.py +++ b/tests/test_grammar.py @@ -1,4 +1,5 @@ import llama_cpp +import json tree = """ leaf ::= "." @@ -6,8 +7,46 @@ root ::= node """ + def test_grammar_from_string(): grammar = llama_cpp.LlamaGrammar.from_string(tree) assert grammar._n_rules == 3 assert grammar._start_rule_index == 2 assert grammar.grammar is not None + + +def test_composed_pydantic_grammar(): + """ + from pydantic import BaseModel + + class A(BaseModel): + a: int + + class B(BaseModel): + a: A + b: int + """ + + # This schema corresponds to the grammar in the comment above. + # We don't use the pydantic models directly to avoid the dependency. + schema = { + "$defs": { + "A": { + "properties": {"a": {"title": "A", "type": "integer"}}, + "required": ["a"], + "title": "A", + "type": "object", + } + }, + "properties": { + "a": {"$ref": "#/$defs/A"}, + "b": {"title": "B", "type": "integer"}, + }, + "required": ["a", "b"], + "title": "B", + "type": "object", + } + + grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(schema)) + + assert grammar.grammar is not None From 1eaace8ea394423df6185f2cc83e7cb98a4bc35b Mon Sep 17 00:00:00 2001 From: anil <412569+aniljava@users.noreply.github.com> Date: Mon, 15 Jan 2024 09:46:35 -0600 Subject: [PATCH 018/624] Fix low_level_api_chat_cpp example to match current API (#1086) * Fix low_level_api_chat_cpp to match current API * Fix low_level_api_chat_cpp to match current API * Using None instead of empty string to so that default prompt template can be used if no prompt provided --------- Co-authored-by: Anil Pathak --- examples/low_level_api/common.py | 2 +- .../low_level_api/low_level_api_chat_cpp.py | 50 +++++++++++++------ 2 files changed, 37 insertions(+), 15 deletions(-) diff --git a/examples/low_level_api/common.py b/examples/low_level_api/common.py index 55d08db5f..1a5152530 100644 --- a/examples/low_level_api/common.py +++ b/examples/low_level_api/common.py @@ -106,7 +106,7 @@ def gpt_params_parse(argv = None): parser.add_argument("--mirostat_lr", type=float, default=0.1, help="Mirostat learning rate, parameter eta",dest="mirostat_eta") parser.add_argument("-m", "--model", type=str, default="./models/llama-7B/ggml-model.bin", help="model path",dest="model") - parser.add_argument("-p", "--prompt", type=str, default="", help="initial prompt",dest="prompt") + parser.add_argument("-p", "--prompt", type=str, default=None, help="initial prompt",dest="prompt") parser.add_argument("-f", "--file", type=str, default=None, help="file containing initial prompt to load",dest="file") parser.add_argument("--session", type=str, default=None, help="file to cache model state in (may be large!)",dest="path_session") parser.add_argument("--in-prefix", type=str, default="", help="string to prefix user inputs with", dest="input_prefix") diff --git a/examples/low_level_api/low_level_api_chat_cpp.py b/examples/low_level_api/low_level_api_chat_cpp.py index 44b6d4a35..02c09afb0 100644 --- a/examples/low_level_api/low_level_api_chat_cpp.py +++ b/examples/low_level_api/low_level_api_chat_cpp.py @@ -62,7 +62,7 @@ def __init__(self, params: GptParams) -> None: self.multibyte_fix = [] # model load - self.lparams = llama_cpp.llama_context_default_params() + self.lparams = llama_cpp.llama_model_default_params() self.lparams.n_ctx = self.params.n_ctx self.lparams.n_parts = self.params.n_parts self.lparams.seed = self.params.seed @@ -72,7 +72,11 @@ def __init__(self, params: GptParams) -> None: self.model = llama_cpp.llama_load_model_from_file( self.params.model.encode("utf8"), self.lparams) - self.ctx = llama_cpp.llama_new_context_with_model(self.model, self.lparams) + + # Context Params. + self.cparams = llama_cpp.llama_context_default_params() + + self.ctx = llama_cpp.llama_new_context_with_model(self.model, self.cparams) if (not self.ctx): raise RuntimeError(f"error: failed to load model '{self.params.model}'") @@ -244,7 +248,7 @@ def __init__(self, params: GptParams) -> None: # tokenize a prompt def _tokenize(self, prompt, bos=True): _arr = (llama_cpp.llama_token * ((len(prompt) + 1) * 4))() - _n = llama_cpp.llama_tokenize(self.ctx, prompt.encode("utf8", errors="ignore"), _arr, len(_arr), bos) + _n = llama_cpp.llama_tokenize(self.model, prompt.encode("utf8", errors="ignore"), len(prompt), _arr, len(_arr), bos, False) return _arr[:_n] def set_color(self, c): @@ -304,7 +308,7 @@ def generate(self): self.n_past += n_eval""" if (llama_cpp.llama_eval( - self.ctx, (llama_cpp.llama_token * len(self.embd))(*self.embd), len(self.embd), self.n_past, self.params.n_threads + self.ctx, (llama_cpp.llama_token * len(self.embd))(*self.embd), len(self.embd), self.n_past ) != 0): raise Exception("Failed to llama_eval!") @@ -332,7 +336,7 @@ def generate(self): id = 0 logits = llama_cpp.llama_get_logits(self.ctx) - n_vocab = llama_cpp.llama_n_vocab(self.ctx) + n_vocab = llama_cpp.llama_n_vocab(self.model) # Apply params.logit_bias map for key, value in self.params.logit_bias.items(): @@ -349,12 +353,20 @@ def generate(self): last_n_repeat = min(len(self.last_n_tokens), repeat_last_n, self.n_ctx) _arr = (llama_cpp.llama_token * last_n_repeat)(*self.last_n_tokens[len(self.last_n_tokens) - last_n_repeat:]) - llama_cpp.llama_sample_repetition_penalty(self.ctx, candidates_p, - _arr, - last_n_repeat, llama_cpp.c_float(self.params.repeat_penalty)) - llama_cpp.llama_sample_frequency_and_presence_penalties(self.ctx, candidates_p, - _arr, - last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty)) + llama_cpp.llama_sample_repetition_penalties( + ctx=self.ctx, + candidates=candidates_p, + last_tokens_data = _arr, + penalty_last_n = last_n_repeat, + penalty_repeat = llama_cpp.c_float(self.params.repeat_penalty), + penalty_freq = llama_cpp.c_float(self.params.frequency_penalty), + penalty_present = llama_cpp.c_float(self.params.presence_penalty), + ) + + # NOT PRESENT IN CURRENT VERSION ? + # llama_cpp.llama_sample_frequency_and_presence_penalti(self.ctx, candidates_p, + # _arr, + # last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty)) if not self.params.penalize_nl: logits[llama_cpp.llama_token_nl()] = nl_logit @@ -473,7 +485,7 @@ def exit(self): def token_to_str(self, token_id: int) -> bytes: size = 32 buffer = (ctypes.c_char * size)() - n = llama_cpp.llama_token_to_piece_with_model( + n = llama_cpp.llama_token_to_piece( self.model, llama_cpp.llama_token(token_id), buffer, size) assert n <= size return bytes(buffer[:n]) @@ -532,6 +544,9 @@ def interact(self): print(i,end="",flush=True) self.params.input_echo = False + # Using string instead of tokens to check for antiprompt, + # It is more reliable than tokens for interactive mode. + generated_str = "" while self.params.interactive: self.set_color(util.CONSOLE_COLOR_USER_INPUT) if (self.params.instruct): @@ -546,6 +561,10 @@ def interact(self): try: for i in self.output(): print(i,end="",flush=True) + generated_str += i + for ap in self.params.antiprompt: + if generated_str.endswith(ap): + raise KeyboardInterrupt except KeyboardInterrupt: self.set_color(util.CONSOLE_COLOR_DEFAULT) if not self.params.instruct: @@ -561,7 +580,7 @@ def interact(self): time_now = datetime.now() prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. {AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer {USER_NAME}’s requests immediately and with details and precision. -There are no annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. +Transcript below contains only the recorded dialog between two, without any annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown. @@ -575,8 +594,11 @@ def interact(self): {AI_NAME}: A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae. {USER_NAME}: Name a color. {AI_NAME}: Blue -{USER_NAME}:""" +{USER_NAME}: """ + params = gpt_params_parse() + if params.prompt is None and params.file is None: + params.prompt = prompt with LLaMAInteract(params) as m: m.interact() From 7eff42c2397e0d084544a39870cf07266c371970 Mon Sep 17 00:00:00 2001 From: yieldthought Date: Mon, 15 Jan 2024 16:52:10 +0100 Subject: [PATCH 019/624] Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor (#1012) The existing code often causes "LookupError: unknown encoding: ascii" when open() called in a destructor. Saving open in self.open is not enough to avoid this. Instead, we can avoid reopening /dev/null every time by doing it once when the module is loaded. --- llama_cpp/_utils.py | 20 +++++++++----------- 1 file changed, 9 insertions(+), 11 deletions(-) diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index 171f357f5..f7b6ba6b1 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -1,11 +1,15 @@ import os import sys +import sys, traceback + +# Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor +outnull_file = open(os.devnull, "w") +errnull_file = open(os.devnull, "w") class suppress_stdout_stderr(object): # NOTE: these must be "saved" here to avoid exceptions when using # this context manager inside of a __del__ method - open = open sys = sys os = os @@ -21,9 +25,6 @@ def __enter__(self): if not hasattr(self.sys.stdout, 'fileno') or not hasattr(self.sys.stderr, 'fileno'): return self # Return the instance without making changes - self.outnull_file = self.open(self.os.devnull, "w") - self.errnull_file = self.open(self.os.devnull, "w") - self.old_stdout_fileno_undup = self.sys.stdout.fileno() self.old_stderr_fileno_undup = self.sys.stderr.fileno() @@ -33,11 +34,11 @@ def __enter__(self): self.old_stdout = self.sys.stdout self.old_stderr = self.sys.stderr - self.os.dup2(self.outnull_file.fileno(), self.old_stdout_fileno_undup) - self.os.dup2(self.errnull_file.fileno(), self.old_stderr_fileno_undup) + self.os.dup2(outnull_file.fileno(), self.old_stdout_fileno_undup) + self.os.dup2(errnull_file.fileno(), self.old_stderr_fileno_undup) - self.sys.stdout = self.outnull_file - self.sys.stderr = self.errnull_file + self.sys.stdout = outnull_file + self.sys.stderr = errnull_file return self def __exit__(self, *_): @@ -54,6 +55,3 @@ def __exit__(self, *_): self.os.close(self.old_stdout_fileno) self.os.close(self.old_stderr_fileno) - - self.outnull_file.close() - self.errnull_file.close() From 76aafa61497ce4c9eda22c9892ae93ccb8f7d814 Mon Sep 17 00:00:00 2001 From: Phil H <5756783+phiharri@users.noreply.github.com> Date: Mon, 15 Jan 2024 17:29:29 +0000 Subject: [PATCH 020/624] Implement GGUF metadata KV overrides (#1011) * Implement GGUF metadata overrides * whitespace fix * Fix kv overrides. * Fix pointer and pickle * Match llama.cpp kv_overrides cli argument --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 32 ++++++++++++++++++++++++++++++++ llama_cpp/server/model.py | 20 +++++++++++++++++++- llama_cpp/server/settings.py | 4 ++++ 3 files changed, 55 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 7c819b0d6..6443b6de2 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -735,6 +735,7 @@ def __init__( vocab_only: bool = False, use_mmap: bool = True, use_mlock: bool = False, + kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None, # Context Params seed: int = llama_cpp.LLAMA_DEFAULT_SEED, n_ctx: int = 512, @@ -803,6 +804,7 @@ def __init__( vocab_only: Only load the vocabulary no weights. use_mmap: Use mmap if possible. use_mlock: Force the system to keep the model in RAM. + kv_overrides: Key-value overrides for the model. seed: RNG seed, -1 for random n_ctx: Text context, 0 = from model n_batch: Prompt processing maximum batch size @@ -866,6 +868,34 @@ def __init__( self.model_params.use_mmap = use_mmap if lora_path is None else False self.model_params.use_mlock = use_mlock + self.kv_overrides = kv_overrides + if kv_overrides is not None: + n_overrides = len(kv_overrides) + self._kv_overrides_array = llama_cpp.llama_model_kv_override * (n_overrides + 1) + self._kv_overrides_array_keys = [] + + for k, v in kv_overrides.items(): + key_buf = ctypes.create_string_buffer(k.encode("utf-8")) + self._kv_overrides_array_keys.append(key_buf) + self._kv_overrides_array[i].key = key_buf + if isinstance(v, int): + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT + self._kv_overrides_array[i].value.int_value = v + elif isinstance(v, float): + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_FLOAT + self._kv_overrides_array[i].value.float_value = v + elif isinstance(v, bool): + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL + self._kv_overrides_array[i].value.bool_value = v + else: + raise ValueError(f"Unknown value type for {k}: {v}") + + self._kv_overrides_array_sentinel_key = b'\0' + + # null array sentinel + self._kv_overrides_array[n_overrides].key = self._kv_overrides_array_sentinel_key + self.model_params.kv_overrides = self._kv_overrides_array + self.n_batch = min(n_ctx, n_batch) # ??? self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1) self.n_threads_batch = n_threads_batch or max( @@ -2148,6 +2178,7 @@ def __getstate__(self): vocab_only=self.model_params.vocab_only, use_mmap=self.model_params.use_mmap, use_mlock=self.model_params.use_mlock, + kv_overrides=self.kv_overrides, # Context Params seed=self.context_params.seed, n_ctx=self.context_params.n_ctx, @@ -2190,6 +2221,7 @@ def __setstate__(self, state): vocab_only=state["vocab_only"], use_mmap=state["use_mmap"], use_mlock=state["use_mlock"], + kv_overrides=state["kv_overrides"], # Context Params seed=state["seed"], n_ctx=state["n_ctx"], diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index b9373b7ac..f9be3237d 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -1,6 +1,6 @@ from __future__ import annotations -from typing import Optional, Union, List +from typing import Dict, Optional, Union, List import llama_cpp @@ -71,6 +71,23 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler( clip_model_path=settings.clip_model_path, verbose=settings.verbose ) + + kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None + if settings.kv_overrides is not None: + assert isinstance(settings.kv_overrides, list) + kv_overrides = {} + for kv in settings.kv_overrides: + key, value = kv.split("=") + if ":" in value: + value_type, value = value.split(":") + if value_type == "bool": + kv_overrides[key] = value.lower() in ["true", "1"] + elif value_type == "int": + kv_overrides[key] = int(value) + elif value_type == "float": + kv_overrides[key] = float(value) + else: + raise ValueError(f"Unknown value type {value_type}") _model = llama_cpp.Llama( model_path=settings.model, @@ -81,6 +98,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: vocab_only=settings.vocab_only, use_mmap=settings.use_mmap, use_mlock=settings.use_mlock, + kv_overrides=kv_overrides, # Context Params seed=settings.seed, n_ctx=settings.n_ctx, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 346b4631e..3195d1dbd 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -48,6 +48,10 @@ class ModelSettings(BaseSettings): default=llama_cpp.llama_mlock_supported(), description="Use mlock.", ) + kv_overrides: Optional[List[str]] = Field( + default=None, + description="List of model kv overrides in the format key=type:value where type is one of (bool, int, float). Valid true values are (true, TRUE, 1), otherwise false.", + ) # Context Params seed: int = Field( default=llama_cpp.LLAMA_DEFAULT_SEED, description="Random seed. -1 for random." From 84615adbc6855c8384807c42f0130f9a1763f99d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 15 Jan 2024 12:49:20 -0500 Subject: [PATCH 021/624] Add split_mode option. Closes #1085 --- llama_cpp/llama.py | 7 ++++++- llama_cpp/server/settings.py | 4 ++++ 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 6443b6de2..e4be9d1c9 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -730,6 +730,7 @@ def __init__( *, # Model Params n_gpu_layers: int = 0, + split_mode: int = llama_cpp.LLAMA_SPLIT_LAYER, main_gpu: int = 0, tensor_split: Optional[List[float]] = None, vocab_only: bool = False, @@ -799,7 +800,8 @@ def __init__( Args: model_path: Path to the model. n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded. - main_gpu: The GPU that is used for scratch and small tensors. + split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options. + main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_LAYER: ignored tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split. vocab_only: Only load the vocabulary no weights. use_mmap: Use mmap if possible. @@ -850,6 +852,7 @@ def __init__( self.model_params.n_gpu_layers = ( 0x7FFFFFFF if n_gpu_layers == -1 else n_gpu_layers ) # 0x7FFFFFFF is INT32 max, will be auto set to all layers + self.model_params.split_mode = split_mode self.model_params.main_gpu = main_gpu self.tensor_split = tensor_split self._c_tensor_split = None @@ -2173,6 +2176,7 @@ def __getstate__(self): model_path=self.model_path, # Model Params n_gpu_layers=self.model_params.n_gpu_layers, + split_mode=self.model_params.split_mode, main_gpu=self.model_params.main_gpu, tensor_split=self.tensor_split, vocab_only=self.model_params.vocab_only, @@ -2216,6 +2220,7 @@ def __setstate__(self, state): model_path=state["model_path"], # Model Params n_gpu_layers=state["n_gpu_layers"], + split_mode=state["split_mode"], main_gpu=state["main_gpu"], tensor_split=state["tensor_split"], vocab_only=state["vocab_only"], diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 3195d1dbd..902a43919 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -28,6 +28,10 @@ class ModelSettings(BaseSettings): ge=-1, description="The number of layers to put on the GPU. The rest will be on the CPU. Set -1 to move all to GPU.", ) + split_mode: int = Field( + default=llama_cpp.LLAMA_SPLIT_LAYER, + description="The split mode to use.", + ) main_gpu: int = Field( default=0, ge=0, From 4b11fa83c00a3c04cfb47775ffcd226167d52044 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 15 Jan 2024 12:54:51 -0500 Subject: [PATCH 022/624] Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 5061247a1..c0748ee14 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.29] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4483396751c79dea540808b9cb9238245d06da2b +- feat: Add split_mode option by @abetlen in 84615adbc6855c8384807c42f0130f9a1763f99d +- feat: Implement GGUF metadata KV overrides by @phiharri in #1011 +- fix: Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor by @yieldthought in #1012 +- fix: Fix low_level_api_chat_cpp example to match current API by @aniljava in #1086 +- fix: Fix Pydantic model parsing by @DeNeutoy in #1087 + ## [0.2.28] - feat: Update llama.cpp to ggerganov/llama.cpp@6efb8eb30e7025b168f3fda3ff83b9b386428ad6 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 33234fb5f..65206bf28 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.28" \ No newline at end of file +__version__ = "0.2.29" \ No newline at end of file From e39778f8ebaa9f68df992de38c13487d239e14a4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 16 Jan 2024 11:56:44 -0500 Subject: [PATCH 023/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 448339675..862f5e41a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4483396751c79dea540808b9cb9238245d06da2b +Subproject commit 862f5e41ab1fdf12d6f59455aad3f5dd8258f805 From cfb7da98ed64feb882065222b17ca261356a7f59 Mon Sep 17 00:00:00 2001 From: anil <412569+aniljava@users.noreply.github.com> Date: Tue, 16 Jan 2024 11:52:52 -0600 Subject: [PATCH 024/624] Support Accept text/event-stream in chat and completion endpoints, resolves #1083 (#1088) Co-authored-by: Anil Pathak Co-authored-by: Andrei Betlen --- llama_cpp/server/app.py | 59 +++++++++++++++++++++++++++++++++++++++-- 1 file changed, 57 insertions(+), 2 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index c54e4eb5c..fed0a6deb 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -197,7 +197,36 @@ async def authenticate( @router.post( - "/v1/completions", summary="Completion", dependencies=[Depends(authenticate)] + "/v1/completions", + summary="Completion", + dependencies=[Depends(authenticate)], + response_model= Union[ + llama_cpp.CreateCompletionResponse, + str, + ], + responses={ + "200": { + "description": "Successful Response", + "content": { + "application/json": { + "schema": { + "anyOf": [ + {"$ref": "#/components/schemas/CreateCompletionResponse"} + ], + "title": "Completion response, when stream=False", + } + }, + "text/event-stream":{ + "schema": { + "type": "string", + "title": "Server Side Streaming response, when stream=True. " + + "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 + "example": """data: {... see CreateCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""" + } + } + }, + } + }, ) @router.post( "/v1/engines/copilot-codex/completions", @@ -280,7 +309,33 @@ async def create_embedding( @router.post( - "/v1/chat/completions", summary="Chat", dependencies=[Depends(authenticate)] + "/v1/chat/completions", summary="Chat", dependencies=[Depends(authenticate)], + response_model= Union[ + llama_cpp.ChatCompletion, str + ], + responses={ + "200": { + "description": "Successful Response", + "content": { + "application/json": { + "schema": { + "anyOf": [ + {"$ref": "#/components/schemas/CreateChatCompletionResponse"} + ], + "title": "Completion response, when stream=False", + } + }, + "text/event-stream":{ + "schema": { + "type": "string", + "title": "Server Side Streaming response, when stream=True" + + "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 + "example": """data: {... see CreateChatCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""" + } + } + }, + } + }, ) async def create_chat_completion( request: Request, From 9c36688b33d32f77c501f4815047e4fc5581cf83 Mon Sep 17 00:00:00 2001 From: Kyle Mistele Date: Tue, 16 Jan 2024 17:54:06 -0600 Subject: [PATCH 025/624] fix(cli): allow passing n_ctx=0 to openAI API server args to use model n_ctx_train field per #1015 (#1093) --- llama_cpp/server/settings.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 902a43919..a10390c75 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -60,7 +60,7 @@ class ModelSettings(BaseSettings): seed: int = Field( default=llama_cpp.LLAMA_DEFAULT_SEED, description="Random seed. -1 for random." ) - n_ctx: int = Field(default=2048, ge=1, description="The context size.") + n_ctx: int = Field(default=2048, ge=0, description="The context size.") n_batch: int = Field( default=512, ge=1, description="The batch size to use per eval." ) From 84380fe9a64f784bd8f9d280185c4b0977efb449 Mon Sep 17 00:00:00 2001 From: Jerry Liu Date: Tue, 16 Jan 2024 16:10:50 -0800 Subject: [PATCH 026/624] Add llamaindex integration to readme (#1092) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index b2e879ec4..ad5d0f1e8 100644 --- a/README.md +++ b/README.md @@ -14,6 +14,7 @@ This package provides: - High-level Python API for text completion - OpenAI-like API - [LangChain compatibility](https://python.langchain.com/docs/integrations/llms/llamacpp) + - [LlamaIndex compatibility](https://docs.llamaindex.ai/en/stable/examples/llm/llama_2_llama_cpp.html) - OpenAI compatible web server - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) From d5dbb3f8de147efd00f00bdb1d44d4cb4c83ecbc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 16 Jan 2024 19:35:57 -0500 Subject: [PATCH 027/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 862f5e41a..5c9996090 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 862f5e41ab1fdf12d6f59455aad3f5dd8258f805 +Subproject commit 5c999609013a30c06e6fd28be8db5c2074bcc196 From 3b92419132700259acb4690d0ce2e2ee979e00bc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Jan 2024 09:09:12 -0500 Subject: [PATCH 028/624] Move cache classes to llama_cache submodule. --- llama_cpp/llama.py | 147 ++------------------------------------ llama_cpp/llama_cache.py | 150 +++++++++++++++++++++++++++++++++++++++ 2 files changed, 157 insertions(+), 140 deletions(-) create mode 100644 llama_cpp/llama_cache.py diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index e4be9d1c9..c6b55abae 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -3,7 +3,6 @@ import uuid import time import multiprocessing -from abc import ABC, abstractmethod from typing import ( List, Optional, @@ -12,16 +11,20 @@ Sequence, Iterator, Deque, - Tuple, Callable, ) -from collections import deque, OrderedDict +from collections import deque -import diskcache import ctypes from .llama_types import * from .llama_grammar import LlamaGrammar +from .llama_cache import ( + BaseLlamaCache, + LlamaCache, # type: ignore + LlamaDiskCache, # type: ignore + LlamaRAMCache, # type: ignore +) import llama_cpp.llama_cpp as llama_cpp import llama_cpp.llama_chat_format as llama_chat_format @@ -31,142 +34,6 @@ from ._utils import suppress_stdout_stderr -class BaseLlamaCache(ABC): - """Base cache class for a llama.cpp model.""" - - def __init__(self, capacity_bytes: int = (2 << 30)): - self.capacity_bytes = capacity_bytes - - @property - @abstractmethod - def cache_size(self) -> int: - raise NotImplementedError - - def _find_longest_prefix_key( - self, - key: Tuple[int, ...], - ) -> Optional[Tuple[int, ...]]: - pass - - @abstractmethod - def __getitem__(self, key: Sequence[int]) -> "LlamaState": - raise NotImplementedError - - @abstractmethod - def __contains__(self, key: Sequence[int]) -> bool: - raise NotImplementedError - - @abstractmethod - def __setitem__(self, key: Sequence[int], value: "LlamaState") -> None: - raise NotImplementedError - - -class LlamaRAMCache(BaseLlamaCache): - """Cache for a llama.cpp model using RAM.""" - - def __init__(self, capacity_bytes: int = (2 << 30)): - super().__init__(capacity_bytes) - self.capacity_bytes = capacity_bytes - self.cache_state: OrderedDict[Tuple[int, ...], "LlamaState"] = OrderedDict() - - @property - def cache_size(self): - return sum([state.llama_state_size for state in self.cache_state.values()]) - - def _find_longest_prefix_key( - self, - key: Tuple[int, ...], - ) -> Optional[Tuple[int, ...]]: - min_len = 0 - min_key = None - keys = ( - (k, Llama.longest_token_prefix(k, key)) for k in self.cache_state.keys() - ) - for k, prefix_len in keys: - if prefix_len > min_len: - min_len = prefix_len - min_key = k - return min_key - - def __getitem__(self, key: Sequence[int]) -> "LlamaState": - key = tuple(key) - _key = self._find_longest_prefix_key(key) - if _key is None: - raise KeyError("Key not found") - value = self.cache_state[_key] - self.cache_state.move_to_end(_key) - return value - - def __contains__(self, key: Sequence[int]) -> bool: - return self._find_longest_prefix_key(tuple(key)) is not None - - def __setitem__(self, key: Sequence[int], value: "LlamaState"): - key = tuple(key) - if key in self.cache_state: - del self.cache_state[key] - self.cache_state[key] = value - while self.cache_size > self.capacity_bytes and len(self.cache_state) > 0: - self.cache_state.popitem(last=False) - - -# Alias for backwards compatibility -LlamaCache = LlamaRAMCache - - -class LlamaDiskCache(BaseLlamaCache): - """Cache for a llama.cpp model using disk.""" - - def __init__( - self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30) - ): - super().__init__(capacity_bytes) - self.cache = diskcache.Cache(cache_dir) - - @property - def cache_size(self): - return int(self.cache.volume()) # type: ignore - - def _find_longest_prefix_key( - self, - key: Tuple[int, ...], - ) -> Optional[Tuple[int, ...]]: - min_len = 0 - min_key: Optional[Tuple[int, ...]] = None - for k in self.cache.iterkeys(): # type: ignore - prefix_len = Llama.longest_token_prefix(k, key) - if prefix_len > min_len: - min_len = prefix_len - min_key = k # type: ignore - return min_key - - def __getitem__(self, key: Sequence[int]) -> "LlamaState": - key = tuple(key) - _key = self._find_longest_prefix_key(key) - if _key is None: - raise KeyError("Key not found") - value: "LlamaState" = self.cache.pop(_key) # type: ignore - # NOTE: This puts an integer as key in cache, which breaks, - # Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens - # self.cache.push(_key, side="front") # type: ignore - return value - - def __contains__(self, key: Sequence[int]) -> bool: - return self._find_longest_prefix_key(tuple(key)) is not None - - def __setitem__(self, key: Sequence[int], value: "LlamaState"): - print("LlamaDiskCache.__setitem__: called", file=sys.stderr) - key = tuple(key) - if key in self.cache: - print("LlamaDiskCache.__setitem__: delete", file=sys.stderr) - del self.cache[key] - self.cache[key] = value - print("LlamaDiskCache.__setitem__: set", file=sys.stderr) - while self.cache_size > self.capacity_bytes and len(self.cache) > 0: - key_to_remove = next(iter(self.cache)) - del self.cache[key_to_remove] - print("LlamaDiskCache.__setitem__: trim", file=sys.stderr) - - class LlamaState: def __init__( self, diff --git a/llama_cpp/llama_cache.py b/llama_cpp/llama_cache.py new file mode 100644 index 000000000..9e9870a52 --- /dev/null +++ b/llama_cpp/llama_cache.py @@ -0,0 +1,150 @@ +import sys +from abc import ABC, abstractmethod +from typing import ( + Optional, + Sequence, + Tuple, +) +from collections import OrderedDict + +import diskcache + +import llama_cpp.llama + +from .llama_types import * + + +class BaseLlamaCache(ABC): + """Base cache class for a llama.cpp model.""" + + def __init__(self, capacity_bytes: int = (2 << 30)): + self.capacity_bytes = capacity_bytes + + @property + @abstractmethod + def cache_size(self) -> int: + raise NotImplementedError + + def _find_longest_prefix_key( + self, + key: Tuple[int, ...], + ) -> Optional[Tuple[int, ...]]: + pass + + @abstractmethod + def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState": + raise NotImplementedError + + @abstractmethod + def __contains__(self, key: Sequence[int]) -> bool: + raise NotImplementedError + + @abstractmethod + def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState") -> None: + raise NotImplementedError + + +class LlamaRAMCache(BaseLlamaCache): + """Cache for a llama.cpp model using RAM.""" + + def __init__(self, capacity_bytes: int = (2 << 30)): + super().__init__(capacity_bytes) + self.capacity_bytes = capacity_bytes + self.cache_state: OrderedDict[Tuple[int, ...], "llama_cpp.llama.LlamaState"] = OrderedDict() + + @property + def cache_size(self): + return sum([state.llama_state_size for state in self.cache_state.values()]) + + def _find_longest_prefix_key( + self, + key: Tuple[int, ...], + ) -> Optional[Tuple[int, ...]]: + min_len = 0 + min_key = None + keys = ( + (k, llama_cpp.llama.Llama.longest_token_prefix(k, key)) for k in self.cache_state.keys() + ) + for k, prefix_len in keys: + if prefix_len > min_len: + min_len = prefix_len + min_key = k + return min_key + + def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState": + key = tuple(key) + _key = self._find_longest_prefix_key(key) + if _key is None: + raise KeyError("Key not found") + value = self.cache_state[_key] + self.cache_state.move_to_end(_key) + return value + + def __contains__(self, key: Sequence[int]) -> bool: + return self._find_longest_prefix_key(tuple(key)) is not None + + def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"): + key = tuple(key) + if key in self.cache_state: + del self.cache_state[key] + self.cache_state[key] = value + while self.cache_size > self.capacity_bytes and len(self.cache_state) > 0: + self.cache_state.popitem(last=False) + + +# Alias for backwards compatibility +LlamaCache = LlamaRAMCache + + +class LlamaDiskCache(BaseLlamaCache): + """Cache for a llama.cpp model using disk.""" + + def __init__( + self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30) + ): + super().__init__(capacity_bytes) + self.cache = diskcache.Cache(cache_dir) + + @property + def cache_size(self): + return int(self.cache.volume()) # type: ignore + + def _find_longest_prefix_key( + self, + key: Tuple[int, ...], + ) -> Optional[Tuple[int, ...]]: + min_len = 0 + min_key: Optional[Tuple[int, ...]] = None + for k in self.cache.iterkeys(): # type: ignore + prefix_len = llama_cpp.llama.Llama.longest_token_prefix(k, key) + if prefix_len > min_len: + min_len = prefix_len + min_key = k # type: ignore + return min_key + + def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState": + key = tuple(key) + _key = self._find_longest_prefix_key(key) + if _key is None: + raise KeyError("Key not found") + value: "llama_cpp.llama.LlamaState" = self.cache.pop(_key) # type: ignore + # NOTE: This puts an integer as key in cache, which breaks, + # Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens + # self.cache.push(_key, side="front") # type: ignore + return value + + def __contains__(self, key: Sequence[int]) -> bool: + return self._find_longest_prefix_key(tuple(key)) is not None + + def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"): + print("LlamaDiskCache.__setitem__: called", file=sys.stderr) + key = tuple(key) + if key in self.cache: + print("LlamaDiskCache.__setitem__: delete", file=sys.stderr) + del self.cache[key] + self.cache[key] = value + print("LlamaDiskCache.__setitem__: set", file=sys.stderr) + while self.cache_size > self.capacity_bytes and len(self.cache) > 0: + key_to_remove = next(iter(self.cache)) + del self.cache[key_to_remove] + print("LlamaDiskCache.__setitem__: trim", file=sys.stderr) From cc4630e66f2a7f12c3287aece85779ab499c1c9d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Jan 2024 09:14:00 -0500 Subject: [PATCH 029/624] Move helper classes to _internals submodule --- llama_cpp/_internals.py | 770 ++++++++++++++++++++++++++++++++++++++++ llama_cpp/llama.py | 518 +-------------------------- 2 files changed, 776 insertions(+), 512 deletions(-) create mode 100644 llama_cpp/_internals.py diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py new file mode 100644 index 000000000..208de8c2a --- /dev/null +++ b/llama_cpp/_internals.py @@ -0,0 +1,770 @@ +from __future__ import annotations + +import os +import ctypes + +from typing import ( + List, + Optional, + Sequence, +) +from dataclasses import dataclass, field + +import numpy as np +import numpy.typing as npt + +from .llama_types import * +from .llama_grammar import LlamaGrammar + +import llama_cpp.llama_cpp as llama_cpp + +from ._utils import suppress_stdout_stderr + + +# Python wrappers over llama.h structs + + +class _LlamaModel: + """Intermediate Python wrapper for a llama.cpp llama_model. + NOTE: For stability it's recommended you use the Llama class instead.""" + + _llama_free_model = None + # NOTE: this must be "saved" here to avoid exceptions when calling __del__ + _suppress_stdout_stderr = suppress_stdout_stderr + + def __init__( + self, + *, + path_model: str, + params: llama_cpp.llama_model_params, + verbose: bool = True, + ): + self.path_model = path_model + self.params = params + self.verbose = verbose + + self._llama_free_model = llama_cpp._lib.llama_free_model # type: ignore + + if not os.path.exists(path_model): + raise ValueError(f"Model path does not exist: {path_model}") + + with self._suppress_stdout_stderr(disable=self.verbose): + self.model = llama_cpp.llama_load_model_from_file( + self.path_model.encode("utf-8"), self.params + ) + + def __del__(self): + with self._suppress_stdout_stderr(disable=self.verbose): + if self.model is not None and self._llama_free_model is not None: + self._llama_free_model(self.model) + self.model = None + + def vocab_type(self) -> int: + assert self.model is not None + return llama_cpp.llama_vocab_type(self.model) + + def n_vocab(self) -> int: + assert self.model is not None + return llama_cpp.llama_n_vocab(self.model) + + def n_ctx_train(self) -> int: + assert self.model is not None + return llama_cpp.llama_n_ctx_train(self.model) + + def n_embd(self) -> int: + assert self.model is not None + return llama_cpp.llama_n_embd(self.model) + + def rope_freq_scale_train(self) -> float: + assert self.model is not None + return llama_cpp.llama_rope_freq_scale_train(self.model) + + def desc(self) -> str: + assert self.model is not None + buf = ctypes.create_string_buffer(1024) + llama_cpp.llama_model_desc(self.model, buf, 1024) # type: ignore + return buf.value.decode("utf-8") + + def size(self) -> int: + assert self.model is not None + return llama_cpp.llama_model_size(self.model) + + def n_params(self) -> int: + assert self.model is not None + return llama_cpp.llama_model_n_params(self.model) + + def get_tensor(self, name: str) -> ctypes.c_void_p: + assert self.model is not None + return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) + + def apply_lora_from_file( + self, + lora_path: str, + scale: float, + path_base_model: Optional[str], + n_threads: int, + ): + assert self.model is not None + return llama_cpp.llama_model_apply_lora_from_file( + self.model, + lora_path.encode("utf-8"), + scale, + path_base_model.encode("utf-8") + if path_base_model is not None + else llama_cpp.c_char_p(0), + n_threads, + ) + + # Vocab + + def token_get_text(self, token: int) -> str: + # TODO: Fix + assert self.model is not None + return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8") + + def token_get_score(self, token: int) -> float: + assert self.model is not None + return llama_cpp.llama_token_get_score(self.model, token) + + def token_get_type(self, token: int) -> int: + assert self.model is not None + return llama_cpp.llama_token_get_type(self.model, token) + + # Special tokens + + def token_bos(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_bos(self.model) + + def token_eos(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_eos(self.model) + + def token_nl(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_nl(self.model) + + def token_prefix(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_prefix(self.model) + + def token_middle(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_middle(self.model) + + def token_suffix(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_suffix(self.model) + + def token_eot(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_eot(self.model) + + # Tokenization + + def tokenize(self, text: bytes, add_bos: bool, special: bool): + assert self.model is not None + n_ctx = self.n_ctx_train() + tokens = (llama_cpp.llama_token * n_ctx)() + n_tokens = llama_cpp.llama_tokenize( + self.model, text, len(text), tokens, n_ctx, add_bos, special + ) + if n_tokens < 0: + n_tokens = abs(n_tokens) + tokens = (llama_cpp.llama_token * n_tokens)() + n_tokens = llama_cpp.llama_tokenize( + self.model, text, len(text), tokens, n_tokens, add_bos, special + ) + if n_tokens < 0: + raise RuntimeError( + f'Failed to tokenize: text="{text}" n_tokens={n_tokens}' + ) + return list(tokens[:n_tokens]) + + def token_to_piece(self, token: int) -> bytes: + assert self.model is not None + buf = ctypes.create_string_buffer(32) + llama_cpp.llama_token_to_piece(self.model, token, buf, 32) # type: ignore + return bytes(buf) + + def detokenize(self, tokens: List[int]) -> bytes: + assert self.model is not None + output = b"" + size = 32 + buffer = (ctypes.c_char * size)() + for token in tokens: + n = llama_cpp.llama_token_to_piece( + self.model, llama_cpp.llama_token(token), buffer, size + ) + assert n <= size + output += bytes(buffer[:n]) + # NOTE: Llama1 models automatically added a space at the start of the prompt + # this line removes a leading space if the first token is a beginning of sentence token + return ( + output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output + ) + + @staticmethod + def default_params(): + """Get the default llama_model_params.""" + return llama_cpp.llama_model_default_params() + + +class _LlamaContext: + """Intermediate Python wrapper for a llama.cpp llama_context. + NOTE: For stability it's recommended you use the Llama class instead.""" + + _llama_free = None + # NOTE: this must be "saved" here to avoid exceptions when calling __del__ + _suppress_stdout_stderr = suppress_stdout_stderr + + def __init__( + self, + *, + model: _LlamaModel, + params: llama_cpp.llama_context_params, + verbose: bool = True, + ): + self.model = model + self.params = params + self.verbose = verbose + + self._llama_free = llama_cpp._lib.llama_free # type: ignore + + with self._suppress_stdout_stderr(disable=self.verbose): + self.ctx = llama_cpp.llama_new_context_with_model( + self.model.model, self.params + ) + + def __del__(self): + with self._suppress_stdout_stderr(disable=self.verbose): + if self.ctx is not None and self._llama_free is not None: + self._llama_free(self.ctx) + self.ctx = None + + def n_ctx(self) -> int: + assert self.ctx is not None + return llama_cpp.llama_n_ctx(self.ctx) + + def kv_cache_clear(self): + assert self.ctx is not None + llama_cpp.llama_kv_cache_clear(self.ctx) + + def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int): + assert self.ctx is not None + llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1) + + def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int): + assert self.ctx is not None + llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1) + + def kv_cache_seq_keep(self, seq_id: int): + assert self.ctx is not None + llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id) + + def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): + assert self.ctx is not None + llama_cpp.llama_kv_cache_seq_shift(self.ctx, seq_id, p0, p1, shift) + + def get_state_size(self) -> int: + assert self.ctx is not None + return llama_cpp.llama_get_state_size(self.ctx) + + # TODO: copy_state_data + + # TODO: set_state_data + + # TODO: llama_load_session_file + + # TODO: llama_save_session_file + + def decode(self, batch: "_LlamaBatch"): + assert self.ctx is not None + assert batch.batch is not None + return_code = llama_cpp.llama_decode( + ctx=self.ctx, + batch=batch.batch, + ) + if return_code != 0: + raise RuntimeError(f"llama_decode returned {return_code}") + + def set_n_threads(self, n_threads: int, n_threads_batch: int): + assert self.ctx is not None + llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch) + + def get_logits(self): + assert self.ctx is not None + return llama_cpp.llama_get_logits(self.ctx) + + def get_logits_ith(self, i: int): + assert self.ctx is not None + return llama_cpp.llama_get_logits_ith(self.ctx, i) + + def get_embeddings(self): + assert self.ctx is not None + return llama_cpp.llama_get_embeddings(self.ctx) + + # Sampling functions + + def set_rng_seed(self, seed: int): + assert self.ctx is not None + llama_cpp.llama_set_rng_seed(self.ctx, seed) + + def sample_repetition_penalties( + self, + candidates: "_LlamaTokenDataArray", + last_tokens_data: "llama_cpp.Array[llama_cpp.llama_token]", + penalty_last_n: int, + penalty_repeat: float, + penalty_freq: float, + penalty_present: float, + ): + assert self.ctx is not None + llama_cpp.llama_sample_repetition_penalties( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + last_tokens_data, + penalty_last_n, + penalty_repeat, + penalty_freq, + penalty_present, + ) + + def sample_classifier_free_guidance( + self, + candidates: "_LlamaTokenDataArray", + guidance_ctx: "_LlamaContext", + scale: float, + ): + assert self.ctx is not None + assert guidance_ctx.ctx is not None + llama_cpp.llama_sample_classifier_free_guidance( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + guidance_ctx.ctx, + scale, + ) + + def sample_softmax(self, candidates: "_LlamaTokenDataArray"): + assert self.ctx is not None + llama_cpp.llama_sample_softmax( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + ) + + def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): + assert self.ctx is not None + llama_cpp.llama_sample_top_k( + self.ctx, ctypes.byref(candidates.candidates), k, min_keep # type: ignore + ) + + def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): + assert self.ctx is not None + llama_cpp.llama_sample_top_p( + self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore + ) + + def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): + assert self.ctx is not None + llama_cpp.llama_sample_min_p( + self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore + ) + + def sample_tail_free( + self, candidates: "_LlamaTokenDataArray", z: float, min_keep: int + ): + assert self.ctx is not None + llama_cpp.llama_sample_tail_free( + self.ctx, ctypes.byref(candidates.candidates), z, min_keep # type: ignore + ) + + def sample_typical( + self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int + ): + assert self.ctx is not None + llama_cpp.llama_sample_typical( + self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore + ) + + def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): + assert self.ctx is not None + llama_cpp.llama_sample_temp( + self.ctx, ctypes.byref(candidates.candidates), temp # type: ignore + ) + + def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): + assert self.ctx is not None + assert grammar.grammar is not None + llama_cpp.llama_sample_grammar( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + grammar.grammar, + ) + + def sample_token_mirostat( + self, + candidates: "_LlamaTokenDataArray", + tau: float, + eta: float, + m: int, + mu: ctypes._Pointer[ctypes.c_float], # type: ignore + ) -> int: + assert self.ctx is not None + return llama_cpp.llama_sample_token_mirostat( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + tau, + eta, + m, + mu, + ) + + def sample_token_mirostat_v2( + self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: ctypes._Pointer[ctypes.c_float] # type: ignore + ) -> int: + assert self.ctx is not None + return llama_cpp.llama_sample_token_mirostat_v2( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + tau, + eta, + mu, + ) + + def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: + assert self.ctx is not None + return llama_cpp.llama_sample_token_greedy( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + ) + + def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: + assert self.ctx is not None + return llama_cpp.llama_sample_token( + self.ctx, + ctypes.byref(candidates.candidates), # type: ignore + ) + + # Grammar + def grammar_accept_token(self, grammar: LlamaGrammar, token: int): + assert self.ctx is not None + assert grammar.grammar is not None + llama_cpp.llama_grammar_accept_token(self.ctx, grammar.grammar, token) + + def reset_timings(self): + assert self.ctx is not None + llama_cpp.llama_reset_timings(self.ctx) + + def print_timings(self): + assert self.ctx is not None + llama_cpp.llama_print_timings(self.ctx) + + # Utility functions + @staticmethod + def default_params(): + """Get the default llama_context_params.""" + return llama_cpp.llama_context_default_params() + + +class _LlamaBatch: + _llama_batch_free = None + # NOTE: this must be "saved" here to avoid exceptions when calling __del__ + _suppress_stdout_stderr = suppress_stdout_stderr + + def __init__( + self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True + ): + self.n_tokens = n_tokens + self.embd = embd + self.n_seq_max = n_seq_max + self.verbose = verbose + + self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore + + with self._suppress_stdout_stderr(disable=self.verbose): + self.batch = llama_cpp.llama_batch_init( + self.n_tokens, self.embd, self.n_seq_max + ) + + def __del__(self): + with self._suppress_stdout_stderr(disable=self.verbose): + if self.batch is not None and self._llama_batch_free is not None: + self._llama_batch_free(self.batch) + self.batch = None + + def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): + assert self.batch is not None + n_tokens = len(batch) + self.batch.n_tokens = n_tokens + for i in range(n_tokens): + self.batch.token[i] = batch[i] + self.batch.pos[i] = n_past + i + self.batch.seq_id[i][0] = 0 + self.batch.n_seq_id[i] = 1 + self.batch.logits[i] = logits_all + self.batch.logits[n_tokens - 1] = True + + +class _LlamaTokenDataArray: + def __init__(self, *, n_vocab: int): + self.n_vocab = n_vocab + self.candidates_data = np.array( + [], + dtype=np.dtype( + [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True + ), + ) + self.candidates_data.resize(3, self.n_vocab, refcheck=False) + self.candidates = llama_cpp.llama_token_data_array( + data=self.candidates_data.ctypes.data_as(llama_cpp.llama_token_data_p), + size=self.n_vocab, + sorted=False, + ) + self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) + self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) + + def copy_logits(self, logits: npt.NDArray[np.single]): + self.candidates_data["id"][:] = self.default_candidates_data_id + self.candidates_data["logit"][:] = logits + self.candidates_data["p"][:] = self.default_candidates_data_p + self.candidates.data = self.candidates_data.ctypes.data_as( + llama_cpp.llama_token_data_p + ) + self.candidates.sorted = llama_cpp.c_bool(False) + self.candidates.size = llama_cpp.c_size_t(self.n_vocab) + + +# Python wrappers over common/common +def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> list[int]: + n_tokens = len(text) + 1 if add_bos else len(text) + result = (llama_cpp.llama_token * n_tokens)() + n_tokens = llama_cpp.llama_tokenize( + model.model, + text.encode("utf-8"), + len(text), + result, + n_tokens, + add_bos, + special, + ) + if n_tokens < 0: + result = (llama_cpp.llama_token * -n_tokens)() + check = llama_cpp.llama_tokenize( + model.model, + text.encode("utf-8"), + len(text), + result, + len(result), + add_bos, + special, + ) + if check != -n_tokens: + raise RuntimeError(f'Failed to tokenize: text="{text}" n_tokens={n_tokens}') + else: + result = result[:n_tokens] + return list(result) + + +def _token_to_piece(model: _LlamaModel, token: int) -> str: + assert model.model is not None + result = (ctypes.c_char * 8)(0) + n_tokens = llama_cpp.llama_token_to_piece(model.model, token, result, len(result)) + if n_tokens < 0: + result = (ctypes.c_char * -n_tokens)(0) + check = llama_cpp.llama_token_to_piece(model.model, token, result, len(result)) + if check != -n_tokens: + raise RuntimeError(f"Failed to get piece: token={token}") + else: + result = result[:n_tokens] + return bytes(result).decode("utf-8") + + +def _detokenize_spm(model: _LlamaModel, tokens: List[int]) -> str: + bos_id = model.token_bos() + result = "" + for i, token in enumerate(tokens): + piece = _token_to_piece(model, token) + if ( + (tokens[0] == bos_id and i == 1) or (tokens[0] != bos_id and i == 0) + ) and piece[0] == " ": + piece = piece[1:] + result += piece + return result + + +def _detokenize_bpe(model: _LlamaModel, tokens: List[int]) -> str: + result = "" + for token in tokens: + piece = _token_to_piece(model, token) + result += piece + return result + + +def _should_add_bos(model: _LlamaModel) -> bool: + assert model.model is not None + add_bos = llama_cpp.llama_add_bos_token(model.model) + if add_bos != -1: + return add_bos != 0 + else: + return llama_cpp.llama_vocab_type(model.model) == llama_cpp.LLAMA_VOCAB_TYPE_SPM + + +# Python wrappers over common/sampling structs + + +@dataclass +class _LlamaSamplingParams: + n_prev: int = 64 + n_probs: int = 0 + top_k: int = 40 + top_p: float = 0.95 + min_p: float = 0.05 + tfs_z: float = 1.00 + typical_p: float = 1.00 + temp: float = 0.80 + penalty_last_n: int = 64 + penalty_repeat: float = 1.10 + penalty_freq: float = 0.00 + penalty_present: float = 0.00 + mirostat: int = 0 + mirostat_tau: float = 5.00 + mirostat_eta: float = 0.10 + penalize_nl: bool = True + + grammar: str = "" + + cfg_negative_prompt: str = "" + cfg_scale: float = 1.00 + + logit_bias: dict[int, float] = field(default_factory=dict) + + +@dataclass +class _LlamaSamplingContext: + params: _LlamaSamplingParams = field(default_factory=_LlamaSamplingParams) + mirostat_mu: ctypes.c_float = field(default_factory=ctypes.c_float) + grammar: Optional[LlamaGrammar] = None + # NOTE: Missing parsed_grammar + prev: list[int] = field(default_factory=list) + cur: list[llama_cpp.llama_token_data] = field(default_factory=list) + + def reset(self): + self.prev = [] + self.cur = [] + if self.grammar is not None: + self.grammar.reset() + + def cp(self): + return _LlamaSamplingContext( + params=self.params, + mirostat_mu=self.mirostat_mu, + grammar=self.grammar, + prev=self.prev.copy(), + cur=self.cur.copy(), + ) + + def last(self) -> Optional[int]: + if len(self.prev) > 0: + return self.prev[-1] + else: + return None + + def prev_str(self, ctx_main: _LlamaContext, n: int) -> str: + return ctx_main.model.detokenize(self.prev[-n:]).decode("utf-8") + + def sample( + self, ctx_main: _LlamaContext, ctx_cfg: Optional[_LlamaContext] = None, idx: int = 0, logits_array: Optional[npt.NDArray[np.single]] = None + ): + n_vocab = ctx_main.model.n_vocab() + id: int = 0 + + if logits_array is None: + logits = ctx_main.get_logits_ith(idx) + logits_array = np.array( + ctypes.cast(logits, ctypes.POINTER(ctypes.c_float * n_vocab)).contents, + dtype=np.single, + ) + + # apply logit_bias + for token, logit_bias in self.params.logit_bias.items(): + logits_array[token] += logit_bias + + token_data_array = _LlamaTokenDataArray( + n_vocab=n_vocab + ) # TODO: Only create this once + token_data_array.copy_logits(logits_array) + + if ctx_cfg is not None: + ctx_main.sample_classifier_free_guidance( + token_data_array, ctx_cfg, self.params.cfg_scale + ) + + # apply penalties + if len(self.prev) > 0: + nl_token = ctx_main.model.token_nl() + nl_logit = logits_array[nl_token] + if self.params.penalty_last_n > 0: + ctx_main.sample_repetition_penalties( + token_data_array, + # TODO: Only create this once + (llama_cpp.llama_token * len(self.prev))(*self.prev), # type: ignore + self.params.penalty_last_n, + self.params.penalty_repeat, + self.params.penalty_freq, + self.params.penalty_present, + ) + if not self.params.penalize_nl: + token_data_array.candidates_data["logit"][nl_token] = nl_logit + + if self.grammar is not None: + ctx_main.sample_grammar(token_data_array, self.grammar) + + if self.params.temp < 0: + ctx_main.sample_softmax(token_data_array) + id = token_data_array.candidates_data["id"][0] + elif self.params.temp == 0: + id = ctx_main.sample_token_greedy(token_data_array) + else: + if self.params.mirostat == 1: + mirostat_m = 100 + ctx_main.sample_temp(token_data_array, self.params.temp) + id = ctx_main.sample_token_mirostat( + token_data_array, + self.params.mirostat_tau, + self.params.mirostat_eta, + mirostat_m, + ctypes.pointer(self.mirostat_mu), + ) + elif self.params.mirostat == 2: + ctx_main.sample_temp(token_data_array, self.params.temp) + id = ctx_main.sample_token_mirostat_v2( + token_data_array, + self.params.mirostat_tau, + self.params.mirostat_eta, + ctypes.pointer(self.mirostat_mu), + ) + else: + min_keep = max(1, self.params.n_probs) + ctx_main.sample_top_k( + token_data_array, self.params.top_k, min_keep=min_keep + ) + ctx_main.sample_tail_free( + token_data_array, self.params.tfs_z, min_keep=min_keep + ) + ctx_main.sample_typical( + token_data_array, self.params.typical_p, min_keep=min_keep + ) + ctx_main.sample_top_p( + token_data_array, self.params.top_p, min_keep=min_keep + ) + ctx_main.sample_min_p( + token_data_array, self.params.min_p, min_keep=min_keep + ) + ctx_main.sample_temp(token_data_array, self.params.temp) + id = ctx_main.sample_token(token_data_array) + return id + + def accept(self, ctx_main: _LlamaContext, id: int, apply_grammar: bool): + if apply_grammar and self.grammar is not None: + ctx_main.grammar_accept_token(self.grammar, id) + self.prev.append(id) \ No newline at end of file diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index c6b55abae..f4e5dcd2b 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -32,6 +32,12 @@ import numpy.typing as npt from ._utils import suppress_stdout_stderr +from ._internals import ( + _LlamaModel, # type: ignore + _LlamaContext, # type: ignore + _LlamaBatch, # type: ignore + _LlamaTokenDataArray, # type: ignore +) class LlamaState: @@ -74,518 +80,6 @@ def __call__( return any([stopping_criteria(input_ids, logits) for stopping_criteria in self]) -class _LlamaModel: - """Intermediate Python wrapper for a llama.cpp llama_model. - - NOTE: For stability it's recommended you use the Llama class instead.""" - - _llama_free_model = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - suppress_stdout_stderr = suppress_stdout_stderr - - def __init__( - self, - *, - path_model: str, - params: llama_cpp.llama_model_params, - verbose: bool = True, - ): - self.path_model = path_model - self.params = params - self.verbose = verbose - - self._llama_free_model = llama_cpp._lib.llama_free_model # type: ignore - - if not os.path.exists(path_model): - raise ValueError(f"Model path does not exist: {path_model}") - - with suppress_stdout_stderr(disable=self.verbose): - self.model = llama_cpp.llama_load_model_from_file( - self.path_model.encode("utf-8"), self.params - ) - - def __del__(self): - with self.suppress_stdout_stderr(disable=self.verbose): - if self.model is not None and self._llama_free_model is not None: - self._llama_free_model(self.model) - self.model = None - - def vocab_type(self) -> int: - assert self.model is not None - return llama_cpp.llama_vocab_type(self.model) - - def n_vocab(self) -> int: - assert self.model is not None - return llama_cpp.llama_n_vocab(self.model) - - def n_ctx_train(self) -> int: - assert self.model is not None - return llama_cpp.llama_n_ctx_train(self.model) - - def n_embd(self) -> int: - assert self.model is not None - return llama_cpp.llama_n_embd(self.model) - - def rope_freq_scale_train(self) -> float: - assert self.model is not None - return llama_cpp.llama_rope_freq_scale_train(self.model) - - def desc(self) -> str: - assert self.model is not None - buf = ctypes.create_string_buffer(1024) - llama_cpp.llama_model_desc(self.model, buf, 1024) # type: ignore - return buf.value.decode("utf-8") - - def size(self) -> int: - assert self.model is not None - return llama_cpp.llama_model_size(self.model) - - def n_params(self) -> int: - assert self.model is not None - return llama_cpp.llama_model_n_params(self.model) - - def get_tensor(self, name: str) -> ctypes.c_void_p: - assert self.model is not None - return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) - - def apply_lora_from_file( - self, - lora_path: str, - scale: float, - path_base_model: Optional[str], - n_threads: int, - ): - assert self.model is not None - return llama_cpp.llama_model_apply_lora_from_file( - self.model, - lora_path.encode("utf-8"), - scale, - path_base_model.encode("utf-8") - if path_base_model is not None - else llama_cpp.c_char_p(0), - n_threads, - ) - - # Vocab - - def token_get_text(self, token: int) -> str: - # TODO: Fix - assert self.model is not None - return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8") - - def token_get_score(self, token: int) -> float: - assert self.model is not None - return llama_cpp.llama_token_get_score(self.model, token) - - def token_get_type(self, token: int) -> int: - assert self.model is not None - return llama_cpp.llama_token_get_type(self.model, token) - - # Special tokens - - def token_bos(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_bos(self.model) - - def token_eos(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_eos(self.model) - - def token_nl(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_nl(self.model) - - def token_prefix(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_prefix(self.model) - - def token_middle(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_middle(self.model) - - def token_suffix(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_suffix(self.model) - - def token_eot(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_eot(self.model) - - # Tokenization - - def tokenize(self, text: bytes, add_bos: bool, special: bool): - assert self.model is not None - n_ctx = self.n_ctx_train() - tokens = (llama_cpp.llama_token * n_ctx)() - n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_ctx, add_bos, special - ) - if n_tokens < 0: - n_tokens = abs(n_tokens) - tokens = (llama_cpp.llama_token * n_tokens)() - n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_tokens, add_bos, special - ) - if n_tokens < 0: - raise RuntimeError( - f'Failed to tokenize: text="{text}" n_tokens={n_tokens}' - ) - return list(tokens[:n_tokens]) - - def token_to_piece(self, token: int) -> bytes: - assert self.model is not None - buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32) # type: ignore - return bytes(buf) - - def detokenize(self, tokens: List[int]) -> bytes: - assert self.model is not None - output = b"" - size = 32 - buffer = (ctypes.c_char * size)() - for token in tokens: - n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token), buffer, size - ) - assert n <= size - output += bytes(buffer[:n]) - # NOTE: Llama1 models automatically added a space at the start of the prompt - # this line removes a leading space if the first token is a beginning of sentence token - return ( - output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output - ) - - @staticmethod - def default_params(): - """Get the default llama_model_params.""" - return llama_cpp.llama_model_default_params() - - -class _LlamaContext: - """Intermediate Python wrapper for a llama.cpp llama_context. - - NOTE: For stability it's recommended you use the Llama class instead.""" - - _llama_free = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - suppress_stdout_stderr = suppress_stdout_stderr - - def __init__( - self, - *, - model: _LlamaModel, - params: llama_cpp.llama_context_params, - verbose: bool = True, - ): - self.model = model - self.params = params - self.verbose = verbose - - self._llama_free = llama_cpp._lib.llama_free # type: ignore - - with suppress_stdout_stderr(disable=self.verbose): - self.ctx = llama_cpp.llama_new_context_with_model( - self.model.model, self.params - ) - - def __del__(self): - with self.suppress_stdout_stderr(disable=self.verbose): - if self.ctx is not None and self._llama_free is not None: - self._llama_free(self.ctx) - self.ctx = None - - def n_ctx(self) -> int: - assert self.ctx is not None - return llama_cpp.llama_n_ctx(self.ctx) - - def kv_cache_clear(self): - assert self.ctx is not None - llama_cpp.llama_kv_cache_clear(self.ctx) - - def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1) - - def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1) - - def kv_cache_seq_keep(self, seq_id: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id) - - def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_shift(self.ctx, seq_id, p0, p1, shift) - - def get_state_size(self) -> int: - assert self.ctx is not None - return llama_cpp.llama_get_state_size(self.ctx) - - # TODO: copy_state_data - - # TODO: set_state_data - - # TODO: llama_load_session_file - - # TODO: llama_save_session_file - - def decode(self, batch: "_LlamaBatch"): - assert self.ctx is not None - assert batch.batch is not None - return_code = llama_cpp.llama_decode( - ctx=self.ctx, - batch=batch.batch, - ) - if return_code != 0: - raise RuntimeError(f"llama_decode returned {return_code}") - - def set_n_threads(self, n_threads: int, n_threads_batch: int): - assert self.ctx is not None - llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch) - - def get_logits(self): - assert self.ctx is not None - return llama_cpp.llama_get_logits(self.ctx) - - def get_logits_ith(self, i: int): - assert self.ctx is not None - return llama_cpp.llama_get_logits_ith(self.ctx, i) - - def get_embeddings(self): - assert self.ctx is not None - return llama_cpp.llama_get_embeddings(self.ctx) - - # Sampling functions - - def set_rng_seed(self, seed: int): - assert self.ctx is not None - llama_cpp.llama_set_rng_seed(self.ctx, seed) - - def sample_repetition_penalties( - self, - candidates: "_LlamaTokenDataArray", - last_tokens_data: "llama_cpp.Array[llama_cpp.llama_token]", - penalty_last_n: int, - penalty_repeat: float, - penalty_freq: float, - penalty_present: float, - ): - assert self.ctx is not None - llama_cpp.llama_sample_repetition_penalties( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - last_tokens_data, - penalty_last_n, - penalty_repeat, - penalty_freq, - penalty_present, - ) - - def sample_classifier_free_guidance( - self, - candidates: "_LlamaTokenDataArray", - guidance_ctx: "_LlamaContext", - scale: float, - ): - assert self.ctx is not None - assert guidance_ctx.ctx is not None - llama_cpp.llama_sample_classifier_free_guidance( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - guidance_ctx.ctx, - scale, - ) - - def sample_softmax(self, candidates: "_LlamaTokenDataArray"): - assert self.ctx is not None - llama_cpp.llama_sample_softmax( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - ) - - def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): - assert self.ctx is not None - llama_cpp.llama_sample_top_k( - self.ctx, ctypes.byref(candidates.candidates), k, min_keep # type: ignore - ) - - def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - assert self.ctx is not None - llama_cpp.llama_sample_top_p( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore - ) - - def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - assert self.ctx is not None - llama_cpp.llama_sample_min_p( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore - ) - - def sample_tail_free( - self, candidates: "_LlamaTokenDataArray", z: float, min_keep: int - ): - assert self.ctx is not None - llama_cpp.llama_sample_tail_free( - self.ctx, ctypes.byref(candidates.candidates), z, min_keep # type: ignore - ) - - def sample_typical( - self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int - ): - assert self.ctx is not None - llama_cpp.llama_sample_typical( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore - ) - - def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): - assert self.ctx is not None - llama_cpp.llama_sample_temp( - self.ctx, ctypes.byref(candidates.candidates), temp # type: ignore - ) - - def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): - assert self.ctx is not None - assert grammar.grammar is not None - llama_cpp.llama_sample_grammar( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - grammar.grammar, - ) - - def sample_token_mirostat( - self, - candidates: "_LlamaTokenDataArray", - tau: float, - eta: float, - m: int, - mu: float, - ) -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token_mirostat( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - tau, - eta, - m, - ctypes.pointer(ctypes.c_float(mu)), - ) - - def sample_token_mirostat_v2( - self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: float - ) -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token_mirostat_v2( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - tau, - eta, - ctypes.pointer(ctypes.c_float(mu)), - ) - - def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token_greedy( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - ) - - def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - ) - - # Grammar - def grammar_accept_token(self, grammar: LlamaGrammar, token: int): - assert self.ctx is not None - assert grammar.grammar is not None - llama_cpp.llama_grammar_accept_token(self.ctx, grammar.grammar, token) - - def reset_timings(self): - assert self.ctx is not None - llama_cpp.llama_reset_timings(self.ctx) - - def print_timings(self): - assert self.ctx is not None - llama_cpp.llama_print_timings(self.ctx) - - # Utility functions - @staticmethod - def default_params(): - """Get the default llama_context_params.""" - return llama_cpp.llama_context_default_params() - - -class _LlamaBatch: - _llama_batch_free = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - suppress_stdout_stderr = suppress_stdout_stderr - - def __init__( - self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True - ): - self.n_tokens = n_tokens - self.embd = embd - self.n_seq_max = n_seq_max - self.verbose = verbose - - self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore - - with suppress_stdout_stderr(disable=self.verbose): - self.batch = llama_cpp.llama_batch_init( - self.n_tokens, self.embd, self.n_seq_max - ) - - def __del__(self): - with self.suppress_stdout_stderr(disable=self.verbose): - if self.batch is not None and self._llama_batch_free is not None: - self._llama_batch_free(self.batch) - self.batch = None - - def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): - assert self.batch is not None - n_tokens = len(batch) - self.batch.n_tokens = n_tokens - for i in range(n_tokens): - self.batch.token[i] = batch[i] - self.batch.pos[i] = n_past + i - self.batch.seq_id[i][0] = 0 - self.batch.n_seq_id[i] = 1 - self.batch.logits[i] = logits_all - self.batch.logits[n_tokens - 1] = True - - -class _LlamaTokenDataArray: - def __init__(self, *, n_vocab: int): - self.n_vocab = n_vocab - self.candidates_data = np.array( - [], - dtype=np.dtype( - [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True - ), - ) - self.candidates_data.resize(3, self.n_vocab, refcheck=False) - self.candidates = llama_cpp.llama_token_data_array( - data=self.candidates_data.ctypes.data_as(llama_cpp.llama_token_data_p), - size=self.n_vocab, - sorted=False, - ) - self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) - self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) - - def copy_logits(self, logits: npt.NDArray[np.single]): - self.candidates_data["id"][:] = self.default_candidates_data_id - self.candidates_data["logit"][:] = logits - self.candidates_data["p"][:] = self.default_candidates_data_p - self.candidates.data = self.candidates_data.ctypes.data_as( - llama_cpp.llama_token_data_p - ) - self.candidates.sorted = llama_cpp.c_bool(False) - self.candidates.size = llama_cpp.c_size_t(self.n_vocab) - - class Llama: """High-level Python wrapper for a llama.cpp model.""" From 7b46bb5a786a0460d06adc053d3e494e92b75b39 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Jan 2024 09:16:13 -0500 Subject: [PATCH 030/624] Re-order classes in llama.py --- llama_cpp/llama.py | 82 ++++++++++++++++++++++++---------------------- 1 file changed, 42 insertions(+), 40 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index f4e5dcd2b..25abf36cb 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import os import sys import uuid @@ -40,46 +42,6 @@ ) -class LlamaState: - def __init__( - self, - input_ids: npt.NDArray[np.intc], - scores: npt.NDArray[np.single], - n_tokens: int, - llama_state: bytes, - llama_state_size: int, - ): - self.input_ids = input_ids - self.scores = scores - self.n_tokens = n_tokens - self.llama_state = llama_state - self.llama_state_size = llama_state_size - - -LogitsProcessor = Callable[ - [npt.NDArray[np.intc], npt.NDArray[np.single]], npt.NDArray[np.single] -] - - -class LogitsProcessorList(List[LogitsProcessor]): - def __call__( - self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single] - ) -> npt.NDArray[np.single]: - for processor in self: - scores = processor(input_ids, scores) - return scores - - -StoppingCriteria = Callable[[npt.NDArray[np.intc], npt.NDArray[np.single]], bool] - - -class StoppingCriteriaList(List[StoppingCriteria]): - def __call__( - self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single] - ) -> bool: - return any([stopping_criteria(input_ids, logits) for stopping_criteria in self]) - - class Llama: """High-level Python wrapper for a llama.cpp model.""" @@ -1733,3 +1695,43 @@ def decode(self, tokens: List[int]) -> str: @classmethod def from_ggml_file(cls, path: str) -> "LlamaTokenizer": return cls(Llama(model_path=path, vocab_only=True)) + + +class LlamaState: + def __init__( + self, + input_ids: npt.NDArray[np.intc], + scores: npt.NDArray[np.single], + n_tokens: int, + llama_state: bytes, + llama_state_size: int, + ): + self.input_ids = input_ids + self.scores = scores + self.n_tokens = n_tokens + self.llama_state = llama_state + self.llama_state_size = llama_state_size + + +LogitsProcessor = Callable[ + [npt.NDArray[np.intc], npt.NDArray[np.single]], npt.NDArray[np.single] +] + + +class LogitsProcessorList(List[LogitsProcessor]): + def __call__( + self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single] + ) -> npt.NDArray[np.single]: + for processor in self: + scores = processor(input_ids, scores) + return scores + + +StoppingCriteria = Callable[[npt.NDArray[np.intc], npt.NDArray[np.single]], bool] + + +class StoppingCriteriaList(List[StoppingCriteria]): + def __call__( + self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single] + ) -> bool: + return any([stopping_criteria(input_ids, logits) for stopping_criteria in self]) From 52adc231153624d7a1c949c97c5a0c5b3e4eabf7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Jan 2024 09:27:40 -0500 Subject: [PATCH 031/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5c9996090..4f4bf35f4 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5c999609013a30c06e6fd28be8db5c2074bcc196 +Subproject commit 4f4bf35f46600441dec2f941e667291eeb9a18d8 From 6bfe98bd801262b68c7cb4761e67a626d91534c0 Mon Sep 17 00:00:00 2001 From: Austin <77757836+teleprint-me@users.noreply.github.com> Date: Wed, 17 Jan 2024 09:47:52 -0500 Subject: [PATCH 032/624] Integration of Jinja2 Templating (#875) * feat: Add support for jinja templating Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * fix: Refactor chat formatter and update interface for jinja templates - Simplify the `llama2_template` in `llama_jinja_format.py` by removing unnecessary line breaks for readability without affecting functionality. - Update `ChatFormatterInterface` constructor to accept a more generic `Optional[object]` type for the template parameter, enhancing flexibility. - Introduce a `template` property to `ChatFormatterInterface` for standardized access to the template string. - Replace `MetaSingleton` metaclass with `Singleton` for the `ChatFormatterFactory` to streamline the singleton implementation. These changes enhance code readability, maintain usability, and ensure consistency in the chat formatter's design pattern usage. * Add outline for Jinja2 templating integration documentation Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * Add jinja2 as a dependency with version range for Hugging Face transformers compatibility Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * Update jinja2 version constraint for mkdocs-material compatibility Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * Fix attribute name in AutoChatFormatter - Changed attribute name from `self._renderer` to `self._environment` --------- Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> --- docs/templates.md | 52 ++++++++++++ llama_cpp/llama_jinja_format.py | 138 ++++++++++++++++++++++++++++++++ pyproject.toml | 4 +- tests/test_llama_chat_format.py | 50 ++++++++++++ 4 files changed, 243 insertions(+), 1 deletion(-) create mode 100644 docs/templates.md create mode 100644 llama_cpp/llama_jinja_format.py create mode 100644 tests/test_llama_chat_format.py diff --git a/docs/templates.md b/docs/templates.md new file mode 100644 index 000000000..5acdaa196 --- /dev/null +++ b/docs/templates.md @@ -0,0 +1,52 @@ +# Templates + +This document provides a comprehensive guide to the integration of Jinja2 templating into the `llama-cpp-python` project, with a focus on enhancing the chat functionality of the `llama-2` model. + +## Introduction + +- Brief explanation of the `llama-cpp-python` project's need for a templating system. +- Overview of the `llama-2` model's interaction with templating. + +## Jinja2 Dependency Integration + +- Rationale for choosing Jinja2 as the templating engine. + - Compatibility with Hugging Face's `transformers`. + - Desire for advanced templating features and simplicity. +- Detailed steps for adding `jinja2` to `pyproject.toml` for dependency management. + +## Template Management Refactor + +- Summary of the refactor and the motivation behind it. +- Description of the new chat handler selection logic: + 1. Preference for a user-specified `chat_handler`. + 2. Fallback to a user-specified `chat_format`. + 3. Defaulting to a chat format from a `.gguf` file if available. + 4. Utilizing the `llama2` default chat format as the final fallback. +- Ensuring backward compatibility throughout the refactor. + +## Implementation Details + +- In-depth look at the new `AutoChatFormatter` class. +- Example code snippets showing how to utilize the Jinja2 environment and templates. +- Guidance on how to provide custom templates or use defaults. + +## Testing and Validation + +- Outline of the testing strategy to ensure seamless integration. +- Steps for validating backward compatibility with existing implementations. + +## Benefits and Impact + +- Analysis of the expected benefits, including consistency, performance gains, and improved developer experience. +- Discussion of the potential impact on current users and contributors. + +## Future Work + +- Exploration of how templating can evolve within the project. +- Consideration of additional features or optimizations for the templating engine. +- Mechanisms for community feedback on the templating system. + +## Conclusion + +- Final thoughts on the integration of Jinja2 templating. +- Call to action for community involvement and feedback. diff --git a/llama_cpp/llama_jinja_format.py b/llama_cpp/llama_jinja_format.py new file mode 100644 index 000000000..68faaf6ca --- /dev/null +++ b/llama_cpp/llama_jinja_format.py @@ -0,0 +1,138 @@ +""" +llama_cpp/llama_jinja_format.py +""" +import dataclasses +from typing import Any, Callable, Dict, List, Optional, Protocol, Union + +import jinja2 +from jinja2 import Template + +# NOTE: We sacrifice readability for usability. +# It will fail to work as expected if we attempt to format it in a readable way. +llama2_template = """{% for message in messages %}{% if message['role'] == 'user' %}[INST] {{ message['content'] }} [/INST]\n{% elif message['role'] == 'assistant' %}{{ message['content'] }}\n{% elif message['role'] == 'system' %}<> {{ message['content'] }} <>\n{% endif %}{% endfor %}""" + + +class MetaSingleton(type): + """ + Metaclass for implementing the Singleton pattern. + """ + + _instances = {} + + def __call__(cls, *args, **kwargs): + if cls not in cls._instances: + cls._instances[cls] = super(MetaSingleton, cls).__call__(*args, **kwargs) + return cls._instances[cls] + + +class Singleton(object, metaclass=MetaSingleton): + """ + Base class for implementing the Singleton pattern. + """ + + def __init__(self): + super(Singleton, self).__init__() + + +@dataclasses.dataclass +class ChatFormatterResponse: + prompt: str + stop: Optional[Union[str, List[str]]] = None + + +# Base Chat Formatter Protocol +class ChatFormatterInterface(Protocol): + def __init__(self, template: Optional[object] = None): + ... + + def __call__( + self, + messages: List[Dict[str, str]], + **kwargs, + ) -> ChatFormatterResponse: + ... + + @property + def template(self) -> str: + ... + + +class AutoChatFormatter(ChatFormatterInterface): + def __init__( + self, + template: Optional[str] = None, + template_class: Optional[Template] = None, + ): + if template is not None: + self._template = template + else: + self._template = llama2_template # default template + + self._environment = jinja2.Environment( + loader=jinja2.BaseLoader(), + trim_blocks=True, + lstrip_blocks=True, + ).from_string( + self._template, + template_class=template_class, + ) + + def __call__( + self, + messages: List[Dict[str, str]], + **kwargs: Any, + ) -> ChatFormatterResponse: + formatted_sequence = self._environment.render(messages=messages, **kwargs) + return ChatFormatterResponse(prompt=formatted_sequence) + + @property + def template(self) -> str: + return self._template + + +class FormatterNotFoundException(Exception): + pass + + +class ChatFormatterFactory(Singleton): + _chat_formatters: Dict[str, Callable[[], ChatFormatterInterface]] = {} + + def register_formatter( + self, + name: str, + formatter_callable: Callable[[], ChatFormatterInterface], + overwrite=False, + ): + if not overwrite and name in self._chat_formatters: + raise ValueError( + f"Formatter with name '{name}' is already registered. Use `overwrite=True` to overwrite it." + ) + self._chat_formatters[name] = formatter_callable + + def unregister_formatter(self, name: str): + if name in self._chat_formatters: + del self._chat_formatters[name] + else: + raise ValueError(f"No formatter registered under the name '{name}'.") + + def get_formatter_by_name(self, name: str) -> ChatFormatterInterface: + try: + formatter_callable = self._chat_formatters[name] + return formatter_callable() + except KeyError: + raise FormatterNotFoundException( + f"Invalid chat format: {name} (valid formats: {list(self._chat_formatters.keys())})" + ) + + +# Define a chat format class +class Llama2Formatter(AutoChatFormatter): + def __init__(self): + super().__init__(llama2_template) + + +# With the Singleton pattern applied, regardless of where or how many times +# ChatFormatterFactory() is called, it will always return the same instance +# of the factory, ensuring that the factory's state is consistent throughout +# the application. +ChatFormatterFactory().register_formatter("llama-2", Llama2Formatter) diff --git a/pyproject.toml b/pyproject.toml index b5affaa9d..806127d89 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -11,10 +11,13 @@ license = { text = "MIT" } authors = [ { name = "Andrei Betlen", email = "abetlen@gmail.com" }, ] +# mkdocs-martiral requires "jinja2~=3.0" +# transformers requires "jinja2>=2.11.3" dependencies = [ "typing-extensions>=4.5.0", "numpy>=1.20.0", "diskcache>=5.6.1", + "jinja2>=2.11.3", ] requires-python = ">=3.8" classifiers = [ @@ -72,4 +75,3 @@ Changelog = "https://llama-cpp-python.readthedocs.io/en/latest/changelog/" [tool.pytest.ini_options] addopts = "--ignore=vendor" - diff --git a/tests/test_llama_chat_format.py b/tests/test_llama_chat_format.py new file mode 100644 index 000000000..4eebcb674 --- /dev/null +++ b/tests/test_llama_chat_format.py @@ -0,0 +1,50 @@ +from typing import List + +import pytest + +from llama_cpp import ChatCompletionMessage +from llama_cpp.llama_jinja_format import Llama2Formatter + + +@pytest.fixture +def sequence_of_messages() -> List[ChatCompletionMessage]: + return [ + ChatCompletionMessage(role="system", content="Welcome to CodeHelp Bot!"), + ChatCompletionMessage( + role="user", content="Hi there! I need some help with Python." + ), + ChatCompletionMessage( + role="assistant", content="Of course! What do you need help with in Python?" + ), + ChatCompletionMessage( + role="user", + content="I'm trying to write a function to find the factorial of a number, but I'm stuck.", + ), + ChatCompletionMessage( + role="assistant", + content="I can help with that! Would you like a recursive or iterative solution?", + ), + ChatCompletionMessage( + role="user", content="Let's go with a recursive solution." + ), + ] + + +def test_llama2_formatter(sequence_of_messages): + expected_prompt = ( + "<> Welcome to CodeHelp Bot! <>\n" + "[INST] Hi there! I need some help with Python. [/INST]\n" + "Of course! What do you need help with in Python?\n" + "[INST] I'm trying to write a function to find the factorial of a number, but I'm stuck. [/INST]\n" + "I can help with that! Would you like a recursive or iterative solution?\n" + "[INST] Let's go with a recursive solution. [/INST]\n" + ) + + llama2_formatter_instance = Llama2Formatter() + formatter_response = llama2_formatter_instance(sequence_of_messages) + assert ( + expected_prompt == formatter_response.prompt + ), "The formatted prompt does not match the expected output." + + +# Optionally, include a test for the 'stop' if it's part of the functionality. From 48c3b77e6f558a9899de0e1155c7dc0c7958d8e8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 11:08:57 -0500 Subject: [PATCH 033/624] Offload KQV by default --- llama_cpp/llama.py | 2 +- llama_cpp/server/settings.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 25abf36cb..6cdc1eb76 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -77,7 +77,7 @@ def __init__( mul_mat_q: bool = True, logits_all: bool = False, embedding: bool = False, - offload_kqv: bool = False, + offload_kqv: bool = True, # Sampling Params last_n_tokens_size: int = 64, # LoRA Params diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index a10390c75..dc5be209d 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -90,7 +90,7 @@ class ModelSettings(BaseSettings): logits_all: bool = Field(default=True, description="Whether to return logits.") embedding: bool = Field(default=True, description="Whether to use embeddings.") offload_kqv: bool = Field( - default=False, description="Whether to offload kqv to the GPU." + default=True, description="Whether to offload kqv to the GPU." ) # Sampling Params last_n_tokens_size: int = Field( From b8fc1c7d83ad4a9207c707ba1d954fe580286a01 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 21:21:37 -0500 Subject: [PATCH 034/624] feat: Add ability to load chat format from huggingface autotokenizer or tokenizer_config.json files. --- llama_cpp/_utils.py | 25 ++- llama_cpp/llama_chat_format.py | 375 +++++++++++++++++++++----------- llama_cpp/llama_jinja_format.py | 138 ------------ llama_cpp/server/model.py | 23 +- llama_cpp/server/settings.py | 9 + tests/test_llama_chat_format.py | 101 +++++---- 6 files changed, 355 insertions(+), 316 deletions(-) delete mode 100644 llama_cpp/llama_jinja_format.py diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index f7b6ba6b1..4a106470b 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -1,7 +1,8 @@ import os import sys -import sys, traceback +import sys +from typing import Any, Dict # Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor outnull_file = open(os.devnull, "w") @@ -55,3 +56,25 @@ def __exit__(self, *_): self.os.close(self.old_stdout_fileno) self.os.close(self.old_stderr_fileno) + + +class MetaSingleton(type): + """ + Metaclass for implementing the Singleton pattern. + """ + + _instances: Dict[type, Any] = {} + + def __call__(cls, *args: Any, **kwargs: Any) -> Any: + if cls not in cls._instances: + cls._instances[cls] = super(MetaSingleton, cls).__call__(*args, **kwargs) + return cls._instances[cls] + + +class Singleton(object, metaclass=MetaSingleton): + """ + Base class for implementing the Singleton pattern. + """ + + def __init__(self): + super(Singleton, self).__init__() diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 0ef7bd4a8..3d18d904f 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -6,18 +6,28 @@ import dataclasses from typing import Any, Dict, Iterator, List, Optional, Tuple, Union, Protocol +import jinja2 + import llama_cpp.llama as llama import llama_cpp.llama_types as llama_types import llama_cpp.llama_grammar as llama_grammar -from ._utils import suppress_stdout_stderr +from ._utils import suppress_stdout_stderr, Singleton class LlamaChatCompletionHandler(Protocol): + """Base Protocol for a llama chat completion handler. + + Very generic protocol that can be used to implement any chat format. + The only hard requirement is that it must return a ChatCompletion when + stream=False and an iterator of ChatCompletionChunks when stream=True.""" + def __call__( self, *, + # llama.cpp instance llama: llama.Llama, + # openai api parameters messages: List[llama_types.ChatCompletionRequestMessage], functions: Optional[List[llama_types.ChatCompletionFunction]] = None, function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, @@ -26,8 +36,6 @@ def __call__( temperature: float = 0.2, top_p: float = 0.95, top_k: int = 40, - min_p: float = 0.05, - typical_p: float = 1.0, stream: bool = False, stop: Optional[Union[str, List[str]]] = [], seed: Optional[int] = None, @@ -38,14 +46,17 @@ def __call__( presence_penalty: float = 0.0, frequency_penalty: float = 0.0, repeat_penalty: float = 1.1, + model: Optional[str] = None, + logit_bias: Optional[Dict[str, float]] = None, + # llama.cpp parameters + min_p: float = 0.05, + typical_p: float = 1.0, tfs_z: float = 1.0, mirostat_mode: int = 0, mirostat_tau: float = 5.0, mirostat_eta: float = 0.1, - model: Optional[str] = None, logits_processor: Optional[llama.LogitsProcessorList] = None, grammar: Optional[llama.LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, **kwargs, # type: ignore ) -> Union[ llama_types.CreateChatCompletionResponse, @@ -54,21 +65,83 @@ def __call__( ... -CHAT_HANDLERS: Dict[str, LlamaChatCompletionHandler] = {} +class LlamaChatCompletionHandlerNotFoundException(Exception): + pass + + +class LlamaChatCompletionHandlerRegistry(Singleton): + _chat_handlers: Dict[str, LlamaChatCompletionHandler] = {} + + def register_chat_completion_handler( + self, + name: str, + chat_handler: LlamaChatCompletionHandler, + overwrite: bool = False, + ): + if not overwrite and name in self._chat_handlers: + raise ValueError( + f"Formatter with name '{name}' is already registered. Use `overwrite=True` to overwrite it." + ) + self._chat_handlers[name] = chat_handler + + def unregister_chat_handler(self, name: str): + if name in self._chat_handlers: + del self._chat_handlers[name] + else: + raise ValueError(f"No formatter registered under the name '{name}'.") + + def get_chat_completion_handler_by_name( + self, name: str + ) -> LlamaChatCompletionHandler: + try: + chat_handler = self._chat_handlers[name] + return chat_handler + except KeyError: + raise LlamaChatCompletionHandlerNotFoundException( + f"Invalid chat handler: {name} (valid formats: {list(self._chat_handlers.keys())})" + ) def get_chat_completion_handler(name: str) -> LlamaChatCompletionHandler: - return CHAT_HANDLERS[name] + return LlamaChatCompletionHandlerRegistry().get_chat_completion_handler_by_name( + name + ) def register_chat_completion_handler(name: str): def decorator(f: LlamaChatCompletionHandler): - CHAT_HANDLERS[name] = f + LlamaChatCompletionHandlerRegistry().register_chat_completion_handler(name, f) return f return decorator +### Chat Formatter ### + + +@dataclasses.dataclass +class ChatFormatterResponse: + prompt: str + stop: Optional[Union[str, List[str]]] = None + + +class ChatFormatter(Protocol): + """Base Protocol for a chat formatter. A chat formatter is a function that + takes a list of messages and returns a formatted prompt. It can also return + a stop token or list of stop tokens to use for the completion.""" + + def __call__( + self, + *, + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, + ) -> ChatFormatterResponse: + ... + + +### Utility functions for formatting chat prompts ### + + def _get_system_message( messages: List[llama_types.ChatCompletionRequestMessage], ) -> str: @@ -80,14 +153,18 @@ def _get_system_message( def _map_roles( - messages: List[llama_types.ChatCompletionRequestMessage], role_map: Dict[str, str] + messages: List[llama_types.ChatCompletionRequestMessage], + role_map: Dict[str, str], ) -> List[Tuple[str, Optional[str]]]: """Map the message roles.""" output: List[Tuple[str, Optional[str]]] = [] for message in messages: role = message["role"] if role in role_map: - output.append((role_map[role], message["content"])) + content: str | None = ( + message["content"] if isinstance(message["content"], str) else None + ) + output.append((role_map[role], content)) return output @@ -99,7 +176,8 @@ def _format_llama2( ret = system_message + sep for i, (role, message) in enumerate(messages): if system_message and i == 0: - ret += message + seps[i % 2] + m = message or "" + ret += m + seps[i % 2] elif message: ret += role + message + " " + seps[i % 2] else: @@ -172,6 +250,7 @@ def _format_chatml( ret += role + "\n" return ret + def _format_chatglm3( system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str ) -> str: @@ -187,30 +266,10 @@ def _format_chatglm3( return ret -@dataclasses.dataclass -class ChatFormatterResponse: - prompt: str - stop: Optional[Union[str, List[str]]] = None - - -class ChatFormatter(Protocol): - def __call__( - self, - *, - messages: List[llama_types.ChatCompletionRequestMessage], - **kwargs: Any, - ) -> ChatFormatterResponse: - ... - - -class BasicChatHandler: - def __init__(self, chat_format: str): - self.chat_format = chat_format - - def _convert_text_completion_to_chat( completion: llama_types.Completion, ) -> llama_types.ChatCompletion: + assert "usage" in completion return { "id": "chat" + completion["id"], "object": "chat.completion", @@ -286,127 +345,175 @@ def _convert_completion_to_chat( return _convert_text_completion_to_chat(completion) -_CHAT_FORMATS: Dict[str, ChatFormatter] = {} +def chat_formatter_to_chat_completion_handler( + chat_formatter: ChatFormatter, +) -> LlamaChatCompletionHandler: + def chat_completion_handler( + *, + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + seed: Optional[int] = None, + response_format: Optional[ + llama_types.ChatCompletionRequestResponseFormat + ] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + logit_bias: Optional[Dict[str, float]] = None, + **kwargs, # type: ignore + ) -> Union[ + llama_types.CreateChatCompletionResponse, + Iterator[llama_types.CreateChatCompletionStreamResponse], + ]: + result = chat_formatter( + messages=messages, + functions=functions, + function_call=function_call, + ) + prompt = result.prompt + if result.stop is not None: + stop = [] if stop is None else [stop] if isinstance(stop, str) else stop + rstop = result.stop if isinstance(result.stop, list) else [result.stop] + stop = stop + rstop + if response_format is not None and response_format["type"] == "json_object": + grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) -def register_chat_format(name: str): - def decorator(f: ChatFormatter): - def basic_create_chat_completion( - *, - llama: llama.Llama, - messages: List[llama_types.ChatCompletionRequestMessage], - functions: Optional[List[llama_types.ChatCompletionFunction]] = None, - function_call: Optional[ - llama_types.ChatCompletionRequestFunctionCall - ] = None, - tools: Optional[List[llama_types.ChatCompletionTool]] = None, - tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, - temperature: float = 0.2, - top_p: float = 0.95, - top_k: int = 40, - min_p: float = 0.05, - typical_p: float = 1.0, - stream: bool = False, - stop: Optional[Union[str, List[str]]] = [], - seed: Optional[int] = None, - response_format: Optional[ - llama_types.ChatCompletionRequestResponseFormat - ] = None, - max_tokens: Optional[int] = None, - presence_penalty: float = 0.0, - frequency_penalty: float = 0.0, - repeat_penalty: float = 1.1, - tfs_z: float = 1.0, - mirostat_mode: int = 0, - mirostat_tau: float = 5.0, - mirostat_eta: float = 0.1, - model: Optional[str] = None, - logits_processor: Optional[llama.LogitsProcessorList] = None, - grammar: Optional[llama.LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, - **kwargs, # type: ignore - ) -> Union[ - llama_types.CreateChatCompletionResponse, - Iterator[llama_types.CreateChatCompletionStreamResponse], - ]: - result = f( - messages=messages, - functions=functions, - function_call=function_call, - ) - prompt = result.prompt - if result.stop is not None: - stop = [] if stop is None else [stop] if isinstance(stop, str) else stop - rstop = result.stop if isinstance(result.stop, list) else [result.stop] - stop = stop + rstop + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + seed=seed, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + logit_bias=logit_bias, + ) + return _convert_completion_to_chat(completion_or_chunks, stream=stream) - if response_format is not None and response_format["type"] == "json_object": - grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF - ) + return chat_completion_handler - completion_or_chunks = llama.create_completion( - prompt=prompt, - temperature=temperature, - top_p=top_p, - top_k=top_k, - min_p=min_p, - typical_p=typical_p, - stream=stream, - stop=stop, - seed=seed, - max_tokens=max_tokens, - presence_penalty=presence_penalty, - frequency_penalty=frequency_penalty, - repeat_penalty=repeat_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - model=model, - logits_processor=logits_processor, - grammar=grammar, - logit_bias=logit_bias, - ) - return _convert_completion_to_chat(completion_or_chunks, stream=stream) - register_chat_completion_handler(name)(basic_create_chat_completion) +def register_chat_format(name: str): + def decorator(f: ChatFormatter): + chat_completion_handler = chat_formatter_to_chat_completion_handler(f) + LlamaChatCompletionHandlerRegistry().register_chat_completion_handler( + name, chat_completion_handler + ) return f - return decorator -def get_chat_format(name: str): - try: - return _CHAT_FORMATS[name] - except KeyError: - raise ValueError( - f"Invalid chat format: {name} (valid formats: {list(_CHAT_FORMATS.keys())})" - ) - - def hf_autotokenizer_to_chat_formatter( pretrained_model_name_or_path: Union[str, os.PathLike[str]] ) -> ChatFormatter: # https://huggingface.co/docs/transformers/main/chat_templating # https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1#instruction-format # https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/blob/main/tokenizer_config.json - from transformers import AutoTokenizer + from transformers import AutoTokenizer # type: ignore - tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path) + tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path) # type: ignore def format_autotokenizer( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - tokenizer.use_default_system_prompt = False - _prompt = tokenizer.apply_chat_template(messages, tokenize=False) + tokenizer.use_default_system_prompt = False # type: ignore + prompt: str = tokenizer.apply_chat_template(messages, tokenize=False) # type: ignore + assert isinstance(prompt, str) # Return formatted prompt and eos token by default - return ChatFormatterResponse(prompt=_prompt, stop=tokenizer.eos_token) + return ChatFormatterResponse(prompt=prompt, stop=tokenizer.eos_token) return format_autotokenizer +def hf_autotokenizer_to_chat_completion_handler( + pretrained_model_name_or_path: Union[str, os.PathLike[str]] +) -> LlamaChatCompletionHandler: + chat_formatter = hf_autotokenizer_to_chat_formatter(pretrained_model_name_or_path) + return chat_formatter_to_chat_completion_handler(chat_formatter) + + +def hf_tokenizer_config_to_chat_formatter(tokenizer_config: Dict[str, Any]) -> ChatFormatter: + assert isinstance(tokenizer_config, dict) + + assert "chat_template" in tokenizer_config + assert isinstance(tokenizer_config["chat_template"], str) + chat_template = tokenizer_config["chat_template"] + + assert "bos_token" in tokenizer_config + assert isinstance(tokenizer_config["bos_token"], str) + bos_token = tokenizer_config["bos_token"] + + assert "eos_token" in tokenizer_config + assert isinstance(tokenizer_config["eos_token"], str) + eos_token = tokenizer_config["eos_token"] + + env = jinja2.Environment( + loader=jinja2.BaseLoader(), + trim_blocks=True, + lstrip_blocks=True, + ).from_string(chat_template) + + def format_autotokenizer( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, + ) -> ChatFormatterResponse: + # TODO: veryify this is correct + # Add a blank assistant message to the end of the messages to prompt the model to generate a response + prompt = env.render( + messages=[ + *messages, + llama_types.ChatCompletionRequestAssistantMessage( + role="assistant", content="" + ), + ], + bos_token=bos_token, + eos_token=eos_token, + ) + return ChatFormatterResponse(prompt=prompt, stop=eos_token) + return format_autotokenizer + + +def hf_tokenizer_config_to_chat_completion_handler( + tokenizer_config: Dict[str, Any], +) -> LlamaChatCompletionHandler: + chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config) + return chat_formatter_to_chat_completion_handler(chat_formatter) + + # see https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/tokenization_llama.py # system prompt is "embedded" in the first message @register_chat_format("llama-2") @@ -437,21 +544,23 @@ def format_alpaca( _prompt = _format_add_colon_two(system_message, _messages, _sep, _sep2) return ChatFormatterResponse(prompt=_prompt) + @register_chat_format("qwen") def format_qwen( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: _roles = dict(user="<|im_start|>user", assistant="<|im_start|>assistant") - system_message="You are a helpful assistant." - system_template="<|im_start|>system\n{system_message}" - system_message=system_template.format(system_message=system_message) + system_message = "You are a helpful assistant." + system_template = "<|im_start|>system\n{system_message}" + system_message = system_template.format(system_message=system_message) _messages = _map_roles(messages, _roles) _messages.append((_roles["assistant"], None)) _sep = "<|im_end|>" _prompt = _format_chatml(system_message, _messages, _sep) _sep2 = "<|endoftext|>" - return ChatFormatterResponse(prompt=_prompt,stop=_sep2) + return ChatFormatterResponse(prompt=_prompt, stop=_sep2) + @register_chat_format("vicuna") def format( @@ -650,6 +759,7 @@ def format_mistrallite( _prompt = _format_no_colon_single(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt) + @register_chat_format("zephyr") def format_zephyr( messages: List[llama_types.ChatCompletionRequestMessage], @@ -699,6 +809,7 @@ def format_chatml( _prompt = _format_chatml(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt, stop=_sep) + @register_chat_format("chatglm3") def format_chatglm3( messages: List[llama_types.ChatCompletionRequestMessage], @@ -739,7 +850,7 @@ def format_openchat( @register_chat_format("saiga") def format_saiga( messages: list[llama_types.ChatCompletionRequestMessage], - **kwargs, + **kwargs: Any, ) -> ChatFormatterResponse: _message_template = "{role}\n{content}" _roles = dict(user="user", bot="bot", system="system") diff --git a/llama_cpp/llama_jinja_format.py b/llama_cpp/llama_jinja_format.py deleted file mode 100644 index 68faaf6ca..000000000 --- a/llama_cpp/llama_jinja_format.py +++ /dev/null @@ -1,138 +0,0 @@ -""" -llama_cpp/llama_jinja_format.py -""" -import dataclasses -from typing import Any, Callable, Dict, List, Optional, Protocol, Union - -import jinja2 -from jinja2 import Template - -# NOTE: We sacrifice readability for usability. -# It will fail to work as expected if we attempt to format it in a readable way. -llama2_template = """{% for message in messages %}{% if message['role'] == 'user' %}[INST] {{ message['content'] }} [/INST]\n{% elif message['role'] == 'assistant' %}{{ message['content'] }}\n{% elif message['role'] == 'system' %}<> {{ message['content'] }} <>\n{% endif %}{% endfor %}""" - - -class MetaSingleton(type): - """ - Metaclass for implementing the Singleton pattern. - """ - - _instances = {} - - def __call__(cls, *args, **kwargs): - if cls not in cls._instances: - cls._instances[cls] = super(MetaSingleton, cls).__call__(*args, **kwargs) - return cls._instances[cls] - - -class Singleton(object, metaclass=MetaSingleton): - """ - Base class for implementing the Singleton pattern. - """ - - def __init__(self): - super(Singleton, self).__init__() - - -@dataclasses.dataclass -class ChatFormatterResponse: - prompt: str - stop: Optional[Union[str, List[str]]] = None - - -# Base Chat Formatter Protocol -class ChatFormatterInterface(Protocol): - def __init__(self, template: Optional[object] = None): - ... - - def __call__( - self, - messages: List[Dict[str, str]], - **kwargs, - ) -> ChatFormatterResponse: - ... - - @property - def template(self) -> str: - ... - - -class AutoChatFormatter(ChatFormatterInterface): - def __init__( - self, - template: Optional[str] = None, - template_class: Optional[Template] = None, - ): - if template is not None: - self._template = template - else: - self._template = llama2_template # default template - - self._environment = jinja2.Environment( - loader=jinja2.BaseLoader(), - trim_blocks=True, - lstrip_blocks=True, - ).from_string( - self._template, - template_class=template_class, - ) - - def __call__( - self, - messages: List[Dict[str, str]], - **kwargs: Any, - ) -> ChatFormatterResponse: - formatted_sequence = self._environment.render(messages=messages, **kwargs) - return ChatFormatterResponse(prompt=formatted_sequence) - - @property - def template(self) -> str: - return self._template - - -class FormatterNotFoundException(Exception): - pass - - -class ChatFormatterFactory(Singleton): - _chat_formatters: Dict[str, Callable[[], ChatFormatterInterface]] = {} - - def register_formatter( - self, - name: str, - formatter_callable: Callable[[], ChatFormatterInterface], - overwrite=False, - ): - if not overwrite and name in self._chat_formatters: - raise ValueError( - f"Formatter with name '{name}' is already registered. Use `overwrite=True` to overwrite it." - ) - self._chat_formatters[name] = formatter_callable - - def unregister_formatter(self, name: str): - if name in self._chat_formatters: - del self._chat_formatters[name] - else: - raise ValueError(f"No formatter registered under the name '{name}'.") - - def get_formatter_by_name(self, name: str) -> ChatFormatterInterface: - try: - formatter_callable = self._chat_formatters[name] - return formatter_callable() - except KeyError: - raise FormatterNotFoundException( - f"Invalid chat format: {name} (valid formats: {list(self._chat_formatters.keys())})" - ) - - -# Define a chat format class -class Llama2Formatter(AutoChatFormatter): - def __init__(self): - super().__init__(llama2_template) - - -# With the Singleton pattern applied, regardless of where or how many times -# ChatFormatterFactory() is called, it will always return the same instance -# of the factory, ensuring that the factory's state is consistent throughout -# the application. -ChatFormatterFactory().register_formatter("llama-2", Llama2Formatter) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index f9be3237d..c2d6b6d1a 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -1,5 +1,7 @@ from __future__ import annotations +import json + from typing import Dict, Optional, Union, List import llama_cpp @@ -71,7 +73,25 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler( clip_model_path=settings.clip_model_path, verbose=settings.verbose ) - + elif settings.chat_format == "hf-autotokenizer": + assert ( + settings.hf_pretrained_model_name_or_path is not None + ), "hf_pretrained_model_name_or_path must be set for hf-autotokenizer" + chat_handler = ( + llama_cpp.llama_chat_format.hf_autotokenizer_to_chat_formatter( + settings.hf_pretrained_model_name_or_path + ) + ) + elif settings.chat_format == "hf-tokenizer-config": + assert ( + settings.hf_tokenizer_config_path is not None + ), "hf_tokenizer_config_path must be set for hf-tokenizer-config" + chat_handler = ( + llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_formatter( + json.load(open(settings.hf_tokenizer_config_path)) + ) + ) + kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None if settings.kv_overrides is not None: assert isinstance(settings.kv_overrides, list) @@ -141,4 +161,3 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: cache = llama_cpp.LlamaRAMCache(capacity_bytes=settings.cache_size) _model.set_cache(cache) return _model - diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index dc5be209d..9f0dc8a73 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -134,6 +134,15 @@ class ModelSettings(BaseSettings): default=2 << 30, description="The size of the cache in bytes. Only used if cache is True.", ) + # Tokenizer Options + hf_tokenizer_config_path: Optional[str] = Field( + default=None, + description="The path to a HuggingFace tokenizer_config.json file.", + ) + hf_pretrained_model_name_or_path: Optional[str] = Field( + default=None, + description="The model name or path to a pretrained HuggingFace tokenizer model. Same as you would pass to AutoTokenizer.from_pretrained().", + ) # Misc verbose: bool = Field( default=True, description="Whether to print debug information." diff --git a/tests/test_llama_chat_format.py b/tests/test_llama_chat_format.py index 4eebcb674..1ef18d927 100644 --- a/tests/test_llama_chat_format.py +++ b/tests/test_llama_chat_format.py @@ -1,50 +1,65 @@ -from typing import List +import json -import pytest +from llama_cpp import ( + ChatCompletionRequestUserMessage, +) +from llama_cpp.llama_chat_format import hf_tokenizer_config_to_chat_formatter -from llama_cpp import ChatCompletionMessage -from llama_cpp.llama_jinja_format import Llama2Formatter +mistral_7b_tokenizer_config = """{ + "add_bos_token": true, + "add_eos_token": false, + "added_tokens_decoder": { + "0": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "1": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "2": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + } + }, + "additional_special_tokens": [], + "bos_token": "", + "clean_up_tokenization_spaces": false, + "eos_token": "", + "legacy": true, + "model_max_length": 1000000000000000019884624838656, + "pad_token": null, + "sp_model_kwargs": {}, + "spaces_between_special_tokens": false, + "tokenizer_class": "LlamaTokenizer", + "unk_token": "", + "use_default_system_prompt": false, + "chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" +}""" -@pytest.fixture -def sequence_of_messages() -> List[ChatCompletionMessage]: - return [ - ChatCompletionMessage(role="system", content="Welcome to CodeHelp Bot!"), - ChatCompletionMessage( - role="user", content="Hi there! I need some help with Python." - ), - ChatCompletionMessage( - role="assistant", content="Of course! What do you need help with in Python?" - ), - ChatCompletionMessage( - role="user", - content="I'm trying to write a function to find the factorial of a number, but I'm stuck.", - ), - ChatCompletionMessage( - role="assistant", - content="I can help with that! Would you like a recursive or iterative solution?", - ), - ChatCompletionMessage( - role="user", content="Let's go with a recursive solution." - ), - ] - -def test_llama2_formatter(sequence_of_messages): - expected_prompt = ( - "<> Welcome to CodeHelp Bot! <>\n" - "[INST] Hi there! I need some help with Python. [/INST]\n" - "Of course! What do you need help with in Python?\n" - "[INST] I'm trying to write a function to find the factorial of a number, but I'm stuck. [/INST]\n" - "I can help with that! Would you like a recursive or iterative solution?\n" - "[INST] Let's go with a recursive solution. [/INST]\n" +def test_hf_tokenizer_config_str_to_chat_formatter(): + tokenizer_config = json.loads(mistral_7b_tokenizer_config) + chat_formatter = hf_tokenizer_config_to_chat_formatter( + tokenizer_config + ) + chat_formatter_respoonse = chat_formatter( + messages=[ + ChatCompletionRequestUserMessage(role="user", content="Hello, world!"), + ] ) - llama2_formatter_instance = Llama2Formatter() - formatter_response = llama2_formatter_instance(sequence_of_messages) - assert ( - expected_prompt == formatter_response.prompt - ), "The formatted prompt does not match the expected output." - - -# Optionally, include a test for the 'stop' if it's part of the functionality. + assert chat_formatter_respoonse.prompt == ("[INST] Hello, world! [/INST]" "") From 89cce50f8c332cdb72636d2f61e37a1309feafca Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 21:21:49 -0500 Subject: [PATCH 035/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 13 +++++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 14 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 9e8e3cec7..ef1627230 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -91,6 +91,12 @@ def _load_shared_library(lib_base_name: str): c_uint8_p = POINTER(c_uint8) c_size_t_p = POINTER(c_size_t) +# from ggml-backend.h +# typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); +ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE( + c_bool, c_void_p, c_bool, c_void_p +) + # llama.h bindings _lib.llama_max_devices.argtypes = [] @@ -448,6 +454,9 @@ class llama_model_params(Structure): # float yarn_beta_slow; // YaRN high correction dim # uint32_t yarn_orig_ctx; // YaRN original context size +# ggml_backend_sched_eval_callback cb_eval; +# void * cb_eval_user_data; + # enum ggml_type type_k; // data type for K cache # enum ggml_type type_v; // data type for V cache @@ -475,6 +484,8 @@ class llama_context_params(Structure): yarn_beta_fast (float): YaRN low correction dim yarn_beta_slow (float): YaRN high correction dim yarn_orig_ctx (int): YaRN original context size + cb_eval (ggml_backend_sched_eval_callback): callback for scheduling eval + cb_eval_user_data (ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache type_v (int): data type for V cache mul_mat_q (bool): if true, use experimental mul_mat_q kernels (DEPRECATED - always true) @@ -497,6 +508,8 @@ class llama_context_params(Structure): ("yarn_beta_fast", c_float), ("yarn_beta_slow", c_float), ("yarn_orig_ctx", c_uint32), + ("cb_eval", ggml_backend_sched_eval_callback), + ("cb_eval_user_data", c_void_p), ("type_k", c_int), ("type_v", c_int), ("mul_mat_q", c_bool), diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4f4bf35f4..2d5419d08 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4f4bf35f46600441dec2f941e667291eeb9a18d8 +Subproject commit 2d5419d08ab1131623e6a1d554607b7663435e87 From be23404ed43b807406a6db55a0cbb830b47b11e3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 21:22:19 -0500 Subject: [PATCH 036/624] Cleanup pyproject --- pyproject.toml | 2 -- 1 file changed, 2 deletions(-) diff --git a/pyproject.toml b/pyproject.toml index 806127d89..413097201 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -11,8 +11,6 @@ license = { text = "MIT" } authors = [ { name = "Andrei Betlen", email = "abetlen@gmail.com" }, ] -# mkdocs-martiral requires "jinja2~=3.0" -# transformers requires "jinja2>=2.11.3" dependencies = [ "typing-extensions>=4.5.0", "numpy>=1.20.0", From 3ca86ab3909444917a272607aa62b105584a50a8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 21:22:45 -0500 Subject: [PATCH 037/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 2d5419d08..57e2a7a52 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 2d5419d08ab1131623e6a1d554607b7663435e87 +Subproject commit 57e2a7a52a819883f40dada8a2edc24ecf48186b From 03ed547bfd6907039e6d594e7203dca155280492 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 21:23:26 -0500 Subject: [PATCH 038/624] Remove templates doc --- docs/templates.md | 52 ----------------------------------------------- 1 file changed, 52 deletions(-) delete mode 100644 docs/templates.md diff --git a/docs/templates.md b/docs/templates.md deleted file mode 100644 index 5acdaa196..000000000 --- a/docs/templates.md +++ /dev/null @@ -1,52 +0,0 @@ -# Templates - -This document provides a comprehensive guide to the integration of Jinja2 templating into the `llama-cpp-python` project, with a focus on enhancing the chat functionality of the `llama-2` model. - -## Introduction - -- Brief explanation of the `llama-cpp-python` project's need for a templating system. -- Overview of the `llama-2` model's interaction with templating. - -## Jinja2 Dependency Integration - -- Rationale for choosing Jinja2 as the templating engine. - - Compatibility with Hugging Face's `transformers`. - - Desire for advanced templating features and simplicity. -- Detailed steps for adding `jinja2` to `pyproject.toml` for dependency management. - -## Template Management Refactor - -- Summary of the refactor and the motivation behind it. -- Description of the new chat handler selection logic: - 1. Preference for a user-specified `chat_handler`. - 2. Fallback to a user-specified `chat_format`. - 3. Defaulting to a chat format from a `.gguf` file if available. - 4. Utilizing the `llama2` default chat format as the final fallback. -- Ensuring backward compatibility throughout the refactor. - -## Implementation Details - -- In-depth look at the new `AutoChatFormatter` class. -- Example code snippets showing how to utilize the Jinja2 environment and templates. -- Guidance on how to provide custom templates or use defaults. - -## Testing and Validation - -- Outline of the testing strategy to ensure seamless integration. -- Steps for validating backward compatibility with existing implementations. - -## Benefits and Impact - -- Analysis of the expected benefits, including consistency, performance gains, and improved developer experience. -- Discussion of the potential impact on current users and contributors. - -## Future Work - -- Exploration of how templating can evolve within the project. -- Consideration of additional features or optimizations for the templating engine. -- Mechanisms for community feedback on the templating system. - -## Conclusion - -- Final thoughts on the integration of Jinja2 templating. -- Call to action for community involvement and feedback. From 656f3d896845d428f994dd205b28d8f524cbff67 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Jan 2024 21:30:36 -0500 Subject: [PATCH 039/624] Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index c0748ee14..92636baed 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.30] + +- feat: Update llama.cpp to ggerganov/llama.cpp@57e2a7a52a819883f40dada8a2edc24ecf48186b +- feat(server): Add ability to load chat format from huggingface autotokenizer or tokenizer_config.json files by @abetlen in b8fc1c7d83ad4a9207c707ba1d954fe580286a01 +- feat: Integration of Jinja2 Templating for chat formats by @teleprint-me in #875 +- fix: Offload KQV by default by @abetlen in 48c3b77e6f558a9899de0e1155c7dc0c7958d8e8 +- fix: Support Accept text/event-stream in chat and completion endpoints, resolves #1083 by @aniljava in #1088 +- fix(cli): allow passing n_ctx=0 to openAI API server args to use model n_ctx_train field per #1015 by @K-Mistele in #1093 + ## [0.2.29] - feat: Update llama.cpp to ggerganov/llama.cpp@4483396751c79dea540808b9cb9238245d06da2b diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 65206bf28..210404a8e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.29" \ No newline at end of file +__version__ = "0.2.30" \ No newline at end of file From 141293a75b564a8699e0acba1da24d9aa1cf0ab1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 08:17:49 -0500 Subject: [PATCH 040/624] Fix python3.8 support --- llama_cpp/server/cli.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/server/cli.py b/llama_cpp/server/cli.py index 8e32d2c0e..3dd007676 100644 --- a/llama_cpp/server/cli.py +++ b/llama_cpp/server/cli.py @@ -55,7 +55,7 @@ def _parse_bool_arg(arg: str | bytes | bool) -> bool: raise ValueError(f"Invalid boolean argument: {arg}") -def add_args_from_model(parser: argparse.ArgumentParser, model: type[BaseModel]): +def add_args_from_model(parser: argparse.ArgumentParser, model: Type[BaseModel]): """Add arguments from a pydantic model to an argparse parser.""" for name, field in model.model_fields.items(): @@ -83,7 +83,7 @@ def add_args_from_model(parser: argparse.ArgumentParser, model: type[BaseModel]) ) -T = TypeVar("T", bound=type[BaseModel]) +T = TypeVar("T", bound=Type[BaseModel]) def parse_model_from_args(model: T, args: argparse.Namespace) -> T: From 3babe3512cb95743108f2b595210c38ed6f1b904 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 08:31:59 -0500 Subject: [PATCH 041/624] Fix mirostat sampling --- llama_cpp/llama.py | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 6cdc1eb76..32eb3fe28 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -329,6 +329,8 @@ def __init__( (n_ctx, self._n_vocab), dtype=np.single ) + self._mirostat_mu = ctypes.c_float(2.0 * 5.0) # TODO: Move this to sampling context + @property def ctx(self) -> llama_cpp.llama_context_p: assert self._ctx.ctx is not None @@ -516,7 +518,7 @@ def sample( candidates=self._candidates, tau=mirostat_tau, eta=mirostat_eta, - mu=2.0 * mirostat_tau, + mu=ctypes.pointer(self._mirostat_mu), m=100, ) elif mirostat_mode == 2: @@ -525,7 +527,7 @@ def sample( candidates=self._candidates, tau=mirostat_tau, eta=mirostat_eta, - mu=2.0 * mirostat_tau, + mu=ctypes.pointer(self._mirostat_mu) ) else: self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1) @@ -581,6 +583,10 @@ def generate( Yields: The generated tokens. """ + # Reset mirostat sampling + self._mirostat_mu = ctypes.c_float(2.0 * mirostat_tau) + + # Check for kv cache prefix match if reset and self.n_tokens > 0: longest_prefix = 0 for a, b in zip(self._input_ids, tokens[:-1]): @@ -595,12 +601,15 @@ def generate( tokens = tokens[longest_prefix:] self.n_tokens = longest_prefix + # Reset the model state if reset: self.reset() + # Reset the grammar if grammar is not None: grammar.reset() + # Eval and sample while True: self.eval(tokens) token = self.sample( From 0f54948482fc7e2ad71f511c1536b7a77c5c1ba7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 08:41:52 -0500 Subject: [PATCH 042/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 57e2a7a52..a5cacb22b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 57e2a7a52a819883f40dada8a2edc24ecf48186b +Subproject commit a5cacb22b2114fd9adf61c00cbb237384d86bced From e21c3c7a91f354cdb139f4bd6a5d48193ea806f0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 08:47:56 -0500 Subject: [PATCH 043/624] Update makefile --- Makefile | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/Makefile b/Makefile index e930609ff..5ed3fa262 100644 --- a/Makefile +++ b/Makefile @@ -10,22 +10,22 @@ deps: python3 -m pip install -e ".[all]" build: - python3 -m pip install -e . + python3 -m pip install --verbose -e . build.cuda: - CMAKE_ARGS="-DLLAMA_CUBLAS=on" python3 -m pip install -e . + CMAKE_ARGS="-DLLAMA_CUBLAS=on" python3 -m pip install --verbose -e . build.opencl: - CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install -e . + CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e . build.openblas: - CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install -e . + CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e . build.blis: - CMAKE_ARGS="-DLLAMA_OPENBLAS=on -DLLAMA_OPENBLAS_VENDOR=blis" python3 -m pip install -e . + CMAKE_ARGS="-DLLAMA_OPENBLAS=on -DLLAMA_OPENBLAS_VENDOR=blis" python3 -m pip install --verbose -e . build.metal: - CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install -e . + CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install --verbose -e . build.sdist: python3 -m build --sdist From 833a7f1a86f2136df5f75c1bd62d2e4d5adaa439 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 09:03:35 -0500 Subject: [PATCH 044/624] Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 92636baed..797785e83 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.31] + +- feat: Update llama.cpp to ggerganov/llama.cpp@a5cacb22b2114fd9adf61c00cbb237384d86bced +- fix: Mirostat sampling now passes correct type to ctypes and tracks state during generation by @abetlen in 3babe3512cb95743108f2b595210c38ed6f1b904 +- fix: Python3.8 support in server by @abetlen in 141293a75b564a8699e0acba1da24d9aa1cf0ab1 + ## [0.2.30] - feat: Update llama.cpp to ggerganov/llama.cpp@57e2a7a52a819883f40dada8a2edc24ecf48186b diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 210404a8e..186983855 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.30" \ No newline at end of file +__version__ = "0.2.31" \ No newline at end of file From 5a34c57e5479e50c99aba9b38218cc48e6560b81 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 10:46:03 -0500 Subject: [PATCH 045/624] feat: Expose gguf model metadata in metadata property --- llama_cpp/_internals.py | 25 +++++++++++++++++++++++++ llama_cpp/llama.py | 10 ++++++++++ 2 files changed, 35 insertions(+) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 208de8c2a..ec47c4216 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -204,6 +204,31 @@ def detokenize(self, tokens: List[int]) -> bytes: output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output ) + # Extra + def metadata(self) -> Dict[str, str]: + assert self.model is not None + metadata: Dict[str, str] = {} + buffer_size = 1024 + buffer = ctypes.create_string_buffer(buffer_size) + # zero the buffer + buffer.value = b'\0' * buffer_size + # iterate over model keys + for i in range(llama_cpp.llama_model_meta_count(self.model)): + nbytes = llama_cpp.llama_model_meta_key_by_index(self.model, i, buffer, buffer_size) + if nbytes > buffer_size: + buffer_size = nbytes + buffer = ctypes.create_string_buffer(buffer_size) + nbytes = llama_cpp.llama_model_meta_key_by_index(self.model, i, buffer, buffer_size) + key = buffer.value.decode("utf-8") + nbytes = llama_cpp.llama_model_meta_val_str_by_index(self.model, i, buffer, buffer_size) + if nbytes > buffer_size: + buffer_size = nbytes + buffer = ctypes.create_string_buffer(buffer_size) + nbytes = llama_cpp.llama_model_meta_val_str_by_index(self.model, i, buffer, buffer_size) + value = buffer.value.decode("utf-8") + metadata[key] = value + return metadata + @staticmethod def default_params(): """Get the default llama_model_params.""" diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 32eb3fe28..5c66bcf09 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -331,6 +331,16 @@ def __init__( self._mirostat_mu = ctypes.c_float(2.0 * 5.0) # TODO: Move this to sampling context + try: + self.metadata = self._model.metadata() + except Exception as e: + self.metadata = {} + if self.verbose: + print(f"Failed to load metadata: {e}", file=sys.stderr) + + if self.verbose: + print(f"Model metadata: {self.metadata}", file=sys.stderr) + @property def ctx(self) -> llama_cpp.llama_context_p: assert self._ctx.ctx is not None From be09318c26add8674ce494ae7cc480cce72a4146 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 15:04:42 -0500 Subject: [PATCH 046/624] feat: Add Jinja2ChatFormatter --- llama_cpp/llama_chat_format.py | 323 +++++++++++++++++++-------------- 1 file changed, 188 insertions(+), 135 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 3d18d904f..4e1b174a5 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -121,14 +121,21 @@ def decorator(f: LlamaChatCompletionHandler): @dataclasses.dataclass class ChatFormatterResponse: + """Dataclass that stores completion parameters for a given chat format and + create_chat_completion request. + + prompt contains the formatted prompt generated from the chat format and messages. + stop contains the stop token or list of stop tokens to use for the chat format.""" + prompt: str stop: Optional[Union[str, List[str]]] = None class ChatFormatter(Protocol): """Base Protocol for a chat formatter. A chat formatter is a function that - takes a list of messages and returns a formatted prompt. It can also return - a stop token or list of stop tokens to use for the completion.""" + takes a list of messages and returns a chat format response which can be used + to generate a completion. The response can also include a stop token or list + of stop tokens to use for the completion.""" def __call__( self, @@ -139,131 +146,43 @@ def __call__( ... -### Utility functions for formatting chat prompts ### - - -def _get_system_message( - messages: List[llama_types.ChatCompletionRequestMessage], -) -> str: - """Get the first system message.""" - for message in messages: - if message["role"] == "system": - return message["content"] or "" - return "" - - -def _map_roles( - messages: List[llama_types.ChatCompletionRequestMessage], - role_map: Dict[str, str], -) -> List[Tuple[str, Optional[str]]]: - """Map the message roles.""" - output: List[Tuple[str, Optional[str]]] = [] - for message in messages: - role = message["role"] - if role in role_map: - content: str | None = ( - message["content"] if isinstance(message["content"], str) else None - ) - output.append((role_map[role], content)) - return output - - -def _format_llama2( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str -) -> str: - """Format the prompt with the llama2 style.""" - seps = [sep, sep2] - ret = system_message + sep - for i, (role, message) in enumerate(messages): - if system_message and i == 0: - m = message or "" - ret += m + seps[i % 2] - elif message: - ret += role + message + " " + seps[i % 2] - else: - ret += role + " " - return ret - - -def _format_add_colon_single( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the add-colon-single style.""" - ret = system_message + sep - for role, message in messages: - if message: - ret += role + ": " + message + sep - else: - ret += role + ":" - return ret - - -def _format_add_colon_two( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str -) -> str: - """Format the prompt with the add-colon-two style.""" - seps = [sep, sep2] - ret = system_message + seps[0] - for i, (role, message) in enumerate(messages): - if message: - ret += role + ": " + message + seps[i % 2] - else: - ret += role + ":" - return ret - - -def _format_no_colon_single( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the no-colon-single style.""" - ret = system_message - for role, message in messages: - if message: - ret += role + message + sep - else: - ret += role - return ret - - -def _format_add_colon_space_single( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the add-colon-space-single style.""" - ret = system_message + sep - for role, message in messages: - if message: - ret += role + ": " + message + sep - else: - ret += role + ": " # must be end with a space - return ret - +class Jinja2ChatFormatter(ChatFormatter): + def __init__( + self, + template: str, + eos_token: str, + bos_token: str, + ): + """A chat formatter that uses jinja2 templates to format the prompt.""" + self.template = template + self.eos_token = eos_token + self.bos_token = bos_token -def _format_chatml( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the chatml style.""" - ret = "" if system_message == "" else system_message + sep + "\n" - for role, message in messages: - if message: - ret += role + "\n" + message + sep + "\n" - else: - ret += role + "\n" - return ret + self._environment = jinja2.Environment( + loader=jinja2.BaseLoader(), + trim_blocks=True, + lstrip_blocks=True, + ).from_string(self.template) + def __call__( + self, + *, + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, + ) -> ChatFormatterResponse: + messages = [ + *messages, + llama_types.ChatCompletionRequestAssistantMessage( + role="assistant", content="" + ), + ] + prompt = self._environment.render( + messages=messages, eos_token=self.eos_token, bos_token=self.bos_token + ) + return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token]) -def _format_chatglm3( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the chatglm3 style.""" - ret = "" - if system_message: - ret += system_message - for role, message in messages: - if message: - ret += role + "\n" + " " + message - else: - ret += role - return ret + def to_chat_handler(self) -> LlamaChatCompletionHandler: + return chat_formatter_to_chat_completion_handler(self) def _convert_text_completion_to_chat( @@ -426,16 +345,6 @@ def chat_completion_handler( return chat_completion_handler -def register_chat_format(name: str): - def decorator(f: ChatFormatter): - chat_completion_handler = chat_formatter_to_chat_completion_handler(f) - LlamaChatCompletionHandlerRegistry().register_chat_completion_handler( - name, chat_completion_handler - ) - return f - return decorator - - def hf_autotokenizer_to_chat_formatter( pretrained_model_name_or_path: Union[str, os.PathLike[str]] ) -> ChatFormatter: @@ -466,7 +375,9 @@ def hf_autotokenizer_to_chat_completion_handler( return chat_formatter_to_chat_completion_handler(chat_formatter) -def hf_tokenizer_config_to_chat_formatter(tokenizer_config: Dict[str, Any]) -> ChatFormatter: +def hf_tokenizer_config_to_chat_formatter( + tokenizer_config: Dict[str, Any] +) -> ChatFormatter: assert isinstance(tokenizer_config, dict) assert "chat_template" in tokenizer_config @@ -504,6 +415,7 @@ def format_autotokenizer( eos_token=eos_token, ) return ChatFormatterResponse(prompt=prompt, stop=eos_token) + return format_autotokenizer @@ -514,6 +426,147 @@ def hf_tokenizer_config_to_chat_completion_handler( return chat_formatter_to_chat_completion_handler(chat_formatter) +### Utility functions for formatting chat prompts ### + + +def _get_system_message( + messages: List[llama_types.ChatCompletionRequestMessage], +) -> str: + """Get the first system message.""" + for message in messages: + if message["role"] == "system": + return message["content"] or "" + return "" + + +def _map_roles( + messages: List[llama_types.ChatCompletionRequestMessage], + role_map: Dict[str, str], +) -> List[Tuple[str, Optional[str]]]: + """Map the message roles.""" + output: List[Tuple[str, Optional[str]]] = [] + for message in messages: + role = message["role"] + if role in role_map: + content: str | None = ( + message["content"] if isinstance(message["content"], str) else None + ) + output.append((role_map[role], content)) + return output + + +def _format_llama2( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str +) -> str: + """Format the prompt with the llama2 style.""" + seps = [sep, sep2] + ret = system_message + sep + for i, (role, message) in enumerate(messages): + if system_message and i == 0: + m = message or "" + ret += m + seps[i % 2] + elif message: + ret += role + message + " " + seps[i % 2] + else: + ret += role + " " + return ret + + +def _format_add_colon_single( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the add-colon-single style.""" + ret = system_message + sep + for role, message in messages: + if message: + ret += role + ": " + message + sep + else: + ret += role + ":" + return ret + + +def _format_add_colon_two( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str +) -> str: + """Format the prompt with the add-colon-two style.""" + seps = [sep, sep2] + ret = system_message + seps[0] + for i, (role, message) in enumerate(messages): + if message: + ret += role + ": " + message + seps[i % 2] + else: + ret += role + ":" + return ret + + +def _format_no_colon_single( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the no-colon-single style.""" + ret = system_message + for role, message in messages: + if message: + ret += role + message + sep + else: + ret += role + return ret + + +def _format_add_colon_space_single( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the add-colon-space-single style.""" + ret = system_message + sep + for role, message in messages: + if message: + ret += role + ": " + message + sep + else: + ret += role + ": " # must be end with a space + return ret + + +def _format_chatml( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the chatml style.""" + ret = "" if system_message == "" else system_message + sep + "\n" + for role, message in messages: + if message: + ret += role + "\n" + message + sep + "\n" + else: + ret += role + "\n" + return ret + + +def _format_chatglm3( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the chatglm3 style.""" + ret = "" + if system_message: + ret += system_message + for role, message in messages: + if message: + ret += role + "\n" + " " + message + else: + ret += role + return ret + + +### Chat Formats ### + + +def register_chat_format(name: str): + def decorator(f: ChatFormatter): + chat_completion_handler = chat_formatter_to_chat_completion_handler(f) + LlamaChatCompletionHandlerRegistry().register_chat_completion_handler( + name, chat_completion_handler + ) + return f + + return decorator + + # see https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/tokenization_llama.py # system prompt is "embedded" in the first message @register_chat_format("llama-2") From ac2e96d4b4610cb9cd9b0c978c76ece6567f5c02 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Jan 2024 15:33:43 -0500 Subject: [PATCH 047/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index a5cacb22b..381ee1957 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit a5cacb22b2114fd9adf61c00cbb237384d86bced +Subproject commit 381ee195721d8e747ee31a60c0751822b3072f02 From 7f3209b1eb4ad3260ba063801fab80a8c25a2f4c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Jan 2024 18:37:24 -0500 Subject: [PATCH 048/624] feat: Add add_generation_prompt option for jinja2chatformatter. --- llama_cpp/llama_chat_format.py | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 4e1b174a5..02bdbcfa7 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -152,11 +152,13 @@ def __init__( template: str, eos_token: str, bos_token: str, + add_generation_prompt: bool = True, ): """A chat formatter that uses jinja2 templates to format the prompt.""" self.template = template self.eos_token = eos_token self.bos_token = bos_token + self.add_generation_prompt = add_generation_prompt self._environment = jinja2.Environment( loader=jinja2.BaseLoader(), @@ -170,12 +172,13 @@ def __call__( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - messages = [ - *messages, - llama_types.ChatCompletionRequestAssistantMessage( - role="assistant", content="" - ), - ] + if self.add_generation_prompt: + messages = [ + *messages, + llama_types.ChatCompletionRequestAssistantMessage( + role="assistant", content="" + ), + ] prompt = self._environment.render( messages=messages, eos_token=self.eos_token, bos_token=self.bos_token ) From 24f39454e91cf5dddbc4b6041aead4accc7c7a2d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Jan 2024 18:38:04 -0500 Subject: [PATCH 049/624] fix: pass chat handler not chat formatter for huggingface autotokenizer and tokenizer_config formats. --- llama_cpp/server/model.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index c2d6b6d1a..bbb68069d 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -78,7 +78,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: settings.hf_pretrained_model_name_or_path is not None ), "hf_pretrained_model_name_or_path must be set for hf-autotokenizer" chat_handler = ( - llama_cpp.llama_chat_format.hf_autotokenizer_to_chat_formatter( + llama_cpp.llama_chat_format.hf_autotokenizer_to_chat_completion_handler( settings.hf_pretrained_model_name_or_path ) ) @@ -87,7 +87,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: settings.hf_tokenizer_config_path is not None ), "hf_tokenizer_config_path must be set for hf-tokenizer-config" chat_handler = ( - llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_formatter( + llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_completion_handler( json.load(open(settings.hf_tokenizer_config_path)) ) ) From 88fbccaaa3416f38552d80d84c71fdb40c7c477a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Jan 2024 18:38:44 -0500 Subject: [PATCH 050/624] docs: Add macosx wrong arch fix to README --- README.md | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index ad5d0f1e8..f97ea0f77 100644 --- a/README.md +++ b/README.md @@ -113,6 +113,10 @@ See the above instructions and set `CMAKE_ARGS` to the BLAS backend you want to ### MacOS Notes +Detailed MacOS Metal GPU install documentation is available at [docs/install/macos.md](https://llama-cpp-python.readthedocs.io/en/latest/install/macos/) + +#### M1 Mac Performance Issue + Note: If you are using Apple Silicon (M1) Mac, make sure you have installed a version of Python that supports arm64 architecture. For example: ``` wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh @@ -120,7 +124,13 @@ bash Miniforge3-MacOSX-arm64.sh ``` Otherwise, while installing it will build the llama.cpp x86 version which will be 10x slower on Apple Silicon (M1) Mac. -Detailed MacOS Metal GPU install documentation is available at [docs/install/macos.md](https://llama-cpp-python.readthedocs.io/en/latest/install/macos/) +#### M Series Mac Error: `(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64'))` + +Try installing with + +``` +CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on" pip install --upgrade --verbose --force-reinstall --no-cache-dir llama-cpp-python +``` ### Upgrading and Reinstalling From 8eefdbca03d005095f6645d4e5b42b982af9daf0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Jan 2024 19:01:27 -0500 Subject: [PATCH 051/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 381ee1957..504dc37be 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 381ee195721d8e747ee31a60c0751822b3072f02 +Subproject commit 504dc37be8446fb09b1ede70300250ad41be32a2 From d3f5528ca8bcb9d69d4f27e21631e911f1fb9bfe Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Jan 2024 19:06:53 -0500 Subject: [PATCH 052/624] fix: from_json_schema oneof/anyof bug. Closes #1097 --- llama_cpp/llama_grammar.py | 23 +++++++++++++---------- tests/test_grammar.py | 26 ++++++++++++++++++++++++++ 2 files changed, 39 insertions(+), 10 deletions(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index c02e65642..d8ef563c2 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1432,7 +1432,6 @@ def _add_rule(self, name: str, rule: str): return key def visit(self, schema: Dict[str, Any], name: str) -> str: - schema_type: Optional[str] = schema.get("type") # type: ignore rule_name = name or "root" if "$defs" in schema: @@ -1458,7 +1457,19 @@ def visit(self, schema: Dict[str, Any], name: str) -> str: rule = " | ".join((self._format_literal(v) for v in schema["enum"])) return self._add_rule(rule_name, rule) - elif schema_type == "object" and "properties" in schema: + elif "$ref" in schema: + ref = schema["$ref"] + assert ref.startswith("#/$defs/"), f"Unrecognized schema: {schema}" + # inline $defs + def_name = ref[len("#/$defs/") :] + def_schema = self._defs[def_name] + return self.visit(def_schema, f'{name}{"-" if name else ""}{def_name}') + + + schema_type: Optional[str] = schema.get("type") # type: ignore + assert isinstance(schema_type, str), f"Unrecognized schema: {schema}" + + if schema_type == "object" and "properties" in schema: # TODO: `required` keyword prop_order = self._prop_order prop_pairs = sorted( @@ -1489,14 +1500,6 @@ def visit(self, schema: Dict[str, Any], name: str) -> str: ) return self._add_rule(rule_name, rule) - elif "$ref" in schema: - ref = schema["$ref"] - assert ref.startswith("#/$defs/"), f"Unrecognized schema: {schema}" - # inline $defs - def_name = ref[len("#/$defs/") :] - def_schema = self._defs[def_name] - return self.visit(def_schema, f'{name}{"-" if name else ""}{def_name}') - else: assert schema_type in PRIMITIVE_RULES, f"Unrecognized schema: {schema}" return self._add_rule( diff --git a/tests/test_grammar.py b/tests/test_grammar.py index ef9392b7a..cb221880a 100644 --- a/tests/test_grammar.py +++ b/tests/test_grammar.py @@ -50,3 +50,29 @@ class B(BaseModel): grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(schema)) assert grammar.grammar is not None + + +def test_grammar_anyof(): + sch = { + "properties": { + "temperature": { + "description": "The temperature mentioned", + "type": "number", + }, + "unit": { + "anyOf": [ + { + "description": "Unit for temperature", + "enum": ["celsius", "fahrenheit"], + "type": "string", + }, + {"type": "null"}, + ], + }, + }, + "type": "object", + } + + grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(sch)) + + assert grammar.grammar is not None \ No newline at end of file From 2ce0b8aa2c2f81d999bbb2a7246a9f221f9d52d0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Jan 2024 20:30:24 -0500 Subject: [PATCH 053/624] Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 797785e83..4fff9194d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.32] + +- feat: Update llama.cpp to ggerganov/llama.cpp@504dc37be8446fb09b1ede70300250ad41be32a2 +- fix: from_json_schema oneof/anyof bug by @jndiogo in d3f5528ca8bcb9d69d4f27e21631e911f1fb9bfe +- fix: pass chat handler not chat formatter for huggingface autotokenizer and tokenizer_config formats by @abetlen in 24f39454e91cf5dddbc4b6041aead4accc7c7a2d +- feat: Add add_generation_prompt option for jinja2chatformatter by @abetlen in 7f3209b1eb4ad3260ba063801fab80a8c25a2f4c +- feat: Add Jinja2ChatFormatter by @abetlen in be09318c26add8674ce494ae7cc480cce72a4146 +- feat: Expose gguf model metadata in metadata property by @abetlen in 5a34c57e5479e50c99aba9b38218cc48e6560b81 + ## [0.2.31] - feat: Update llama.cpp to ggerganov/llama.cpp@a5cacb22b2114fd9adf61c00cbb237384d86bced diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 186983855..dda8335b5 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.31" \ No newline at end of file +__version__ = "0.2.32" \ No newline at end of file From 5b982d0f8c6f35242c8862ffdce00e17cea0b44f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 22 Jan 2024 08:32:48 -0500 Subject: [PATCH 054/624] fix: use both eos and bos tokens as stop sequences for hf-tokenizer-config chat format. --- llama_cpp/llama_chat_format.py | 20 ++++++++++++-------- 1 file changed, 12 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 02bdbcfa7..6c274aa82 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -379,7 +379,8 @@ def hf_autotokenizer_to_chat_completion_handler( def hf_tokenizer_config_to_chat_formatter( - tokenizer_config: Dict[str, Any] + tokenizer_config: Dict[str, Any], + add_generation_prompt: bool = True, ) -> ChatFormatter: assert isinstance(tokenizer_config, dict) @@ -401,31 +402,34 @@ def hf_tokenizer_config_to_chat_formatter( lstrip_blocks=True, ).from_string(chat_template) - def format_autotokenizer( + def format_tokenizer_config( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: # TODO: veryify this is correct # Add a blank assistant message to the end of the messages to prompt the model to generate a response - prompt = env.render( - messages=[ + if add_generation_prompt: + messages = [ *messages, llama_types.ChatCompletionRequestAssistantMessage( role="assistant", content="" ), - ], + ] + prompt = env.render( + messages=messages, bos_token=bos_token, eos_token=eos_token, ) - return ChatFormatterResponse(prompt=prompt, stop=eos_token) + return ChatFormatterResponse(prompt=prompt, stop=[eos_token, bos_token]) - return format_autotokenizer + return format_tokenizer_config def hf_tokenizer_config_to_chat_completion_handler( tokenizer_config: Dict[str, Any], + add_generation_prompt: bool = True, ) -> LlamaChatCompletionHandler: - chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config) + chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config, add_generation_prompt=add_generation_prompt) return chat_formatter_to_chat_completion_handler(chat_formatter) From fcdf337d84591c46c0a26b8660b2e537af1fd353 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 22 Jan 2024 11:25:11 -0500 Subject: [PATCH 055/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index ef1627230..5de837f87 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -188,6 +188,7 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -211,6 +212,7 @@ def _load_shared_library(lib_base_name: str): LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19 LLAMA_FTYPE_MOSTLY_IQ2_XS = 20 LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 +LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 504dc37be..6f9939d11 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 504dc37be8446fb09b1ede70300250ad41be32a2 +Subproject commit 6f9939d119b2d004c264952eb510bd106455531e From 7e63928bc9eb87170cbf3d856a318e0794f2505f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Jan 2024 18:42:39 -0500 Subject: [PATCH 056/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6f9939d11..26d607608 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6f9939d119b2d004c264952eb510bd106455531e +Subproject commit 26d607608d794efa56df3bdb6043a2f94c1d632c From fe5d6ea6484a06021a2531c6f66c9873c1ee753d Mon Sep 17 00:00:00 2001 From: Phil H <5756783+phiharri@users.noreply.github.com> Date: Wed, 24 Jan 2024 03:00:38 +0000 Subject: [PATCH 057/624] fix: GGUF metadata KV overrides, re #1011 (#1116) * kv overrides another attempt * add sentinel element, simplify array population * ensure sentinel element is zeroed --- llama_cpp/llama.py | 21 +++++++++------------ 1 file changed, 9 insertions(+), 12 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 5c66bcf09..538781dc0 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -194,16 +194,16 @@ def __init__( self.model_params.use_mmap = use_mmap if lora_path is None else False self.model_params.use_mlock = use_mlock + # kv_overrides is the original python dict self.kv_overrides = kv_overrides if kv_overrides is not None: - n_overrides = len(kv_overrides) - self._kv_overrides_array = llama_cpp.llama_model_kv_override * (n_overrides + 1) - self._kv_overrides_array_keys = [] - - for k, v in kv_overrides.items(): - key_buf = ctypes.create_string_buffer(k.encode("utf-8")) - self._kv_overrides_array_keys.append(key_buf) - self._kv_overrides_array[i].key = key_buf + + # _kv_overrides_array is a ctypes.Array of llama_model_kv_override Structs + kvo_array_len = len(kv_overrides) + 1 # for sentinel element + self._kv_overrides_array = (llama_cpp.llama_model_kv_override * kvo_array_len)() + + for i, (k, v) in enumerate(kv_overrides.items()): + self._kv_overrides_array[i].key = k.encode('utf-8'); if isinstance(v, int): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT self._kv_overrides_array[i].value.int_value = v @@ -216,10 +216,7 @@ def __init__( else: raise ValueError(f"Unknown value type for {k}: {v}") - self._kv_overrides_array_sentinel_key = b'\0' - - # null array sentinel - self._kv_overrides_array[n_overrides].key = self._kv_overrides_array_sentinel_key + self._kv_overrides_array[-1].key = b'\0' # ensure sentinel element is zeroed self.model_params.kv_overrides = self._kv_overrides_array self.n_batch = min(n_ctx, n_batch) # ??? From 4d6b2f7b91a8dfd4b4e283ea73492772b3471afe Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Jan 2024 22:08:27 -0500 Subject: [PATCH 058/624] fix: format --- llama_cpp/llama.py | 21 +++++++++++++-------- 1 file changed, 13 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 538781dc0..3d1580013 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -197,13 +197,14 @@ def __init__( # kv_overrides is the original python dict self.kv_overrides = kv_overrides if kv_overrides is not None: - # _kv_overrides_array is a ctypes.Array of llama_model_kv_override Structs - kvo_array_len = len(kv_overrides) + 1 # for sentinel element - self._kv_overrides_array = (llama_cpp.llama_model_kv_override * kvo_array_len)() + kvo_array_len = len(kv_overrides) + 1 # for sentinel element + self._kv_overrides_array = ( + llama_cpp.llama_model_kv_override * kvo_array_len + )() for i, (k, v) in enumerate(kv_overrides.items()): - self._kv_overrides_array[i].key = k.encode('utf-8'); + self._kv_overrides_array[i].key = k.encode("utf-8") if isinstance(v, int): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT self._kv_overrides_array[i].value.int_value = v @@ -216,7 +217,9 @@ def __init__( else: raise ValueError(f"Unknown value type for {k}: {v}") - self._kv_overrides_array[-1].key = b'\0' # ensure sentinel element is zeroed + self._kv_overrides_array[ + -1 + ].key = b"\0" # ensure sentinel element is zeroed self.model_params.kv_overrides = self._kv_overrides_array self.n_batch = min(n_ctx, n_batch) # ??? @@ -326,7 +329,9 @@ def __init__( (n_ctx, self._n_vocab), dtype=np.single ) - self._mirostat_mu = ctypes.c_float(2.0 * 5.0) # TODO: Move this to sampling context + self._mirostat_mu = ctypes.c_float( + 2.0 * 5.0 + ) # TODO: Move this to sampling context try: self.metadata = self._model.metadata() @@ -334,7 +339,7 @@ def __init__( self.metadata = {} if self.verbose: print(f"Failed to load metadata: {e}", file=sys.stderr) - + if self.verbose: print(f"Model metadata: {self.metadata}", file=sys.stderr) @@ -534,7 +539,7 @@ def sample( candidates=self._candidates, tau=mirostat_tau, eta=mirostat_eta, - mu=ctypes.pointer(self._mirostat_mu) + mu=ctypes.pointer(self._mirostat_mu), ) else: self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1) From 9677a1f2c895979c8a8fc2b68002fca4bac8c5fb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Jan 2024 22:28:03 -0500 Subject: [PATCH 059/624] fix: Check order --- llama_cpp/llama.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3d1580013..74739cbde 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -205,15 +205,15 @@ def __init__( for i, (k, v) in enumerate(kv_overrides.items()): self._kv_overrides_array[i].key = k.encode("utf-8") - if isinstance(v, int): + if isinstance(v, bool): + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL + self._kv_overrides_array[i].value.bool_value = v + elif isinstance(v, int): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT self._kv_overrides_array[i].value.int_value = v elif isinstance(v, float): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_FLOAT self._kv_overrides_array[i].value.float_value = v - elif isinstance(v, bool): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL - self._kv_overrides_array[i].value.bool_value = v else: raise ValueError(f"Unknown value type for {k}: {v}") From c970d41a85381fd55235136f123422df0bf0c7e7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 24 Jan 2024 10:38:30 -0500 Subject: [PATCH 060/624] fix: llama_log_set should be able to accept null pointer --- llama_cpp/llama_cpp.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 5de837f87..d31a5da33 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2528,7 +2528,7 @@ def llama_print_system_info() -> bytes: # // If this is not called, or NULL is supplied, everything is output on stderr. # LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); def llama_log_set( - log_callback: "ctypes._FuncPointer", user_data: c_void_p # type: ignore + log_callback: Union["ctypes._FuncPointer", c_void_p], user_data: c_void_p # type: ignore ): """Set callback for all future logging events. @@ -2536,7 +2536,7 @@ def llama_log_set( return _lib.llama_log_set(log_callback, user_data) -_lib.llama_log_set.argtypes = [llama_log_callback, c_void_p] +_lib.llama_log_set.argtypes = [ctypes.c_void_p, c_void_p] _lib.llama_log_set.restype = None From c343baaba83595fb7c6aa90e6aa8c3a6507d9430 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 24 Jan 2024 10:40:50 -0500 Subject: [PATCH 061/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 26d607608..c9b316c78 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 26d607608d794efa56df3bdb6043a2f94c1d632c +Subproject commit c9b316c78fba31e65879a2ec91cbafd341b88cce From 5b258bf840992f20ab6748ada2c2d122247169dd Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 24 Jan 2024 10:51:15 -0500 Subject: [PATCH 062/624] docs: Update README with more param common examples --- README.md | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index f97ea0f77..f602b1739 100644 --- a/README.md +++ b/README.md @@ -104,6 +104,7 @@ CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python ### Windows Notes If you run into issues where it complains it can't find `'nmake'` `'?'` or CMAKE_C_COMPILER, you can extract w64devkit as [mentioned in llama.cpp repo](https://github.com/ggerganov/llama.cpp#openblas) and add those manually to CMAKE_ARGS before running `pip` install: + ```ps $env:CMAKE_GENERATOR = "MinGW Makefiles" $env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" @@ -118,17 +119,19 @@ Detailed MacOS Metal GPU install documentation is available at [docs/install/mac #### M1 Mac Performance Issue Note: If you are using Apple Silicon (M1) Mac, make sure you have installed a version of Python that supports arm64 architecture. For example: -``` + +```bash wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh bash Miniforge3-MacOSX-arm64.sh ``` + Otherwise, while installing it will build the llama.cpp x86 version which will be 10x slower on Apple Silicon (M1) Mac. #### M Series Mac Error: `(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64'))` Try installing with -``` +```bash CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on" pip install --upgrade --verbose --force-reinstall --no-cache-dir llama-cpp-python ``` @@ -152,7 +155,12 @@ Below is a short example demonstrating how to use the high-level API to for basi ```python >>> from llama_cpp import Llama ->>> llm = Llama(model_path="./models/7B/llama-model.gguf") +>>> llm = Llama( + model_path="./models/7B/llama-model.gguf", + # n_gpu_layers=-1, # Uncomment to use GPU acceleration + # seed=1337, # Uncomment to set a specific seed + # n_ctx=2048, # Uncomment to increase the context window +) >>> output = llm( "Q: Name the planets in the solar system? A: ", # Prompt max_tokens=32, # Generate up to 32 tokens @@ -191,7 +199,10 @@ Note that `chat_format` option must be set for the particular model you are usin ```python >>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/llama-2/llama-model.gguf", chat_format="llama-2") +>>> llm = Llama( + model_path="path/to/llama-2/llama-model.gguf", + chat_format="llama-2" +) >>> llm.create_chat_completion( messages = [ {"role": "system", "content": "You are an assistant who perfectly describes images."}, From d6fb16e05524eb99a28436e73d7f45b1ca6f8b6d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Jan 2024 10:51:48 -0500 Subject: [PATCH 063/624] docs: Update README --- README.md | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index f602b1739..7813c96b0 100644 --- a/README.md +++ b/README.md @@ -163,7 +163,7 @@ Below is a short example demonstrating how to use the high-level API to for basi ) >>> output = llm( "Q: Name the planets in the solar system? A: ", # Prompt - max_tokens=32, # Generate up to 32 tokens + max_tokens=32, # Generate up to 32 tokens, set to None to generate up to the end of the context window stop=["Q:", "\n"], # Stop generating just before the model would generate a new question echo=True # Echo the prompt back in the output ) # Generate a completion, can also call create_completion @@ -425,6 +425,9 @@ pip install -e .[all] make clean ``` +You can also test out specific commits of `lama.cpp` by checking out the desired commit in the `vendor/llama.cpp` submodule and then running `make clean` and `pip install -e .` again. Any changes in the `llama.h` API will require +changes to the `llama_cpp/llama_cpp.py` file to match the new API (additional changes may be required elsewhere). + ## FAQ ### Are there pre-built binaries / binary wheels available? From dc5a436224c15c2a985b4e5d43b37d0df2d63b6e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Jan 2024 11:19:34 -0500 Subject: [PATCH 064/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c9b316c78..ddc5a5033 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c9b316c78fba31e65879a2ec91cbafd341b88cce +Subproject commit ddc5a5033f948dc7ab0a3a6ec2d914d13c274077 From 2588f34a22076525ea62e727fb292d07d20f463a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Jan 2024 11:22:42 -0500 Subject: [PATCH 065/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ddc5a5033..faa3526a1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ddc5a5033f948dc7ab0a3a6ec2d914d13c274077 +Subproject commit faa3526a1eba458120987ed8269e5616385a76f4 From cde7514c3d28e6d52f272614e9957208c344dde5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Jan 2024 11:23:18 -0500 Subject: [PATCH 066/624] feat(server): include llama-cpp-python version in openapi spec --- llama_cpp/server/app.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index fed0a6deb..368022c45 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -118,7 +118,7 @@ def create_app( app = FastAPI( middleware=middleware, title="🦙 llama.cpp Python API", - version="0.0.1", + version=llama_cpp.__version__, ) app.add_middleware( CORSMiddleware, From f5cc6b30538c34192227d3c46f08f213dce6443c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Jan 2024 11:28:16 -0500 Subject: [PATCH 067/624] Bump version --- CHANGELOG.md | 8 ++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 4fff9194d..8a94ef554 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.33] + +- feat: Update llama.cpp to ggerganov/llama.cpp@faa3526a1eba458120987ed8269e5616385a76f4 +- feat(server): include llama-cpp-python version in openapi spec by @abetlen in cde7514c3d28e6d52f272614e9957208c344dde5 +- fix: use both eos and bos tokens as stop sequences for hf-tokenizer-config chat format. by @abetlen in 5b982d0f8c6f35242c8862ffdce00e17cea0b44f +- fix: GGUF metadata KV overrides, re #1011 by @phiharri in #1116 +- fix: llama_log_set should be able to accept null pointer by @abetlen in c970d41a85381fd55235136f123422df0bf0c7e7 + ## [0.2.32] - feat: Update llama.cpp to ggerganov/llama.cpp@504dc37be8446fb09b1ede70300250ad41be32a2 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index dda8335b5..55f695e4a 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.32" \ No newline at end of file +__version__ = "0.2.33" \ No newline at end of file From 35918873b4010a230a9aa478fd16f35127d7eb9a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 26 Jan 2024 11:45:48 -0500 Subject: [PATCH 068/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 32 +++++++++++++++++++++++++++++--- vendor/llama.cpp | 2 +- 2 files changed, 30 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d31a5da33..c4256dd4c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -93,9 +93,7 @@ def _load_shared_library(lib_base_name: str): # from ggml-backend.h # typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); -ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE( - c_bool, c_void_p, c_bool, c_void_p -) +ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE(c_bool, c_void_p, c_bool, c_void_p) # llama.h bindings @@ -2174,6 +2172,34 @@ def llama_sample_typical( _lib.llama_sample_typical.restype = None +# /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772. +# LLAMA_API void llama_sample_entropy( +# struct llama_context * ctx, +# llama_token_data_array * candidates_p, +# float min_temp, +# float max_temp, +# float exponent_val); +def llama_sample_entropy( + ctx: llama_context_p, + candidates, # type: _Pointer[llama_token_data_array] + min_temp: Union[c_float, float], + max_temp: Union[c_float, float], + exponent_val: Union[c_float, float], +): + """Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.""" + return _lib.llama_sample_entropy(ctx, candidates, min_temp, max_temp, exponent_val) + + +_lib.llama_sample_entropy.argtypes = [ + llama_context_p, + llama_token_data_array_p, + c_float, + c_float, + c_float, +] +_lib.llama_sample_entropy.restype = None + + # LLAMA_API void llama_sample_temp( # struct llama_context * ctx, # llama_token_data_array * candidates, diff --git a/vendor/llama.cpp b/vendor/llama.cpp index faa3526a1..5f1925a8c 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit faa3526a1eba458120987ed8269e5616385a76f4 +Subproject commit 5f1925a8cef81eb9b372faaae34b0dd76d5361d4 From c6d3bd62e8db86d7c9234d763f7b7f94e6aa8cc0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 27 Jan 2024 16:22:46 -0500 Subject: [PATCH 069/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5f1925a8c..6db2b41a7 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5f1925a8cef81eb9b372faaae34b0dd76d5361d4 +Subproject commit 6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 From d8f6914f459c1a8536ace5b379492a482fa0db16 Mon Sep 17 00:00:00 2001 From: Andrei Date: Sat, 27 Jan 2024 16:52:18 -0500 Subject: [PATCH 070/624] Add json schema mode (#1122) * Add json schema mode * Add llava chat format support --- llama_cpp/llama_chat_format.py | 21 ++++++++++++++++----- llama_cpp/llama_types.py | 1 + 2 files changed, 17 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 6c274aa82..e418d40bc 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -318,7 +318,14 @@ def chat_completion_handler( stop = stop + rstop if response_format is not None and response_format["type"] == "json_object": - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + try: + # create grammar from json schema + if "schema" in response_format: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(response_format["schema"]) + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) completion_or_chunks = llama.create_completion( prompt=prompt, @@ -1434,10 +1441,14 @@ def __call__( prompt = llama.input_ids[: llama.n_tokens].tolist() if response_format is not None and response_format["type"] == "json_object": - with suppress_stdout_stderr(disable=self.verbose): - grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF - ) + try: + # create grammar from json schema + if "schema" in response_format: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(response_format["schema"]) + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) return _convert_completion_to_chat( llama.create_completion( diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 5b51e98ce..c3deba87b 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -154,6 +154,7 @@ class ChatCompletionFunctionCallOption(TypedDict): class ChatCompletionRequestResponseFormat(TypedDict): type: Literal["text", "json_object"] + schema: NotRequired[JsonType] # https://docs.endpoints.anyscale.com/guides/json_mode/ class ChatCompletionRequestMessageContentPartText(TypedDict): From c1d0fff8a990b1e3ebede2f2abf0f6350b2c30a3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 27 Jan 2024 18:36:56 -0500 Subject: [PATCH 071/624] Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 8a94ef554..ddf1d0d6d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.34] + +- feat: Update llama.cpp to ggerganov/llama.cpp@6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 +- feat: Add json schema mode by @abetlen in #1122 + ## [0.2.33] - feat: Update llama.cpp to ggerganov/llama.cpp@faa3526a1eba458120987ed8269e5616385a76f4 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 55f695e4a..36eacabdc 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.33" \ No newline at end of file +__version__ = "0.2.34" \ No newline at end of file From 399fa1e03b3b966105669cb2b698ef7e440051f9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 27 Jan 2024 19:36:33 -0500 Subject: [PATCH 072/624] docs: Add JSON and JSON schema mode examples to README --- README.md | 53 +++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 53 insertions(+) diff --git a/README.md b/README.md index 7813c96b0..f5cd50e88 100644 --- a/README.md +++ b/README.md @@ -216,6 +216,59 @@ Note that `chat_format` option must be set for the particular model you are usin Chat completion is available through the [`create_chat_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion) method of the [`Llama`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama) class. +### JSON and JSON Schema Mode + +If you want to constrain chat responses to only valid JSON or a specific JSON Schema you can use the `response_format` argument to the `create_chat_completion` method. + +#### Json Mode + +The following example will constrain the response to be valid JSON. + +```python +>>> from llama_cpp import Llama +>>> llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") +>>> llm.create_chat_completion( + messages=[ + { + "role": "system", + "content": "You are a helpful assistant that outputs in JSON.", + }, + {"role": "user", "content": "Who won the world series in 2020"}, + ], + response_format={ + "type": "json_object", + }, + temperature=0.7, +) +``` + +#### Json Mode + +To constrain the response to a specific JSON Schema, you can use the `schema` property of the `response_format` argument. + +```python +>>> from llama_cpp import Llama +>>> llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") +>>> llm.create_chat_completion( + messages=[ + { + "role": "system", + "content": "You are a helpful assistant that outputs in JSON.", + }, + {"role": "user", "content": "Who won the world series in 2020"}, + ], + response_format={ + "type": "json_object", + "schema": { + "type": "object", + "properties": {"team_name": {"type": "string"}}, + "required": ["team_name"], + }, + }, + temperature=0.7, +) +``` + ### Function Calling The high-level API also provides a simple interface for function calling. From 8c592100623ec08583c34190a15d4c1d07945b25 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 27 Jan 2024 19:37:59 -0500 Subject: [PATCH 073/624] docs: Fix typo --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index f5cd50e88..faf4a87a2 100644 --- a/README.md +++ b/README.md @@ -220,7 +220,7 @@ Chat completion is available through the [`create_chat_completion`](https://llam If you want to constrain chat responses to only valid JSON or a specific JSON Schema you can use the `response_format` argument to the `create_chat_completion` method. -#### Json Mode +#### JSON Mode The following example will constrain the response to be valid JSON. @@ -242,7 +242,7 @@ The following example will constrain the response to be valid JSON. ) ``` -#### Json Mode +#### JSON Schema Mode To constrain the response to a specific JSON Schema, you can use the `schema` property of the `response_format` argument. From ccf4908bfd8a3cefa0f554a11a3d16fe0d979980 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 28 Jan 2024 12:55:32 -0500 Subject: [PATCH 074/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6db2b41a7..35dec26cc 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 +Subproject commit 35dec26cc25a9ff7d8c3ed52326b94f772b911ce From 31e0288a410b054a0b40aaa2b54b635d9a261f8e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 28 Jan 2024 19:34:27 -0500 Subject: [PATCH 075/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 35dec26cc..d2f650cb5 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 35dec26cc25a9ff7d8c3ed52326b94f772b911ce +Subproject commit d2f650cb5b04ee2726663e79b47da5efe196ce00 From 52c4a84faf07d92565e0c4f77ba884a48fc24b52 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 28 Jan 2024 19:35:37 -0500 Subject: [PATCH 076/624] Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index ddf1d0d6d..20671f845 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.35] + +- feat: Update llama.cpp to ggerganov/llama.cpp@d2f650cb5b04ee2726663e79b47da5efe196ce00 + ## [0.2.34] - feat: Update llama.cpp to ggerganov/llama.cpp@6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 36eacabdc..639a5a507 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.34" \ No newline at end of file +__version__ = "0.2.35" \ No newline at end of file From ce38dbdf0730aa2fb3a659e0a5bda2e8ad07da17 Mon Sep 17 00:00:00 2001 From: Rafaelblsilva Date: Mon, 29 Jan 2024 02:34:42 -0300 Subject: [PATCH 077/624] Add mistral instruct chat format as "mistral-instruct" (#799) * Added mistral instruct chat format as "mistral" * Fix stop sequence (merge issue) * Update chat format name to `mistral-instruct` --------- Co-authored-by: Andrei --- llama_cpp/llama_chat_format.py | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index e418d40bc..989275a27 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -877,6 +877,22 @@ def format_chatml( return ChatFormatterResponse(prompt=_prompt, stop=_sep) +@register_chat_format("mistral-instruct") +def format_mistral( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + _roles = dict(user="[INST] ", assistant="[/INST]") + _sep = " " + system_template = """{system_message}""" + system_message = _get_system_message(messages) + system_message = system_template.format(system_message=system_message) + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_no_colon_single(system_message, _messages, _sep) + return ChatFormatterResponse(prompt=_prompt) + + @register_chat_format("chatglm3") def format_chatglm3( messages: List[llama_types.ChatCompletionRequestMessage], From 9ae5819ee4a3d892c70e6a2f52de76f1882edcc0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Jan 2024 00:59:01 -0500 Subject: [PATCH 078/624] Add chat format test. --- llama_cpp/llama_chat_format.py | 22 ++++++++++++---------- tests/test_llama_chat_format.py | 23 +++++++++++++++++++++++ 2 files changed, 35 insertions(+), 10 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 989275a27..5466de3a4 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -878,19 +878,21 @@ def format_chatml( @register_chat_format("mistral-instruct") -def format_mistral( +def format_mistral_instruct( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - _roles = dict(user="[INST] ", assistant="[/INST]") - _sep = " " - system_template = """{system_message}""" - system_message = _get_system_message(messages) - system_message = system_template.format(system_message=system_message) - _messages = _map_roles(messages, _roles) - _messages.append((_roles["assistant"], None)) - _prompt = _format_no_colon_single(system_message, _messages, _sep) - return ChatFormatterResponse(prompt=_prompt) + bos = "" + eos = "" + stop = eos + prompt = bos + for message in messages: + if message["role"] == "user" and message["content"] is not None and isinstance(message["content"], str): + prompt += "[INST] " + message["content"] + elif message["role"] == "assistant" and message["content"] is not None and isinstance(message["content"], str): + prompt += " [/INST]" + message["content"] + eos + prompt += " [/INST]" + return ChatFormatterResponse(prompt=prompt, stop=stop) @register_chat_format("chatglm3") diff --git a/tests/test_llama_chat_format.py b/tests/test_llama_chat_format.py index 1ef18d927..c10aee42e 100644 --- a/tests/test_llama_chat_format.py +++ b/tests/test_llama_chat_format.py @@ -1,10 +1,33 @@ import json +import jinja2 + from llama_cpp import ( ChatCompletionRequestUserMessage, ) +import llama_cpp.llama_types as llama_types +import llama_cpp.llama_chat_format as llama_chat_format + from llama_cpp.llama_chat_format import hf_tokenizer_config_to_chat_formatter +def test_mistral_instruct(): + chat_template = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" + chat_formatter = jinja2.Template(chat_template) + messages = [ + llama_types.ChatCompletionRequestUserMessage(role="user", content="Instruction"), + llama_types.ChatCompletionRequestAssistantMessage(role="assistant", content="Model answer"), + llama_types.ChatCompletionRequestUserMessage(role="user", content="Follow-up instruction"), + ] + response = llama_chat_format.format_mistral_instruct( + messages=messages, + ) + reference = chat_formatter.render( + messages=messages, + bos_token="", + eos_token="", + ) + assert response.prompt == reference + mistral_7b_tokenizer_config = """{ "add_bos_token": true, From 85f8c4c06e33d5ea7df35e662412d36441a65456 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Jan 2024 10:39:08 -0500 Subject: [PATCH 079/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index d2f650cb5..2aed77eb0 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit d2f650cb5b04ee2726663e79b47da5efe196ce00 +Subproject commit 2aed77eb06a329f0d82bb1c467f4244904d4073f From 9f7852acfafee8aee9dd19baf3528ede7b45b881 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Jan 2024 10:39:23 -0500 Subject: [PATCH 080/624] misc: Add vulkan target --- Makefile | 3 +++ 1 file changed, 3 insertions(+) diff --git a/Makefile b/Makefile index 5ed3fa262..806b120c5 100644 --- a/Makefile +++ b/Makefile @@ -27,6 +27,9 @@ build.blis: build.metal: CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install --verbose -e . +build.vulkan: + CMAKE_ARGS="-DLLAMA_VULKAN=on" python3 -m pip install --verbose -e . + build.sdist: python3 -m build --sdist From 464af5b39fea3cf1ba16e755a9df85f09bbb25ac Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Jan 2024 10:46:04 -0500 Subject: [PATCH 081/624] Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 20671f845..e99dd1767 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.36] + +- feat: Update llama.cpp to ggerganov/llama.cpp@2aed77eb06a329f0d82bb1c467f4244904d4073f +- feat: Add mistral instruct chat format as "mistral-instruct" by @Rafaelblsilva in #799 + ## [0.2.35] - feat: Update llama.cpp to ggerganov/llama.cpp@d2f650cb5b04ee2726663e79b47da5efe196ce00 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 639a5a507..f73f3d42f 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.35" \ No newline at end of file +__version__ = "0.2.36" \ No newline at end of file From 843e77e3e2310b1cfdc6b41982aec9007312e142 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Jan 2024 11:01:26 -0500 Subject: [PATCH 082/624] docs: Add Vulkan build instructions --- README.md | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/README.md b/README.md index faf4a87a2..05326ac90 100644 --- a/README.md +++ b/README.md @@ -101,6 +101,14 @@ To install with hipBLAS / ROCm support for AMD cards, set the `LLAMA_HIPBLAS=on` CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python ``` +#### Vulkan + +To install with Vulkan support, set the `LLAMA_VULKAN=on` environment variable before installing: + +```bash +CMAKE_ARGS="-DLLAMA_VULKAN=on" pip install llama-cpp-python +``` + ### Windows Notes If you run into issues where it complains it can't find `'nmake'` `'?'` or CMAKE_C_COMPILER, you can extract w64devkit as [mentioned in llama.cpp repo](https://github.com/ggerganov/llama.cpp#openblas) and add those manually to CMAKE_ARGS before running `pip` install: From 059f6b3ac8bcc95389d487fecbab0fbe38eb798e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Jan 2024 11:02:25 -0500 Subject: [PATCH 083/624] docs: fix typos --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 05326ac90..35c0c478a 100644 --- a/README.md +++ b/README.md @@ -71,7 +71,7 @@ CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp- #### cuBLAS -To install with cuBLAS, set the `LLAMA_CUBLAS=1` environment variable before installing: +To install with cuBLAS, set the `LLAMA_CUBLAS=on` environment variable before installing: ```bash CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python @@ -87,7 +87,7 @@ CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python #### CLBlast -To install with CLBlast, set the `LLAMA_CLBLAST=1` environment variable before installing: +To install with CLBlast, set the `LLAMA_CLBLAST=on` environment variable before installing: ```bash CMAKE_ARGS="-DLLAMA_CLBLAST=on" pip install llama-cpp-python From da003d87681f02475eedb6937443e5f07db889b0 Mon Sep 17 00:00:00 2001 From: Andrei Date: Mon, 29 Jan 2024 14:22:23 -0500 Subject: [PATCH 084/624] Automatically set chat format from gguf (#1110) * Use jinja formatter to load chat format from gguf * Fix off-by-one error in metadata loader * Implement chat format auto-detection --- llama_cpp/_internals.py | 4 ++-- llama_cpp/llama.py | 37 +++++++++++++++++++++++++++++++++- llama_cpp/llama_chat_format.py | 30 +++++++++++++++++++++++++-- llama_cpp/server/settings.py | 4 ++-- 4 files changed, 68 insertions(+), 7 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index ec47c4216..651cd4ccf 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -216,13 +216,13 @@ def metadata(self) -> Dict[str, str]: for i in range(llama_cpp.llama_model_meta_count(self.model)): nbytes = llama_cpp.llama_model_meta_key_by_index(self.model, i, buffer, buffer_size) if nbytes > buffer_size: - buffer_size = nbytes + buffer_size = nbytes + 1 buffer = ctypes.create_string_buffer(buffer_size) nbytes = llama_cpp.llama_model_meta_key_by_index(self.model, i, buffer, buffer_size) key = buffer.value.decode("utf-8") nbytes = llama_cpp.llama_model_meta_val_str_by_index(self.model, i, buffer, buffer_size) if nbytes > buffer_size: - buffer_size = nbytes + buffer_size = nbytes + 1 buffer = ctypes.create_string_buffer(buffer_size) nbytes = llama_cpp.llama_model_meta_val_str_by_index(self.model, i, buffer, buffer_size) value = buffer.value.decode("utf-8") diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 74739cbde..b5618c10d 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -87,7 +87,7 @@ def __init__( # Backend Params numa: bool = False, # Chat Format Params - chat_format: str = "llama-2", + chat_format: Optional[str] = None, chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None, # Misc verbose: bool = True, @@ -343,6 +343,41 @@ def __init__( if self.verbose: print(f"Model metadata: {self.metadata}", file=sys.stderr) + if self.chat_format is None and self.chat_handler is None and "tokenizer.chat_template" in self.metadata: + chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata(self.metadata) + + if chat_format is not None: + self.chat_format = chat_format + if self.verbose: + print(f"Guessed chat format: {chat_format}", file=sys.stderr) + else: + template = self.metadata["tokenizer.chat_template"] + try: + eos_token_id = int(self.metadata["tokenizer.ggml.eos_token_id"]) + except: + eos_token_id = self.token_eos() + try: + bos_token_id = int(self.metadata["tokenizer.ggml.bos_token_id"]) + except: + bos_token_id = self.token_bos() + + eos_token = self.detokenize([eos_token_id]).decode("utf-8") + bos_token = self.detokenize([bos_token_id]).decode("utf-8") + + if self.verbose: + print(f"Using chat template: {template}", file=sys.stderr) + print(f"Using chat eos_token: {eos_token}", file=sys.stderr) + print(f"Using chat bos_token: {bos_token}", file=sys.stderr) + + self.chat_handler = llama_chat_format.Jinja2ChatFormatter( + template=template, + eos_token=eos_token, + bos_token=bos_token + ).to_chat_handler() + + if self.chat_format is None and self.chat_handler is None: + self.chat_format = "llama-2" + @property def ctx(self) -> llama_cpp.llama_context_p: assert self._ctx.ctx is not None diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 5466de3a4..4bc4a6c97 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -14,6 +14,20 @@ from ._utils import suppress_stdout_stderr, Singleton +### Common Chat Templates and Special Tokens ### + +# Source: https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/tokenizer_config.json +CHATML_CHAT_TEMPLATE = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" +CHATML_BOS_TOKEN = "" +CHATML_EOS_TOKEN = "<|im_end|>" + +# Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/blob/main/tokenizer_config.json +MISTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token + ' ' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" +MISTRAL_INSTRUCT_BOS_TOKEN = "" +MISTRAL_INSTRUCT_EOS_TOKEN = "" + + +### Chat Completion Handler ### class LlamaChatCompletionHandler(Protocol): """Base Protocol for a llama chat completion handler. @@ -118,7 +132,6 @@ def decorator(f: LlamaChatCompletionHandler): ### Chat Formatter ### - @dataclasses.dataclass class ChatFormatterResponse: """Dataclass that stores completion parameters for a given chat format and @@ -440,7 +453,20 @@ def hf_tokenizer_config_to_chat_completion_handler( return chat_formatter_to_chat_completion_handler(chat_formatter) +def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[str]: + if "tokenizer.chat_template" not in metadata: + return None + + if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE: + return "chatml" + + if metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE: + return "mistral-instruct" + + return None + ### Utility functions for formatting chat prompts ### +# TODO: Replace these with jinja2 templates def _get_system_message( @@ -929,7 +955,6 @@ def format_openchat( _prompt = _format_chatml(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt, stop=_sep) - # Chat format for Saiga models, see more details and available models: # https://huggingface.co/collections/IlyaGusev/saiga2-saigamistral-6505d4ccc3d1e53166b636cd @register_chat_format("saiga") @@ -951,6 +976,7 @@ def format_saiga( _prompt += "bot" return ChatFormatterResponse(prompt=_prompt.strip()) +# Tricky chat formats that require custom chat handlers @register_chat_completion_handler("functionary") def functionary_chat_handler( diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 9f0dc8a73..9fe1a7bfd 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -113,8 +113,8 @@ class ModelSettings(BaseSettings): description="Enable NUMA support.", ) # Chat Format Params - chat_format: str = Field( - default="llama-2", + chat_format: Optional[str] = Field( + default=None, description="Chat format to use.", ) clip_model_path: Optional[str] = Field( From 011cd84ded5a2150dd66e9a1bdcecdc5142112eb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Jan 2024 09:48:09 -0500 Subject: [PATCH 085/624] Update llama.cpp --- Makefile | 6 ++++++ llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 3 files changed, 9 insertions(+), 1 deletion(-) diff --git a/Makefile b/Makefile index 806b120c5..ff1484c2b 100644 --- a/Makefile +++ b/Makefile @@ -30,6 +30,12 @@ build.metal: build.vulkan: CMAKE_ARGS="-DLLAMA_VULKAN=on" python3 -m pip install --verbose -e . +build.kompute: + CMAKE_ARGS="-DLLAMA_KOMPUTE=on" python3 -m pip install --verbose -e . + +build.sycl: + CMAKE_ARGS="-DLLAMA_SYCL=on" python3 -m pip install --verbose -e . + build.sdist: python3 -m build --sdist diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index c4256dd4c..2168579c8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -187,6 +187,7 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -211,6 +212,7 @@ def _load_shared_library(lib_base_name: str): LLAMA_FTYPE_MOSTLY_IQ2_XS = 20 LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22 +LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 2aed77eb0..8f8ddfcfa 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 2aed77eb06a329f0d82bb1c467f4244904d4073f +Subproject commit 8f8ddfcfadc830b936318c3ea9fe2e8e3365aa85 From 13b7ced7dab618b729a486750c2e18d64d34bbd1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Jan 2024 12:21:41 -0500 Subject: [PATCH 086/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8f8ddfcfa..fea4fd4ba 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8f8ddfcfadc830b936318c3ea9fe2e8e3365aa85 +Subproject commit fea4fd4ba7f6b754ac795387b275e1a014a77bde From 247a16de66112b5c4924c934eb40abd139ec7fac Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Jan 2024 12:23:07 -0500 Subject: [PATCH 087/624] docs: Update README --- README.md | 47 ++++++++++++++++++++++++++++------------------- 1 file changed, 28 insertions(+), 19 deletions(-) diff --git a/README.md b/README.md index 35c0c478a..0a77bbdaa 100644 --- a/README.md +++ b/README.md @@ -12,20 +12,17 @@ This package provides: - Low-level access to C API via `ctypes` interface. - High-level Python API for text completion - - OpenAI-like API - - [LangChain compatibility](https://python.langchain.com/docs/integrations/llms/llamacpp) - - [LlamaIndex compatibility](https://docs.llamaindex.ai/en/stable/examples/llm/llama_2_llama_cpp.html) + - OpenAI-like API + - [LangChain compatibility](https://python.langchain.com/docs/integrations/llms/llamacpp) + - [LlamaIndex compatibility](https://docs.llamaindex.ai/en/stable/examples/llm/llama_2_llama_cpp.html) - OpenAI compatible web server - - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) - - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) - - [Vision API support](https://llama-cpp-python.readthedocs.io/en/latest/server/#multimodal-models) - - [Multiple Models](https://llama-cpp-python.readthedocs.io/en/latest/server/#configuration-and-multi-model-support) + - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) + - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) + - [Vision API support](https://llama-cpp-python.readthedocs.io/en/latest/server/#multimodal-models) + - [Multiple Models](https://llama-cpp-python.readthedocs.io/en/latest/server/#configuration-and-multi-model-support) Documentation is available at [https://llama-cpp-python.readthedocs.io/en/latest](https://llama-cpp-python.readthedocs.io/en/latest). - - - ## Installation `llama-cpp-python` can be installed directly from PyPI as a source distribution by running: @@ -38,7 +35,6 @@ This will build `llama.cpp` from source using cmake and your system's c compiler If you run into issues during installation add the `--verbose` flag to the `pip install` command to see the full cmake build log. - ### Installation with Specific Hardware Acceleration (BLAS, CUDA, Metal, etc) The default pip install behaviour is to build `llama.cpp` for CPU only on Linux and Windows and use Metal on MacOS. @@ -109,13 +105,29 @@ To install with Vulkan support, set the `LLAMA_VULKAN=on` environment variable b CMAKE_ARGS="-DLLAMA_VULKAN=on" pip install llama-cpp-python ``` +#### Kompute + +To install with Kompute support, set the `LLAMA_KOMPUTE=on` environment variable before installing: + +```bash +CMAKE_ARGS="-DLLAMA_KOMPUTE=on" pip install llama-cpp-python +``` + +#### SYCL + +To install with SYCL support, set the `LLAMA_SYCL=on` environment variable before installing: + +```bash +CMAKE_ARGS="-DLLAMA_SYCL=on" pip install llama-cpp-python +``` + ### Windows Notes If you run into issues where it complains it can't find `'nmake'` `'?'` or CMAKE_C_COMPILER, you can extract w64devkit as [mentioned in llama.cpp repo](https://github.com/ggerganov/llama.cpp#openblas) and add those manually to CMAKE_ARGS before running `pip` install: ```ps $env:CMAKE_GENERATOR = "MinGW Makefiles" -$env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" +$env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" ``` See the above instructions and set `CMAKE_ARGS` to the BLAS backend you want to use. @@ -165,7 +177,7 @@ Below is a short example demonstrating how to use the high-level API to for basi >>> from llama_cpp import Llama >>> llm = Llama( model_path="./models/7B/llama-model.gguf", - # n_gpu_layers=-1, # Uncomment to use GPU acceleration + # n_gpu_layers=-1, # Uncomment to use GPU acceleration # seed=1337, # Uncomment to set a specific seed # n_ctx=2048, # Uncomment to increase the context window ) @@ -284,7 +296,6 @@ The high-level API also provides a simple interface for function calling. Note that the only model that supports full function calling at this time is "functionary". The gguf-converted files for this model can be found here: [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) - ```python >>> from llama_cpp import Llama >>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", chat_format="functionary") @@ -293,7 +304,7 @@ The gguf-converted files for this model can be found here: [functionary-7b-v1](h { "role": "system", "content": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary" - + }, { "role": "user", @@ -332,7 +343,6 @@ The gguf-converted files for this model can be found here: [functionary-7b-v1](h ### Multi-modal Models - `llama-cpp-python` supports the llava1.5 family of multi-modal models which allow the language model to read information from both text and images. @@ -378,7 +388,6 @@ For instance, if you want to work with larger contexts, you can expand the conte llm = Llama(model_path="./models/7B/llama-model.gguf", n_ctx=2048) ``` - ## OpenAI Compatible Web Server `llama-cpp-python` offers a web server which aims to act as a drop-in replacement for the OpenAI API. @@ -426,7 +435,8 @@ A Docker image is available on [GHCR](https://ghcr.io/abetlen/llama-cpp-python). ```bash docker run --rm -it -p 8000:8000 -v /path/to/models:/models -e MODEL=/models/llama-model.gguf ghcr.io/abetlen/llama-cpp-python:latest ``` -[Docker on termux (requires root)](https://gist.github.com/FreddieOliveira/efe850df7ff3951cb62d74bd770dce27) is currently the only known way to run this on phones, see [termux support issue](https://github.com/abetlen/llama-cpp-python/issues/389) + +[Docker on termux (requires root)](https://gist.github.com/FreddieOliveira/efe850df7ff3951cb62d74bd770dce27) is currently the only known way to run this on phones, see [termux support issue](https://github.com/abetlen/llama-cpp-python/issues/389) ## Low-level API @@ -454,7 +464,6 @@ Below is a short example demonstrating how to use the low-level API to tokenize Check out the [examples folder](examples/low_level_api) for more examples of using the low-level API. - ## Documentation Documentation is available via [https://llama-cpp-python.readthedocs.io/](https://llama-cpp-python.readthedocs.io/). From bf9e824922a3fa95b336ad441eca7e42f9b33358 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Jan 2024 12:27:27 -0500 Subject: [PATCH 088/624] Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index e99dd1767..435af43a8 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.37] + +- feat: Update llama.cpp to ggerganov/llama.cpp@fea4fd4ba7f6b754ac795387b275e1a014a77bde +- feat: Automatically set chat format from gguf by @abetlen in #1110 + ## [0.2.36] - feat: Update llama.cpp to ggerganov/llama.cpp@2aed77eb06a329f0d82bb1c467f4244904d4073f diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index f73f3d42f..4ce899ccd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.36" \ No newline at end of file +__version__ = "0.2.37" \ No newline at end of file From 411494706a29ee4cf3b2a9da26b970794eb18fba Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jan 2024 08:35:21 -0500 Subject: [PATCH 089/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index fea4fd4ba..15606309a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit fea4fd4ba7f6b754ac795387b275e1a014a77bde +Subproject commit 15606309a05ccf7fadbaad5538cb7c32acb1e06b From 078cca0361bf5a94d2cf52ed04980d20e32d6f95 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jan 2024 08:42:21 -0500 Subject: [PATCH 090/624] fix: Pass raise_exception and add_generation_prompt to jinja2 chat template --- llama_cpp/llama_chat_format.py | 17 +++++++++-------- 1 file changed, 9 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 4bc4a6c97..08f991b67 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -185,16 +185,17 @@ def __call__( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - if self.add_generation_prompt: - messages = [ - *messages, - llama_types.ChatCompletionRequestAssistantMessage( - role="assistant", content="" - ), - ] + def raise_exception(message: str): + raise ValueError(message) + prompt = self._environment.render( - messages=messages, eos_token=self.eos_token, bos_token=self.bos_token + messages=messages, + eos_token=self.eos_token, + bos_token=self.bos_token, + raise_exception=raise_exception, + add_generation_prompt=self.add_generation_prompt ) + return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token]) def to_chat_handler(self) -> LlamaChatCompletionHandler: From 2b37d8e438ff6285b586ca92cc165f52d1f5afdc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jan 2024 10:37:19 -0500 Subject: [PATCH 091/624] fix: Run server command. Closes #1143 --- examples/high_level_api/fastapi_server.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/high_level_api/fastapi_server.py b/examples/high_level_api/fastapi_server.py index 4b3189dd1..9421db57b 100644 --- a/examples/high_level_api/fastapi_server.py +++ b/examples/high_level_api/fastapi_server.py @@ -9,7 +9,7 @@ Then run: ``` -uvicorn llama_cpp.server.app:app --reload +uvicorn --factory llama_cpp.server.app:create_app --reload ``` or From 71e3e4c435826677ff671f1d82748fe1dd4d64e1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jan 2024 10:41:42 -0500 Subject: [PATCH 092/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 41 ++++++++++++++++++++++++++++++++++------- vendor/llama.cpp | 2 +- 2 files changed, 35 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 2168579c8..431a99f55 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -98,7 +98,7 @@ def _load_shared_library(lib_base_name: str): # llama.h bindings _lib.llama_max_devices.argtypes = [] -_lib.llama_max_devices.restype = ctypes.c_int32 +_lib.llama_max_devices.restype = ctypes.c_size_t LLAMA_MAX_DEVICES = _lib.llama_max_devices() @@ -390,7 +390,7 @@ class llama_model_kv_override(Structure): # // LLAMA_SPLIT_LAYER: ignored # int32_t main_gpu; -# // proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES +# // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() # const float * tensor_split; # // Called with a progress value between 0.0 and 1.0. Pass NULL to disable. @@ -417,7 +417,7 @@ class llama_model_params(Structure): n_gpu_layers (int): number of layers to store in VRAM split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored - tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES + tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. progress_callback_user_data (ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data @@ -760,16 +760,43 @@ def llama_time_us() -> int: _lib.llama_time_us.restype = ctypes.c_int64 -# LLAMA_API int32_t llama_max_devices(void); +# LLAMA_API size_t llama_max_devices(void); def llama_max_devices() -> int: return _lib.llama_max_devices() _lib.llama_max_devices.argtypes = [] -_lib.llama_max_devices.restype = ctypes.c_int32 +_lib.llama_max_devices.restype = ctypes.c_size_t -# LLAMA_API bool llama_mmap_supported (void); +# LLAMA_API bool llama_supports_mmap (void); +def llama_supports_mmap() -> bool: + return _lib.llama_supports_mmap() + + +_lib.llama_supports_mmap.argtypes = [] +_lib.llama_supports_mmap.restype = c_bool + + +# LLAMA_API bool llama_supports_mlock (void); +def llama_supports_mlock() -> bool: + return _lib.llama_supports_mlock() + + +_lib.llama_supports_mlock.argtypes = [] +_lib.llama_supports_mlock.restype = c_bool + + +# LLAMA_API bool llama_supports_gpu_offload(void); +def llama_supports_gpu_offload() -> bool: + return _lib.llama_supports_gpu_offload() + + +_lib.llama_supports_gpu_offload.argtypes = [] +_lib.llama_supports_gpu_offload.restype = c_bool + + +# LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); def llama_mmap_supported() -> bool: return _lib.llama_mmap_supported() @@ -778,7 +805,7 @@ def llama_mmap_supported() -> bool: _lib.llama_mmap_supported.restype = c_bool -# LLAMA_API bool llama_mlock_supported(void); +# LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); def llama_mlock_supported() -> bool: return _lib.llama_mlock_supported() diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 15606309a..5cb04dbc1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 15606309a05ccf7fadbaad5538cb7c32acb1e06b +Subproject commit 5cb04dbc16d1da38c8fdcc0111b40e67d00dd1c3 From fb762a60411f53278454b4e9888c5bd9712d3779 Mon Sep 17 00:00:00 2001 From: Andrei Date: Wed, 31 Jan 2024 14:08:14 -0500 Subject: [PATCH 093/624] Add speculative decoding (#1120) * Add draft model param to llama class, implement basic prompt lookup decoding draft model * Use samplingcontext for sampling * Use 1d array * Use draft model for sampling * Fix dumb mistake * Allow for later extensions to the LlamaDraftModel api * Cleanup * Adaptive candidate prediction * Update implementation to match hf transformers * Tuning * Fix bug where last token was not used for ngram prediction * Remove heuristic for num_pred_tokens (no benefit) * fix: n_candidates bug. * Add draft_model_num_pred_tokens server setting * Cleanup * Update README --- README.md | 18 ++++ llama_cpp/llama.py | 181 ++++++++++++++++---------------- llama_cpp/llama_speculative.py | 64 +++++++++++ llama_cpp/server/model.py | 9 ++ llama_cpp/server/settings.py | 9 ++ tests/test_llama_speculative.py | 16 +++ 6 files changed, 207 insertions(+), 90 deletions(-) create mode 100644 llama_cpp/llama_speculative.py create mode 100644 tests/test_llama_speculative.py diff --git a/README.md b/README.md index 0a77bbdaa..4131bb3d4 100644 --- a/README.md +++ b/README.md @@ -378,6 +378,24 @@ Then you'll need to use a custom chat handler to load the clip model and process ) ``` +### Speculative Decoding + +`llama-cpp-python` supports speculative decoding which allows the model to generate completions based on a draft model. + +The fastest way to use speculative decoding is through the `LlamaPromptLookupDecoding` class. + +Just pass this as a draft model to the `Llama` class during initialization. + +```python +from llama_cpp import Llama +from llama_cpp.llama_speculative import LlamaPromptLookupDecoding + +llama = Llama( + model_path="path/to/model.gguf", + draft_model=LlamaPromptLookupDecoding(num_pred_tokens=10) # num_pred_tokens is the number of tokens to predict 10 is the default and generally good for gpu, 2 performs better for cpu-only machines. +) +``` + ### Adjusting the Context Window The context window of the Llama models determines the maximum number of tokens that can be processed at once. By default, this is set to 512 tokens, but can be adjusted based on your requirements. diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index b5618c10d..f00fd4fc5 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -30,6 +30,8 @@ import llama_cpp.llama_cpp as llama_cpp import llama_cpp.llama_chat_format as llama_chat_format +from llama_cpp.llama_speculative import LlamaDraftModel + import numpy as np import numpy.typing as npt @@ -39,6 +41,8 @@ _LlamaContext, # type: ignore _LlamaBatch, # type: ignore _LlamaTokenDataArray, # type: ignore + _LlamaSamplingParams, # type: ignore + _LlamaSamplingContext, # type: ignore ) @@ -89,6 +93,8 @@ def __init__( # Chat Format Params chat_format: Optional[str] = None, chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None, + # Speculative Decoding + draft_model: Optional[LlamaDraftModel] = None, # Misc verbose: bool = True, # Extra Params @@ -152,6 +158,7 @@ def __init__( numa: Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init) chat_format: String specifying the chat format to use when calling create_chat_completion. chat_handler: Optional chat handler to use when calling create_chat_completion. + draft_model: Optional draft model to use for speculative decoding. verbose: Print verbose output to stderr. Raises: @@ -315,6 +322,8 @@ def __init__( self.chat_format = chat_format self.chat_handler = chat_handler + self.draft_model = draft_model + self._n_vocab = self.n_vocab() self._n_ctx = self.n_ctx() @@ -503,6 +512,7 @@ def sample( penalize_nl: bool = True, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, + idx: Optional[int] = None, ): """Sample a token from the model. @@ -517,77 +527,46 @@ def sample( """ assert self._ctx is not None assert self.n_tokens > 0 - last_n_tokens_data = [llama_cpp.llama_token(0)] * max( - 0, self.last_n_tokens_size - self.n_tokens - ) + self._input_ids[-self.last_n_tokens_size :].tolist() - last_n_tokens_size = len(last_n_tokens_data) - n_vocab = self._n_vocab - n_ctx = self._n_ctx - top_k = n_vocab if top_k <= 0 else top_k - last_n_tokens_size = n_ctx if last_n_tokens_size < 0 else last_n_tokens_size - last_n_tokens_data_c = (llama_cpp.llama_token * last_n_tokens_size)( - *last_n_tokens_data - ) - logits: npt.NDArray[np.single] = self._scores[-1, :] + + if idx is None: + logits: npt.NDArray[np.single] = self._scores[-1, :] + else: + logits = self._scores[idx, :] if logits_processor is not None: - logits[:] = logits_processor(self._input_ids, logits) - - nl_logit = logits[self._token_nl] - self._candidates.copy_logits(logits) - self._ctx.sample_repetition_penalties( - candidates=self._candidates, - last_tokens_data=last_n_tokens_data_c, - penalty_last_n=last_n_tokens_size, + logits[:] = ( + logits_processor(self._input_ids, logits) + if idx is None + else logits_processor(self._input_ids[:idx], logits) + ) + + sampling_params = _LlamaSamplingParams( + top_k=top_k, + top_p=top_p, + min_p=min_p, + tfs_z=tfs_z, + typical_p=typical_p, + temp=temp, + penalty_last_n=self.last_n_tokens_size, penalty_repeat=repeat_penalty, penalty_freq=frequency_penalty, penalty_present=presence_penalty, + mirostat=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + penalize_nl=penalize_nl, + ) + sampling_context = _LlamaSamplingContext( + params=sampling_params, + grammar=grammar, + ) + sampling_context.prev = list(self.eval_tokens) + id = sampling_context.sample(ctx_main=self._ctx, logits_array=logits) + sampling_context.accept( + ctx_main=self._ctx, + id=id, + apply_grammar=grammar is not None, ) - if not penalize_nl: - self._candidates.candidates.data[self._token_nl].logit = llama_cpp.c_float( - nl_logit - ) - - if grammar is not None: - self._ctx.sample_grammar( - candidates=self._candidates, - grammar=grammar, - ) - - if temp < 0.0: - self._ctx.sample_softmax(candidates=self._candidates) - id = self._candidates.candidates.data[0].id - elif temp == 0.0: - id = self._ctx.sample_token_greedy(candidates=self._candidates) - elif mirostat_mode == 1: - self._ctx.sample_temp(candidates=self._candidates, temp=temp) - id = self._ctx.sample_token_mirostat( - candidates=self._candidates, - tau=mirostat_tau, - eta=mirostat_eta, - mu=ctypes.pointer(self._mirostat_mu), - m=100, - ) - elif mirostat_mode == 2: - self._ctx.sample_temp(candidates=self._candidates, temp=temp) - id = self._ctx.sample_token_mirostat_v2( - candidates=self._candidates, - tau=mirostat_tau, - eta=mirostat_eta, - mu=ctypes.pointer(self._mirostat_mu), - ) - else: - self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1) - self._ctx.sample_tail_free(candidates=self._candidates, z=tfs_z, min_keep=1) - self._ctx.sample_typical( - candidates=self._candidates, p=typical_p, min_keep=1 - ) - self._ctx.sample_top_p(candidates=self._candidates, p=top_p, min_keep=1) - self._ctx.sample_min_p(candidates=self._candidates, p=min_p, min_keep=1) - self._ctx.sample_temp(candidates=self._candidates, temp=temp) - id = self._ctx.sample_token(candidates=self._candidates) - if grammar is not None: - self._ctx.grammar_accept_token(grammar=grammar, token=id) return id def generate( @@ -656,34 +635,56 @@ def generate( if grammar is not None: grammar.reset() + sample_idx = self.n_tokens + len(tokens) - 1 + tokens = list(tokens) + # Eval and sample while True: self.eval(tokens) - token = self.sample( - top_k=top_k, - top_p=top_p, - min_p=min_p, - typical_p=typical_p, - temp=temp, - repeat_penalty=repeat_penalty, - frequency_penalty=frequency_penalty, - presence_penalty=presence_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - logits_processor=logits_processor, - grammar=grammar, - penalize_nl=penalize_nl, - ) - if stopping_criteria is not None and stopping_criteria( - self._input_ids, self._scores[-1, :] - ): - return - tokens_or_none = yield token - tokens = [token] - if tokens_or_none is not None: - tokens.extend(tokens_or_none) + while sample_idx < self.n_tokens: + token = self.sample( + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + logits_processor=logits_processor, + grammar=grammar, + penalize_nl=penalize_nl, + idx=sample_idx, + ) + + sample_idx += 1 + if stopping_criteria is not None and stopping_criteria( + self._input_ids, self._scores[-1, :] + ): + return + tokens_or_none = yield token + tokens.clear() + tokens.append(token) + if tokens_or_none is not None: + tokens.extend(tokens_or_none) + + if sample_idx < self.n_tokens and token != self._input_ids[sample_idx]: + self.n_tokens = sample_idx + self._ctx.kv_cache_seq_rm(-1, self.n_tokens, -1) + break + + if self.draft_model is not None: + self.input_ids[self.n_tokens : self.n_tokens + len(tokens)] = tokens + draft_tokens = self.draft_model(self.input_ids[:self.n_tokens + len(tokens)]) + tokens.extend( + draft_tokens.astype(int)[ + : self._n_ctx - self.n_tokens - len(tokens) + ] + ) def create_embedding( self, input: Union[str, List[str]], model: Optional[str] = None diff --git a/llama_cpp/llama_speculative.py b/llama_cpp/llama_speculative.py new file mode 100644 index 000000000..39dfb903b --- /dev/null +++ b/llama_cpp/llama_speculative.py @@ -0,0 +1,64 @@ +import abc + +from typing import Any + +import numpy as np +import numpy.typing as npt + + +class LlamaDraftModel(abc.ABC): + @abc.abstractmethod + def __call__( + self, input_ids: npt.NDArray[np.intc], /, **kwargs: Any + ) -> npt.NDArray[np.intc]: + raise NotImplementedError() + + +class LlamaPromptLookupDecoding(LlamaDraftModel): + """Based on https://github.com/apoorvumang/prompt-lookup-decoding""" + + def __init__(self, max_ngram_size: int = 2, num_pred_tokens: int = 10): + self.max_ngram_size = max_ngram_size + self.num_pred_tokens = num_pred_tokens + + @staticmethod + def find_candidate_pred_tokens( + input_ids: npt.NDArray[np.intc], + max_ngram_size: int, + num_pred_tokens: int, + ): + input_length = input_ids.shape[0] + + for ngram_size in range(min(max_ngram_size, input_length - 1), 0, -1): + # Create sliding windows of size ngram_size + windows = np.lib.stride_tricks.sliding_window_view(input_ids, (ngram_size,)) + + # Convert ngram to an array for comparison + ngram_array = input_ids[-ngram_size:] + + # Find where the windows match the ngram + matches = np.all(windows == ngram_array, axis=1) + + # Get the indices of matches + match_indices = np.nonzero(matches)[0] + + # Iterate through match indices to find a valid continuation + for idx in match_indices: + start_idx = idx + ngram_size + end_idx = start_idx + num_pred_tokens + end_idx = min(end_idx, input_length) + + if start_idx < end_idx: + return input_ids[start_idx:end_idx] + + # If no match is found, return an empty array + return np.array([], dtype=np.intc) + + def __call__( + self, input_ids: npt.NDArray[np.intc], /, **kwargs: Any + ) -> npt.NDArray[np.intc]: + return self.find_candidate_pred_tokens( + input_ids=input_ids, + max_ngram_size=self.max_ngram_size, + num_pred_tokens=self.num_pred_tokens, + ) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index bbb68069d..925ab99b7 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -5,6 +5,7 @@ from typing import Dict, Optional, Union, List import llama_cpp +import llama_cpp.llama_speculative as llama_speculative from llama_cpp.server.settings import ModelSettings @@ -92,6 +93,12 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: ) ) + draft_model = None + if settings.draft_model is not None: + draft_model = llama_speculative.LlamaPromptLookupDecoding( + num_pred_tokens=settings.draft_model_num_pred_tokens + ) + kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None if settings.kv_overrides is not None: assert isinstance(settings.kv_overrides, list) @@ -147,6 +154,8 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: # Chat Format Params chat_format=settings.chat_format, chat_handler=chat_handler, + # Speculative Decoding + draft_model=draft_model, # Misc verbose=settings.verbose, ) diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 9fe1a7bfd..60f3eeca2 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -143,6 +143,15 @@ class ModelSettings(BaseSettings): default=None, description="The model name or path to a pretrained HuggingFace tokenizer model. Same as you would pass to AutoTokenizer.from_pretrained().", ) + # Speculative Decoding + draft_model: Optional[str] = Field( + default=None, + description="Method to use for speculative decoding. One of (prompt-lookup-decoding).", + ) + draft_model_num_pred_tokens: int = Field( + default=10, + description="Number of tokens to predict using the draft model.", + ) # Misc verbose: bool = Field( default=True, description="Whether to print debug information." diff --git a/tests/test_llama_speculative.py b/tests/test_llama_speculative.py new file mode 100644 index 000000000..b5d450567 --- /dev/null +++ b/tests/test_llama_speculative.py @@ -0,0 +1,16 @@ +import numpy as np + +from llama_cpp.llama_speculative import LlamaPromptLookupDecoding + +def test_find_candidate_pred_tokens(): + find_candidate_pred_tokens = LlamaPromptLookupDecoding.find_candidate_pred_tokens + + # Test Case 1: Matching ngram is found + input_ids1 = np.array([1, 2, 3, 1, 2, 3, 1, 2, 3]) + result1 = find_candidate_pred_tokens(input_ids1, max_ngram_size=3, num_pred_tokens=2) + assert np.array_equal(result1, np.array([1, 2])) + + # Test Case 2: Matching ngram is not found + input_ids2 = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9]) + result2 = find_candidate_pred_tokens(input_ids2, max_ngram_size=3, num_pred_tokens=2) + assert np.array_equal(result2, np.array([])) From a8cb34eacdab7e20553e82632c4c9bd1bdebe54b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jan 2024 15:05:51 -0500 Subject: [PATCH 094/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5cb04dbc1..1cfb5372c 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5cb04dbc16d1da38c8fdcc0111b40e67d00dd1c3 +Subproject commit 1cfb5372cf5707c8ec6dde7c874f4a44a6c4c915 From 3322eadbf30a68731f6aafe0b4d055255b46d8f7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jan 2024 15:10:18 -0500 Subject: [PATCH 095/624] Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 435af43a8..96322109d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.38] + +- feat: Update llama.cpp to ggerganov/llama.cpp@1cfb5372cf5707c8ec6dde7c874f4a44a6c4c915 +- feat: Add speculative decoding by @abetlen in #1120 +- fix: Pass raise_exception and add_generation_prompt to jinja2 chat template 078cca0361bf5a94d2cf52ed04980d20e32d6f95 + ## [0.2.37] - feat: Update llama.cpp to ggerganov/llama.cpp@fea4fd4ba7f6b754ac795387b275e1a014a77bde diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 4ce899ccd..94cd401d2 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.37" \ No newline at end of file +__version__ = "0.2.38" \ No newline at end of file From de526d02143b8057286e2e10de67512c2ad3480a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 1 Feb 2024 12:35:31 -0500 Subject: [PATCH 096/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1cfb5372c..8ca511cad 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1cfb5372cf5707c8ec6dde7c874f4a44a6c4c915 +Subproject commit 8ca511cadee2c67f0bd8c7034a2513778ee9a1b7 From 8a5911bd5d3eaab67217cad6a6dce7a041f6137b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 2 Feb 2024 09:41:27 -0500 Subject: [PATCH 097/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8ca511cad..191221178 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8ca511cadee2c67f0bd8c7034a2513778ee9a1b7 +Subproject commit 191221178f51b6e81122c5bda0fd79620e547d07 From bebfba0f08b3198cee2f1652393f032af4bf2016 Mon Sep 17 00:00:00 2001 From: Dulsara Date: Fri, 2 Feb 2024 22:35:46 +0530 Subject: [PATCH 098/624] Fix: fileno error google colab (#729) (#1156) Instead of using a devnull just made a dummy class with a 'write()' method that does nothing. --- llama_cpp/_utils.py | 38 ++++++++------------------------------ 1 file changed, 8 insertions(+), 30 deletions(-) diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index 4a106470b..4990c117c 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -4,9 +4,9 @@ import sys from typing import Any, Dict -# Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor -outnull_file = open(os.devnull, "w") -errnull_file = open(os.devnull, "w") +class NullDevice(): + def write(self, s): + pass class suppress_stdout_stderr(object): # NOTE: these must be "saved" here to avoid exceptions when using @@ -21,41 +21,19 @@ def __init__(self, disable: bool = True): def __enter__(self): if self.disable: return self - - # Check if sys.stdout and sys.stderr have fileno method - if not hasattr(self.sys.stdout, 'fileno') or not hasattr(self.sys.stderr, 'fileno'): - return self # Return the instance without making changes - - self.old_stdout_fileno_undup = self.sys.stdout.fileno() - self.old_stderr_fileno_undup = self.sys.stderr.fileno() - - self.old_stdout_fileno = self.os.dup(self.old_stdout_fileno_undup) - self.old_stderr_fileno = self.os.dup(self.old_stderr_fileno_undup) - self.old_stdout = self.sys.stdout self.old_stderr = self.sys.stderr - self.os.dup2(outnull_file.fileno(), self.old_stdout_fileno_undup) - self.os.dup2(errnull_file.fileno(), self.old_stderr_fileno_undup) - - self.sys.stdout = outnull_file - self.sys.stderr = errnull_file + self.sys.stdout = NullDevice() + self.sys.stderr = NullDevice() return self def __exit__(self, *_): if self.disable: return - - # Check if sys.stdout and sys.stderr have fileno method - if hasattr(self.sys.stdout, 'fileno') and hasattr(self.sys.stderr, 'fileno'): - self.sys.stdout = self.old_stdout - self.sys.stderr = self.old_stderr - - self.os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) - self.os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) - - self.os.close(self.old_stdout_fileno) - self.os.close(self.old_stderr_fileno) + + self.sys.stdout = self.old_stdout + self.sys.stderr = self.old_stderr class MetaSingleton(type): From 7467f129e5cfb61115184402af6645fd9cb4401b Mon Sep 17 00:00:00 2001 From: Andrei Date: Fri, 2 Feb 2024 12:18:55 -0500 Subject: [PATCH 099/624] Revert "Fix: fileno error google colab (#729) (#1156)" (#1157) This reverts commit bebfba0f08b3198cee2f1652393f032af4bf2016. --- llama_cpp/_utils.py | 38 ++++++++++++++++++++++++++++++-------- 1 file changed, 30 insertions(+), 8 deletions(-) diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index 4990c117c..4a106470b 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -4,9 +4,9 @@ import sys from typing import Any, Dict -class NullDevice(): - def write(self, s): - pass +# Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor +outnull_file = open(os.devnull, "w") +errnull_file = open(os.devnull, "w") class suppress_stdout_stderr(object): # NOTE: these must be "saved" here to avoid exceptions when using @@ -21,19 +21,41 @@ def __init__(self, disable: bool = True): def __enter__(self): if self.disable: return self + + # Check if sys.stdout and sys.stderr have fileno method + if not hasattr(self.sys.stdout, 'fileno') or not hasattr(self.sys.stderr, 'fileno'): + return self # Return the instance without making changes + + self.old_stdout_fileno_undup = self.sys.stdout.fileno() + self.old_stderr_fileno_undup = self.sys.stderr.fileno() + + self.old_stdout_fileno = self.os.dup(self.old_stdout_fileno_undup) + self.old_stderr_fileno = self.os.dup(self.old_stderr_fileno_undup) + self.old_stdout = self.sys.stdout self.old_stderr = self.sys.stderr - self.sys.stdout = NullDevice() - self.sys.stderr = NullDevice() + self.os.dup2(outnull_file.fileno(), self.old_stdout_fileno_undup) + self.os.dup2(errnull_file.fileno(), self.old_stderr_fileno_undup) + + self.sys.stdout = outnull_file + self.sys.stderr = errnull_file return self def __exit__(self, *_): if self.disable: return - - self.sys.stdout = self.old_stdout - self.sys.stderr = self.old_stderr + + # Check if sys.stdout and sys.stderr have fileno method + if hasattr(self.sys.stdout, 'fileno') and hasattr(self.sys.stderr, 'fileno'): + self.sys.stdout = self.old_stdout + self.sys.stderr = self.old_stderr + + self.os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) + self.os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) + + self.os.close(self.old_stdout_fileno) + self.os.close(self.old_stderr_fileno) class MetaSingleton(type): From 3553b146701b88f12d96575f55a8e6ef67b9c1a6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 5 Feb 2024 13:26:50 -0500 Subject: [PATCH 100/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++-- vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 431a99f55..da2a7f32a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -445,7 +445,7 @@ class llama_model_params(Structure): # uint32_t n_batch; // prompt processing maximum batch size # uint32_t n_threads; // number of threads to use for generation # uint32_t n_threads_batch; // number of threads to use for batch processing -# int8_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` +# int32_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` # // ref: https://github.com/ggerganov/llama.cpp/pull/2054 # float rope_freq_base; // RoPE base frequency, 0 = from model @@ -502,7 +502,7 @@ class llama_context_params(Structure): ("n_batch", c_uint32), ("n_threads", c_uint32), ("n_threads_batch", c_uint32), - ("rope_scaling_type", c_int8), + ("rope_scaling_type", c_int32), ("rope_freq_base", c_float), ("rope_freq_scale", c_float), ("yarn_ext_factor", c_float), diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 191221178..78b00dda6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 191221178f51b6e81122c5bda0fd79620e547d07 +Subproject commit 78b00dda6c0d62c34f5371d47718defff6ed2b22 From 59760c85eddc72dfcc1839f43760ef72c23d6874 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 5 Feb 2024 21:52:12 -0500 Subject: [PATCH 101/624] fix: Use llama_log_callback to avoid suppress_stdout_stderr --- llama_cpp/_internals.py | 51 ++++++++++++++++------------------------- llama_cpp/_logger.py | 37 ++++++++++++++++++++++++++++++ llama_cpp/llama.py | 7 +++--- 3 files changed, 61 insertions(+), 34 deletions(-) create mode 100644 llama_cpp/_logger.py diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 651cd4ccf..3a71ef0fa 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -18,8 +18,6 @@ import llama_cpp.llama_cpp as llama_cpp -from ._utils import suppress_stdout_stderr - # Python wrappers over llama.h structs @@ -30,7 +28,6 @@ class _LlamaModel: _llama_free_model = None # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - _suppress_stdout_stderr = suppress_stdout_stderr def __init__( self, @@ -48,16 +45,14 @@ def __init__( if not os.path.exists(path_model): raise ValueError(f"Model path does not exist: {path_model}") - with self._suppress_stdout_stderr(disable=self.verbose): - self.model = llama_cpp.llama_load_model_from_file( - self.path_model.encode("utf-8"), self.params - ) + self.model = llama_cpp.llama_load_model_from_file( + self.path_model.encode("utf-8"), self.params + ) def __del__(self): - with self._suppress_stdout_stderr(disable=self.verbose): - if self.model is not None and self._llama_free_model is not None: - self._llama_free_model(self.model) - self.model = None + if self.model is not None and self._llama_free_model is not None: + self._llama_free_model(self.model) + self.model = None def vocab_type(self) -> int: assert self.model is not None @@ -240,8 +235,6 @@ class _LlamaContext: NOTE: For stability it's recommended you use the Llama class instead.""" _llama_free = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - _suppress_stdout_stderr = suppress_stdout_stderr def __init__( self, @@ -256,16 +249,16 @@ def __init__( self._llama_free = llama_cpp._lib.llama_free # type: ignore - with self._suppress_stdout_stderr(disable=self.verbose): - self.ctx = llama_cpp.llama_new_context_with_model( - self.model.model, self.params - ) + assert self.model.model is not None + + self.ctx = llama_cpp.llama_new_context_with_model( + self.model.model, self.params + ) def __del__(self): - with self._suppress_stdout_stderr(disable=self.verbose): - if self.ctx is not None and self._llama_free is not None: - self._llama_free(self.ctx) - self.ctx = None + if self.ctx is not None and self._llama_free is not None: + self._llama_free(self.ctx) + self.ctx = None def n_ctx(self) -> int: assert self.ctx is not None @@ -493,8 +486,6 @@ def default_params(): class _LlamaBatch: _llama_batch_free = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - _suppress_stdout_stderr = suppress_stdout_stderr def __init__( self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True @@ -506,16 +497,14 @@ def __init__( self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore - with self._suppress_stdout_stderr(disable=self.verbose): - self.batch = llama_cpp.llama_batch_init( - self.n_tokens, self.embd, self.n_seq_max - ) + self.batch = llama_cpp.llama_batch_init( + self.n_tokens, self.embd, self.n_seq_max + ) def __del__(self): - with self._suppress_stdout_stderr(disable=self.verbose): - if self.batch is not None and self._llama_batch_free is not None: - self._llama_batch_free(self.batch) - self.batch = None + if self.batch is not None and self._llama_batch_free is not None: + self._llama_batch_free(self.batch) + self.batch = None def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): assert self.batch is not None diff --git a/llama_cpp/_logger.py b/llama_cpp/_logger.py new file mode 100644 index 000000000..7638170a9 --- /dev/null +++ b/llama_cpp/_logger.py @@ -0,0 +1,37 @@ +import sys +import ctypes +import logging + +import llama_cpp + +# enum ggml_log_level { +# GGML_LOG_LEVEL_ERROR = 2, +# GGML_LOG_LEVEL_WARN = 3, +# GGML_LOG_LEVEL_INFO = 4, +# GGML_LOG_LEVEL_DEBUG = 5 +# }; +GGML_LOG_LEVEL_TO_LOGGING_LEVEL = { + 2: logging.ERROR, + 3: logging.WARNING, + 4: logging.INFO, + 5: logging.DEBUG, +} + +logger = logging.getLogger("llama-cpp-python") + + +@llama_cpp.llama_log_callback +def llama_log_callback( + level: int, + text: bytes, + user_data: ctypes.c_void_p, +): + if logger.level <= GGML_LOG_LEVEL_TO_LOGGING_LEVEL[level]: + print(text.decode("utf-8"), end="", flush=True, file=sys.stderr) + + +llama_cpp.llama_log_set(llama_log_callback, ctypes.c_void_p(0)) + + +def set_verbose(verbose: bool): + logger.setLevel(logging.DEBUG if verbose else logging.ERROR) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index f00fd4fc5..85943dbcd 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -35,7 +35,6 @@ import numpy as np import numpy.typing as npt -from ._utils import suppress_stdout_stderr from ._internals import ( _LlamaModel, # type: ignore _LlamaContext, # type: ignore @@ -44,6 +43,7 @@ _LlamaSamplingParams, # type: ignore _LlamaSamplingContext, # type: ignore ) +from ._logger import set_verbose class Llama: @@ -169,10 +169,11 @@ def __init__( """ self.verbose = verbose + set_verbose(verbose) + self.numa = numa if not Llama.__backend_initialized: - with suppress_stdout_stderr(disable=self.verbose): - llama_cpp.llama_backend_init(self.numa) + llama_cpp.llama_backend_init(self.numa) Llama.__backend_initialized = True self.model_path = model_path From 310fbf4e494e30d6a78be1b8a6b9be81f61204e7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 5 Feb 2024 22:07:14 -0500 Subject: [PATCH 102/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 78b00dda6..098f6d737 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 78b00dda6c0d62c34f5371d47718defff6ed2b22 +Subproject commit 098f6d737b65134cf220d12b9b706e8cfc5e4610 From 5e3e67af47908919968f06b738321c32e646a97b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 6 Feb 2024 12:44:07 -0500 Subject: [PATCH 103/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 098f6d737..b08f22c88 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 098f6d737b65134cf220d12b9b706e8cfc5e4610 +Subproject commit b08f22c882a1443e6b97081f3ce718a4d1a741f8 From 34f31040f610925552a66b3a033e31320b6f6ad8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 6 Feb 2024 12:47:59 -0500 Subject: [PATCH 104/624] Bump version --- CHANGELOG.md | 7 ++++++- llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 2 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 96322109d..5ce0b4313 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,11 +7,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.39] + +- feat: Update llama.cpp to ggerganov/llama.cpp@b08f22c882a1443e6b97081f3ce718a4d1a741f8 +- fix: Fix destructor logging bugs by using llama_log_callback to avoid suppress_stdout_stderr by @abetlen in 59760c85eddc72dfcc1839f43760ef72c23d6874 + ## [0.2.38] - feat: Update llama.cpp to ggerganov/llama.cpp@1cfb5372cf5707c8ec6dde7c874f4a44a6c4c915 - feat: Add speculative decoding by @abetlen in #1120 -- fix: Pass raise_exception and add_generation_prompt to jinja2 chat template 078cca0361bf5a94d2cf52ed04980d20e32d6f95 +- fix: Pass raise_exception and add_generation_prompt to jinja2 chat template by @abetlen in 078cca0361bf5a94d2cf52ed04980d20e32d6f95 ## [0.2.37] diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 94cd401d2..837e3c993 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.38" \ No newline at end of file +__version__ = "0.2.39" \ No newline at end of file From ce1277549012a33e5c2360f42bf53aaf1b95e528 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 6 Feb 2024 18:50:56 -0500 Subject: [PATCH 105/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index b08f22c88..213d1439f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit b08f22c882a1443e6b97081f3ce718a4d1a741f8 +Subproject commit 213d1439fadefe182f69c5f7e8dd3b4b6572ebcb From 901827013b732d74f1f67033062d13a6204a62bd Mon Sep 17 00:00:00 2001 From: Jeffrey Fong Date: Thu, 8 Feb 2024 09:07:03 +0800 Subject: [PATCH 106/624] feat: Integrate functionary v1.4 and v2 models + add custom tokenizer support to Llama class (#1078) * convert functionary-v1 chat handler to use hf autotokenizer * add hf_tokenizer + inteegrate functionary-v1.4 prompt template * integrate functionary v2 prompt template * update readme * set up parallel function calling wip * set up parallel function calling * Update README.md * Update README.md * refactor tokenizers * include old functionary handler for backward compatibility * add hf_tokenizer_path in server ModelSettings * convert functionary-v1 chat handler to use hf autotokenizer * add hf_tokenizer + inteegrate functionary-v1.4 prompt template * integrate functionary v2 prompt template * update readme * set up parallel function calling wip * resolve merge conflict * Update README.md * Update README.md * refactor tokenizers * include old functionary handler for backward compatibility * add hf_tokenizer_path in server ModelSettings * Cleanup PR, fix breaking changes * Use hf_pretrained_model_name_or_path for tokenizer * fix hf tokenizer in streaming * update README * refactor offset mapping --------- Co-authored-by: Andrei --- README.md | 19 +- llama_cpp/llama.py | 101 ++++++-- llama_cpp/llama_chat_format.py | 433 ++++++++++++++++++++++++++++++++- llama_cpp/server/model.py | 6 + 4 files changed, 525 insertions(+), 34 deletions(-) diff --git a/README.md b/README.md index 4131bb3d4..bddef6459 100644 --- a/README.md +++ b/README.md @@ -293,19 +293,16 @@ To constrain the response to a specific JSON Schema, you can use the `schema` pr The high-level API also provides a simple interface for function calling. -Note that the only model that supports full function calling at this time is "functionary". -The gguf-converted files for this model can be found here: [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) +The only set of models that supports full function calling at this time is [functionary](https://github.com/MeetKai/functionary). The various gguf-converted files for this set of models can be found [here](https://huggingface.co/meetkai). Functionary is able to intelligently call functions and also analyze any provided function outputs to generate coherent responses. All v2 models of functionary supports **parallel function calling**. You can provide either `functionary-v1` or `functionary-v2` for the `chat_format` when initializing the Llama class. + +Note that due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is required to provide HF Tokenizer for functionary. The `LlamaHFTokenizer` class can be initialized and passed into the Llama class. This will override the default llama.cpp tokenizer used in Llama class. The tokenizer files are already included in the respective HF repositories hosting the gguf files. ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", chat_format="functionary") +>>> from llama_cpp import Llama, LlamaHFTokenizer +>>> tokenizer = LlamaHFTokenizer.from_pretrained("path/to/functionary/") +>>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", tokenizer=tokenizer, chat_format="functionary-v2") >>> llm.create_chat_completion( messages = [ - { - "role": "system", - "content": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary" - - }, { "role": "user", "content": "Extract Jason is 25 years old" @@ -332,12 +329,12 @@ The gguf-converted files for this model can be found here: [functionary-7b-v1](h } } }], - tool_choice=[{ + tool_choice={ "type": "function", "function": { "name": "UserDetail" } - }] + }, ) ``` diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 85943dbcd..bad75dfaf 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2,6 +2,7 @@ import os import sys +import abc import uuid import time import multiprocessing @@ -14,11 +15,14 @@ Iterator, Deque, Callable, + Any, ) from collections import deque import ctypes +from llama_cpp.llama_types import List + from .llama_types import * from .llama_grammar import LlamaGrammar from .llama_cache import ( @@ -95,6 +99,8 @@ def __init__( chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None, # Speculative Decoding draft_model: Optional[LlamaDraftModel] = None, + # Tokenizer Override + tokenizer: Optional[BaseLlamaTokenizer] = None, # Misc verbose: bool = True, # Extra Params @@ -159,6 +165,7 @@ def __init__( chat_format: String specifying the chat format to use when calling create_chat_completion. chat_handler: Optional chat handler to use when calling create_chat_completion. draft_model: Optional draft model to use for speculative decoding. + tokenizer: Optional tokenizer to override the default tokenizer from llama.cpp. verbose: Print verbose output to stderr. Raises: @@ -235,6 +242,7 @@ def __init__( self.n_threads_batch = n_threads_batch or max( multiprocessing.cpu_count() // 2, 1 ) + # Context Params self.context_params = llama_cpp.llama_context_default_params() self.context_params.seed = seed @@ -286,6 +294,10 @@ def __init__( self._model = _LlamaModel( path_model=self.model_path, params=self.model_params, verbose=self.verbose ) + + # Override tokenizer + self.tokenizer_ = tokenizer or LlamaTokenizer(self) + # Set the default value for the context and correct the batch if n_ctx == 0: n_ctx = self._model.n_ctx_train() @@ -431,18 +443,19 @@ def tokenize( Returns: A list of tokens. """ - return self._model.tokenize(text, add_bos, special) + return self.tokenizer_.tokenize(text, add_bos, special) - def detokenize(self, tokens: List[int]) -> bytes: + def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: """Detokenize a list of tokens. Args: tokens: The list of tokens to detokenize. + prev_tokens: The list of previous tokens. Offset mapping will be performed if provided Returns: The detokenized string. """ - return self._model.detokenize(tokens) + return self.tokenizer_.detokenize(tokens, prev_tokens) def set_cache(self, cache: Optional[BaseLlamaCache]): """Set the cache. @@ -935,7 +948,8 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - remaining_text = self.detokenize(remaining_tokens) + prev_tokens = completion_tokens[:returned_tokens] + remaining_text = self.detokenize(completion_tokens, prev_tokens) remaining_length = len(remaining_text) # We want to avoid yielding any characters from @@ -957,13 +971,13 @@ def logit_bias_processor( for token in remaining_tokens: if token == self.token_bos(): continue - token_end_position += len(self.detokenize([token])) + token_end_position += len(remaining_text) # Check if stop sequence is in the token if token_end_position > ( remaining_length - first_stop_position ): break - token_str = self.detokenize([token]).decode( + token_str = remaining_text.decode( "utf-8", errors="ignore" ) text_offset = len(prompt) + len( @@ -988,11 +1002,7 @@ def logit_bias_processor( } top_logprob.update({token_str: current_logprobs[int(token)]}) logprobs_or_none = { - "tokens": [ - self.detokenize([token]).decode( - "utf-8", errors="ignore" - ) - ], + "tokens": [token_str], "text_offset": [text_offset], "token_logprobs": [current_logprobs[int(token)]], "top_logprobs": [top_logprob], @@ -1005,9 +1015,7 @@ def logit_bias_processor( "model": model_name, "choices": [ { - "text": self.detokenize([token]).decode( - "utf-8", errors="ignore" - ), + "text": token_str, "index": 0, "logprobs": logprobs_or_none, "finish_reason": None, @@ -1019,7 +1027,7 @@ def logit_bias_processor( decode_success = False for i in range(1, len(remaining_tokens) + 1): try: - bs = self.detokenize(remaining_tokens[:i]) + bs = remaining_text ts = bs.decode("utf-8") decode_success = True break @@ -1055,6 +1063,7 @@ def logit_bias_processor( if len(completion_tokens) >= max_tokens: text = self.detokenize(completion_tokens) + finish_reason = "length" break @@ -1693,8 +1702,8 @@ def n_vocab(self) -> int: """Return the vocabulary size.""" return self._model.n_vocab() - def tokenizer(self) -> "LlamaTokenizer": - """Return the tokenizer for this model.""" + def tokenizer(self) -> LlamaTokenizer: + """Return the llama tokenizer for this model.""" return LlamaTokenizer(self) def token_eos(self) -> int: @@ -1738,23 +1747,71 @@ def longest_token_prefix(a: Sequence[int], b: Sequence[int]): return longest_prefix -class LlamaTokenizer: +class BaseLlamaTokenizer(abc.ABC): + @abc.abstractmethod + def tokenize(self, text: bytes, add_bos: bool = True, special: bool = True) -> List[int]: + raise NotImplementedError + + @abc.abstractmethod + def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: + raise NotImplementedError + + +class LlamaTokenizer(BaseLlamaTokenizer): def __init__(self, llama: Llama): self.llama = llama + self._model = llama._model # type: ignore + + def tokenize(self, text: bytes, add_bos: bool = True, special: bool = True) -> List[int]: + return self._model.tokenize(text, add_bos=add_bos, special=special) + + def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: + if prev_tokens is not None: + return self._model.detokenize(tokens[len(prev_tokens):]) + else: + return self._model.detokenize(tokens) - def encode(self, text: str, add_bos: bool = True) -> List[int]: - return self.llama.tokenize( - text.encode("utf-8", errors="ignore"), add_bos=add_bos, special=True + def encode(self, text: str, add_bos: bool = True, special: bool = True) -> List[int]: + return self.tokenize( + text.encode("utf-8", errors="ignore"), add_bos=add_bos, special=special ) def decode(self, tokens: List[int]) -> str: - return self.llama.detokenize(tokens).decode("utf-8", errors="ignore") + return self.detokenize(tokens).decode("utf-8", errors="ignore") @classmethod def from_ggml_file(cls, path: str) -> "LlamaTokenizer": return cls(Llama(model_path=path, vocab_only=True)) +class LlamaHFTokenizer(BaseLlamaTokenizer): + def __init__(self, hf_tokenizer: Any): + self.hf_tokenizer = hf_tokenizer + + def tokenize(self, text: bytes, add_bos: bool = True, special: bool = True) -> List[int]: + return self.hf_tokenizer.encode(text.decode("utf-8", errors="ignore"), add_special_tokens=special) + + def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: + if prev_tokens is not None: + text = self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") + prev_text = self.hf_tokenizer.decode(prev_tokens).encode("utf-8", errors="ignore") + return text[len(prev_text):] + else: + return self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path: str) -> "LlamaHFTokenizer": + try: + from transformers import AutoTokenizer + except ImportError: + raise ImportError( + "The `transformers` library is required to use the `HFTokenizer`." + "You can install it with `pip install transformers`." + ) + hf_tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=pretrained_model_name_or_path) + return cls(hf_tokenizer) + + class LlamaState: def __init__( self, diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 08f991b67..2e4204121 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -4,7 +4,9 @@ import json import ctypes import dataclasses -from typing import Any, Dict, Iterator, List, Optional, Tuple, Union, Protocol +import random +import string +from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol import jinja2 @@ -1332,6 +1334,435 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): ) +@register_chat_completion_handler("functionary-v1") +@register_chat_completion_handler("functionary-v2") +def functionary_v1_v2_chat_handler( + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + response_format: Optional[llama_types.ChatCompletionRequestResponseFormat] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + **kwargs, # type: ignore +) -> Union[llama_types.ChatCompletion, Iterator[llama_types.ChatCompletionChunk]]: + SYSTEM_MESSAGE = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary""" + + tokenizer = llama.tokenizer_ + assert hasattr(tokenizer, "hf_tokenizer"), "Please provide a valid hf_tokenizer_path from https://huggingface.co/meetkai when initializing the Llama class" + from transformers import AutoTokenizer + + if "<|START_OF_FUNCTION_CALL|>" in tokenizer.hf_tokenizer.additional_special_tokens: + version = "v1" + END_SYSTEM_TOKEN = "<|END_OF_SYSTEM|>" + END_USER_TOKEN = "<|END_OF_USER|>" + END_ASSISTANT_TOKEN = "<|END_OF_ASSISTANT|>" + END_FUNCTION_RESULT_TOKEN = "<|END_OF_FUNCTION_RESULT|>" + START_FUNCTION_CALL_TOKEN = "<|START_OF_FUNCTION_CALL|>" + END_FUNCTION_CALL_TOKEN = "<|END_OF_FUNCTION_CALL|>" + else: + version = "v2" + RECIPIENT_TOKEN = "<|recipient|>" + FROM_TOKEN = "<|from|>" + STOP_TOKEN = "<|stop|>" + CONTENT_TOKEN = "<|content|>" + + def generate_type_definition( + param: Dict[str, llama_types.JsonType], indent_level: int, shared_defs + ) -> str: + indent = " " * indent_level + if "$ref" in param: + # Reference to a shared definition + ref_name = param["$ref"].split("/")[ + -1 + ] # Extract the type name from the reference + return ref_name + elif param.get("type") == "array": + items = param.get("items", {}) + item_type = generate_type_definition(items, indent_level + 1, shared_defs) + return f"Array<{item_type}>" + elif param.get("type") == "object": + properties = param.get("properties", {}) + nested_schema = "{\n" + for nested_param_name, nested_param in properties.items(): + nested_param_type = generate_type_definition( + nested_param, indent_level + 1, shared_defs + ) + nested_schema += ( + f"{indent} {nested_param_name}: {nested_param_type},\n" + ) + nested_schema += indent + "}" + return nested_schema + elif "enum" in param: + # Enum type + return " | ".join([f'"{enum_value}"' for enum_value in param["enum"]]) + else: + # Simple type + return param.get("type", "any") + + def generate_shared_definitions(shared_defs, indent_level: int) -> str: + indent = " " * indent_level + shared_definitions = "" + for def_name, def_properties in shared_defs.items(): + shared_definitions += f"{indent}type {def_name} = " + if def_properties.get("type") == "object": + shared_definitions += generate_type_definition( + def_properties, indent_level, shared_defs + ) + elif "enum" in def_properties: + # Enum type + shared_definitions += " | ".join( + [f'"{enum_value}"' for enum_value in def_properties["enum"]] + ) + shared_definitions += ";\n" + return shared_definitions + + def generate_schema_from_functions(functions, namespace="functions") -> str: + schema = ( + "// Supported function definitions that should be called when necessary.\n" + ) + schema += f"namespace {namespace} {{\n\n" + + # Generate shared definitions + shared_definitions = {} + for function in functions: + parameters = function.get("parameters", {}) + shared_definitions.update(parameters.get("$defs", {})) + + schema += generate_shared_definitions(shared_definitions, 1) + + for function in functions: + function_name = function["name"] + description = function.get("description", "") + parameters = function.get("parameters", {}) + required_params = parameters.get("required", []) + + schema += f"// {description}\n" + schema += f"type {function_name} = (_: {{\n" + + for param_name, param in parameters.get("properties", {}).items(): + param_description = param.get("description", "") + param_type = generate_type_definition(param, 2, shared_definitions) + optional_indicator = "" if param_name in required_params else "?" + schema += f"// {param_description}\n" + schema += f"{param_name}{optional_indicator}: {param_type},\n" + schema += "}) => any;\n\n" + + schema += "}} // namespace {}".format(namespace) + return schema + + def prepare_messages_for_inference( + messages: List[llama_types.ChatCompletionRequestMessage], + tokenizer: AutoTokenizer, + version: Literal["v1", "v2"], + functions: Optional[List[llama_types.ChatCompletionFunctions]] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + ): + all_messages: List[llama_types.ChatCompletionRequestMessage] = [] + if functions is not None: + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=generate_schema_from_functions(functions) + ) + ) + elif tools is not None: + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", + content=generate_schema_from_functions( + [ + tool["function"] + for tool in tools + if tool["type"] == "function" + ] + ), + ) + ) + + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=SYSTEM_MESSAGE + ) + ) + + for message in messages: + # Function call responses + if message["role"] == "function" and "name" in message: + message["name"] = f"functions.{message['name']}" + # Function call requests by assistant + if "function_call" in message: + message["function_call"][ + "name" + ] = f"functions.{message['function_call']['name']}" + all_messages.append(message) + + if version == "v1": + suffix = "assistant:\n" + else: + suffix = "<|from|>assistant\n<|recipient|>" + + return tokenizer.hf_tokenizer.apply_chat_template(all_messages, tokenize=False) + suffix + + if tools is not None: + functions = [tool["function"] for tool in tools if tool["type"] == "function"] + + if tool_choice is not None: + function_call = ( + tool_choice if isinstance(tool_choice, str) else tool_choice["function"] + ) + + prompt = prepare_messages_for_inference(messages, tokenizer, version, functions, tools) + + # If no tools/functions are provided + if function_call is None and (functions is None or len(functions) == 0): + if version == "v1": + stop = END_ASSISTANT_TOKEN + else: + stop = STOP_TOKEN + prompt += "all\n<|content|>" + + completion_or_completion_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore + + assert stream is False # TODO: support stream mode + + def get_grammar(function_call): + function_body = None + for function in functions or []: + if function["name"] == function_call: + function_body = function["parameters"] + break + for tool in tools or []: + if tool["type"] == "function" and tool["function"]["name"] == function_call: + function_body = tool["function"]["parameters"] + break + + try: + with suppress_stdout_stderr(disable=llama.verbose): + grammar_text = llama_grammar.json_schema_to_gbnf( + json.dumps(function_body) + ) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.json_schema_to_gbnf(json.dumps(function_body)) + ) + print(grammar_text) + except Exception as e: + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + with suppress_stdout_stderr(disable=llama.verbose): + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF + ) + + return grammar + + def create_completion(stop): + completion: llama_types.Completion = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + + return completion + + function_calls, function_bodies = [], [] + + if version == "v1": + # If no or "auto" tool_choice/function_call + if function_call is None or ( + isinstance(function_call, str) and function_call == "auto" + ): + stops = ["\n", END_ASSISTANT_TOKEN] + # If tool_choice/function_call is "none" + elif isinstance(function_call, str) and function_call == "none": + prompt = prepare_messages_for_inference(messages, tokenizer, version, [], []) + stops = END_ASSISTANT_TOKEN + # If tool_choice/function_call is provided + elif isinstance(function_call, dict): + prompt += f"{START_FUNCTION_CALL_TOKEN}{function_call['name']}:\n" + stops = END_FUNCTION_CALL_TOKEN + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + else: + prompt = prompt + stops = ["\n", END_ASSISTANT_TOKEN] + + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + + # If the generation does not involve a function call + if START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN not in completion_text: + return _convert_completion_to_chat(completion, stream=stream) # type: ignore + # If the generation involves a function call in completion, generate the parameters + elif START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN in completion_text: + prompt += completion_text.replace(f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN) + "\n" + function_calls.append(completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip()) + grammar = get_grammar(function_calls[-1]) + completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) + function_bodies.append(completion["choices"][0]["text"].strip()) + # If the prompt involves a function call, just append generated parameters to function_bodies + else: + function_bodies.append(completion_text.strip()) + else: + # Loop until all parallel function calls are generated + while True: + # If no or "auto" tool_choice/function_call + if function_call is None or ( + isinstance(function_call, str) and function_call == "auto" + ): + grammar = None + stops = CONTENT_TOKEN + # If tool_choice/function_call is "none" + elif isinstance(function_call, str) and function_call == "none": + prompt = prepare_messages_for_inference(messages, tokenizer, version, [], []) + "all\n<|content|>" + stops = STOP_TOKEN + # If tool_choice/function_call is provided + elif isinstance(function_call, dict): + prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" + stops = STOP_TOKEN + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + else: + prompt = prompt + stops = STOP_TOKEN + + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + + # If the generation does not involve a function call + if prompt.endswith("all\n<|content|>") and not completion_text.startswith("all"): + return _convert_completion_to_chat(completion, stream=stream) # type: ignore + # Generate model response if the model decides not to call any function + elif (prompt.endswith(RECIPIENT_TOKEN) and completion_text.startswith("all")): + prompt += completion_text + CONTENT_TOKEN + completion = create_completion(stop=STOP_TOKEN) + return _convert_completion_to_chat(completion, stream=stream) # type: ignore + # Generate parameters if model decides to call a function + elif prompt.endswith(RECIPIENT_TOKEN): + function_calls.append(completion_text[:-1]) + grammar = get_grammar(function_calls[-1]) + completion = create_completion(stop=[STOP_TOKEN, "\n"]) + function_bodies.append(completion["choices"][0]["text"].strip()) + prompt += f"{function_calls[-1]}\n{CONTENT_TOKEN}{function_bodies[-1]}" + grammar = None + + # Try to generate the beginning of next turn + # If empty completion, break from loop + next_turn_completion_text = create_completion( + stop=[STOP_TOKEN, RECIPIENT_TOKEN] + )["choices"][0]["text"] + if len(next_turn_completion_text) > 0: + prompt += f"\n{FROM_TOKEN}assistant\n{RECIPIENT_TOKEN}" + else: + break + # Break from loop if tool_choice/function_call is provided as a dict + else: + function_bodies.append(completion_text.strip()) + break + + assert "usage" in completion + assert len(function_calls) > 0 + assert len(function_calls) == len(function_bodies) + + tool_calls = [] + for function_call, function_body in zip(function_calls, function_bodies): + tool_calls.append( + { + "id": "call_" + "".join( + [random.choice(string.ascii_letters + string.digits) for _ in range(24)] + ), + "type": "function", + "function": { + "name": function_call, + "arguments": function_body, + }, + } + ) + + # TODO: support stream mode + return llama_types.CreateChatCompletionResponse( + id="chat" + completion["id"], + object="chat.completion", + created=completion["created"], + model=completion["model"], + choices=[ + { + "index": 0, + "message": { + "role": "assistant", + "content": None, + "function_call": { + "name": tool_calls[0]["function"]["name"], + "arguments": tool_calls[0]["function"]["arguments"], + }, + "tool_calls": tool_calls, + }, + "finish_reason": "tool_calls", + } + ], + usage=completion["usage"], + ) + + class Llava15ChatHandler: _clip_free = None diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 925ab99b7..6d8ec2467 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -93,6 +93,10 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: ) ) + tokenizer: Optional[llama_cpp.BaseLlamaTokenizer] = None + if settings.hf_pretrained_model_name_or_path is not None: + tokenizer = llama_cpp.LlamaHFTokenizer.from_pretrained(settings.hf_pretrained_model_name_or_path) + draft_model = None if settings.draft_model is not None: draft_model = llama_speculative.LlamaPromptLookupDecoding( @@ -156,6 +160,8 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler=chat_handler, # Speculative Decoding draft_model=draft_model, + # Tokenizer + tokenizer=tokenizer, # Misc verbose=settings.verbose, ) From 2ef7ba3aed572609fbf7292adb125e41e5279a15 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Feb 2024 01:07:44 -0500 Subject: [PATCH 107/624] misc: rename grammar test --- tests/{test_grammar.py => test_llama_grammar.py} | 0 1 file changed, 0 insertions(+), 0 deletions(-) rename tests/{test_grammar.py => test_llama_grammar.py} (100%) diff --git a/tests/test_grammar.py b/tests/test_llama_grammar.py similarity index 100% rename from tests/test_grammar.py rename to tests/test_llama_grammar.py From b5fca911b57a23565c55c31802fb9603a0c6497c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Feb 2024 01:08:18 -0500 Subject: [PATCH 108/624] feat: Move tokenizer to own module --- llama_cpp/llama.py | 69 ++------------------------ llama_cpp/llama_tokenizer.py | 96 ++++++++++++++++++++++++++++++++++++ 2 files changed, 100 insertions(+), 65 deletions(-) create mode 100644 llama_cpp/llama_tokenizer.py diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index bad75dfaf..30ae3b56c 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2,7 +2,6 @@ import os import sys -import abc import uuid import time import multiprocessing @@ -15,7 +14,6 @@ Iterator, Deque, Callable, - Any, ) from collections import deque @@ -31,6 +29,10 @@ LlamaDiskCache, # type: ignore LlamaRAMCache, # type: ignore ) +from .llama_tokenizer import ( + BaseLlamaTokenizer, + LlamaTokenizer +) import llama_cpp.llama_cpp as llama_cpp import llama_cpp.llama_chat_format as llama_chat_format @@ -1747,69 +1749,6 @@ def longest_token_prefix(a: Sequence[int], b: Sequence[int]): return longest_prefix -class BaseLlamaTokenizer(abc.ABC): - @abc.abstractmethod - def tokenize(self, text: bytes, add_bos: bool = True, special: bool = True) -> List[int]: - raise NotImplementedError - - @abc.abstractmethod - def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: - raise NotImplementedError - - -class LlamaTokenizer(BaseLlamaTokenizer): - def __init__(self, llama: Llama): - self.llama = llama - self._model = llama._model # type: ignore - - def tokenize(self, text: bytes, add_bos: bool = True, special: bool = True) -> List[int]: - return self._model.tokenize(text, add_bos=add_bos, special=special) - - def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: - if prev_tokens is not None: - return self._model.detokenize(tokens[len(prev_tokens):]) - else: - return self._model.detokenize(tokens) - - def encode(self, text: str, add_bos: bool = True, special: bool = True) -> List[int]: - return self.tokenize( - text.encode("utf-8", errors="ignore"), add_bos=add_bos, special=special - ) - - def decode(self, tokens: List[int]) -> str: - return self.detokenize(tokens).decode("utf-8", errors="ignore") - - @classmethod - def from_ggml_file(cls, path: str) -> "LlamaTokenizer": - return cls(Llama(model_path=path, vocab_only=True)) - - -class LlamaHFTokenizer(BaseLlamaTokenizer): - def __init__(self, hf_tokenizer: Any): - self.hf_tokenizer = hf_tokenizer - - def tokenize(self, text: bytes, add_bos: bool = True, special: bool = True) -> List[int]: - return self.hf_tokenizer.encode(text.decode("utf-8", errors="ignore"), add_special_tokens=special) - - def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: - if prev_tokens is not None: - text = self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") - prev_text = self.hf_tokenizer.decode(prev_tokens).encode("utf-8", errors="ignore") - return text[len(prev_text):] - else: - return self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") - - @classmethod - def from_pretrained(cls, pretrained_model_name_or_path: str) -> "LlamaHFTokenizer": - try: - from transformers import AutoTokenizer - except ImportError: - raise ImportError( - "The `transformers` library is required to use the `HFTokenizer`." - "You can install it with `pip install transformers`." - ) - hf_tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=pretrained_model_name_or_path) - return cls(hf_tokenizer) class LlamaState: diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py new file mode 100644 index 000000000..0ad3c3afe --- /dev/null +++ b/llama_cpp/llama_tokenizer.py @@ -0,0 +1,96 @@ +from __future__ import annotations + +import abc +from typing import ( + List, + Optional, + Any, +) + +import llama_cpp +from llama_cpp.llama_types import List + + +class BaseLlamaTokenizer(abc.ABC): + @abc.abstractmethod + def tokenize( + self, text: bytes, add_bos: bool = True, special: bool = True + ) -> List[int]: + raise NotImplementedError + + @abc.abstractmethod + def detokenize( + self, tokens: List[int], prev_tokens: Optional[List[int]] = None + ) -> bytes: + raise NotImplementedError + + +class LlamaTokenizer(BaseLlamaTokenizer): + def __init__(self, llama: llama_cpp.Llama): + self.llama = llama + self._model = llama._model # type: ignore + + def tokenize( + self, text: bytes, add_bos: bool = True, special: bool = True + ) -> List[int]: + return self._model.tokenize(text, add_bos=add_bos, special=special) + + def detokenize( + self, tokens: List[int], prev_tokens: Optional[List[int]] = None + ) -> bytes: + if prev_tokens is not None: + return self._model.detokenize(tokens[len(prev_tokens) :]) + else: + return self._model.detokenize(tokens) + + def encode( + self, text: str, add_bos: bool = True, special: bool = True + ) -> List[int]: + return self.tokenize( + text.encode("utf-8", errors="ignore"), add_bos=add_bos, special=special + ) + + def decode(self, tokens: List[int]) -> str: + return self.detokenize(tokens).decode("utf-8", errors="ignore") + + @classmethod + def from_ggml_file(cls, path: str) -> "LlamaTokenizer": + return cls(llama_cpp.Llama(model_path=path, vocab_only=True)) + + +class LlamaHFTokenizer(BaseLlamaTokenizer): + def __init__(self, hf_tokenizer: Any): + self.hf_tokenizer = hf_tokenizer + + def tokenize( + self, text: bytes, add_bos: bool = True, special: bool = True + ) -> List[int]: + return self.hf_tokenizer.encode( + text.decode("utf-8", errors="ignore"), add_special_tokens=special + ) + + def detokenize( + self, tokens: List[int], prev_tokens: Optional[List[int]] = None + ) -> bytes: + if prev_tokens is not None: + text = self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") + prev_text = self.hf_tokenizer.decode(prev_tokens).encode( + "utf-8", errors="ignore" + ) + return text[len(prev_text) :] + else: + return self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path: str) -> "LlamaHFTokenizer": + try: + from transformers import AutoTokenizer + except ImportError: + raise ImportError( + "The `transformers` library is required to use the `HFTokenizer`." + "You can install it with `pip install transformers`." + ) + hf_tokenizer = AutoTokenizer.from_pretrained( + pretrained_model_name_or_path=pretrained_model_name_or_path + ) + return cls(hf_tokenizer) From 85d3374b4d5892e51e27b9973f9ce3623e076e2a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Feb 2024 01:13:28 -0500 Subject: [PATCH 109/624] fix: broken import --- llama_cpp/server/model.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 6d8ec2467..5308dc2a8 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -6,6 +6,7 @@ import llama_cpp import llama_cpp.llama_speculative as llama_speculative +import llama_cpp.llama_tokenizer as llama_tokenizer from llama_cpp.server.settings import ModelSettings @@ -95,7 +96,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: tokenizer: Optional[llama_cpp.BaseLlamaTokenizer] = None if settings.hf_pretrained_model_name_or_path is not None: - tokenizer = llama_cpp.LlamaHFTokenizer.from_pretrained(settings.hf_pretrained_model_name_or_path) + tokenizer = llama_tokenizer.LlamaHFTokenizer.from_pretrained(settings.hf_pretrained_model_name_or_path) draft_model = None if settings.draft_model is not None: From dfc1b173414b550f8f5be1b94430af16b53a63cb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Feb 2024 23:38:12 -0500 Subject: [PATCH 110/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 213d1439f..8e6a9d2de 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 213d1439fadefe182f69c5f7e8dd3b4b6572ebcb +Subproject commit 8e6a9d2de0096af7120606c74ee2f26684e87b41 From e16f06e6eb555947f4404c20732921c8ea76c4f7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 9 Feb 2024 02:02:13 -0500 Subject: [PATCH 111/624] fix: revert _create_completions. --- llama_cpp/llama.py | 20 ++++++++++++-------- 1 file changed, 12 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index bad75dfaf..f445fb0e9 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -948,8 +948,7 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - prev_tokens = completion_tokens[:returned_tokens] - remaining_text = self.detokenize(completion_tokens, prev_tokens) + remaining_text = self.detokenize(remaining_tokens) remaining_length = len(remaining_text) # We want to avoid yielding any characters from @@ -971,13 +970,13 @@ def logit_bias_processor( for token in remaining_tokens: if token == self.token_bos(): continue - token_end_position += len(remaining_text) + token_end_position += len(self.detokenize([token])) # Check if stop sequence is in the token if token_end_position > ( remaining_length - first_stop_position ): break - token_str = remaining_text.decode( + token_str = self.detokenize([token]).decode( "utf-8", errors="ignore" ) text_offset = len(prompt) + len( @@ -1002,7 +1001,11 @@ def logit_bias_processor( } top_logprob.update({token_str: current_logprobs[int(token)]}) logprobs_or_none = { - "tokens": [token_str], + "tokens": [ + self.detokenize([token]).decode( + "utf-8", errors="ignore" + ) + ], "text_offset": [text_offset], "token_logprobs": [current_logprobs[int(token)]], "top_logprobs": [top_logprob], @@ -1015,7 +1018,9 @@ def logit_bias_processor( "model": model_name, "choices": [ { - "text": token_str, + "text": self.detokenize([token]).decode( + "utf-8", errors="ignore" + ), "index": 0, "logprobs": logprobs_or_none, "finish_reason": None, @@ -1027,7 +1032,7 @@ def logit_bias_processor( decode_success = False for i in range(1, len(remaining_tokens) + 1): try: - bs = remaining_text + bs = self.detokenize(remaining_tokens[:i]) ts = bs.decode("utf-8") decode_success = True break @@ -1063,7 +1068,6 @@ def logit_bias_processor( if len(completion_tokens) >= max_tokens: text = self.detokenize(completion_tokens) - finish_reason = "length" break From 63b0c37836169baa71c04484e5344294928bd359 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 9 Feb 2024 13:36:58 -0500 Subject: [PATCH 112/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8e6a9d2de..4b7b38bef 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8e6a9d2de0096af7120606c74ee2f26684e87b41 +Subproject commit 4b7b38bef5addbd31f453871d79647fbae6bec8a From 19b55ad3e55cc707938b191ab7779f5fd69cd0c6 Mon Sep 17 00:00:00 2001 From: Douglas Hanley Date: Sun, 11 Feb 2024 12:53:59 -0600 Subject: [PATCH 113/624] feat: use gpu backend for clip if available (#1175) --- CMakeLists.txt | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 795dad726..b4df8ef45 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -46,6 +46,14 @@ if (LLAMA_BUILD) ) if (LLAVA_BUILD) + if (LLAMA_CUBLAS) + add_compile_definitions(GGML_USE_CUBLAS) + endif() + + if (LLAMA_METAL) + add_compile_definitions(GGML_USE_METAL) + endif() + # Building llava add_subdirectory(vendor/llama.cpp/examples/llava) set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") From 918ff27e501f621ab7d511a9c71c0783d870082c Mon Sep 17 00:00:00 2001 From: Akarshan Biswas Date: Mon, 12 Feb 2024 00:25:15 +0530 Subject: [PATCH 114/624] docs: Set the correct command for compiling with syscl support (#1172) --- README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index bddef6459..59a7cd491 100644 --- a/README.md +++ b/README.md @@ -118,7 +118,8 @@ CMAKE_ARGS="-DLLAMA_KOMPUTE=on" pip install llama-cpp-python To install with SYCL support, set the `LLAMA_SYCL=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_SYCL=on" pip install llama-cpp-python +source /opt/intel/oneapi/setvars.sh +CMAKE_ARGS="-DLLAMA_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install llama-cpp-python ``` ### Windows Notes From a05d90446fea83426fe8dc0d6c78afe6c3dd0894 Mon Sep 17 00:00:00 2001 From: Connor Date: Sun, 11 Feb 2024 10:57:57 -0800 Subject: [PATCH 115/624] fix: Circular dependancy preventing early Llama object free (#1176) commit 901827013b732d74f1f67033062d13a6204a62bd introduced a cyclic dependency within Llama objects. That change causes old models to linger in memory longer than necessary, thereby creating memory bloat in most applications attempting to switch between models at runtime. This patch simply removes the problematic line, allowing models to deallocate without relying on GC. One might also consider combining `weakref.ref` with a `@property` if the `llama` attribute is absolutely necessary to expose in the tokenizer class. --- llama_cpp/llama_tokenizer.py | 1 - 1 file changed, 1 deletion(-) diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py index 0ad3c3afe..c2aad4744 100644 --- a/llama_cpp/llama_tokenizer.py +++ b/llama_cpp/llama_tokenizer.py @@ -27,7 +27,6 @@ def detokenize( class LlamaTokenizer(BaseLlamaTokenizer): def __init__(self, llama: llama_cpp.Llama): - self.llama = llama self._model = llama._model # type: ignore def tokenize( From 936867063984dc695be71eab21115f183dc4d33b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 11 Feb 2024 14:02:46 -0500 Subject: [PATCH 116/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4b7b38bef..97a336507 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4b7b38bef5addbd31f453871d79647fbae6bec8a +Subproject commit 97a336507ed9b971d72262bec7e2b8b7016a054a From 69413ce08e7119b0b10ddb1a52eb8eaa2a865f7f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 11 Feb 2024 19:00:17 -0500 Subject: [PATCH 117/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index da2a7f32a..2724eddec 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -139,9 +139,11 @@ def _load_shared_library(lib_base_name: str): # enum llama_vocab_type { # LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece # LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding +# LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece # }; LLAMA_VOCAB_TYPE_SPM = 0 LLAMA_VOCAB_TYPE_BPE = 1 +LLAMA_VOCAB_TYPE_WPM = 2 # enum llama_token_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 97a336507..3bdc4cd0f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 97a336507ed9b971d72262bec7e2b8b7016a054a +Subproject commit 3bdc4cd0f595a6096cca4a64aa75ffa8a3503465 From 153a0049d90329dec2fc44628fdc6fc8c5f31ae4 Mon Sep 17 00:00:00 2001 From: Andrei Date: Mon, 12 Feb 2024 15:56:07 -0500 Subject: [PATCH 118/624] feat: Generic chatml Function Calling (#957) * Add demo notebook * Add initial chat handler * Update OpenAI types * Add generic chatml function calling (wip) * Update chatml generic function calling. * Progress on auto-tool calls * fix streaming functions * Remove print statements * fix: Suppress output from llama.cpp init and grammar creation * Add OpenAI v1 python api compatible chat completion function * Support non-streaming multi-tool calls * Format * Include function_call in response. --- .../notebooks/OpenHermesFunctionCalling.ipynb | 910 ++++++++++++++++++ llama_cpp/llama.py | 38 +- llama_cpp/llama_chat_format.py | 761 +++++++++++++-- llama_cpp/llama_types.py | 11 +- 4 files changed, 1660 insertions(+), 60 deletions(-) create mode 100644 examples/notebooks/OpenHermesFunctionCalling.ipynb diff --git a/examples/notebooks/OpenHermesFunctionCalling.ipynb b/examples/notebooks/OpenHermesFunctionCalling.ipynb new file mode 100644 index 000000000..c0de3fdc2 --- /dev/null +++ b/examples/notebooks/OpenHermesFunctionCalling.ipynb @@ -0,0 +1,910 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 1, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n" + ] + } + ], + "source": [ + "import json\n", + "import inspect\n", + "from typing import get_type_hints\n", + "\n", + "class Article:\n", + " pass\n", + "\n", + "class Weather:\n", + " pass\n", + "\n", + "class Directions:\n", + " pass\n", + "\n", + "def calculate_mortgage_payment(loan_amount: int, interest_rate: float, loan_term: int) -> float:\n", + " \"\"\"Get the monthly mortgage payment given an interest rate percentage.\"\"\"\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_article_details(title: str, authors: list[str], short_summary: str, date_published: str, tags: list[str]) -> Article:\n", + " '''Get article details from unstructured article text.\n", + "date_published: formatted as \"MM/DD/YYYY\"'''\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_weather(zip_code: str) -> Weather:\n", + " \"\"\"Get the current weather given a zip code.\"\"\"\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_directions(start: str, destination: str) -> Directions:\n", + " \"\"\"Get directions from Google Directions API.\n", + "start: start address as a string including zipcode (if any)\n", + "destination: end address as a string including zipcode (if any)\"\"\"\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_type_name(t):\n", + " name = str(t)\n", + " if \"list\" in name or \"dict\" in name:\n", + " return name\n", + " else:\n", + " return t.__name__\n", + "\n", + "def serialize_function_to_json(func):\n", + " signature = inspect.signature(func)\n", + " type_hints = get_type_hints(func)\n", + "\n", + " function_info = {\n", + " \"name\": func.__name__,\n", + " \"description\": func.__doc__,\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {}\n", + " },\n", + " \"returns\": type_hints.get('return', 'void').__name__\n", + " }\n", + "\n", + " for name, _ in signature.parameters.items():\n", + " param_type = get_type_name(type_hints.get(name, type(None)))\n", + " function_info[\"parameters\"][\"properties\"][name] = {\"type\": param_type}\n", + "\n", + " return json.dumps(function_info, indent=2)\n", + "\n", + "print(serialize_function_to_json(get_article_details))" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "metadata": {}, + "outputs": [], + "source": [ + "import xml.etree.ElementTree as ET\n", + "import re\n", + "\n", + "def extract_function_calls(completion):\n", + " completion = completion.strip()\n", + " pattern = r\"((.*?))\"\n", + " match = re.search(pattern, completion, re.DOTALL)\n", + " if not match:\n", + " return None\n", + " \n", + " multiplefn = match.group(1)\n", + " root = ET.fromstring(multiplefn)\n", + " functions = root.findall(\"functioncall\")\n", + " return [json.loads(fn.text) for fn in functions]" + ] + }, + { + "cell_type": "code", + "execution_count": 12, + "metadata": {}, + "outputs": [], + "source": [ + "def generate_hermes_prompt(prompt, functions):\n", + " functions = \"\\n\\n\".join([serialize_function_to_json(fn) for fn in functions])\n", + " prompt = f\"\"\"<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{functions}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {{\"name\": \"function_name\", \"arguments\": {{\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}}}} \n", + " {{\"name\": \"function_name\", \"arguments\": {{\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}}}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "{prompt}<|im_end|>\n", + "<|im_start|>assistant\"\"\"\n", + " return prompt" + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "What's the weather in 10001?<|im_end|>\n", + "<|im_start|>assistant\n", + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.<|im_end|>\n", + "<|im_start|>assistant\n", + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "What's the current exchange rate for USD to EUR?<|im_end|>\n", + "<|im_start|>assistant\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 10001?\",\n", + " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", + " \"What's the current exchange rate for USD to EUR?\"\n", + "]\n", + "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", + "\n", + "for prompt in prompts:\n", + " print(generate_hermes_prompt(prompt, functions))" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no\n", + "ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes\n", + "ggml_init_cublas: found 1 CUDA devices:\n", + " Device 0: NVIDIA GeForce RTX 2060, compute capability 7.5\n", + "llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from ../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf (version GGUF V3 (latest))\n", + "llama_model_loader: - tensor 0: token_embd.weight q4_K [ 4096, 32002, 1, 1 ]\n", + "llama_model_loader: - tensor 1: blk.0.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 2: blk.0.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 3: blk.0.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 4: blk.0.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 5: blk.0.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 6: blk.0.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 7: blk.0.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 8: blk.0.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 9: blk.0.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 10: blk.1.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 11: blk.1.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 12: blk.1.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 13: blk.1.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 14: blk.1.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 15: blk.1.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 16: blk.1.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 17: blk.1.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 18: blk.1.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 19: blk.2.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 20: blk.2.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 21: blk.2.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 22: blk.2.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 23: blk.2.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 24: blk.2.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 25: blk.2.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 26: blk.2.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 27: blk.2.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 28: blk.3.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 29: blk.3.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 30: blk.3.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 31: blk.3.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 32: blk.3.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 33: blk.3.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 34: blk.3.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 35: blk.3.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 36: blk.3.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 37: blk.4.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 38: blk.4.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 39: blk.4.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 40: blk.4.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 41: blk.4.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 42: blk.4.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 43: blk.4.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 44: blk.4.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 45: blk.4.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 46: blk.5.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 47: blk.5.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 48: blk.5.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 49: blk.5.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 50: blk.5.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 51: blk.5.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 52: blk.5.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 53: blk.5.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 54: blk.5.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 55: blk.6.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 56: blk.6.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 57: blk.6.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 58: blk.6.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 59: blk.6.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 60: blk.6.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 61: blk.6.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 62: blk.6.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 63: blk.6.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 64: blk.7.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 65: blk.7.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 66: blk.7.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 67: blk.7.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 68: blk.7.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 69: blk.7.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 70: blk.7.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 71: blk.7.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 72: blk.7.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 73: blk.8.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 74: blk.8.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 75: blk.8.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 76: blk.8.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 77: blk.8.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 78: blk.8.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 79: blk.8.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 80: blk.8.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 81: blk.8.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 82: blk.9.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 83: blk.9.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 84: blk.9.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 85: blk.9.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 86: blk.9.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 87: blk.9.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 88: blk.9.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 89: blk.9.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 90: blk.9.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 91: blk.10.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 92: blk.10.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 93: blk.10.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 94: blk.10.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 95: blk.10.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 96: blk.10.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 97: blk.10.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 98: blk.10.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 99: blk.10.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 100: blk.11.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 101: blk.11.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 102: blk.11.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 103: blk.11.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 104: blk.11.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 105: blk.11.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 106: blk.11.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 107: blk.11.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 108: blk.11.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 109: blk.12.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 110: blk.12.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 111: blk.12.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 112: blk.12.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 113: blk.12.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 114: blk.12.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 115: blk.12.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 116: blk.12.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 117: blk.12.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 118: blk.13.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 119: blk.13.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 120: blk.13.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 121: blk.13.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 122: blk.13.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 123: blk.13.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 124: blk.13.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 125: blk.13.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 126: blk.13.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 127: blk.14.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 128: blk.14.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 129: blk.14.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 130: blk.14.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 131: blk.14.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 132: blk.14.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 133: blk.14.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 134: blk.14.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 135: blk.14.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 136: blk.15.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 137: blk.15.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 138: blk.15.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 139: blk.15.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 140: blk.15.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 141: blk.15.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 142: blk.15.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 143: blk.15.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 144: blk.15.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 145: blk.16.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 146: blk.16.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 147: blk.16.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 148: blk.16.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 149: blk.16.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 150: blk.16.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 151: blk.16.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 152: blk.16.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 153: blk.16.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 154: blk.17.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 155: blk.17.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 156: blk.17.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 157: blk.17.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 158: blk.17.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 159: blk.17.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 160: blk.17.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 161: blk.17.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 162: blk.17.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 163: blk.18.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 164: blk.18.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 165: blk.18.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 166: blk.18.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 167: blk.18.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 168: blk.18.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 169: blk.18.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 170: blk.18.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 171: blk.18.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 172: blk.19.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 173: blk.19.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 174: blk.19.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 175: blk.19.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 176: blk.19.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 177: blk.19.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 178: blk.19.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 179: blk.19.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 180: blk.19.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 181: blk.20.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 182: blk.20.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 183: blk.20.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 184: blk.20.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 185: blk.20.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 186: blk.20.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 187: blk.20.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 188: blk.20.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 189: blk.20.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 190: blk.21.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 191: blk.21.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 192: blk.21.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 193: blk.21.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 194: blk.21.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 195: blk.21.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 196: blk.21.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 197: blk.21.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 198: blk.21.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 199: blk.22.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 200: blk.22.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 201: blk.22.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 202: blk.22.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 203: blk.22.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 204: blk.22.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 205: blk.22.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 206: blk.22.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 207: blk.22.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 208: blk.23.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 209: blk.23.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 210: blk.23.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 211: blk.23.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 212: blk.23.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 213: blk.23.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 214: blk.23.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 215: blk.23.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 216: blk.23.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 217: blk.24.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 218: blk.24.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 219: blk.24.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 220: blk.24.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 221: blk.24.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 222: blk.24.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 223: blk.24.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 224: blk.24.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 225: blk.24.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 226: blk.25.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 227: blk.25.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 228: blk.25.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 229: blk.25.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 230: blk.25.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 231: blk.25.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 232: blk.25.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 233: blk.25.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 234: blk.25.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 235: blk.26.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 236: blk.26.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 237: blk.26.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 238: blk.26.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 239: blk.26.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 240: blk.26.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 241: blk.26.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 242: blk.26.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 243: blk.26.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 244: blk.27.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 245: blk.27.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 246: blk.27.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 247: blk.27.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 248: blk.27.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 249: blk.27.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 250: blk.27.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 251: blk.27.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 252: blk.27.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 253: blk.28.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 254: blk.28.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 255: blk.28.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 256: blk.28.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 257: blk.28.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 258: blk.28.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 259: blk.28.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 260: blk.28.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 261: blk.28.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 262: blk.29.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 263: blk.29.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 264: blk.29.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 265: blk.29.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 266: blk.29.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 267: blk.29.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 268: blk.29.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 269: blk.29.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 270: blk.29.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 271: blk.30.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 272: blk.30.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 273: blk.30.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 274: blk.30.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 275: blk.30.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 276: blk.30.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 277: blk.30.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 278: blk.30.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 279: blk.30.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 280: blk.31.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 281: blk.31.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 282: blk.31.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 283: blk.31.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 284: blk.31.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 285: blk.31.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 286: blk.31.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 287: blk.31.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 288: blk.31.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 289: output_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 290: output.weight q6_K [ 4096, 32002, 1, 1 ]\n", + "llama_model_loader: - kv 0: general.architecture str = llama\n", + "llama_model_loader: - kv 1: general.name str = teknium_openhermes-2.5-mistral-7b\n", + "llama_model_loader: - kv 2: llama.context_length u32 = 32768\n", + "llama_model_loader: - kv 3: llama.embedding_length u32 = 4096\n", + "llama_model_loader: - kv 4: llama.block_count u32 = 32\n", + "llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336\n", + "llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128\n", + "llama_model_loader: - kv 7: llama.attention.head_count u32 = 32\n", + "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 8\n", + "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010\n", + "llama_model_loader: - kv 10: llama.rope.freq_base f32 = 10000.000000\n", + "llama_model_loader: - kv 11: general.file_type u32 = 15\n", + "llama_model_loader: - kv 12: tokenizer.ggml.model str = llama\n", + "llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32002] = [\"\", \"\", \"\", \"<0x00>\", \"<...\n", + "llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32002] = [0.000000, 0.000000, 0.000000, 0.0000...\n", + "llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32002] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...\n", + "llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1\n", + "llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 32000\n", + "llama_model_loader: - kv 18: tokenizer.ggml.padding_token_id u32 = 0\n", + "llama_model_loader: - kv 19: general.quantization_version u32 = 2\n", + "llama_model_loader: - type f32: 65 tensors\n", + "llama_model_loader: - type q4_K: 193 tensors\n", + "llama_model_loader: - type q6_K: 33 tensors\n", + "llm_load_vocab: special tokens definition check successful ( 261/32002 ).\n", + "llm_load_print_meta: format = GGUF V3 (latest)\n", + "llm_load_print_meta: arch = llama\n", + "llm_load_print_meta: vocab type = SPM\n", + "llm_load_print_meta: n_vocab = 32002\n", + "llm_load_print_meta: n_merges = 0\n", + "llm_load_print_meta: n_ctx_train = 32768\n", + "llm_load_print_meta: n_embd = 4096\n", + "llm_load_print_meta: n_head = 32\n", + "llm_load_print_meta: n_head_kv = 8\n", + "llm_load_print_meta: n_layer = 32\n", + "llm_load_print_meta: n_rot = 128\n", + "llm_load_print_meta: n_gqa = 4\n", + "llm_load_print_meta: f_norm_eps = 0.0e+00\n", + "llm_load_print_meta: f_norm_rms_eps = 1.0e-05\n", + "llm_load_print_meta: f_clamp_kqv = 0.0e+00\n", + "llm_load_print_meta: f_max_alibi_bias = 0.0e+00\n", + "llm_load_print_meta: n_ff = 14336\n", + "llm_load_print_meta: rope scaling = linear\n", + "llm_load_print_meta: freq_base_train = 10000.0\n", + "llm_load_print_meta: freq_scale_train = 1\n", + "llm_load_print_meta: n_yarn_orig_ctx = 32768\n", + "llm_load_print_meta: rope_finetuned = unknown\n", + "llm_load_print_meta: model type = 7B\n", + "llm_load_print_meta: model ftype = mostly Q4_K - Medium\n", + "llm_load_print_meta: model params = 7.24 B\n", + "llm_load_print_meta: model size = 4.07 GiB (4.83 BPW) \n", + "llm_load_print_meta: general.name = teknium_openhermes-2.5-mistral-7b\n", + "llm_load_print_meta: BOS token = 1 ''\n", + "llm_load_print_meta: EOS token = 32000 '<|im_end|>'\n", + "llm_load_print_meta: UNK token = 0 ''\n", + "llm_load_print_meta: PAD token = 0 ''\n", + "llm_load_print_meta: LF token = 13 '<0x0A>'\n", + "llm_load_tensors: ggml ctx size = 0.11 MiB\n", + "llm_load_tensors: using CUDA for GPU acceleration\n", + "llm_load_tensors: mem required = 70.42 MiB\n", + "llm_load_tensors: offloading 32 repeating layers to GPU\n", + "llm_load_tensors: offloading non-repeating layers to GPU\n", + "llm_load_tensors: offloaded 35/35 layers to GPU\n", + "llm_load_tensors: VRAM used: 4095.06 MiB\n", + "...............................................................................................\n", + "llama_new_context_with_model: n_ctx = 2048\n", + "llama_new_context_with_model: freq_base = 10000.0\n", + "llama_new_context_with_model: freq_scale = 1\n", + "llama_kv_cache_init: offloading v cache to GPU\n", + "llama_kv_cache_init: offloading k cache to GPU\n", + "llama_kv_cache_init: VRAM kv self = 256.00 MiB\n", + "llama_new_context_with_model: kv self size = 256.00 MiB\n", + "llama_build_graph: non-view tensors processed: 740/740\n", + "llama_new_context_with_model: compute buffer total size = 159.07 MiB\n", + "llama_new_context_with_model: VRAM scratch buffer: 156.00 MiB\n", + "llama_new_context_with_model: total VRAM used: 4507.07 MiB (model: 4095.06 MiB, context: 412.00 MiB)\n" + ] + } + ], + "source": [ + "import llama_cpp\n", + "\n", + "llama = llama_cpp.Llama(model_path=\"../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf\", n_gpu_layers=-1, n_ctx=2048, verbose=False)" + ] + }, + { + "cell_type": "code", + "execution_count": 22, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "[{'name': 'get_weather', 'arguments': {'zip_code': '10001'}}]\n", + "====================================================================================================\n", + "[{'name': 'calculate_mortgage_payment', 'arguments': {'loan_amount': 200000, 'interest_rate': 0.04, 'loan_term': 30}}]\n", + "====================================================================================================\n", + "Unfortunately, I do not have a built-in function to check currency exchange rates. However, you can use third-party APIs or websites like Google Finance or XE to get this information.\n", + "====================================================================================================\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 10001?\",\n", + " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", + " \"What's the current exchange rate for USD to EUR?\"\n", + "]\n", + "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", + "\n", + "for prompt in prompts:\n", + " prompt = generate_hermes_prompt(prompt, functions)\n", + " completion = llama.create_completion(prompt, max_tokens=-1)[\"choices\"][0][\"text\"]\n", + " function_calls = extract_function_calls(completion)\n", + " if function_calls:\n", + " print(function_calls)\n", + " else:\n", + " print(completion.strip())\n", + " print(\"=\"*100)" + ] + }, + { + "cell_type": "code", + "execution_count": 23, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "get_weather\n", + "{'zip_code': '05751'}\n", + "====================================================================================================\n", + "get_weather\n", + "{'zip_code': '05751'}\n", + "get_weather\n", + "{'zip_code': '07030'}\n", + "calculate_mortgage_payment\n", + "{'loan_amount': 250000, 'interest_rate': 4.18, 'loan_term': 30}\n", + "====================================================================================================\n", + "I don't have a function to get exchange rates, but I can provide some resources where you can find this information. You can check websites like Google Finance, XE.com, or Yahoo Finance for up-to-date currency exchange rates.\n", + "====================================================================================================\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 05751?\",\n", + " \"I'm planning a trip to Killington, Vermont (05751) from Hoboken, NJ (07030). Can you get me weather for both locations and directions?\",\n", + " \"What's the current exchange rate for USD to EUR?\"\n", + "]\n", + "\n", + "for prompt in prompts:\n", + " completion = llama.create_completion(generate_hermes_prompt(prompt, functions), max_tokens=-1)[\"choices\"][0][\"text\"]\n", + " function_calls = extract_function_calls(completion)\n", + "\n", + " if function_calls:\n", + " for function in function_calls:\n", + " print(function[\"name\"])\n", + " print(function[\"arguments\"])\n", + " else:\n", + " print(completion.strip())\n", + "\n", + " print(\"=\"*100)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": ".venv", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.11.5+" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3dcb4b58a..3efd95df5 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -50,6 +50,9 @@ _LlamaSamplingContext, # type: ignore ) from ._logger import set_verbose +from ._utils import ( + suppress_stdout_stderr +) class Llama: @@ -182,7 +185,8 @@ def __init__( self.numa = numa if not Llama.__backend_initialized: - llama_cpp.llama_backend_init(self.numa) + with suppress_stdout_stderr(disable=verbose): + llama_cpp.llama_backend_init(self.numa) Llama.__backend_initialized = True self.model_path = model_path @@ -1567,6 +1571,38 @@ def create_chat_completion( logit_bias=logit_bias, ) + def create_chat_completion_openai_v1( + self, + *args: Any, + **kwargs: Any, + ): + """Generate a chat completion with return type based on the the OpenAI v1 API. + + OpenAI python package is required to use this method. + + You can install it with `pip install openai`. + + Args: + *args: Positional arguments to pass to create_chat_completion. + **kwargs: Keyword arguments to pass to create_chat_completion. + + Returns: + Generated chat completion or a stream of chat completion chunks. + """ + try: + from openai.types.chat import ChatCompletion, ChatCompletionChunk + stream = kwargs.get("stream", False) # type: ignore + assert isinstance(stream, bool) + if stream: + return (ChatCompletionChunk(**chunk) for chunk in self.create_chat_completion(*args, **kwargs)) # type: ignore + else: + return ChatCompletion(**self.create_chat_completion(*args, **kwargs)) # type: ignore + except ImportError: + raise ImportError( + "To use create_chat_completion_openai_v1, you must install the openai package." + "You can install it with `pip install openai`." + ) + def __getstate__(self): return dict( model_path=self.model_path, diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 2e4204121..af60d5f26 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -31,6 +31,7 @@ ### Chat Completion Handler ### + class LlamaChatCompletionHandler(Protocol): """Base Protocol for a llama chat completion handler. @@ -77,8 +78,7 @@ def __call__( ) -> Union[ llama_types.CreateChatCompletionResponse, Iterator[llama_types.CreateChatCompletionStreamResponse], - ]: - ... + ]: ... class LlamaChatCompletionHandlerNotFoundException(Exception): @@ -134,6 +134,7 @@ def decorator(f: LlamaChatCompletionHandler): ### Chat Formatter ### + @dataclasses.dataclass class ChatFormatterResponse: """Dataclass that stores completion parameters for a given chat format and @@ -157,8 +158,7 @@ def __call__( *, messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, - ) -> ChatFormatterResponse: - ... + ) -> ChatFormatterResponse: ... class Jinja2ChatFormatter(ChatFormatter): @@ -195,7 +195,7 @@ def raise_exception(message: str): eos_token=self.eos_token, bos_token=self.bos_token, raise_exception=raise_exception, - add_generation_prompt=self.add_generation_prompt + add_generation_prompt=self.add_generation_prompt, ) return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token]) @@ -255,11 +255,13 @@ def _convert_text_completion_chunks_to_chat( "choices": [ { "index": 0, - "delta": { - "content": chunk["choices"][0]["text"], - } - if chunk["choices"][0]["finish_reason"] is None - else {}, + "delta": ( + { + "content": chunk["choices"][0]["text"], + } + if chunk["choices"][0]["finish_reason"] is None + else {} + ), "finish_reason": chunk["choices"][0]["finish_reason"], } ], @@ -338,10 +340,12 @@ def chat_completion_handler( # create grammar from json schema if "schema" in response_format: grammar = llama_grammar.LlamaGrammar.from_json_schema( - json.dumps(response_format["schema"]) + json.dumps(response_format["schema"]), verbose=llama.verbose ) except Exception as e: - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) completion_or_chunks = llama.create_completion( prompt=prompt, @@ -452,7 +456,9 @@ def hf_tokenizer_config_to_chat_completion_handler( tokenizer_config: Dict[str, Any], add_generation_prompt: bool = True, ) -> LlamaChatCompletionHandler: - chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config, add_generation_prompt=add_generation_prompt) + chat_formatter = hf_tokenizer_config_to_chat_formatter( + tokenizer_config, add_generation_prompt=add_generation_prompt + ) return chat_formatter_to_chat_completion_handler(chat_formatter) @@ -463,11 +469,12 @@ def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[s if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE: return "chatml" - if metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE: + if metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE: return "mistral-instruct" return None + ### Utility functions for formatting chat prompts ### # TODO: Replace these with jinja2 templates @@ -916,9 +923,17 @@ def format_mistral_instruct( stop = eos prompt = bos for message in messages: - if message["role"] == "user" and message["content"] is not None and isinstance(message["content"], str): + if ( + message["role"] == "user" + and message["content"] is not None + and isinstance(message["content"], str) + ): prompt += "[INST] " + message["content"] - elif message["role"] == "assistant" and message["content"] is not None and isinstance(message["content"], str): + elif ( + message["role"] == "assistant" + and message["content"] is not None + and isinstance(message["content"], str) + ): prompt += " [/INST]" + message["content"] + eos prompt += " [/INST]" return ChatFormatterResponse(prompt=prompt, stop=stop) @@ -958,6 +973,7 @@ def format_openchat( _prompt = _format_chatml(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt, stop=_sep) + # Chat format for Saiga models, see more details and available models: # https://huggingface.co/collections/IlyaGusev/saiga2-saigamistral-6505d4ccc3d1e53166b636cd @register_chat_format("saiga") @@ -979,8 +995,10 @@ def format_saiga( _prompt += "bot" return ChatFormatterResponse(prompt=_prompt.strip()) + # Tricky chat formats that require custom chat handlers + @register_chat_completion_handler("functionary") def functionary_chat_handler( llama: llama.Llama, @@ -1253,7 +1271,8 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): json.dumps(function_body) ) grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.json_schema_to_gbnf(json.dumps(function_body)) + llama_grammar.json_schema_to_gbnf(json.dumps(function_body)), + verbose=llama.verbose, ) print(grammar_text) except Exception as e: @@ -1264,11 +1283,14 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): print(e) with suppress_stdout_stderr(disable=llama.verbose): grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF + llama_grammar.JSON_GBNF, + verbose=llama.verbose, ) else: with suppress_stdout_stderr(disable=llama.verbose): - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) completion: llama_types.Completion = llama.create_completion( prompt=new_prompt, @@ -1365,11 +1387,13 @@ def functionary_v1_v2_chat_handler( **kwargs, # type: ignore ) -> Union[llama_types.ChatCompletion, Iterator[llama_types.ChatCompletionChunk]]: SYSTEM_MESSAGE = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary""" - + tokenizer = llama.tokenizer_ - assert hasattr(tokenizer, "hf_tokenizer"), "Please provide a valid hf_tokenizer_path from https://huggingface.co/meetkai when initializing the Llama class" + assert hasattr( + tokenizer, "hf_tokenizer" + ), "Please provide a valid hf_tokenizer_path from https://huggingface.co/meetkai when initializing the Llama class" from transformers import AutoTokenizer - + if "<|START_OF_FUNCTION_CALL|>" in tokenizer.hf_tokenizer.additional_special_tokens: version = "v1" END_SYSTEM_TOKEN = "<|END_OF_SYSTEM|>" @@ -1513,13 +1537,16 @@ def prepare_messages_for_inference( "name" ] = f"functions.{message['function_call']['name']}" all_messages.append(message) - + if version == "v1": suffix = "assistant:\n" else: suffix = "<|from|>assistant\n<|recipient|>" - - return tokenizer.hf_tokenizer.apply_chat_template(all_messages, tokenize=False) + suffix + + return ( + tokenizer.hf_tokenizer.apply_chat_template(all_messages, tokenize=False) + + suffix + ) if tools is not None: functions = [tool["function"] for tool in tools if tool["type"] == "function"] @@ -1529,8 +1556,10 @@ def prepare_messages_for_inference( tool_choice if isinstance(tool_choice, str) else tool_choice["function"] ) - prompt = prepare_messages_for_inference(messages, tokenizer, version, functions, tools) - + prompt = prepare_messages_for_inference( + messages, tokenizer, version, functions, tools + ) + # If no tools/functions are provided if function_call is None and (functions is None or len(functions) == 0): if version == "v1": @@ -1538,7 +1567,7 @@ def prepare_messages_for_inference( else: stop = STOP_TOKEN prompt += "all\n<|content|>" - + completion_or_completion_chunks = llama.create_completion( prompt=prompt, temperature=temperature, @@ -1561,9 +1590,9 @@ def prepare_messages_for_inference( grammar=grammar, ) return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore - + assert stream is False # TODO: support stream mode - + def get_grammar(function_call): function_body = None for function in functions or []: @@ -1574,7 +1603,7 @@ def get_grammar(function_call): if tool["type"] == "function" and tool["function"]["name"] == function_call: function_body = tool["function"]["parameters"] break - + try: with suppress_stdout_stderr(disable=llama.verbose): grammar_text = llama_grammar.json_schema_to_gbnf( @@ -1592,11 +1621,11 @@ def get_grammar(function_call): print(e) with suppress_stdout_stderr(disable=llama.verbose): grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF + llama_grammar.JSON_GBNF, verbose=llama.verbose ) - + return grammar - + def create_completion(stop): completion: llama_types.Completion = llama.create_completion( prompt=prompt, @@ -1619,11 +1648,11 @@ def create_completion(stop): logits_processor=logits_processor, grammar=grammar, ) - + return completion - + function_calls, function_bodies = [], [] - + if version == "v1": # If no or "auto" tool_choice/function_call if function_call is None or ( @@ -1632,7 +1661,9 @@ def create_completion(stop): stops = ["\n", END_ASSISTANT_TOKEN] # If tool_choice/function_call is "none" elif isinstance(function_call, str) and function_call == "none": - prompt = prepare_messages_for_inference(messages, tokenizer, version, [], []) + prompt = prepare_messages_for_inference( + messages, tokenizer, version, [], [] + ) stops = END_ASSISTANT_TOKEN # If tool_choice/function_call is provided elif isinstance(function_call, dict): @@ -1647,14 +1678,27 @@ def create_completion(stop): completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] - + # If the generation does not involve a function call - if START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN not in completion_text: + if ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN not in completion_text + ): return _convert_completion_to_chat(completion, stream=stream) # type: ignore # If the generation involves a function call in completion, generate the parameters - elif START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN in completion_text: - prompt += completion_text.replace(f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN) + "\n" - function_calls.append(completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip()) + elif ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN in completion_text + ): + prompt += ( + completion_text.replace( + f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN + ) + + "\n" + ) + function_calls.append( + completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() + ) grammar = get_grammar(function_calls[-1]) completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) function_bodies.append(completion["choices"][0]["text"].strip()) @@ -1672,7 +1716,10 @@ def create_completion(stop): stops = CONTENT_TOKEN # If tool_choice/function_call is "none" elif isinstance(function_call, str) and function_call == "none": - prompt = prepare_messages_for_inference(messages, tokenizer, version, [], []) + "all\n<|content|>" + prompt = ( + prepare_messages_for_inference(messages, tokenizer, version, [], []) + + "all\n<|content|>" + ) stops = STOP_TOKEN # If tool_choice/function_call is provided elif isinstance(function_call, dict): @@ -1684,15 +1731,17 @@ def create_completion(stop): else: prompt = prompt stops = STOP_TOKEN - + completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] - + # If the generation does not involve a function call - if prompt.endswith("all\n<|content|>") and not completion_text.startswith("all"): + if prompt.endswith("all\n<|content|>") and not completion_text.startswith( + "all" + ): return _convert_completion_to_chat(completion, stream=stream) # type: ignore # Generate model response if the model decides not to call any function - elif (prompt.endswith(RECIPIENT_TOKEN) and completion_text.startswith("all")): + elif prompt.endswith(RECIPIENT_TOKEN) and completion_text.startswith("all"): prompt += completion_text + CONTENT_TOKEN completion = create_completion(stop=STOP_TOKEN) return _convert_completion_to_chat(completion, stream=stream) # type: ignore @@ -1704,7 +1753,7 @@ def create_completion(stop): function_bodies.append(completion["choices"][0]["text"].strip()) prompt += f"{function_calls[-1]}\n{CONTENT_TOKEN}{function_bodies[-1]}" grammar = None - + # Try to generate the beginning of next turn # If empty completion, break from loop next_turn_completion_text = create_completion( @@ -1718,17 +1767,21 @@ def create_completion(stop): else: function_bodies.append(completion_text.strip()) break - + assert "usage" in completion assert len(function_calls) > 0 assert len(function_calls) == len(function_bodies) - + tool_calls = [] for function_call, function_body in zip(function_calls, function_bodies): tool_calls.append( { - "id": "call_" + "".join( - [random.choice(string.ascii_letters + string.digits) for _ in range(24)] + "id": "call_" + + "".join( + [ + random.choice(string.ascii_letters + string.digits) + for _ in range(24) + ] ), "type": "function", "function": { @@ -1924,7 +1977,9 @@ def __call__( json.dumps(response_format["schema"]) ) except Exception as e: - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF + ) return _convert_completion_to_chat( llama.create_completion( @@ -1950,3 +2005,601 @@ def __call__( ), stream=stream, ) + + +@register_chat_completion_handler("chatml-function-calling") +def chatml_function_calling( + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + response_format: Optional[llama_types.ChatCompletionRequestResponseFormat] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + **kwargs, # type: ignore +) -> Union[ + llama_types.CreateChatCompletionResponse, + Iterator[llama_types.CreateChatCompletionStreamResponse], +]: + function_calling_template = ( + "{% for message in messages %}" + "<|im_start|>{{ message.role }}\n" + # System message + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% if tool_calls %}" + "\n\nYou have access to the following functions:\n" + "{% for tool in tools %}" + "\nfunctions.{{ tool.function.name }}:\n" + "{{ tool.function.parameters | tojson }}" + "\n{% endfor %}" + "\n\nYou can respond to users messages with either a single message or one or more function calls." + "\n\nTo respond with a message begin the message with 'message:', use the following format:" + "\n\nmessage:" + "\n" + "\n\nTo respond with one or more function calls begin the message with 'functions.:', use the following format:" + "\n\nfunctions.:" + '\n{ "arg1": "value1", "arg2": "value2" }' + "\nfunctions.:" + '\n{ "arg1": "value1", "arg2": "value2" }' + "{% endif %}" + "\n<|im_end|>\n" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "{{ message.content }}" + "\n<|im_end|>\n" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + ## Reglar message + "{% if message.content and message.content | length > 0 %}" + "message:\n" + "{{ message.content }}" + "\n<|im_end|>\n" + "{% endif %}" + ## Function calls + "{% if message.tool_calls %}" + "{% for tool_call in message.tool_calls %}" + "functions.{{ tool_call.function.name }}:\n" + "{{ tool_call.function.arguments }}" + "{% endfor %}" + "\n<|im_end|>\n" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + ) + template_renderer = jinja2.Environment( + loader=jinja2.BaseLoader(), + autoescape=jinja2.select_autoescape(["html", "xml"]), + undefined=jinja2.StrictUndefined, + ).from_string(function_calling_template) + + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + # Case 1: No tool choice by user + if ( + tool_choice is None + or (isinstance(tool_choice, str) and tool_choice == "none") + or tools is None + or len(tools) == 0 + ): + prompt = template_renderer.render( + messages=messages, + tools=[], + tool_calls=None, + ) + if response_format is not None and response_format["type"] == "json_object": + try: + grammar = ( + llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(response_format["schema"]) + ) + if "schema" in response_format + else None + ) + except Exception as e: + if llama.verbose: + print( + "Failed to parse response format as JSON schema, falling back to default grammar" + ) + print(e) + grammar = ( + llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + if grammar is None + else grammar + ) + return _convert_completion_to_chat( + llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ), + stream=stream, + ) + + def _convert_completion_to_chat_function( + tool_name: str, + completion_or_chunks: Union[ + llama_types.CreateCompletionResponse, + Iterator[llama_types.CreateCompletionStreamResponse], + ], + stream: bool, + ): + if not stream: + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + assert "usage" in completion + tool_id = "call_" + "_0_" + tool_name + "_" + completion["id"] + # TODO: Fix for legacy function calls + chat_completion: llama_types.CreateChatCompletionResponse = { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "index": 0, + "message": { + "role": "assistant", + "content": None, + "function_call": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + "tool_calls": [ + { + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + ], + }, + "finish_reason": "tool_calls", + } + ], + "usage": completion["usage"], + } + return chat_completion + else: + chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore + + def _stream_response_to_function_stream( + chunks: Iterator[llama_types.CreateCompletionStreamResponse], + ) -> Iterator[llama_types.CreateChatCompletionStreamResponse]: + # blank first message + first = True + id_ = None + created = None + model = None + tool_id = None + for chunk in chunks: + if first: + id_ = "chat" + chunk["id"] + created = chunk["created"] + model = chunk["model"] + tool_id = "call_" + "_0_" + tool_name + "_" + chunk["id"] + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": "assistant", + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": "", + }, + } + ], + }, + } + ], + } + first = False + continue + assert tool_id is not None + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": chunk["choices"][0][ + "text" + ], + }, + } + ], + }, + } + ], + } + + if id_ is not None and created is not None and model is not None: + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": "tool_calls", + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + + return _stream_response_to_function_stream(chunks) + + # Case 2: Tool choice by user + if isinstance(tool_choice, dict): + tool_name = tool_choice["function"]["name"] + tool = next( + (tool for tool in tools if tool["function"]["name"] == tool_name), None + ) + if tool is None: + raise ValueError(f"Tool with name '{tool_name}' not found in tools") + prompt = template_renderer.render( + messages=messages, + tools=tools, + tool_calls=True, + ) + prompt += f"functions.{tool_name}:\n" + try: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(tool["function"]["parameters"]), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) + + # Case 3: Automatic tool choice + assert isinstance(tool_choice, str) and tool_choice == "auto" + function_names = " | ".join( + [f'''"functions.{tool['function']['name']}:"''' for tool in tools] + ) + initial_gbnf_tool_grammar = ( + """root ::= functions | "message:"\n""" + f"""functions ::= {function_names}\n""" + ) + follow_up_gbnf_tool_grammar = ( + """root ::= functions | "<|im_end|>"\n""" + f"""functions ::= {function_names}\n""" + ) + prompt = template_renderer.render( + messages=messages, + tools=tools, + tool_calls=True, + ) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=0, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=[":"], + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + initial_gbnf_tool_grammar, verbose=llama.verbose + ), + ) + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + text = completion["choices"][0]["text"] + if "message" in text: + return _convert_completion_to_chat( + llama.create_completion( + prompt=prompt + "message:\n", + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=["<|im_end|>"], + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + follow_up_gbnf_tool_grammar, verbose=llama.verbose + ), + ), + stream=stream, + ) + + # One or more function calls + tool_name = text[len("functions.") :] + tool = next((tool for tool in tools if tool["function"]["name"] == tool_name), None) + if not stream: + completions = [] + completions_tool_name = [] + while tool is not None: + prompt += f"functions.{tool_name}:\n" + try: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(tool["function"]["parameters"]), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=stop, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + completions.append(completion_or_chunks) + completions_tool_name.append(tool_name) + prompt += completion_or_chunks["choices"][0]["text"] + prompt += "\n" + + response = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=stop, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + follow_up_gbnf_tool_grammar, verbose=llama.verbose + ), + ) + + tool_name = response["choices"][0]["text"][len("functions.") :] + tool = next( + (tool for tool in tools if tool["function"]["name"] == tool_name), None + ) + + # Merge completions + function_call = { + "function_call": { + "name": tool_name, + "arguments": completions[0]["choices"][0]["text"], + } + } if len(completions) == 1 else {} + return { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "finish_reason": "tool_calls", + "index": 0, + "message": { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_" + + f"_{i}_" + + tool_name + + "_" + + completion["id"], + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + for i, (tool_name, completion) in enumerate( + zip(completions_tool_name, completions) + ) + ], + **function_call + }, + } + ], + "usage": { + "completion_tokens": sum( + completion["usage"]["completion_tokens"] + for completion in completions + ), + "prompt_tokens": sum( + completion["usage"]["prompt_tokens"] for completion in completions + ), + "total_tokens": sum( + completion["usage"]["total_tokens"] for completion in completions + ), + }, + } + + raise ValueError("Automatic streaming tool choice is not supported") diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index c3deba87b..1b1befebe 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -97,7 +97,7 @@ class CreateChatCompletionResponse(TypedDict): class ChatCompletionMessageToolCallChunkFunction(TypedDict): - name: str + name: Optional[str] arguments: str @@ -118,12 +118,12 @@ class ChatCompletionStreamResponseDeltaFunctionCall(TypedDict): class ChatCompletionStreamResponseDelta(TypedDict): - content: NotRequired[str] + content: NotRequired[Optional[str]] function_call: NotRequired[ - ChatCompletionStreamResponseDeltaFunctionCall + Optional[ChatCompletionStreamResponseDeltaFunctionCall] ] # DEPRECATED - tool_calls: NotRequired[List[ChatCompletionMessageToolCallChunk]] - role: NotRequired[Literal["system", "user", "assistant", "tool"]] + tool_calls: NotRequired[Optional[List[ChatCompletionMessageToolCallChunk]]] + role: NotRequired[Optional[Literal["system", "user", "assistant", "tool"]]] class ChatCompletionStreamResponseChoice(TypedDict): @@ -132,6 +132,7 @@ class ChatCompletionStreamResponseChoice(TypedDict): ChatCompletionStreamResponseDelta, ChatCompletionStreamResponseDeltaEmpty ] finish_reason: Optional[Literal["stop", "length", "tool_calls", "function_call"]] + logprobs: NotRequired[Optional[CompletionLogprobs]] class CreateChatCompletionStreamResponse(TypedDict): From cb791716b42eb897acf66b8b78c4a67b6e026a74 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 12 Feb 2024 16:19:05 -0500 Subject: [PATCH 119/624] fix: Always set logits_all = True when using speculative decoding --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3efd95df5..4869a9d3c 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -281,7 +281,7 @@ def __init__( ) self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0 self.context_params.mul_mat_q = mul_mat_q - self.context_params.logits_all = logits_all + self.context_params.logits_all = logits_all if draft_model is None else True # Must be set to True for speculative decoding self.context_params.embedding = embedding self.context_params.offload_kqv = offload_kqv From b82b0e10148659bceb5e79ae221a4bce2c54da79 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 12 Feb 2024 16:27:43 -0500 Subject: [PATCH 120/624] docs: Temporarily revert function calling docs --- README.md | 22 +++++++++++++--------- 1 file changed, 13 insertions(+), 9 deletions(-) diff --git a/README.md b/README.md index 59a7cd491..679c97709 100644 --- a/README.md +++ b/README.md @@ -292,18 +292,22 @@ To constrain the response to a specific JSON Schema, you can use the `schema` pr ### Function Calling -The high-level API also provides a simple interface for function calling. +The high-level API also provides a simple interface for function calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat forma. -The only set of models that supports full function calling at this time is [functionary](https://github.com/MeetKai/functionary). The various gguf-converted files for this set of models can be found [here](https://huggingface.co/meetkai). Functionary is able to intelligently call functions and also analyze any provided function outputs to generate coherent responses. All v2 models of functionary supports **parallel function calling**. You can provide either `functionary-v1` or `functionary-v2` for the `chat_format` when initializing the Llama class. - -Note that due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is required to provide HF Tokenizer for functionary. The `LlamaHFTokenizer` class can be initialized and passed into the Llama class. This will override the default llama.cpp tokenizer used in Llama class. The tokenizer files are already included in the respective HF repositories hosting the gguf files. +The gguf-converted files for functionary can be found here: [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) ```python ->>> from llama_cpp import Llama, LlamaHFTokenizer ->>> tokenizer = LlamaHFTokenizer.from_pretrained("path/to/functionary/") ->>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", tokenizer=tokenizer, chat_format="functionary-v2") +>>> from llama_cpp import Llama +>>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", chat_format="functionary") +>>> # or +>>> llm = Llama(model_path="path/to/chatml/llama-model.gguf", chat_format="chatml-function-calling") >>> llm.create_chat_completion( messages = [ + { + "role": "system", + "content": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary" + + }, { "role": "user", "content": "Extract Jason is 25 years old" @@ -330,12 +334,12 @@ Note that due to discrepancies between llama.cpp and HuggingFace's tokenizers, i } } }], - tool_choice={ + tool_choice=[{ "type": "function", "function": { "name": "UserDetail" } - }, + }] ) ``` From d605875772a381d863b3960d14e8eeb52908b561 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 12 Feb 2024 16:28:30 -0500 Subject: [PATCH 121/624] Bump version --- CHANGELOG.md | 8 ++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 5ce0b4313..d2bb71077 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.40] + +- feat: Update llama.cpp to ggerganov/llama.cpp@3bdc4cd0f595a6096cca4a64aa75ffa8a3503465 +- feat: Generic chatml Function Calling using chat_format="chatml-function-calling"` by @abetlen in #957 +- fix: Circular dependancy preventing early Llama object free by @notwa in #1176 +- docs: Set the correct command for compiling with syscl support by @akarshanbiswas in #1172 +- feat: use gpu backend for clip if available by @iamlemec in #1175 + ## [0.2.39] - feat: Update llama.cpp to ggerganov/llama.cpp@b08f22c882a1443e6b97081f3ce718a4d1a741f8 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 837e3c993..ccafd029e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.39" \ No newline at end of file +__version__ = "0.2.40" \ No newline at end of file From 4348a6cdf057f5746db213867f93ed1359091fa3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 02:04:54 -0500 Subject: [PATCH 122/624] docs: Fix typo --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 679c97709..3d8d4d401 100644 --- a/README.md +++ b/README.md @@ -292,7 +292,7 @@ To constrain the response to a specific JSON Schema, you can use the `schema` pr ### Function Calling -The high-level API also provides a simple interface for function calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat forma. +The high-level API also provides a simple interface for function calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat format. The gguf-converted files for functionary can be found here: [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) From 5efc45bdfde9c37db27dabecc2955ab9863506c9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 02:43:07 -0500 Subject: [PATCH 123/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 3bdc4cd0f..895407f31 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 3bdc4cd0f595a6096cca4a64aa75ffa8a3503465 +Subproject commit 895407f31b358e3d9335e847d13f033491ec8a5b From d1822fed6b706f38bd1ff0de4dec5baaa3cf84fa Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 02:44:00 -0500 Subject: [PATCH 124/624] fix: Don't change order of json schema object properties unless prop_order is passed, Closes #1180 --- llama_cpp/llama_grammar.py | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index d8ef563c2..3eb3b965b 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1471,12 +1471,15 @@ def visit(self, schema: Dict[str, Any], name: str) -> str: if schema_type == "object" and "properties" in schema: # TODO: `required` keyword - prop_order = self._prop_order - prop_pairs = sorted( - schema["properties"].items(), - # sort by position in prop_order (if specified) then by key - key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), - ) + if self._prop_order: + prop_order = self._prop_order + prop_pairs = sorted( + schema["properties"].items(), + # sort by position in prop_order (if specified) then by key + key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), + ) + else: + prop_pairs = schema["properties"].items() rule = '"{" space' for i, (prop_name, prop_schema) in enumerate(prop_pairs): From 6fe8b427e1608782ad29b313130ba2fa3e4220b8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 02:46:52 -0500 Subject: [PATCH 125/624] Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index d2bb71077..e8fcb80bf 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.41] + +- feat: Update llama.cpp to ggerganov/llama.cpp@895407f31b358e3d9335e847d13f033491ec8a5b +- fix: Don't change order of json schema object properties in generated grammar unless prop_order is passed by @abetlen in d1822fed6b706f38bd1ff0de4dec5baaa3cf84fa + ## [0.2.40] - feat: Update llama.cpp to ggerganov/llama.cpp@3bdc4cd0f595a6096cca4a64aa75ffa8a3503465 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index ccafd029e..6bc5e8a43 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.40" \ No newline at end of file +__version__ = "0.2.41" \ No newline at end of file From 4b0e3320bd8c2c209e29978d0b21e2e471cc9ee3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 03:11:35 -0500 Subject: [PATCH 126/624] fix: minor formatting bugs for chatml-function-calling --- llama_cpp/llama_chat_format.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index af60d5f26..66e40ae41 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2072,12 +2072,14 @@ def chatml_function_calling( "{% if message.role == 'assistant' %}" ## Reglar message "{% if message.content and message.content | length > 0 %}" + "{% if tool_calls %}" "message:\n" + "{% endif %}" "{{ message.content }}" "\n<|im_end|>\n" "{% endif %}" ## Function calls - "{% if message.tool_calls %}" + "{% if 'tool_calls' in message %}" "{% for tool_call in message.tool_calls %}" "functions.{{ tool_call.function.name }}:\n" "{{ tool_call.function.arguments }}" From 68fb71b6a26a1e57331868f959b47ab4b87851e1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 03:24:41 -0500 Subject: [PATCH 127/624] fix: missing generation_prompt in chatml-function-calling --- llama_cpp/llama_chat_format.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 66e40ae41..809a827fa 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2088,6 +2088,7 @@ def chatml_function_calling( "{% endif %}" "{% endif %}" "{% endfor %}" + "{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" ) template_renderer = jinja2.Environment( loader=jinja2.BaseLoader(), @@ -2130,6 +2131,7 @@ def chatml_function_calling( messages=messages, tools=[], tool_calls=None, + add_generation_prompt=True, ) if response_format is not None and response_format["type"] == "json_object": try: @@ -2363,6 +2365,7 @@ def _stream_response_to_function_stream( messages=messages, tools=tools, tool_calls=True, + add_generation_prompt=True, ) prompt += f"functions.{tool_name}:\n" try: @@ -2420,6 +2423,7 @@ def _stream_response_to_function_stream( messages=messages, tools=tools, tool_calls=True, + add_generation_prompt=True, ) completion_or_chunks = llama.create_completion( prompt=prompt, From f7cdf78788da3ef33e3d3a482998d756ee47e8e3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 12:24:00 -0500 Subject: [PATCH 128/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 18 ++++++++++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 19 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 2724eddec..9979a67fa 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -470,6 +470,7 @@ class llama_model_params(Structure): # bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) # bool embedding; // embedding mode only # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU +# bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) # }; class llama_context_params(Structure): """Parameters for llama_context @@ -496,6 +497,7 @@ class llama_context_params(Structure): logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) embedding (bool): embedding mode only offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU + do_pooling (bool): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) """ _fields_ = [ @@ -520,6 +522,7 @@ class llama_context_params(Structure): ("logits_all", c_bool), ("embedding", c_bool), ("offload_kqv", c_bool), + ("do_pooling", c_bool), ] @@ -1699,6 +1702,21 @@ def llama_get_embeddings( _lib.llama_get_embeddings.restype = c_float_p +# // Get the embeddings for the ith sequence +# // llama_get_embeddings(ctx) + i*n_embd +# LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); +def llama_get_embeddings_ith( + ctx: llama_context_p, i: Union[c_int32, int] +): # type: (...) -> Array[float] # type: ignore + """Get the embeddings for the ith sequence + llama_get_embeddings(ctx) + i*n_embd""" + return _lib.llama_get_embeddings_ith(ctx, i) + + +_lib.llama_get_embeddings_ith.argtypes = [llama_context_p, c_int32] +_lib.llama_get_embeddings_ith.restype = c_float_p + + # // # // Vocab # // diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 895407f31..ea9c8e114 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 895407f31b358e3d9335e847d13f033491ec8a5b +Subproject commit ea9c8e11436ad50719987fa23a289c74b7b40d40 From d6be5333e1e28dd07cfec5babd6332c7d1f50788 Mon Sep 17 00:00:00 2001 From: Andrew Lapp Date: Tue, 13 Feb 2024 17:26:07 +0000 Subject: [PATCH 129/624] fix: sample idx off-by-one error for logit_processors (#1179) * fix sample_idx off-by-one error * self._scores is indexed differently, only modify the index within self._input_ids --------- Co-authored-by: Andrew Lapp Co-authored-by: Andrei --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 4869a9d3c..8d726d38f 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -557,7 +557,7 @@ def sample( logits[:] = ( logits_processor(self._input_ids, logits) if idx is None - else logits_processor(self._input_ids[:idx], logits) + else logits_processor(self._input_ids[:idx + 1], logits) ) sampling_params = _LlamaSamplingParams( From b1637c2319936df0ecf1b3eb18ca971b346a147e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 12:35:04 -0500 Subject: [PATCH 130/624] Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index e8fcb80bf..dbc4dcafc 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.42] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ea9c8e11436ad50719987fa23a289c74b7b40d40 +- fix: sample idx off-by-one error for logit_processors by @lapp0 in #1179 +- fix: chat formatting bugs in `chatml-function-calling` by @abetlen in 4b0e3320bd8c2c209e29978d0b21e2e471cc9ee3 and 68fb71b6a26a1e57331868f959b47ab4b87851e1 + ## [0.2.41] - feat: Update llama.cpp to ggerganov/llama.cpp@895407f31b358e3d9335e847d13f033491ec8a5b diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 6bc5e8a43..6e71792f5 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.41" \ No newline at end of file +__version__ = "0.2.42" \ No newline at end of file From 345215a76cf57b769474ea5dc1aefc5ccfb06d5c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 23:02:50 -0500 Subject: [PATCH 131/624] fix: more chatml-function-calling fixes --- llama_cpp/llama_chat_format.py | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 809a827fa..7f365e3f0 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2061,12 +2061,12 @@ def chatml_function_calling( "\nfunctions.:" '\n{ "arg1": "value1", "arg2": "value2" }' "{% endif %}" - "\n<|im_end|>\n" + "<|im_end|>\n" "{% endif %}" # User message "{% if message.role == 'user' %}" "{{ message.content }}" - "\n<|im_end|>\n" + "<|im_end|>\n" "{% endif %}" # Assistant message "{% if message.role == 'assistant' %}" @@ -2076,7 +2076,7 @@ def chatml_function_calling( "message:\n" "{% endif %}" "{{ message.content }}" - "\n<|im_end|>\n" + "<|im_end|>\n" "{% endif %}" ## Function calls "{% if 'tool_calls' in message %}" @@ -2084,11 +2084,11 @@ def chatml_function_calling( "functions.{{ tool_call.function.name }}:\n" "{{ tool_call.function.arguments }}" "{% endfor %}" - "\n<|im_end|>\n" + "<|im_end|>\n" "{% endif %}" "{% endif %}" "{% endfor %}" - "{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" + "{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}" ) template_renderer = jinja2.Environment( loader=jinja2.BaseLoader(), @@ -2120,6 +2120,8 @@ def chatml_function_calling( }, } + stop = [stop, "<|im_end|>"] if isinstance(stop, str) else stop + ["<|im_end|>"] if stop else ["<|im_end|>"] + # Case 1: No tool choice by user if ( tool_choice is None From 7dbbfdecadebe7750be650d9409959640ff9a460 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 23:53:56 -0500 Subject: [PATCH 132/624] fix: submodule kompute is not included in sdist. Closes #1165 --- .github/workflows/build-and-release.yaml | 4 ++-- .github/workflows/build-docker.yaml | 2 +- .github/workflows/publish-to-test.yaml | 2 +- .github/workflows/publish.yaml | 2 +- .github/workflows/test.yaml | 10 +++++----- 5 files changed, 10 insertions(+), 10 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 61027ef99..63c81f1e3 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -16,7 +16,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" # Used to host cibuildwheel - uses: actions/setup-python@v3 @@ -48,7 +48,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - uses: actions/setup-python@v3 with: python-version: "3.8" diff --git a/.github/workflows/build-docker.yaml b/.github/workflows/build-docker.yaml index 27a6b1e50..750b91e1f 100644 --- a/.github/workflows/build-docker.yaml +++ b/.github/workflows/build-docker.yaml @@ -14,7 +14,7 @@ jobs: - name: Checkout uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - name: Set up QEMU uses: docker/setup-qemu-action@v2 diff --git a/.github/workflows/publish-to-test.yaml b/.github/workflows/publish-to-test.yaml index 9932d6130..47e7c40b1 100644 --- a/.github/workflows/publish-to-test.yaml +++ b/.github/workflows/publish-to-test.yaml @@ -18,7 +18,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - name: Set up Python uses: actions/setup-python@v4 with: diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index 7d6c97048..1afdd667d 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -12,7 +12,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - name: Set up Python uses: actions/setup-python@v4 with: diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 2cc6fb02d..77df54697 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -19,7 +19,7 @@ jobs: steps: - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v4 with: @@ -42,7 +42,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v4 with: @@ -65,7 +65,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v4 with: @@ -85,7 +85,7 @@ jobs: # steps: # - uses: actions/checkout@v3 # with: - # submodules: "true" + # submodules: "recursive" # - name: Set up Python 3.8 # uses: actions/setup-python@v4 # with: @@ -112,7 +112,7 @@ jobs: steps: - uses: actions/checkout@v3 with: - submodules: "true" + submodules: "recursive" - name: Set up Python 3.8 uses: actions/setup-python@v4 with: From 7a79e5ac493a3e25a38861828d3e0be3b3c71771 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 23:54:05 -0500 Subject: [PATCH 133/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ea9c8e114..f5ca05485 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ea9c8e11436ad50719987fa23a289c74b7b40d40 +Subproject commit f5ca054855dea83f424003162f26de376e5643f6 From 07a783779a62a4aac0b11161c7e0eb983ff215f8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Feb 2024 23:57:10 -0500 Subject: [PATCH 134/624] fix: Update openbuddy prompt format. Closes #1155 --- llama_cpp/llama_chat_format.py | 19 ++++++++----------- 1 file changed, 8 insertions(+), 11 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 7f365e3f0..8dd0ddfd2 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -734,17 +734,14 @@ def format_openbuddy( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - _system_message = """Consider a conversation between User (a human) and Assistant (named Buddy). -Buddy is an INTP-T, a friendly, intelligent and multilingual AI assistant, by OpenBuddy team. GitHub: https://github.com/OpenBuddy/OpenBuddy -Buddy cannot access the Internet. -Buddy can fluently speak the user's language (e.g. English, Chinese). -Buddy can generate poems, stories, code, essays, songs, parodies, and more. -Buddy possesses vast knowledge about the world, history, and culture. -Buddy's responses are always safe, creative, high-quality, human-like, and interesting. -Buddy strictly refuses to discuss political, NSFW, or other unsafe topics. - -User: Hi. -Assistant: Hi, I'm Buddy, your AI assistant. How can I help you today?""" + _system_message = """You are a helpful, respectful and honest INTP-T AI Assistant named Buddy. You are talking to a human User. +Always answer as helpfully and logically as possible, while being safe. Your answers should not include any harmful, political, religious, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. +If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. +You can speak fluently in many languages, for example: English, Chinese. +You cannot access the internet, but you have vast knowledge, cutoff: 2021-09. +You are trained by OpenBuddy team, (https://openbuddy.ai, https://github.com/OpenBuddy/OpenBuddy), you are based on LLaMA and Falcon transformers model, not related to GPT or OpenAI. + +""" _roles = dict(user="User", assistant="Assistant") _sep = "\n" system_message = _system_message From 6943bab6d817bf71927642ab29e25b94a01fd22c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 14 Feb 2024 03:38:41 -0500 Subject: [PATCH 135/624] fix: destructor exception where internal classes are missing some uninitialized attributes --- llama_cpp/_internals.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 3a71ef0fa..9473d3513 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -42,6 +42,8 @@ def __init__( self._llama_free_model = llama_cpp._lib.llama_free_model # type: ignore + self.model = None + if not os.path.exists(path_model): raise ValueError(f"Model path does not exist: {path_model}") @@ -248,6 +250,7 @@ def __init__( self.verbose = verbose self._llama_free = llama_cpp._lib.llama_free # type: ignore + self.ctx = None assert self.model.model is not None @@ -497,6 +500,7 @@ def __init__( self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore + self.batch = None self.batch = llama_cpp.llama_batch_init( self.n_tokens, self.embd, self.n_seq_max ) From 7b9960d1cbeeca2df6cc3ada6614bc12b2b309fc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 14 Feb 2024 03:47:21 -0500 Subject: [PATCH 136/624] Update llama.cpp --- llama_cpp/llava_cpp.py | 71 +----------------------------------------- vendor/llama.cpp | 2 +- 2 files changed, 2 insertions(+), 71 deletions(-) diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index b1f90b915..8195bd40d 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -146,30 +146,8 @@ def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointe ################################################ -# struct clip_vision_hparams { -# int32_t image_size; -# int32_t patch_size; -# int32_t hidden_size; -# int32_t n_intermediate; -# int32_t projection_dim; -# int32_t n_head; -# int32_t n_layer; -# float eps; -# }; -class clip_vision_hparams(Structure): - _fields_ = [ - ("image_size", c_int32), - ("patch_size", c_int32), - ("hidden_size", c_int32), - ("n_intermediate", c_int32), - ("projection_dim", c_int32), - ("n_head", c_int32), - ("n_layer", c_int32), - ("eps", c_float), - ] - # /** load mmproj model */ -# CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity); +# CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); def clip_model_load(fname: bytes, verbosity: Union[c_int, int]) -> clip_ctx_p: return _libllava.clip_model_load(fname, verbosity) @@ -183,50 +161,3 @@ def clip_free(ctx: clip_ctx_p): _libllava.clip_free.argtypes = [clip_ctx_p] _libllava.clip_free.restype = None - -# size_t clip_embd_nbytes(const struct clip_ctx * ctx); -# int clip_n_patches(const struct clip_ctx * ctx); -# int clip_n_mmproj_embd(const struct clip_ctx * ctx); - -# // RGB uint8 image -# struct clip_image_u8 { -# int nx; -# int ny; -# uint8_t * data = NULL; -# size_t size; -# }; - -# // RGB float32 image (NHWC) -# // Memory layout: RGBRGBRGB... -# struct clip_image_f32 { -# int nx; -# int ny; -# float * data = NULL; -# size_t size; -# }; - -# struct clip_image_u8_batch { -# struct clip_image_u8 * data; -# size_t size; -# }; - -# struct clip_image_f32_batch { -# struct clip_image_f32 * data; -# size_t size; -# }; - -# struct clip_image_u8 * make_clip_image_u8(); -# struct clip_image_f32 * make_clip_image_f32(); -# CLIP_API void clip_image_u8_free(clip_image_u8 * img); -# CLIP_API void clip_image_f32_free(clip_image_f32 * img); -# CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img); -# /** interpret bytes as an image file with length bytes_length, and use the result to populate img */ -# CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img); - -# bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square); -# bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec); - -# bool clip_image_batch_encode(const struct clip_ctx * ctx, const int n_threads, const struct clip_image_f32_batch * imgs, -# float * vec); - -# bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype); \ No newline at end of file diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f5ca05485..aa2341298 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f5ca054855dea83f424003162f26de376e5643f6 +Subproject commit aa2341298924ac89778252015efcb792f2df1e20 From 36b843228f04bd09b642d1500bdf2910f6196c8f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 14 Feb 2024 03:47:40 -0500 Subject: [PATCH 137/624] misc: fix makefile build commands --- Makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Makefile b/Makefile index ff1484c2b..e2ce4d070 100644 --- a/Makefile +++ b/Makefile @@ -19,10 +19,10 @@ build.opencl: CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e . build.openblas: - CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" python3 -m pip install --verbose -e . build.blis: - CMAKE_ARGS="-DLLAMA_OPENBLAS=on -DLLAMA_OPENBLAS_VENDOR=blis" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DLLAMA_BLAS=on -DLLAMA_BLAS_VENDOR=FLAME" python3 -m pip install --verbose -e . build.metal: CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install --verbose -e . From d7a67917ba5b601e146377c6d877893dc49bba83 Mon Sep 17 00:00:00 2001 From: Douglas Hanley Date: Wed, 14 Feb 2024 03:26:09 -0600 Subject: [PATCH 138/624] feat: Support batch embeddings (#1186) * handle batched embeddings * fix normalization issue * fix type hints, ensure no breaking changes to embed * Clear kv cache / reset internal state after embedding complete --------- Co-authored-by: Andrei --- llama_cpp/_internals.py | 22 +++++++ llama_cpp/llama.py | 135 ++++++++++++++++++++++++++++++---------- 2 files changed, 123 insertions(+), 34 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 9473d3513..c60fdff7b 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -510,6 +510,14 @@ def __del__(self): self._llama_batch_free(self.batch) self.batch = None + def n_tokens(self) -> int: + assert self.batch is not None + return self.batch.n_tokens + + def reset(self): + assert self.batch is not None + self.batch.n_tokens = 0 + def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): assert self.batch is not None n_tokens = len(batch) @@ -522,6 +530,20 @@ def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): self.batch.logits[i] = logits_all self.batch.logits[n_tokens - 1] = True + def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool): + assert self.batch is not None + n_tokens = len(batch) + n_tokens0 = self.batch.n_tokens + self.batch.n_tokens += n_tokens + for i in range(n_tokens): + j = n_tokens0 + i + self.batch.token[j] = batch[i] + self.batch.pos[j] = i + self.batch.seq_id[j][0] = seq_id + self.batch.n_seq_id[j] = 1 + self.batch.logits[j] = logits_all + self.batch.logits[n_tokens - 1] = True + class _LlamaTokenDataArray: def __init__(self, *, n_vocab: int): diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 8d726d38f..3e09a20b5 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -717,10 +717,53 @@ def create_embedding( Returns: An embedding object. """ - assert self._ctx.ctx is not None assert self._model.model is not None model_name: str = model if model is not None else self.model_path + # get numeric embeddings + embeds: List[List[float]] + total_tokens: int + embeds, total_tokens = self.embed(input, return_count=True) # type: ignore + + # convert to CreateEmbeddingResponse + data: List[Embedding] = [ + { + "object": "embedding", + "embedding": emb, + "index": idx, + } + for idx, emb in enumerate(embeds) + ] + + return { + "object": "list", + "data": data, + "model": model_name, + "usage": { + "prompt_tokens": total_tokens, + "total_tokens": total_tokens, + }, + } + + def embed( + self, + input: Union[str, List[str]], + normalize: bool = True, + truncate: bool = True, + return_count: bool = False, + ): + """Embed a string. + + Args: + input: The utf-8 encoded string to embed. + + Returns: + A list of embeddings + """ + assert self._ctx.ctx is not None + n_embd = self.n_embd() + n_ctx = self.n_ctx() + if self.context_params.embedding == False: raise RuntimeError( "Llama model must be created with embedding=True to call this method" @@ -734,48 +777,72 @@ def create_embedding( else: inputs = input - data: List[Embedding] = [] + # reset batch + self._batch.reset() + + # decode and fetch embeddings + data: List[List[float]] = [] + def decode_batch(sizes: List[int]): + assert self._ctx.ctx is not None + llama_cpp.llama_kv_cache_clear(self._ctx.ctx) + self._ctx.decode(self._batch) + self._batch.reset() + + # store embeddings + for i, s in enumerate(sizes): + embedding = llama_cpp.llama_get_embeddings_ith(self._ctx.ctx, i)[ + :n_embd + ] + norm = np.linalg.norm(embedding) if normalize else s + embedding: List[float] = [v / float(norm) for v in embedding] + data.append(embedding) + + # init state total_tokens = 0 - for index, input in enumerate(inputs): - tokens = self.tokenize(input.encode("utf-8"), special=True) - self.reset() - self.eval(tokens) + t_batch = 0 + s_sizes: List[int] = [] + + # accumulate batches and encode + for text in inputs: + tokens = self.tokenize(text.encode("utf-8")) + if truncate: + tokens = tokens[:n_ctx] + n_tokens = len(tokens) total_tokens += n_tokens - embedding = llama_cpp.llama_get_embeddings(self._ctx.ctx)[ - : llama_cpp.llama_n_embd(self._model.model) - ] - data.append( - { - "object": "embedding", - "embedding": embedding, - "index": index, - } - ) + # check for overrun + if n_tokens > n_ctx: + raise ValueError( + f"Requested tokens ({n_tokens}) exceed context window of {n_ctx}" + ) + + # time to eval batch + if t_batch + n_tokens > self._n_ctx: + decode_batch(s_sizes) + t_batch = 0 + s_sizes = [] + + # add to batch + self._batch.add_sequence(tokens, len(s_sizes), False) + t_batch += n_tokens + s_sizes.append(n_tokens) + + # hanlde last batch + decode_batch(s_sizes) + if self.verbose: llama_cpp.llama_print_timings(self._ctx.ctx) - return { - "object": "list", - "data": data, - "model": model_name, - "usage": { - "prompt_tokens": total_tokens, - "total_tokens": total_tokens, - }, - } - - def embed(self, input: str) -> List[float]: - """Embed a string. + output = data[0] if isinstance(input, str) else data - Args: - input: The utf-8 encoded string to embed. + llama_cpp.llama_kv_cache_clear(self._ctx.ctx) + self.reset() - Returns: - A list of embeddings - """ - return list(map(float, self.create_embedding(input)["data"][0]["embedding"])) + if return_count: + return output, total_tokens + else: + return output def _create_completion( self, From c336f782693c447a13da250ee12facb535708981 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 14 Feb 2024 04:27:30 -0500 Subject: [PATCH 139/624] Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index aa2341298..8084d5544 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit aa2341298924ac89778252015efcb792f2df1e20 +Subproject commit 8084d554406b767d36b3250b3b787462d5dd626f From ae71ad1a147b10c2c3ba99eb086521cddcc4fad4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 14 Feb 2024 04:31:42 -0500 Subject: [PATCH 140/624] Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index dbc4dcafc..39b553f4f 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.43] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8084d554406b767d36b3250b3b787462d5dd626f +- feat: Support batch embeddings by @iamlemec in #1186 +- fix: submodule kompute is not included in sdist by @abetlen in 7dbbfdecadebe7750be650d9409959640ff9a460 +- fix: fix: Update openbuddy prompt format by @abetlen in 07a783779a62a4aac0b11161c7e0eb983ff215f8 + ## [0.2.42] - feat: Update llama.cpp to ggerganov/llama.cpp@ea9c8e11436ad50719987fa23a289c74b7b40d40 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 6e71792f5..e0bd254eb 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.42" \ No newline at end of file +__version__ = "0.2.43" \ No newline at end of file From 7bb91f025f4b676a741fbf5ddfe9feab40209924 Mon Sep 17 00:00:00 2001 From: Douglas Hanley Date: Thu, 15 Feb 2024 14:16:30 -0600 Subject: [PATCH 141/624] fix: Incorporate embedding pooling layer fixes (#1194) * remove division by token count * truncate to n_batch, not n_ctx --- llama_cpp/llama.py | 33 +++++++++++++++++---------------- 1 file changed, 17 insertions(+), 16 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3e09a20b5..f3c7b4fff 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -762,7 +762,7 @@ def embed( """ assert self._ctx.ctx is not None n_embd = self.n_embd() - n_ctx = self.n_ctx() + n_batch = self.n_batch if self.context_params.embedding == False: raise RuntimeError( @@ -782,54 +782,55 @@ def embed( # decode and fetch embeddings data: List[List[float]] = [] - def decode_batch(sizes: List[int]): + def decode_batch(n_seq: int): assert self._ctx.ctx is not None llama_cpp.llama_kv_cache_clear(self._ctx.ctx) self._ctx.decode(self._batch) self._batch.reset() # store embeddings - for i, s in enumerate(sizes): - embedding = llama_cpp.llama_get_embeddings_ith(self._ctx.ctx, i)[ + for i in range(n_seq): + embedding: List[float] = llama_cpp.llama_get_embeddings_ith(self._ctx.ctx, i)[ :n_embd ] - norm = np.linalg.norm(embedding) if normalize else s - embedding: List[float] = [v / float(norm) for v in embedding] + if normalize: + norm = float(np.linalg.norm(embedding)) + embedding = [v / norm for v in embedding] data.append(embedding) # init state total_tokens = 0 t_batch = 0 - s_sizes: List[int] = [] + p_batch = 0 # accumulate batches and encode for text in inputs: tokens = self.tokenize(text.encode("utf-8")) if truncate: - tokens = tokens[:n_ctx] + tokens = tokens[:n_batch] n_tokens = len(tokens) total_tokens += n_tokens # check for overrun - if n_tokens > n_ctx: + if n_tokens > n_batch: raise ValueError( - f"Requested tokens ({n_tokens}) exceed context window of {n_ctx}" + f"Requested tokens ({n_tokens}) exceed batch size of {n_batch}" ) # time to eval batch - if t_batch + n_tokens > self._n_ctx: - decode_batch(s_sizes) + if t_batch + n_tokens > n_batch: + decode_batch(p_batch) t_batch = 0 - s_sizes = [] + p_batch = 0 # add to batch - self._batch.add_sequence(tokens, len(s_sizes), False) + self._batch.add_sequence(tokens, p_batch, False) t_batch += n_tokens - s_sizes.append(n_tokens) + p_batch += 1 # hanlde last batch - decode_batch(s_sizes) + decode_batch(p_batch) if self.verbose: llama_cpp.llama_print_timings(self._ctx.ctx) From a5cfeb7763e7c2c5833fde1095ed2e64466cb6d8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 15 Feb 2024 15:17:30 -0500 Subject: [PATCH 142/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 9 +++++++++ vendor/llama.cpp | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 9979a67fa..13daadb2e 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -230,6 +230,15 @@ def _load_shared_library(lib_base_name: str): LLAMA_ROPE_SCALING_YARN = 2 LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN +# enum llama_pooling_type { +# LLAMA_POOLING_NONE = 0, +# LLAMA_POOLING_MEAN = 1, +# LLAMA_POOLING_CLS = 2, +# }; +LLAMA_POOLING_NONE = 0 +LLAMA_POOLING_MEAN = 1 +LLAMA_POOLING_CLS = 2 + # enum llama_split_mode { # LLAMA_SPLIT_NONE = 0, // single GPU # LLAMA_SPLIT_LAYER = 1, // split layers and KV across GPUs diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8084d5544..4524290e8 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8084d554406b767d36b3250b3b787462d5dd626f +Subproject commit 4524290e87b8e107cc2b56e1251751546f4b9051 From ea1f88dd293c0c961f98297f8d9aba7e8cf48ce3 Mon Sep 17 00:00:00 2001 From: khimaros Date: Thu, 15 Feb 2024 20:20:13 +0000 Subject: [PATCH 143/624] fix: Use '\n' seperator for EventSourceResponse (#1188) this fixes compatibility with some OpenAI clients, including BetterChatGPT (https://github.com/ztjhz/BetterChatGPT/issues/537). Co-authored-by: Andrei --- llama_cpp/server/app.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 368022c45..7a1391df9 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -290,6 +290,7 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: inner_send_chan=send_chan, iterator=iterator(), ), + sep='\n', ) else: return iterator_or_completion @@ -382,6 +383,7 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: inner_send_chan=send_chan, iterator=iterator(), ), + sep='\n', ) else: return iterator_or_completion From 0ce66bc080fe537590b05b24bf442480bf2dd045 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 15 Feb 2024 16:09:48 -0500 Subject: [PATCH 144/624] fix: create_embedding broken response for input type str --- llama_cpp/llama.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index f3c7b4fff..964b0c80e 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -720,6 +720,8 @@ def create_embedding( assert self._model.model is not None model_name: str = model if model is not None else self.model_path + input = input if isinstance(input, list) else [input] + # get numeric embeddings embeds: List[List[float]] total_tokens: int From f736827b9b58e7ac3ddf2e8737127687d54879c8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 15 Feb 2024 23:10:50 -0500 Subject: [PATCH 145/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 39b553f4f..2c8ff8b13 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.44] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4524290e87b8e107cc2b56e1251751546f4b9051 +- fix: create_embedding broken response for input type str by @abetlen in 0ce66bc080fe537590b05b24bf442480bf2dd045 +- fix: Use '\n' seperator for EventSourceResponse by @khimaros in #1188 +- fix: Incorporate embedding pooling layer fixes by @iamlemec in #1194 + ## [0.2.43] - feat: Update llama.cpp to ggerganov/llama.cpp@8084d554406b767d36b3250b3b787462d5dd626f diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index e0bd254eb..000e4dd36 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.43" \ No newline at end of file +__version__ = "0.2.44" \ No newline at end of file From c2a234a0860f5a65ae1a9655ac1038ef294c8357 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 15 Feb 2024 23:15:50 -0500 Subject: [PATCH 146/624] docs: Add embeddings section --- README.md | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/README.md b/README.md index 3d8d4d401..7da8e5fa7 100644 --- a/README.md +++ b/README.md @@ -398,6 +398,22 @@ llama = Llama( ) ``` +### Embeddings + +`llama-cpp-python` supports generating embeddings from the text. + +```python +import llama_cpp + +llm = llama_cpp.Llama(model_path="path/to/model.gguf", embeddings=True) + +embeddings = llm.create_embedding("Hello, world!") + +# or batched + +embeddings = llm.create_embedding(["Hello, world!", "Goodbye, world!"]) +``` + ### Adjusting the Context Window The context window of the Llama models determines the maximum number of tokens that can be processed at once. By default, this is set to 512 tokens, but can be adjusted based on your requirements. From fdce078cb915cecba7df58eeb494ee8493c87672 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 17 Feb 2024 00:37:51 -0500 Subject: [PATCH 147/624] feat: Update llama.cpp --- llama_cpp/llama.py | 14 ++++++++++---- llama_cpp/llama_cpp.py | 34 +++++++++++++++++++++++++++++++--- llama_cpp/server/settings.py | 4 ++-- vendor/llama.cpp | 2 +- 4 files changed, 44 insertions(+), 10 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 964b0c80e..71f968f44 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -98,7 +98,7 @@ def __init__( lora_scale: float = 1.0, lora_path: Optional[str] = None, # Backend Params - numa: bool = False, + numa: Union[bool, int] = False, # Chat Format Params chat_format: Optional[str] = None, chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None, @@ -166,7 +166,7 @@ def __init__( last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque. lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model. lora_path: Path to a LoRA file to apply to the model. - numa: Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init) + numa: numa policy chat_format: String specifying the chat format to use when calling create_chat_completion. chat_handler: Optional chat handler to use when calling create_chat_completion. draft_model: Optional draft model to use for speculative decoding. @@ -183,12 +183,18 @@ def __init__( set_verbose(verbose) - self.numa = numa if not Llama.__backend_initialized: with suppress_stdout_stderr(disable=verbose): - llama_cpp.llama_backend_init(self.numa) + llama_cpp.llama_backend_init() Llama.__backend_initialized = True + if isinstance(numa, bool): + self.numa = llama_cpp.GGML_NUMA_STRATEGY_DISTRIBUTE if numa else llama_cpp.GGML_NUMA_STRATEGY_DISABLED + + if self.numa != llama_cpp.GGML_NUMA_STRATEGY_DISABLED: + with suppress_stdout_stderr(disable=verbose): + llama_cpp.llama_numa_init(self.numa) + self.model_path = model_path # Model Params diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 13daadb2e..20870371e 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -697,17 +697,45 @@ def llama_model_quantize_default_params() -> llama_model_quantize_params: # // If numa is true, use NUMA optimizations # // Call once at the start of the program # LLAMA_API void llama_backend_init(bool numa); -def llama_backend_init(numa: Union[c_bool, bool]): +# LLAMA_API void llama_backend_init(void); +def llama_backend_init(): """Initialize the llama + ggml backend If numa is true, use NUMA optimizations Call once at the start of the program""" - return _lib.llama_backend_init(numa) + return _lib.llama_backend_init() -_lib.llama_backend_init.argtypes = [c_bool] +_lib.llama_backend_init.argtypes = [] _lib.llama_backend_init.restype = None +# // numa strategies +# enum ggml_numa_strategy { +# GGML_NUMA_STRATEGY_DISABLED = 0, +# GGML_NUMA_STRATEGY_DISTRIBUTE = 1, +# GGML_NUMA_STRATEGY_ISOLATE = 2, +# GGML_NUMA_STRATEGY_NUMACTL = 3, +# GGML_NUMA_STRATEGY_MIRROR = 4, +# GGML_NUMA_STRATEGY_COUNT +# }; +GGML_NUMA_STRATEGY_DISABLED = 0 +GGML_NUMA_STRATEGY_DISTRIBUTE = 1 +GGML_NUMA_STRATEGY_ISOLATE = 2 +GGML_NUMA_STRATEGY_NUMACTL = 3 +GGML_NUMA_STRATEGY_MIRROR = 4 +GGML_NUMA_STRATEGY_COUNT = 5 + + +# //optional: +# LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); +def llama_numa_init(numa: int): + return _lib.llama_numa_init(numa) + + +_lib.llama_numa_init.argtypes = [c_int] +_lib.llama_numa_init.restype = None + + # // Call once at the end of the program - currently only used for MPI # LLAMA_API void llama_backend_free(void); def llama_backend_free(): diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 60f3eeca2..790c6b129 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -2,7 +2,7 @@ import multiprocessing -from typing import Optional, List, Literal +from typing import Optional, List, Literal, Union from pydantic import Field from pydantic_settings import BaseSettings @@ -108,7 +108,7 @@ class ModelSettings(BaseSettings): description="Path to a LoRA file to apply to the model.", ) # Backend Params - numa: bool = Field( + numa: Union[bool, int] = Field( default=False, description="Enable NUMA support.", ) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4524290e8..5bf2b94dd 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4524290e87b8e107cc2b56e1251751546f4b9051 +Subproject commit 5bf2b94dd4fb74378b78604023b31512fec55f8f From 53f6f5f41507b230e52e01210196e9d80edca733 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 17 Feb 2024 01:02:33 -0500 Subject: [PATCH 148/624] fix: self.numa missing --- llama_cpp/llama.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 71f968f44..30cab0af9 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -190,7 +190,9 @@ def __init__( if isinstance(numa, bool): self.numa = llama_cpp.GGML_NUMA_STRATEGY_DISTRIBUTE if numa else llama_cpp.GGML_NUMA_STRATEGY_DISABLED - + else: + self.numa = numa + if self.numa != llama_cpp.GGML_NUMA_STRATEGY_DISABLED: with suppress_stdout_stderr(disable=verbose): llama_cpp.llama_numa_init(self.numa) From 748c0ce05726ed370881bf915d32afe32312f4f2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 18 Feb 2024 21:30:36 -0500 Subject: [PATCH 149/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 20870371e..2a1d3f0fd 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -190,6 +190,7 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -215,6 +216,7 @@ def _load_shared_library(lib_base_name: str): LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22 LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 +LLAMA_FTYPE_MOSTLY_IQ1_S = 24 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5bf2b94dd..a0c2dad9d 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5bf2b94dd4fb74378b78604023b31512fec55f8f +Subproject commit a0c2dad9d43456c677e205c6240a5f8afb0121ac From 6225f027e5b2f999aa5b3e0c829cf5064f1bf9d6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 19 Feb 2024 04:11:34 -0500 Subject: [PATCH 150/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 53 ++++++++++++++++++++++++++++++++++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 54 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 2a1d3f0fd..e946adbf9 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -664,6 +664,18 @@ class llama_timings(Structure): ] +# // used in chat template +# typedef struct llama_chat_message { +# const char * role; +# const char * content; +# } llama_chat_message; +class llama_chat_message(Structure): + _fields_ = [ + ("role", c_char_p), + ("content", c_char_p), + ] + + # // Helpers for getting default parameters # LLAMA_API struct llama_model_params llama_model_default_params(void); def llama_model_default_params() -> llama_model_params: @@ -1956,6 +1968,47 @@ def llama_token_to_piece( _lib.llama_token_to_piece.restype = c_int32 +# /// Apply chat template. Inspired by hf apply_chat_template() on python. +# /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" +# /// NOTE: This function only support some known jinja templates. It is not a jinja parser. +# /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. +# /// @param chat Pointer to a list of multiple llama_chat_message +# /// @param n_msg Number of llama_chat_message in this chat +# /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message. +# /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages) +# /// @param length The size of the allocated buffer +# /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template. +# LLAMA_API int32_t llama_chat_apply_template( +# const struct llama_model * model, +# const char * tmpl, +# const struct llama_chat_message * chat, +# size_t n_msg, +# bool add_ass, +# char * buf, +# int32_t length); +def llama_chat_apply_template( + model: llama_model_p, + tmpl: bytes, + chat: "ctypes._Pointer[llama_chat_message]", + n_msg: int, +) -> int: + return _lib.llama_chat_apply_template( + model, + tmpl, + chat, + n_msg + ) + +_lib.llama_chat_apply_template.argtypes = [ + ctypes.c_void_p, + ctypes.c_char_p, + ctypes.POINTER(llama_chat_message), + ctypes.c_size_t +] +_lib.llama_chat_apply_template.restype = ctypes.c_int32 + + + # // # // Grammar # // diff --git a/vendor/llama.cpp b/vendor/llama.cpp index a0c2dad9d..f53119cec 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit a0c2dad9d43456c677e205c6240a5f8afb0121ac +Subproject commit f53119cec4f073b6d214195ecbe1fad3abdf2b34 From d122bd78588c908a504635aa7951f604f78b6a0a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 19 Feb 2024 22:10:16 -0500 Subject: [PATCH 151/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f53119cec..633782b8d 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f53119cec4f073b6d214195ecbe1fad3abdf2b34 +Subproject commit 633782b8d949f24b619e6c68ee37b5cc79167173 From 04fe33b99991e9f28a726a13a021455b5dfd6ca9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 20 Feb 2024 02:59:02 -0500 Subject: [PATCH 152/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 633782b8d..c0a8c6db3 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 633782b8d949f24b619e6c68ee37b5cc79167173 +Subproject commit c0a8c6db371cb3e4379900867b948879f5842201 From f57b01ac9bb0ec076b253f5155a19b0a2dbb2f69 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 11:04:30 -0500 Subject: [PATCH 153/624] ci: add debug build to dev makefile --- Makefile | 3 +++ 1 file changed, 3 insertions(+) diff --git a/Makefile b/Makefile index e2ce4d070..4ae011074 100644 --- a/Makefile +++ b/Makefile @@ -12,6 +12,9 @@ deps: build: python3 -m pip install --verbose -e . +build.debug: + CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Debug" python3 -m pip install --verbose --config-settings=cmake.verbose=true --config-settings=logging.level=INFO --config-settings=install.strip=false --editable . + build.cuda: CMAKE_ARGS="-DLLAMA_CUBLAS=on" python3 -m pip install --verbose -e . From 4edde21b3d68e853ac296aafa08cd90810657ba0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 11:05:58 -0500 Subject: [PATCH 154/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index e946adbf9..3ebe82b9a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -191,6 +191,7 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -217,6 +218,7 @@ def _load_shared_library(lib_base_name: str): LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22 LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 LLAMA_FTYPE_MOSTLY_IQ1_S = 24 +LLAMA_FTYPE_MOSTLY_IQ4_NL = 25 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c0a8c6db3..89febfed9 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c0a8c6db371cb3e4379900867b948879f5842201 +Subproject commit 89febfed9322c8849520dc63c93ee4f5fd72556e From e42f62c247814099f1d4d396f751d813d153a94c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 11:09:40 -0500 Subject: [PATCH 155/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 2c8ff8b13..43b62c89b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.45] + +- feat: Update llama.cpp to ggerganov/llama.cpp@89febfed9322c8849520dc63c93ee4f5fd72556e + ## [0.2.44] - feat: Update llama.cpp to ggerganov/llama.cpp@4524290e87b8e107cc2b56e1251751546f4b9051 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 000e4dd36..197fc68b7 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.44" \ No newline at end of file +__version__ = "0.2.45" \ No newline at end of file From 0f8aa4ab5c63fbb02630242bfe03d0349929b913 Mon Sep 17 00:00:00 2001 From: Andrei Date: Wed, 21 Feb 2024 16:25:10 -0500 Subject: [PATCH 156/624] feat: Pull models directly from huggingface (#1206) * Add from_pretrained method to Llama class * Update docs * Merge filename and pattern --- README.md | 13 ++++ docs/api-reference.md | 1 + llama_cpp/llama.py | 149 ++++++++++++++++++++++++++++++++++-------- 3 files changed, 136 insertions(+), 27 deletions(-) diff --git a/README.md b/README.md index 7da8e5fa7..03c95e84d 100644 --- a/README.md +++ b/README.md @@ -212,6 +212,19 @@ Below is a short example demonstrating how to use the high-level API to for basi Text completion is available through the [`__call__`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.__call__) and [`create_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_completion) methods of the [`Llama`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama) class. +## Pulling models from Hugging Face + +You can pull `Llama` models from Hugging Face using the `from_pretrained` method. +You'll need to install the `huggingface-hub` package to use this feature (`pip install huggingface-hub`). + +```python +llama = Llama.from_pretrained( + repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", + filename="*q8_0.gguf", + verbose=False +) +``` + ### Chat Completion The high-level API also provides a simple interface for chat completion. diff --git a/docs/api-reference.md b/docs/api-reference.md index 562410fe1..c635e1178 100644 --- a/docs/api-reference.md +++ b/docs/api-reference.md @@ -26,6 +26,7 @@ High-level Python bindings for llama.cpp. - load_state - token_bos - token_eos + - from_pretrained show_root_heading: true ::: llama_cpp.LlamaGrammar diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 30cab0af9..65902c8f4 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -4,6 +4,8 @@ import sys import uuid import time +import json +import fnmatch import multiprocessing from typing import ( List, @@ -16,6 +18,7 @@ Callable, ) from collections import deque +from pathlib import Path import ctypes @@ -29,10 +32,7 @@ LlamaDiskCache, # type: ignore LlamaRAMCache, # type: ignore ) -from .llama_tokenizer import ( - BaseLlamaTokenizer, - LlamaTokenizer -) +from .llama_tokenizer import BaseLlamaTokenizer, LlamaTokenizer import llama_cpp.llama_cpp as llama_cpp import llama_cpp.llama_chat_format as llama_chat_format @@ -50,9 +50,7 @@ _LlamaSamplingContext, # type: ignore ) from ._logger import set_verbose -from ._utils import ( - suppress_stdout_stderr -) +from ._utils import suppress_stdout_stderr class Llama: @@ -189,7 +187,11 @@ def __init__( Llama.__backend_initialized = True if isinstance(numa, bool): - self.numa = llama_cpp.GGML_NUMA_STRATEGY_DISTRIBUTE if numa else llama_cpp.GGML_NUMA_STRATEGY_DISABLED + self.numa = ( + llama_cpp.GGML_NUMA_STRATEGY_DISTRIBUTE + if numa + else llama_cpp.GGML_NUMA_STRATEGY_DISABLED + ) else: self.numa = numa @@ -246,9 +248,9 @@ def __init__( else: raise ValueError(f"Unknown value type for {k}: {v}") - self._kv_overrides_array[ - -1 - ].key = b"\0" # ensure sentinel element is zeroed + self._kv_overrides_array[-1].key = ( + b"\0" # ensure sentinel element is zeroed + ) self.model_params.kv_overrides = self._kv_overrides_array self.n_batch = min(n_ctx, n_batch) # ??? @@ -256,7 +258,7 @@ def __init__( self.n_threads_batch = n_threads_batch or max( multiprocessing.cpu_count() // 2, 1 ) - + # Context Params self.context_params = llama_cpp.llama_context_default_params() self.context_params.seed = seed @@ -289,7 +291,9 @@ def __init__( ) self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0 self.context_params.mul_mat_q = mul_mat_q - self.context_params.logits_all = logits_all if draft_model is None else True # Must be set to True for speculative decoding + self.context_params.logits_all = ( + logits_all if draft_model is None else True + ) # Must be set to True for speculative decoding self.context_params.embedding = embedding self.context_params.offload_kqv = offload_kqv @@ -379,8 +383,14 @@ def __init__( if self.verbose: print(f"Model metadata: {self.metadata}", file=sys.stderr) - if self.chat_format is None and self.chat_handler is None and "tokenizer.chat_template" in self.metadata: - chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata(self.metadata) + if ( + self.chat_format is None + and self.chat_handler is None + and "tokenizer.chat_template" in self.metadata + ): + chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata( + self.metadata + ) if chat_format is not None: self.chat_format = chat_format @@ -406,9 +416,7 @@ def __init__( print(f"Using chat bos_token: {bos_token}", file=sys.stderr) self.chat_handler = llama_chat_format.Jinja2ChatFormatter( - template=template, - eos_token=eos_token, - bos_token=bos_token + template=template, eos_token=eos_token, bos_token=bos_token ).to_chat_handler() if self.chat_format is None and self.chat_handler is None: @@ -459,7 +467,9 @@ def tokenize( """ return self.tokenizer_.tokenize(text, add_bos, special) - def detokenize(self, tokens: List[int], prev_tokens: Optional[List[int]] = None) -> bytes: + def detokenize( + self, tokens: List[int], prev_tokens: Optional[List[int]] = None + ) -> bytes: """Detokenize a list of tokens. Args: @@ -565,7 +575,7 @@ def sample( logits[:] = ( logits_processor(self._input_ids, logits) if idx is None - else logits_processor(self._input_ids[:idx + 1], logits) + else logits_processor(self._input_ids[: idx + 1], logits) ) sampling_params = _LlamaSamplingParams( @@ -707,7 +717,9 @@ def generate( if self.draft_model is not None: self.input_ids[self.n_tokens : self.n_tokens + len(tokens)] = tokens - draft_tokens = self.draft_model(self.input_ids[:self.n_tokens + len(tokens)]) + draft_tokens = self.draft_model( + self.input_ids[: self.n_tokens + len(tokens)] + ) tokens.extend( draft_tokens.astype(int)[ : self._n_ctx - self.n_tokens - len(tokens) @@ -792,6 +804,7 @@ def embed( # decode and fetch embeddings data: List[List[float]] = [] + def decode_batch(n_seq: int): assert self._ctx.ctx is not None llama_cpp.llama_kv_cache_clear(self._ctx.ctx) @@ -800,9 +813,9 @@ def decode_batch(n_seq: int): # store embeddings for i in range(n_seq): - embedding: List[float] = llama_cpp.llama_get_embeddings_ith(self._ctx.ctx, i)[ - :n_embd - ] + embedding: List[float] = llama_cpp.llama_get_embeddings_ith( + self._ctx.ctx, i + )[:n_embd] if normalize: norm = float(np.linalg.norm(embedding)) embedding = [v / norm for v in embedding] @@ -1669,12 +1682,13 @@ def create_chat_completion_openai_v1( """ try: from openai.types.chat import ChatCompletion, ChatCompletionChunk - stream = kwargs.get("stream", False) # type: ignore + + stream = kwargs.get("stream", False) # type: ignore assert isinstance(stream, bool) if stream: - return (ChatCompletionChunk(**chunk) for chunk in self.create_chat_completion(*args, **kwargs)) # type: ignore + return (ChatCompletionChunk(**chunk) for chunk in self.create_chat_completion(*args, **kwargs)) # type: ignore else: - return ChatCompletion(**self.create_chat_completion(*args, **kwargs)) # type: ignore + return ChatCompletion(**self.create_chat_completion(*args, **kwargs)) # type: ignore except ImportError: raise ImportError( "To use create_chat_completion_openai_v1, you must install the openai package." @@ -1866,7 +1880,88 @@ def longest_token_prefix(a: Sequence[int], b: Sequence[int]): break return longest_prefix + @classmethod + def from_pretrained( + cls, + repo_id: str, + filename: Optional[str], + local_dir: Optional[Union[str, os.PathLike[str]]] = ".", + local_dir_use_symlinks: Union[bool, Literal["auto"]] = "auto", + **kwargs: Any, + ) -> "Llama": + """Create a Llama model from a pretrained model name or path. + This method requires the huggingface-hub package. + You can install it with `pip install huggingface-hub`. + + Args: + repo_id: The model repo id. + filename: A filename or glob pattern to match the model file in the repo. + local_dir: The local directory to save the model to. + local_dir_use_symlinks: Whether to use symlinks when downloading the model. + **kwargs: Additional keyword arguments to pass to the Llama constructor. + + Returns: + A Llama model.""" + try: + from huggingface_hub import hf_hub_download, HfFileSystem + from huggingface_hub.utils import validate_repo_id + except ImportError: + raise ImportError( + "Llama.from_pretrained requires the huggingface-hub package. " + "You can install it with `pip install huggingface-hub`." + ) + + validate_repo_id(repo_id) + + hffs = HfFileSystem() + + files = [ + file["name"] if isinstance(file, dict) else file + for file in hffs.ls(repo_id) + ] + + # split each file into repo_id, subfolder, filename + file_list: List[str] = [] + for file in files: + rel_path = Path(file).relative_to(repo_id) + file_list.append(str(rel_path)) + matching_files = [file for file in file_list if fnmatch.fnmatch(file, filename)] # type: ignore + + if len(matching_files) == 0: + raise ValueError( + f"No file found in {repo_id} that match {filename}\n\n" + f"Available Files:\n{json.dumps(file_list)}" + ) + + if len(matching_files) > 1: + raise ValueError( + f"Multiple files found in {repo_id} matching {filename}\n\n" + f"Available Files:\n{json.dumps(files)}" + ) + + (matching_file,) = matching_files + + subfolder = str(Path(matching_file).parent) + filename = Path(matching_file).name + + local_dir = "." + + # download the file + hf_hub_download( + repo_id=repo_id, + local_dir=local_dir, + filename=filename, + subfolder=subfolder, + local_dir_use_symlinks=local_dir_use_symlinks, + ) + + model_path = os.path.join(local_dir, filename) + + return cls( + model_path=model_path, + **kwargs, + ) class LlamaState: From 7f51b6071f7e0bb648ace4dc3ca2294fa0865e09 Mon Sep 17 00:00:00 2001 From: Andrei Date: Wed, 21 Feb 2024 16:25:38 -0500 Subject: [PATCH 157/624] feat(low-level-api): Improve API static type-safety and performance (#1205) --- llama_cpp/_internals.py | 15 +- llama_cpp/llama.py | 2 +- llama_cpp/llama_cpp.py | 1502 +++++++++++++++++++++------------------ llama_cpp/llava_cpp.py | 76 +- tests/test_llama.py | 6 +- 5 files changed, 858 insertions(+), 743 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index c60fdff7b..10c1a6fbb 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -108,7 +108,7 @@ def apply_lora_from_file( scale, path_base_model.encode("utf-8") if path_base_model is not None - else llama_cpp.c_char_p(0), + else ctypes.c_char_p(0), n_threads, ) @@ -303,8 +303,8 @@ def decode(self, batch: "_LlamaBatch"): assert self.ctx is not None assert batch.batch is not None return_code = llama_cpp.llama_decode( - ctx=self.ctx, - batch=batch.batch, + self.ctx, + batch.batch, ) if return_code != 0: raise RuntimeError(f"llama_decode returned {return_code}") @@ -493,7 +493,7 @@ class _LlamaBatch: def __init__( self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True ): - self.n_tokens = n_tokens + self._n_tokens = n_tokens self.embd = embd self.n_seq_max = n_seq_max self.verbose = verbose @@ -502,7 +502,7 @@ def __init__( self.batch = None self.batch = llama_cpp.llama_batch_init( - self.n_tokens, self.embd, self.n_seq_max + self._n_tokens, self.embd, self.n_seq_max ) def __del__(self): @@ -570,12 +570,13 @@ def copy_logits(self, logits: npt.NDArray[np.single]): self.candidates.data = self.candidates_data.ctypes.data_as( llama_cpp.llama_token_data_p ) - self.candidates.sorted = llama_cpp.c_bool(False) - self.candidates.size = llama_cpp.c_size_t(self.n_vocab) + self.candidates.sorted = ctypes.c_bool(False) + self.candidates.size = ctypes.c_size_t(self.n_vocab) # Python wrappers over common/common def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> list[int]: + assert model.model is not None n_tokens = len(text) + 1 if add_bos else len(text) result = (llama_cpp.llama_token * n_tokens)() n_tokens = llama_cpp.llama_tokenize( diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 65902c8f4..9fc4ec2d7 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1818,7 +1818,7 @@ def load_state(self, state: LlamaState) -> None: self.input_ids = state.input_ids.copy() self.n_tokens = state.n_tokens state_size = state.llama_state_size - LLamaStateArrayType = llama_cpp.c_uint8 * state_size + LLamaStateArrayType = ctypes.c_uint8 * state_size llama_state = LLamaStateArrayType.from_buffer_copy(state.llama_state) if llama_cpp.llama_set_state_data(self._ctx.ctx, llama_state) != state_size: diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 3ebe82b9a..69dbe097f 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2,26 +2,11 @@ import os import ctypes from ctypes import ( - c_bool, - c_char_p, - c_int, - c_int8, - c_int32, - c_uint8, - c_uint32, - c_int64, - c_size_t, - c_float, - c_double, - c_void_p, - POINTER, _Pointer, # type: ignore - Structure, - Union as CtypesUnion, Array, ) import pathlib -from typing import List, Union +from typing import List, Union, NewType, Optional # Load the library @@ -71,7 +56,7 @@ def _load_shared_library(lib_base_name: str): for _lib_path in _lib_paths: if _lib_path.exists(): try: - return ctypes.CDLL(str(_lib_path), **cdll_args) + return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore except Exception as e: raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") @@ -87,13 +72,13 @@ def _load_shared_library(lib_base_name: str): _lib = _load_shared_library(_lib_base_name) # Misc -c_float_p = POINTER(c_float) -c_uint8_p = POINTER(c_uint8) -c_size_t_p = POINTER(c_size_t) +c_float_p = ctypes.POINTER(ctypes.c_float) +c_uint8_p = ctypes.POINTER(ctypes.c_uint8) +c_size_t_p = ctypes.POINTER(ctypes.c_size_t) # from ggml-backend.h # typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); -ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE(c_bool, c_void_p, c_bool, c_void_p) +ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p, ctypes.c_bool, ctypes.c_void_p) # llama.h bindings @@ -121,19 +106,21 @@ def _load_shared_library(lib_base_name: str): # struct llama_model; -llama_model_p = c_void_p +llama_model_p = NewType("llama_model_p", int) +llama_model_p_ctypes = ctypes.c_void_p # struct llama_context; -llama_context_p = c_void_p +llama_context_p = NewType("llama_context_p", int) +llama_context_p_ctypes = ctypes.c_void_p # typedef int32_t llama_pos; -llama_pos = c_int32 +llama_pos = ctypes.c_int32 # typedef int32_t llama_token; -llama_token = c_int32 -llama_token_p = POINTER(llama_token) +llama_token = ctypes.c_int32 +llama_token_p = ctypes.POINTER(llama_token) # typedef int32_t llama_seq_id; -llama_seq_id = c_int32 +llama_seq_id = ctypes.c_int32 # enum llama_vocab_type { @@ -258,7 +245,7 @@ def _load_shared_library(lib_base_name: str): # float logit; // log-odds of the token # float p; // probability of the token # } llama_token_data; -class llama_token_data(Structure): +class llama_token_data(ctypes.Structure): """Used to store token data Attributes: @@ -268,12 +255,12 @@ class llama_token_data(Structure): _fields_ = [ ("id", llama_token), - ("logit", c_float), - ("p", c_float), + ("logit", ctypes.c_float), + ("p", ctypes.c_float), ] -llama_token_data_p = POINTER(llama_token_data) +llama_token_data_p = ctypes.POINTER(llama_token_data) # typedef struct llama_token_data_array { @@ -281,7 +268,7 @@ class llama_token_data(Structure): # size_t size; # bool sorted; # } llama_token_data_array; -class llama_token_data_array(Structure): +class llama_token_data_array(ctypes.Structure): """Used to sample tokens given logits Attributes: @@ -291,15 +278,15 @@ class llama_token_data_array(Structure): _fields_ = [ ("data", llama_token_data_p), - ("size", c_size_t), - ("sorted", c_bool), + ("size", ctypes.c_size_t), + ("sorted", ctypes.c_bool), ] -llama_token_data_array_p = POINTER(llama_token_data_array) +llama_token_data_array_p = ctypes.POINTER(llama_token_data_array) # typedef bool (*llama_progress_callback)(float progress, void *ctx); -llama_progress_callback = ctypes.CFUNCTYPE(c_bool, c_float, c_void_p) +llama_progress_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_float, ctypes.c_void_p) # // Input data for llama_decode @@ -332,7 +319,7 @@ class llama_token_data_array(Structure): # llama_pos all_pos_1; // used if pos == NULL # llama_seq_id all_seq_id; // used if seq_id == NULL # } llama_batch; -class llama_batch(Structure): +class llama_batch(ctypes.Structure): """Input data for llama_decode A llama_batch object can contain input about one or many sequences @@ -341,19 +328,19 @@ class llama_batch(Structure): Attributes: token (ctypes.Array[llama_token]): the token ids of the input (used when embd is NULL) - embd (ctypes.Array[ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL) + embd (ctypes.Array[ctypes.ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL) pos (ctypes.Array[ctypes.Array[llama_pos]]): the positions of the respective token in the sequence seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs """ _fields_ = [ - ("n_tokens", c_int32), - ("token", POINTER(llama_token)), + ("n_tokens", ctypes.c_int32), + ("token", ctypes.POINTER(llama_token)), ("embd", c_float_p), - ("pos", POINTER(llama_pos)), - ("n_seq_id", POINTER(c_int32)), - ("seq_id", POINTER(POINTER(llama_seq_id))), - ("logits", POINTER(c_int8)), + ("pos", ctypes.POINTER(llama_pos)), + ("n_seq_id", ctypes.POINTER(ctypes.c_int32)), + ("seq_id", ctypes.POINTER(ctypes.POINTER(llama_seq_id))), + ("logits", ctypes.POINTER(ctypes.c_int8)), ("all_pos_0", llama_pos), ("all_pos_1", llama_pos), ("all_seq_id", llama_seq_id), @@ -379,18 +366,18 @@ class llama_batch(Structure): # bool bool_value; # }; # }; -class llama_model_kv_override_value(CtypesUnion): +class llama_model_kv_override_value(ctypes.Union): _fields_ = [ - ("int_value", c_int64), - ("float_value", c_double), - ("bool_value", c_bool), + ("int_value", ctypes.c_int64), + ("float_value", ctypes.c_double), + ("bool_value", ctypes.c_bool), ] -class llama_model_kv_override(Structure): +class llama_model_kv_override(ctypes.Structure): _fields_ = [ ("key", ctypes.c_char * 128), - ("tag", c_int), + ("tag", ctypes.c_int), ("value", llama_model_kv_override_value), ] @@ -425,32 +412,32 @@ class llama_model_kv_override(Structure): # bool use_mmap; // use mmap if possible # bool use_mlock; // force system to keep model in RAM # }; -class llama_model_params(Structure): +class llama_model_params(ctypes.Structure): """Parameters for llama_model Attributes: n_gpu_layers (int): number of layers to store in VRAM split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored - tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() + tensor_split (ctypes.Array[ctypes.ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. - progress_callback_user_data (ctypes.c_void_p): context pointer passed to the progress callback + progress_callback_user_data (ctypes.ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data vocab_only (bool): only load the vocabulary, no weights use_mmap (bool): use mmap if possible use_mlock (bool): force system to keep model in RAM""" _fields_ = [ - ("n_gpu_layers", c_int32), - ("split_mode", c_int), - ("main_gpu", c_int32), + ("n_gpu_layers", ctypes.c_int32), + ("split_mode", ctypes.c_int), + ("main_gpu", ctypes.c_int32), ("tensor_split", c_float_p), ("progress_callback", llama_progress_callback), - ("progress_callback_user_data", c_void_p), - ("kv_overrides", POINTER(llama_model_kv_override)), - ("vocab_only", c_bool), - ("use_mmap", c_bool), - ("use_mlock", c_bool), + ("progress_callback_user_data", ctypes.c_void_p), + ("kv_overrides", ctypes.POINTER(llama_model_kv_override)), + ("vocab_only", ctypes.c_bool), + ("use_mmap", ctypes.c_bool), + ("use_mlock", ctypes.c_bool), ] @@ -485,7 +472,7 @@ class llama_model_params(Structure): # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU # bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) # }; -class llama_context_params(Structure): +class llama_context_params(ctypes.Structure): """Parameters for llama_context Attributes: @@ -503,7 +490,7 @@ class llama_context_params(Structure): yarn_beta_slow (float): YaRN high correction dim yarn_orig_ctx (int): YaRN original context size cb_eval (ggml_backend_sched_eval_callback): callback for scheduling eval - cb_eval_user_data (ctypes.c_void_p): user data for cb_eval + cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache type_v (int): data type for V cache mul_mat_q (bool): if true, use experimental mul_mat_q kernels (DEPRECATED - always true) @@ -514,28 +501,28 @@ class llama_context_params(Structure): """ _fields_ = [ - ("seed", c_uint32), - ("n_ctx", c_uint32), - ("n_batch", c_uint32), - ("n_threads", c_uint32), - ("n_threads_batch", c_uint32), - ("rope_scaling_type", c_int32), - ("rope_freq_base", c_float), - ("rope_freq_scale", c_float), - ("yarn_ext_factor", c_float), - ("yarn_attn_factor", c_float), - ("yarn_beta_fast", c_float), - ("yarn_beta_slow", c_float), - ("yarn_orig_ctx", c_uint32), + ("seed", ctypes.c_uint32), + ("n_ctx", ctypes.c_uint32), + ("n_batch", ctypes.c_uint32), + ("n_threads", ctypes.c_uint32), + ("n_threads_batch", ctypes.c_uint32), + ("rope_scaling_type", ctypes.c_int32), + ("rope_freq_base", ctypes.c_float), + ("rope_freq_scale", ctypes.c_float), + ("yarn_ext_factor", ctypes.c_float), + ("yarn_attn_factor", ctypes.c_float), + ("yarn_beta_fast", ctypes.c_float), + ("yarn_beta_slow", ctypes.c_float), + ("yarn_orig_ctx", ctypes.c_uint32), ("cb_eval", ggml_backend_sched_eval_callback), - ("cb_eval_user_data", c_void_p), - ("type_k", c_int), - ("type_v", c_int), - ("mul_mat_q", c_bool), - ("logits_all", c_bool), - ("embedding", c_bool), - ("offload_kqv", c_bool), - ("do_pooling", c_bool), + ("cb_eval_user_data", ctypes.c_void_p), + ("type_k", ctypes.c_int), + ("type_v", ctypes.c_int), + ("mul_mat_q", ctypes.c_bool), + ("logits_all", ctypes.c_bool), + ("embedding", ctypes.c_bool), + ("offload_kqv", ctypes.c_bool), + ("do_pooling", ctypes.c_bool), ] @@ -545,7 +532,7 @@ class llama_context_params(Structure): # // if it exists. # // It might not exist for progress report where '.' is output repeatedly. # typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data); -llama_log_callback = ctypes.CFUNCTYPE(None, c_int, c_char_p, c_void_p) +llama_log_callback = ctypes.CFUNCTYPE(None, ctypes.c_int, ctypes.c_char_p, ctypes.c_void_p) """Signature for logging events Note that text includes the new line character at the end for most events. If your logging mechanism cannot handle that, check if the last character is '\n' and strip it @@ -563,7 +550,7 @@ class llama_context_params(Structure): # bool pure; // disable k-quant mixtures and quantize all tensors to the same type # void * imatrix; // pointer to importance matrix data # } llama_model_quantize_params; -class llama_model_quantize_params(Structure): +class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize Attributes: @@ -573,23 +560,23 @@ class llama_model_quantize_params(Structure): quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored pure (bool): disable k-quant mixtures and quantize all tensors to the same type - imatrix (ctypes.c_void_p): pointer to importance matrix data + imatrix (ctypes.ctypes.c_void_p): pointer to importance matrix data """ _fields_ = [ - ("nthread", c_int32), - ("ftype", c_int), - ("allow_requantize", c_bool), - ("quantize_output_tensor", c_bool), - ("only_copy", c_bool), - ("pure", c_bool), - ("imatrix", c_void_p), + ("nthread", ctypes.c_int32), + ("ftype", ctypes.c_int), + ("allow_requantize", ctypes.c_bool), + ("quantize_output_tensor", ctypes.c_bool), + ("only_copy", ctypes.c_bool), + ("pure", ctypes.c_bool), + ("imatrix", ctypes.c_void_p), ] # // grammar types # struct llama_grammar; -llama_grammar_p = c_void_p +llama_grammar_p = ctypes.c_void_p # // grammar element type # enum llama_gretype { @@ -629,14 +616,14 @@ class llama_model_quantize_params(Structure): # enum llama_gretype type; # uint32_t value; // Unicode code point or rule ID # } llama_grammar_element; -class llama_grammar_element(Structure): +class llama_grammar_element(ctypes.Structure): _fields_ = [ - ("type", c_int), - ("value", c_uint32), + ("type", ctypes.c_int), + ("value", ctypes.c_uint32), ] -llama_grammar_element_p = POINTER(llama_grammar_element) +llama_grammar_element_p = ctypes.POINTER(llama_grammar_element) # // performance timing information # struct llama_timings { @@ -652,17 +639,17 @@ class llama_grammar_element(Structure): # int32_t n_p_eval; # int32_t n_eval; # }; -class llama_timings(Structure): +class llama_timings(ctypes.Structure): _fields_ = [ - ("t_start_ms", c_double), - ("t_end_ms", c_double), - ("t_load_ms", c_double), - ("t_sample_ms", c_double), - ("t_p_eval_ms", c_double), - ("t_eval_ms", c_double), - ("n_sample", c_int32), - ("n_p_eval", c_int32), - ("n_eval", c_int32), + ("t_start_ms", ctypes.c_double), + ("t_end_ms", ctypes.c_double), + ("t_load_ms", ctypes.c_double), + ("t_sample_ms", ctypes.c_double), + ("t_p_eval_ms", ctypes.c_double), + ("t_eval_ms", ctypes.c_double), + ("n_sample", ctypes.c_int32), + ("n_p_eval", ctypes.c_int32), + ("n_eval", ctypes.c_int32), ] @@ -671,10 +658,10 @@ class llama_timings(Structure): # const char * role; # const char * content; # } llama_chat_message; -class llama_chat_message(Structure): +class llama_chat_message(ctypes.Structure): _fields_ = [ - ("role", c_char_p), - ("content", c_char_p), + ("role", ctypes.c_char_p), + ("content", ctypes.c_char_p), ] @@ -682,31 +669,34 @@ class llama_chat_message(Structure): # LLAMA_API struct llama_model_params llama_model_default_params(void); def llama_model_default_params() -> llama_model_params: """Get default parameters for llama_model""" - return _lib.llama_model_default_params() + ... -_lib.llama_model_default_params.argtypes = [] -_lib.llama_model_default_params.restype = llama_model_params +llama_model_default_params = _lib.llama_model_default_params +llama_model_default_params.argtypes = [] +llama_model_default_params.restype = llama_model_params # LLAMA_API struct llama_context_params llama_context_default_params(void); def llama_context_default_params() -> llama_context_params: """Get default parameters for llama_context""" - return _lib.llama_context_default_params() + ... -_lib.llama_context_default_params.argtypes = [] -_lib.llama_context_default_params.restype = llama_context_params +llama_context_default_params = _lib.llama_context_default_params +llama_context_default_params.argtypes = [] +llama_context_default_params.restype = llama_context_params # LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void); def llama_model_quantize_default_params() -> llama_model_quantize_params: """Get default parameters for llama_model_quantize""" - return _lib.llama_model_quantize_default_params() + ... -_lib.llama_model_quantize_default_params.argtypes = [] -_lib.llama_model_quantize_default_params.restype = llama_model_quantize_params +llama_model_quantize_default_params = _lib.llama_model_quantize_default_params +llama_model_quantize_default_params.argtypes = [] +llama_model_quantize_default_params.restype = llama_model_quantize_params # // Initialize the llama + ggml backend @@ -718,11 +708,12 @@ def llama_backend_init(): """Initialize the llama + ggml backend If numa is true, use NUMA optimizations Call once at the start of the program""" - return _lib.llama_backend_init() + ... -_lib.llama_backend_init.argtypes = [] -_lib.llama_backend_init.restype = None +llama_backend_init = _lib.llama_backend_init +llama_backend_init.argtypes = [] +llama_backend_init.restype = None # // numa strategies @@ -744,206 +735,227 @@ def llama_backend_init(): # //optional: # LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); -def llama_numa_init(numa: int): - return _lib.llama_numa_init(numa) +def llama_numa_init(numa: int, /): + ... -_lib.llama_numa_init.argtypes = [c_int] -_lib.llama_numa_init.restype = None +llama_numa_init = _lib.llama_numa_init +llama_numa_init.argtypes = [ctypes.c_int] +llama_numa_init.restype = None # // Call once at the end of the program - currently only used for MPI # LLAMA_API void llama_backend_free(void); def llama_backend_free(): """Call once at the end of the program - currently only used for MPI""" - return _lib.llama_backend_free() + ... -_lib.llama_backend_free.argtypes = [] -_lib.llama_backend_free.restype = None +llama_backend_free = _lib.llama_backend_free +llama_backend_free.argtypes = [] +llama_backend_free.restype = None # LLAMA_API struct llama_model * llama_load_model_from_file( # const char * path_model, # struct llama_model_params params); def llama_load_model_from_file( - path_model: bytes, params: llama_model_params -) -> llama_model_p: - return _lib.llama_load_model_from_file(path_model, params) + path_model: bytes, params: llama_model_params, / +) -> Optional[llama_model_p]: + ... -_lib.llama_load_model_from_file.argtypes = [c_char_p, llama_model_params] -_lib.llama_load_model_from_file.restype = llama_model_p +llama_load_model_from_file = _lib.llama_load_model_from_file +llama_load_model_from_file.argtypes = [ctypes.c_char_p, llama_model_params] +llama_load_model_from_file.restype = llama_model_p_ctypes # LLAMA_API void llama_free_model(struct llama_model * model); -def llama_free_model(model: llama_model_p): - return _lib.llama_free_model(model) +def llama_free_model(model: llama_model_p, /): + ... -_lib.llama_free_model.argtypes = [llama_model_p] -_lib.llama_free_model.restype = None +llama_free_model = _lib.llama_free_model +llama_free_model.argtypes = [llama_model_p_ctypes] +llama_free_model.restype = None # LLAMA_API struct llama_context * llama_new_context_with_model( # struct llama_model * model, # struct llama_context_params params); def llama_new_context_with_model( - model: llama_model_p, params: llama_context_params -) -> llama_context_p: - return _lib.llama_new_context_with_model(model, params) + model: llama_model_p, params: llama_context_params, / +) -> Optional[llama_context_p]: + ... -_lib.llama_new_context_with_model.argtypes = [llama_model_p, llama_context_params] -_lib.llama_new_context_with_model.restype = llama_context_p +llama_new_context_with_model = _lib.llama_new_context_with_model +llama_new_context_with_model.argtypes = [llama_model_p_ctypes, llama_context_params] +llama_new_context_with_model.restype = llama_context_p_ctypes # // Frees all allocated memory # LLAMA_API void llama_free(struct llama_context * ctx); -def llama_free(ctx: llama_context_p): +def llama_free(ctx: llama_context_p, /): """Frees all allocated memory""" - return _lib.llama_free(ctx) + ... -_lib.llama_free.argtypes = [llama_context_p] -_lib.llama_free.restype = None +llama_free = _lib.llama_free +llama_free.argtypes = [llama_context_p_ctypes] +llama_free.restype = None # LLAMA_API int64_t llama_time_us(void); def llama_time_us() -> int: - return _lib.llama_time_us() + ... -_lib.llama_time_us.argtypes = [] -_lib.llama_time_us.restype = ctypes.c_int64 +llama_time_us = _lib.llama_time_us +llama_time_us.argtypes = [] +llama_time_us.restype = ctypes.c_int64 # LLAMA_API size_t llama_max_devices(void); def llama_max_devices() -> int: - return _lib.llama_max_devices() + ... -_lib.llama_max_devices.argtypes = [] -_lib.llama_max_devices.restype = ctypes.c_size_t +llama_max_devices = _lib.llama_max_devices +llama_max_devices.argtypes = [] +llama_max_devices.restype = ctypes.c_size_t # LLAMA_API bool llama_supports_mmap (void); def llama_supports_mmap() -> bool: - return _lib.llama_supports_mmap() + ... -_lib.llama_supports_mmap.argtypes = [] -_lib.llama_supports_mmap.restype = c_bool +llama_supports_mmap = _lib.llama_supports_mmap +llama_supports_mmap.argtypes = [] +llama_supports_mmap.restype = ctypes.c_bool # LLAMA_API bool llama_supports_mlock (void); def llama_supports_mlock() -> bool: - return _lib.llama_supports_mlock() + ... -_lib.llama_supports_mlock.argtypes = [] -_lib.llama_supports_mlock.restype = c_bool +llama_supports_mlock = _lib.llama_supports_mlock +llama_supports_mlock.argtypes = [] +llama_supports_mlock.restype = ctypes.c_bool # LLAMA_API bool llama_supports_gpu_offload(void); def llama_supports_gpu_offload() -> bool: - return _lib.llama_supports_gpu_offload() + ... -_lib.llama_supports_gpu_offload.argtypes = [] -_lib.llama_supports_gpu_offload.restype = c_bool +llama_supports_gpu_offload = _lib.llama_supports_gpu_offload +llama_supports_gpu_offload.argtypes = [] +llama_supports_gpu_offload.restype = ctypes.c_bool # LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); def llama_mmap_supported() -> bool: - return _lib.llama_mmap_supported() + ... -_lib.llama_mmap_supported.argtypes = [] -_lib.llama_mmap_supported.restype = c_bool +llama_mmap_supported = _lib.llama_mmap_supported +llama_mmap_supported.argtypes = [] +llama_mmap_supported.restype = ctypes.c_bool # LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); def llama_mlock_supported() -> bool: - return _lib.llama_mlock_supported() + ... -_lib.llama_mlock_supported.argtypes = [] -_lib.llama_mlock_supported.restype = c_bool +llama_mlock_supported = _lib.llama_mlock_supported +llama_mlock_supported.argtypes = [] +llama_mlock_supported.restype = ctypes.c_bool # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); -def llama_get_model(ctx: llama_context_p) -> llama_model_p: - return _lib.llama_get_model(ctx) +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: + ... -_lib.llama_get_model.argtypes = [llama_context_p] -_lib.llama_get_model.restype = llama_model_p +llama_get_model = _lib.llama_get_model +llama_get_model.argtypes = [llama_context_p_ctypes] +llama_get_model.restype = llama_model_p_ctypes # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); -def llama_n_ctx(ctx: llama_context_p) -> int: - return _lib.llama_n_ctx(ctx) +def llama_n_ctx(ctx: llama_context_p, /) -> int: + ... -_lib.llama_n_ctx.argtypes = [llama_context_p] -_lib.llama_n_ctx.restype = c_uint32 +llama_n_ctx = _lib.llama_n_ctx +llama_n_ctx.argtypes = [llama_context_p_ctypes] +llama_n_ctx.restype = ctypes.c_uint32 # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); -def llama_n_batch(ctx: llama_context_p) -> int: - return _lib.llama_n_batch(ctx) +def llama_n_batch(ctx: llama_context_p, /) -> int: + ... -_lib.llama_n_batch.argtypes = [llama_context_p] -_lib.llama_n_batch.restype = c_uint32 +llama_n_batch = _lib.llama_n_batch +llama_n_batch.argtypes = [llama_context_p_ctypes] +llama_n_batch.restype = ctypes.c_uint32 # LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); -def llama_vocab_type(model: llama_model_p) -> int: - return _lib.llama_vocab_type(model) +def llama_vocab_type(model: llama_model_p, /) -> int: + ... -_lib.llama_vocab_type.argtypes = [llama_model_p] -_lib.llama_vocab_type.restype = c_int +llama_vocab_type = _lib.llama_vocab_type +llama_vocab_type.argtypes = [llama_model_p_ctypes] +llama_vocab_type.restype = ctypes.c_int # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); -def llama_n_vocab(model: llama_model_p) -> int: - return _lib.llama_n_vocab(model) +def llama_n_vocab(model: llama_model_p, /) -> int: + ... -_lib.llama_n_vocab.argtypes = [llama_model_p] -_lib.llama_n_vocab.restype = c_int32 +llama_n_vocab = _lib.llama_n_vocab +llama_n_vocab.argtypes = [llama_model_p_ctypes] +llama_n_vocab.restype = ctypes.c_int32 # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); -def llama_n_ctx_train(model: llama_model_p) -> int: - return _lib.llama_n_ctx_train(model) +def llama_n_ctx_train(model: llama_model_p, /) -> int: + ... -_lib.llama_n_ctx_train.argtypes = [llama_model_p] -_lib.llama_n_ctx_train.restype = c_int32 +llama_n_ctx_train = _lib.llama_n_ctx_train +llama_n_ctx_train.argtypes = [llama_model_p_ctypes] +llama_n_ctx_train.restype = ctypes.c_int32 # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); -def llama_n_embd(model: llama_model_p) -> int: - return _lib.llama_n_embd(model) +def llama_n_embd(model: llama_model_p, /) -> int: + ... -_lib.llama_n_embd.argtypes = [llama_model_p] -_lib.llama_n_embd.restype = c_int32 +llama_n_embd = _lib.llama_n_embd +llama_n_embd.argtypes = [llama_model_p_ctypes] +llama_n_embd.restype = ctypes.c_int32 # // Get the model's RoPE frequency scaling factor # LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); -def llama_rope_freq_scale_train(model: llama_model_p) -> float: +def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: """Get the model's RoPE frequency scaling factor""" - return _lib.llama_rope_freq_scale_train(model) + ... -_lib.llama_rope_freq_scale_train.argtypes = [llama_model_p] -_lib.llama_rope_freq_scale_train.restype = c_float +llama_rope_freq_scale_train = _lib.llama_rope_freq_scale_train +llama_rope_freq_scale_train.argtypes = [llama_model_p_ctypes] +llama_rope_freq_scale_train.restype = ctypes.c_float # // Functions to access the model's GGUF metadata scalar values # // - The functions return the length of the string on success, or -1 on failure @@ -954,109 +966,117 @@ def llama_rope_freq_scale_train(model: llama_model_p) -> float: # // Get metadata value as a string by key name # LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); def llama_model_meta_val_str( - model: llama_model_p, key: Union[c_char_p, bytes], buf: bytes, buf_size: int + model: llama_model_p, key: Union[ctypes.c_char_p, bytes], buf: bytes, buf_size: int, / ) -> int: """Get metadata value as a string by key name""" - return _lib.llama_model_meta_val_str(model, key, buf, buf_size) + ... -_lib.llama_model_meta_val_str.argtypes = [llama_model_p, c_char_p, c_char_p, c_size_t] -_lib.llama_model_meta_val_str.restype = c_int32 +llama_model_meta_val_str = _lib.llama_model_meta_val_str +llama_model_meta_val_str.argtypes = [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_size_t] +llama_model_meta_val_str.restype = ctypes.c_int32 # // Get the number of metadata key/value pairs # LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model); -def llama_model_meta_count(model: llama_model_p) -> int: +def llama_model_meta_count(model: llama_model_p, /) -> int: """Get the number of metadata key/value pairs""" - return _lib.llama_model_meta_count(model) + ... -_lib.llama_model_meta_count.argtypes = [llama_model_p] -_lib.llama_model_meta_count.restype = c_int32 +llama_model_meta_count = _lib.llama_model_meta_count +llama_model_meta_count.argtypes = [llama_model_p_ctypes] +llama_model_meta_count.restype = ctypes.c_int32 # // Get metadata key name by index # LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); def llama_model_meta_key_by_index( - model: llama_model_p, i: Union[c_int, int], buf: bytes, buf_size: int + model: llama_model_p, i: Union[ctypes.c_int, int], buf: bytes, buf_size: int, / ) -> int: """Get metadata key name by index""" - return _lib.llama_model_meta_key_by_index(model, i, buf, buf_size) + ... -_lib.llama_model_meta_key_by_index.argtypes = [ - llama_model_p, - c_int32, - c_char_p, - c_size_t, +llama_model_meta_key_by_index = _lib.llama_model_meta_key_by_index +llama_model_meta_key_by_index.argtypes = [ + llama_model_p_ctypes, + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_size_t, ] -_lib.llama_model_meta_key_by_index.restype = c_int32 +llama_model_meta_key_by_index.restype = ctypes.c_int32 # // Get metadata value as a string by index # LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); def llama_model_meta_val_str_by_index( - model: llama_model_p, i: Union[c_int, int], buf: bytes, buf_size: int + model: llama_model_p, i: Union[ctypes.c_int, int], buf: bytes, buf_size: int, / ) -> int: """Get metadata value as a string by index""" - return _lib.llama_model_meta_val_str_by_index(model, i, buf, buf_size) + ... -_lib.llama_model_meta_val_str_by_index.argtypes = [ - llama_model_p, - c_int32, - c_char_p, - c_size_t, +llama_model_meta_val_str_by_index = _lib.llama_model_meta_val_str_by_index +llama_model_meta_val_str_by_index.argtypes = [ + llama_model_p_ctypes, + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_size_t, ] -_lib.llama_model_meta_val_str_by_index.restype = c_int32 +llama_model_meta_val_str_by_index.restype = ctypes.c_int32 # // Get a string describing the model type # LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); def llama_model_desc( - model: llama_model_p, buf: bytes, buf_size: Union[c_size_t, int] + model: llama_model_p, buf: bytes, buf_size: Union[ctypes.c_size_t, int], / ) -> int: """Get a string describing the model type""" - return _lib.llama_model_desc(model, buf, buf_size) + ... -_lib.llama_model_desc.argtypes = [llama_model_p, c_char_p, c_size_t] -_lib.llama_model_desc.restype = c_int32 +llama_model_desc = _lib.llama_model_desc +llama_model_desc.argtypes = [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_size_t] +llama_model_desc.restype = ctypes.c_int32 # // Returns the total size of all the tensors in the model in bytes # LLAMA_API uint64_t llama_model_size(const struct llama_model * model); -def llama_model_size(model: llama_model_p) -> int: +def llama_model_size(model: llama_model_p, /) -> int: """Returns the total size of all the tensors in the model in bytes""" - return _lib.llama_model_size(model) + ... -_lib.llama_model_size.argtypes = [llama_model_p] -_lib.llama_model_size.restype = ctypes.c_uint64 +llama_model_size = _lib.llama_model_size +llama_model_size.argtypes = [llama_model_p_ctypes] +llama_model_size.restype = ctypes.c_uint64 # // Returns the total number of parameters in the model # LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); -def llama_model_n_params(model: llama_model_p) -> int: +def llama_model_n_params(model: llama_model_p, /) -> int: """Returns the total number of parameters in the model""" - return _lib.llama_model_n_params(model) + ... -_lib.llama_model_n_params.argtypes = [llama_model_p] -_lib.llama_model_n_params.restype = ctypes.c_uint64 +llama_model_n_params = _lib.llama_model_n_params +llama_model_n_params.argtypes = [llama_model_p_ctypes] +llama_model_n_params.restype = ctypes.c_uint64 # // Get a llama model tensor # LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); def llama_get_model_tensor( - model: llama_model_p, name: Union[c_char_p, bytes] -) -> c_void_p: + model: llama_model_p, name: Union[ctypes.c_char_p, bytes], / +) -> ctypes.c_void_p: """Get a llama model tensor""" - return _lib.llama_get_model_tensor(model, name) + ... -_lib.llama_get_model_tensor.argtypes = [llama_model_p, c_char_p] -_lib.llama_get_model_tensor.restype = c_void_p +llama_get_model_tensor = _lib.llama_get_model_tensor +llama_get_model_tensor.argtypes = [llama_model_p_ctypes, ctypes.c_char_p] +llama_get_model_tensor.restype = ctypes.c_void_p # // Returns 0 on success @@ -1067,18 +1087,20 @@ def llama_get_model_tensor( def llama_model_quantize( fname_inp: bytes, fname_out: bytes, - params, # type: POINTER(llama_model_quantize_params) # type: ignore + params, # type: ctypes.POINTER(llama_model_quantize_params) # type: ignore + / ) -> int: """Returns 0 on success""" - return _lib.llama_model_quantize(fname_inp, fname_out, params) + ... -_lib.llama_model_quantize.argtypes = [ - c_char_p, - c_char_p, - POINTER(llama_model_quantize_params), +llama_model_quantize = _lib.llama_model_quantize +llama_model_quantize.argtypes = [ + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(llama_model_quantize_params), ] -_lib.llama_model_quantize.restype = c_uint32 +llama_model_quantize.restype = ctypes.c_uint32 # // Apply a LoRA adapter to a loaded model @@ -1096,10 +1118,11 @@ def llama_model_quantize( # "use llama_model_apply_lora_from_file instead"); def llama_apply_lora_from_file( ctx: llama_context_p, - path_lora: Union[c_char_p, bytes], - scale: Union[c_float, float], - path_base_model: Union[c_char_p, bytes], - n_threads: Union[c_int, int], + path_lora: Union[ctypes.c_char_p, bytes], + scale: Union[ctypes.c_float, float], + path_base_model: Union[ctypes.c_char_p, bytes], + n_threads: Union[ctypes.c_int, int], + / ) -> int: """Apply a LoRA adapter to a loaded model path_base_model is the path to a higher quality model to use as a base for @@ -1107,19 +1130,18 @@ def llama_apply_lora_from_file( The model needs to be reloaded before applying a new adapter, otherwise the adapter will be applied on top of the previous one Returns 0 on success""" - return _lib.llama_apply_lora_from_file( - ctx, path_lora, scale, path_base_model, n_threads - ) + ... -_lib.llama_apply_lora_from_file.argtypes = [ - llama_context_p, - c_char_p, - c_float, - c_char_p, - c_int32, +llama_apply_lora_from_file = _lib.llama_apply_lora_from_file +llama_apply_lora_from_file.argtypes = [ + llama_context_p_ctypes, + ctypes.c_char_p, + ctypes.c_float, + ctypes.c_char_p, + ctypes.c_int32, ] -_lib.llama_apply_lora_from_file.restype = c_int32 +llama_apply_lora_from_file.restype = ctypes.c_int32 # LLAMA_API int32_t llama_model_apply_lora_from_file( @@ -1130,24 +1152,24 @@ def llama_apply_lora_from_file( # int32_t n_threads); def llama_model_apply_lora_from_file( model: llama_model_p, - path_lora: Union[c_char_p, bytes], - scale: Union[c_float, float], - path_base_model: Union[c_char_p, bytes], - n_threads: Union[c_int, int], + path_lora: Union[ctypes.c_char_p, bytes], + scale: Union[ctypes.c_float, float], + path_base_model: Union[ctypes.c_char_p, bytes], + n_threads: Union[ctypes.c_int, int], + / ) -> int: - return _lib.llama_model_apply_lora_from_file( - model, path_lora, scale, path_base_model, n_threads - ) + ... -_lib.llama_model_apply_lora_from_file.argtypes = [ - llama_model_p, - c_char_p, - c_float, - c_char_p, - c_int32, +llama_model_apply_lora_from_file = _lib.llama_model_apply_lora_from_file +llama_model_apply_lora_from_file.argtypes = [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_float, + ctypes.c_char_p, + ctypes.c_int32, ] -_lib.llama_model_apply_lora_from_file.restype = c_int32 +llama_model_apply_lora_from_file.restype = ctypes.c_int32 # // # // KV cache @@ -1160,7 +1182,7 @@ def llama_model_apply_lora_from_file( # // May be negative if the cell is not populated. # llama_pos pos; # }; -class llama_kv_cache_view_cell(Structure): +class llama_kv_cache_view_cell(ctypes.Structure): _fields_ = [("pos", llama_pos)] @@ -1196,92 +1218,98 @@ class llama_kv_cache_view_cell(Structure): # // The sequences for each cell. There will be n_max_seq items per cell. # llama_seq_id * cells_sequences; # }; -class llama_kv_cache_view(Structure): +class llama_kv_cache_view(ctypes.Structure): _fields_ = [ - ("n_cells", c_int32), - ("n_max_seq", c_int32), - ("token_count", c_int32), - ("used_cells", c_int32), - ("max_contiguous", c_int32), - ("max_contiguous_idx", c_int32), - ("cells", POINTER(llama_kv_cache_view_cell)), - ("cells_sequences", POINTER(llama_seq_id)), + ("n_cells", ctypes.c_int32), + ("n_max_seq", ctypes.c_int32), + ("token_count", ctypes.c_int32), + ("used_cells", ctypes.c_int32), + ("max_contiguous", ctypes.c_int32), + ("max_contiguous_idx", ctypes.c_int32), + ("cells", ctypes.POINTER(llama_kv_cache_view_cell)), + ("cells_sequences", ctypes.POINTER(llama_seq_id)), ] -llama_kv_cache_view_p = POINTER(llama_kv_cache_view) +llama_kv_cache_view_p = ctypes.POINTER(llama_kv_cache_view) # // Create an empty KV cache view. (use only for debugging purposes) # LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq); def llama_kv_cache_view_init( - ctx: llama_context_p, n_max_seq: Union[c_int32, int] + ctx: llama_context_p, n_max_seq: Union[ctypes.c_int32, int], / ) -> llama_kv_cache_view: """Create an empty KV cache view. (use only for debugging purposes)""" - return _lib.llama_kv_cache_view_init(ctx, n_max_seq) + ... -_lib.llama_kv_cache_view_init.argtypes = [llama_context_p, c_int32] -_lib.llama_kv_cache_view_init.restype = llama_kv_cache_view +llama_kv_cache_view_init = _lib.llama_kv_cache_view_init +llama_kv_cache_view_init.argtypes = [llama_context_p_ctypes, ctypes.c_int32] +llama_kv_cache_view_init.restype = llama_kv_cache_view # // Free a KV cache view. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); -def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore +def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore """Free a KV cache view. (use only for debugging purposes)""" - return _lib.llama_kv_cache_view_free(view) + ... -_lib.llama_kv_cache_view_free.argtypes = [llama_kv_cache_view_p] -_lib.llama_kv_cache_view_free.restype = None +llama_kv_cache_view_free = _lib.llama_kv_cache_view_free +llama_kv_cache_view_free.argtypes = [llama_kv_cache_view_p] +llama_kv_cache_view_free.restype = None # // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); -def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore +def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore """Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)""" - return _lib.llama_kv_cache_view_update(ctx, view) + ... -_lib.llama_kv_cache_view_update.argtypes = [llama_context_p, llama_kv_cache_view_p] -_lib.llama_kv_cache_view_update.restype = None +llama_kv_cache_view_update = _lib.llama_kv_cache_view_update +llama_kv_cache_view_update.argtypes = [llama_context_p_ctypes, llama_kv_cache_view_p] +llama_kv_cache_view_update.restype = None # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times # LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); -def llama_get_kv_cache_token_count(ctx: llama_context_p) -> int: +def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: """Returns the number of tokens in the KV cache (slow, use only for debug) If a KV cell has multiple sequences assigned to it, it will be counted multiple times """ - return _lib.llama_get_kv_cache_token_count(ctx) + ... -_lib.llama_get_kv_cache_token_count.argtypes = [llama_context_p] -_lib.llama_get_kv_cache_token_count.restype = c_int32 +llama_get_kv_cache_token_count = _lib.llama_get_kv_cache_token_count +llama_get_kv_cache_token_count.argtypes = [llama_context_p_ctypes] +llama_get_kv_cache_token_count.restype = ctypes.c_int32 # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) # LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); -def llama_get_kv_cache_used_cells(ctx: llama_context_p) -> int: +def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" - return _lib.llama_get_kv_cache_used_cells(ctx) + ... -_lib.llama_get_kv_cache_used_cells.argtypes = [llama_context_p] -_lib.llama_get_kv_cache_used_cells.restype = c_int32 +llama_get_kv_cache_used_cells = _lib.llama_get_kv_cache_used_cells +llama_get_kv_cache_used_cells.argtypes = [llama_context_p_ctypes] +llama_get_kv_cache_used_cells.restype = ctypes.c_int32 # // Clear the KV cache # LLAMA_API void llama_kv_cache_clear( # struct llama_context * ctx); -def llama_kv_cache_clear(ctx: llama_context_p): +def llama_kv_cache_clear(ctx: llama_context_p, /): """Clear the KV cache""" - return _lib.llama_kv_cache_clear(ctx) + ... -_lib.llama_kv_cache_clear.argtypes = [llama_context_p] -_lib.llama_kv_cache_clear.restype = None +llama_kv_cache_clear = _lib.llama_kv_cache_clear +llama_kv_cache_clear.argtypes = [llama_context_p_ctypes] +llama_kv_cache_clear.restype = None # // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) @@ -1298,21 +1326,23 @@ def llama_kv_cache_seq_rm( seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], + / ): """Removes all tokens that belong to the specified sequence and have positions in [p0, p1) seq_id < 0 : match any sequence p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_rm(ctx, seq_id, p0, p1) + ... -_lib.llama_kv_cache_seq_rm.argtypes = [ - llama_context_p, +llama_kv_cache_seq_rm = _lib.llama_kv_cache_seq_rm +llama_kv_cache_seq_rm.argtypes = [ + llama_context_p_ctypes, llama_seq_id, llama_pos, llama_pos, ] -_lib.llama_kv_cache_seq_rm.restype = None +llama_kv_cache_seq_rm.restype = None # // Copy all tokens that belong to the specified sequence to another sequence @@ -1331,22 +1361,24 @@ def llama_kv_cache_seq_cp( seq_id_dst: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], + / ): """Copy all tokens that belong to the specified sequence to another sequence Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1) + ... -_lib.llama_kv_cache_seq_cp.argtypes = [ - llama_context_p, +llama_kv_cache_seq_cp = _lib.llama_kv_cache_seq_cp +llama_kv_cache_seq_cp.argtypes = [ + llama_context_p_ctypes, llama_seq_id, llama_seq_id, llama_pos, llama_pos, ] -_lib.llama_kv_cache_seq_cp.restype = None +llama_kv_cache_seq_cp.restype = None # // Removes all tokens that do not belong to the specified sequence @@ -1356,13 +1388,15 @@ def llama_kv_cache_seq_cp( def llama_kv_cache_seq_keep( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], + / ): """Removes all tokens that do not belong to the specified sequence""" - return _lib.llama_kv_cache_seq_keep(ctx, seq_id) + ... -_lib.llama_kv_cache_seq_keep.argtypes = [llama_context_p, llama_seq_id] -_lib.llama_kv_cache_seq_keep.restype = None +llama_kv_cache_seq_keep = _lib.llama_kv_cache_seq_keep +llama_kv_cache_seq_keep.argtypes = [llama_context_p_ctypes, llama_seq_id] +llama_kv_cache_seq_keep.restype = None # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) @@ -1381,22 +1415,24 @@ def llama_kv_cache_seq_shift( p0: Union[llama_pos, int], p1: Union[llama_pos, int], delta: Union[llama_pos, int], + / ): """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) If the KV cache is RoPEd, the KV data is updated accordingly p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_shift(ctx, seq_id, p0, p1, delta) + ... -_lib.llama_kv_cache_seq_shift.argtypes = [ - llama_context_p, +llama_kv_cache_seq_shift = _lib.llama_kv_cache_seq_shift +llama_kv_cache_seq_shift.argtypes = [ + llama_context_p_ctypes, llama_seq_id, llama_pos, llama_pos, llama_pos, ] -_lib.llama_kv_cache_seq_shift.restype = None +llama_kv_cache_seq_shift.restype = None # // Integer division of the positions by factor of `d > 1` @@ -1414,23 +1450,25 @@ def llama_kv_cache_seq_div( seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], - d: Union[c_int, int], + d: Union[ctypes.c_int, int], + / ): """Integer division of the positions by factor of `d > 1` If the KV cache is RoPEd, the KV data is updated accordingly p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_div(ctx, seq_id, p0, p1, d) + ... -_lib.llama_kv_cache_seq_div.argtypes = [ - llama_context_p, +llama_kv_cache_seq_div = _lib.llama_kv_cache_seq_div +llama_kv_cache_seq_div.argtypes = [ + llama_context_p_ctypes, llama_seq_id, llama_pos, llama_pos, - c_int, + ctypes.c_int, ] -_lib.llama_kv_cache_seq_div.restype = None +llama_kv_cache_seq_div.restype = None # // # // State / sessions @@ -1440,14 +1478,15 @@ def llama_kv_cache_seq_div( # Returns the maximum size in bytes of the state (rng, logits, embedding # and kv_cache) - will often be smaller after compacting tokens # LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); -def llama_get_state_size(ctx: llama_context_p) -> int: +def llama_get_state_size(ctx: llama_context_p, /) -> int: """Returns the maximum size in bytes of the state (rng, logits, embedding and kv_cache) - will often be smaller after compacting tokens""" - return _lib.llama_get_state_size(ctx) + ... -_lib.llama_get_state_size.argtypes = [llama_context_p] -_lib.llama_get_state_size.restype = c_size_t +llama_get_state_size = _lib.llama_get_state_size +llama_get_state_size.argtypes = [llama_context_p_ctypes] +llama_get_state_size.restype = ctypes.c_size_t # Copies the state to the specified destination address. @@ -1457,16 +1496,18 @@ def llama_get_state_size(ctx: llama_context_p) -> int: # struct llama_context * ctx, # uint8_t * dst); def llama_copy_state_data( - ctx: llama_context_p, dst # type: Array[c_uint8] + ctx: llama_context_p, dst, # type: Array[ctypes.c_uint8] + / ) -> int: """Copies the state to the specified destination address. Destination needs to have allocated enough memory. Returns the number of bytes copied""" - return _lib.llama_copy_state_data(ctx, dst) + ... -_lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p] -_lib.llama_copy_state_data.restype = c_size_t +llama_copy_state_data = _lib.llama_copy_state_data +llama_copy_state_data.argtypes = [llama_context_p_ctypes, c_uint8_p] +llama_copy_state_data.restype = ctypes.c_size_t # Set the state reading from the specified address @@ -1475,14 +1516,16 @@ def llama_copy_state_data( # struct llama_context * ctx, # uint8_t * src); def llama_set_state_data( - ctx: llama_context_p, src # type: Array[c_uint8] + ctx: llama_context_p, src, # type: Array[ctypes.c_uint8] + / ) -> int: """Set the state reading from the specified address""" - return _lib.llama_set_state_data(ctx, src) + ... -_lib.llama_set_state_data.argtypes = [llama_context_p, c_uint8_p] -_lib.llama_set_state_data.restype = c_size_t +llama_set_state_data = _lib.llama_set_state_data +llama_set_state_data.argtypes = [llama_context_p_ctypes, c_uint8_p] +llama_set_state_data.restype = ctypes.c_size_t # Save/load session file @@ -1496,22 +1539,22 @@ def llama_load_session_file( ctx: llama_context_p, path_session: bytes, tokens_out, # type: Array[llama_token] - n_token_capacity: Union[c_size_t, int], - n_token_count_out, # type: _Pointer[c_size_t] + n_token_capacity: Union[ctypes.c_size_t, int], + n_token_count_out, # type: _Pointer[ctypes.c_size_t] + / ) -> int: - return _lib.llama_load_session_file( - ctx, path_session, tokens_out, n_token_capacity, n_token_count_out - ) + ... -_lib.llama_load_session_file.argtypes = [ - llama_context_p, - c_char_p, +llama_load_session_file = _lib.llama_load_session_file +llama_load_session_file.argtypes = [ + llama_context_p_ctypes, + ctypes.c_char_p, llama_token_p, - c_size_t, + ctypes.c_size_t, c_size_t_p, ] -_lib.llama_load_session_file.restype = c_size_t +llama_load_session_file.restype = ctypes.c_size_t # LLAMA_API bool llama_save_session_file( @@ -1523,18 +1566,20 @@ def llama_save_session_file( ctx: llama_context_p, path_session: bytes, tokens, # type: Array[llama_token] - n_token_count: Union[c_size_t, int], + n_token_count: Union[ctypes.c_size_t, int], + / ) -> int: - return _lib.llama_save_session_file(ctx, path_session, tokens, n_token_count) + ... -_lib.llama_save_session_file.argtypes = [ - llama_context_p, - c_char_p, +llama_save_session_file = _lib.llama_save_session_file +llama_save_session_file.argtypes = [ + llama_context_p_ctypes, + ctypes.c_char_p, llama_token_p, - c_size_t, + ctypes.c_size_t, ] -_lib.llama_save_session_file.restype = c_size_t +llama_save_session_file.restype = ctypes.c_size_t # // # // Decoding @@ -1555,19 +1600,21 @@ def llama_save_session_file( def llama_eval( ctx: llama_context_p, tokens, # type: Array[llama_token] - n_tokens: Union[c_int, int], - n_past: Union[c_int, int], + n_tokens: Union[ctypes.c_int, int], + n_past: Union[ctypes.c_int, int], + / ) -> int: """Run the llama inference to obtain the logits and probabilities for the next token(s). tokens + n_tokens is the provided batch of new tokens to process n_past is the number of tokens to use from previous eval calls Returns 0 on success DEPRECATED: use llama_decode() instead""" - return _lib.llama_eval(ctx, tokens, n_tokens, n_past) + ... -_lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int32, c_int32] -_lib.llama_eval.restype = c_int +llama_eval = _lib.llama_eval +llama_eval.argtypes = [llama_context_p_ctypes, llama_token_p, ctypes.c_int32, ctypes.c_int32] +llama_eval.restype = ctypes.c_int # // Same as llama_eval, but use float matrix input directly. @@ -1580,17 +1627,19 @@ def llama_eval( # "use llama_decode() instead"); def llama_eval_embd( ctx: llama_context_p, - embd, # type: Array[c_float] - n_tokens: Union[c_int, int], - n_past: Union[c_int, int], + embd, # type: Array[ctypes.c_float] + n_tokens: Union[ctypes.c_int, int], + n_past: Union[ctypes.c_int, int], + / ) -> int: """Same as llama_eval, but use float matrix input directly. DEPRECATED: use llama_decode() instead""" - return _lib.llama_eval_embd(ctx, embd, n_tokens, n_past) + ... -_lib.llama_eval_embd.argtypes = [llama_context_p, c_float_p, c_int32, c_int32] -_lib.llama_eval_embd.restype = c_int +llama_eval_embd = _lib.llama_eval_embd +llama_eval_embd.argtypes = [llama_context_p_ctypes, c_float_p, ctypes.c_int32, ctypes.c_int32] +llama_eval_embd.restype = ctypes.c_int # // Return batch for single sequence of tokens starting at pos_0 @@ -1604,24 +1653,26 @@ def llama_eval_embd( # llama_seq_id seq_id); def llama_batch_get_one( tokens, # type: Array[llama_token] - n_tokens: Union[c_int, int], + n_tokens: Union[ctypes.c_int, int], pos_0: Union[llama_pos, int], seq_id: llama_seq_id, + / ) -> llama_batch: """Return batch for single sequence of tokens starting at pos_0 NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it """ - return _lib.llama_batch_get_one(tokens, n_tokens, pos_0, seq_id) + ... -_lib.llama_batch_get_one.argtypes = [ +llama_batch_get_one = _lib.llama_batch_get_one +llama_batch_get_one.argtypes = [ llama_token_p, - c_int, + ctypes.c_int, llama_pos, llama_seq_id, ] -_lib.llama_batch_get_one.restype = llama_batch +llama_batch_get_one.restype = llama_batch # // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens @@ -1636,9 +1687,10 @@ def llama_batch_get_one( # int32_t embd, # int32_t n_seq_max); def llama_batch_init( - n_tokens: Union[c_int32, int], - embd: Union[c_int32, int], - n_seq_max: Union[c_int32, int], + n_tokens: Union[ctypes.c_int32, int], + embd: Union[ctypes.c_int32, int], + n_seq_max: Union[ctypes.c_int32, int], + / ) -> llama_batch: """Allocates a batch of tokens on the heap that can hold a maximum of n_tokens Each token can be assigned up to n_seq_max sequence ids @@ -1647,22 +1699,24 @@ def llama_batch_init( Otherwise, llama_batch.token will be allocated to store n_tokens llama_token The rest of the llama_batch members are allocated with size n_tokens All members are left uninitialized""" - return _lib.llama_batch_init(n_tokens, embd, n_seq_max) + ... -_lib.llama_batch_init.argtypes = [c_int32, c_int32, c_int32] -_lib.llama_batch_init.restype = llama_batch +llama_batch_init = _lib.llama_batch_init +llama_batch_init.argtypes = [ctypes.c_int32, ctypes.c_int32, ctypes.c_int32] +llama_batch_init.restype = llama_batch # // Frees a batch of tokens allocated with llama_batch_init() # LLAMA_API void llama_batch_free(struct llama_batch batch); -def llama_batch_free(batch: llama_batch): +def llama_batch_free(batch: llama_batch, /): """Frees a batch of tokens allocated with llama_batch_init()""" - return _lib.llama_batch_free(batch) + ... -_lib.llama_batch_free.argtypes = [llama_batch] -_lib.llama_batch_free.restype = None +llama_batch_free = _lib.llama_batch_free +llama_batch_free.argtypes = [llama_batch] +llama_batch_free.restype = None # // Positive return values does not mean a fatal error, but rather a warning. @@ -1672,16 +1726,17 @@ def llama_batch_free(batch: llama_batch): # LLAMA_API int32_t llama_decode( # struct llama_context * ctx, # struct llama_batch batch); -def llama_decode(ctx: llama_context_p, batch: llama_batch) -> int: +def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: """Positive return values does not mean a fatal error, but rather a warning. 0 - success 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) < 0 - error""" - return _lib.llama_decode(ctx, batch) + ... -_lib.llama_decode.argtypes = [llama_context_p, llama_batch] -_lib.llama_decode.restype = c_int32 +llama_decode = _lib.llama_decode +llama_decode.argtypes = [llama_context_p_ctypes, llama_batch] +llama_decode.restype = ctypes.c_int32 # // Set the number of threads used for decoding @@ -1690,18 +1745,20 @@ def llama_decode(ctx: llama_context_p, batch: llama_batch) -> int: # LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); def llama_set_n_threads( ctx: llama_context_p, - n_threads: Union[c_uint32, int], - n_threads_batch: Union[c_uint32, int], + n_threads: Union[ctypes.c_uint32, int], + n_threads_batch: Union[ctypes.c_uint32, int], + / ): """Set the number of threads used for decoding n_threads is the number of threads used for generation (single token) n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) """ - return _lib.llama_set_n_threads(ctx, n_threads, n_threads_batch) + ... -_lib.llama_set_n_threads.argtypes = [llama_context_p, c_uint32, c_uint32] -_lib.llama_set_n_threads.restype = None +llama_set_n_threads = _lib.llama_set_n_threads +llama_set_n_threads.argtypes = [llama_context_p_ctypes, ctypes.c_uint32, ctypes.c_uint32] +llama_set_n_threads.restype = None # // Token logits obtained from the last call to llama_eval() @@ -1712,62 +1769,68 @@ def llama_set_n_threads( # LLAMA_API float * llama_get_logits(struct llama_context * ctx); def llama_get_logits( ctx: llama_context_p, + / ): # type: (...) -> Array[float] # type: ignore """Token logits obtained from the last call to llama_eval() The logits for the last token are stored in the last row Logits for which llama_batch.logits[i] == 0 are undefined Rows: n_tokens provided with llama_batch Cols: n_vocab""" - return _lib.llama_get_logits(ctx) + ... -_lib.llama_get_logits.argtypes = [llama_context_p] -_lib.llama_get_logits.restype = c_float_p +llama_get_logits = _lib.llama_get_logits +llama_get_logits.argtypes = [llama_context_p_ctypes] +llama_get_logits.restype = c_float_p # // Logits for the ith token. Equivalent to: # // llama_get_logits(ctx) + i*n_vocab # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); def llama_get_logits_ith( - ctx: llama_context_p, i: Union[c_int32, int] + ctx: llama_context_p, i: Union[ctypes.c_int32, int] + , / ): # type: (...) -> Array[float] # type: ignore """Logits for the ith token. Equivalent to: llama_get_logits(ctx) + i*n_vocab""" - return _lib.llama_get_logits_ith(ctx, i) + ... -_lib.llama_get_logits_ith.argtypes = [llama_context_p, c_int32] -_lib.llama_get_logits_ith.restype = c_float_p +llama_get_logits_ith = _lib.llama_get_logits_ith +llama_get_logits_ith.argtypes = [llama_context_p_ctypes, ctypes.c_int32] +llama_get_logits_ith.restype = c_float_p # Get the embeddings for the input # shape: [n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); def llama_get_embeddings( - ctx: llama_context_p, + ctx: llama_context_p, / ): # type: (...) -> Array[float] # type: ignore """Get the embeddings for the input shape: [n_embd] (1-dimensional)""" - return _lib.llama_get_embeddings(ctx) + ... -_lib.llama_get_embeddings.argtypes = [llama_context_p] -_lib.llama_get_embeddings.restype = c_float_p +llama_get_embeddings = _lib.llama_get_embeddings +llama_get_embeddings.argtypes = [llama_context_p_ctypes] +llama_get_embeddings.restype = c_float_p # // Get the embeddings for the ith sequence # // llama_get_embeddings(ctx) + i*n_embd # LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); def llama_get_embeddings_ith( - ctx: llama_context_p, i: Union[c_int32, int] + ctx: llama_context_p, i: Union[ctypes.c_int32, int], / ): # type: (...) -> Array[float] # type: ignore """Get the embeddings for the ith sequence llama_get_embeddings(ctx) + i*n_embd""" - return _lib.llama_get_embeddings_ith(ctx, i) + ... -_lib.llama_get_embeddings_ith.argtypes = [llama_context_p, c_int32] -_lib.llama_get_embeddings_ith.restype = c_float_p +llama_get_embeddings_ith = _lib.llama_get_embeddings_ith +llama_get_embeddings_ith.argtypes = [llama_context_p_ctypes, ctypes.c_int32] +llama_get_embeddings_ith.restype = c_float_p # // @@ -1776,125 +1839,137 @@ def llama_get_embeddings_ith( # LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); -def llama_token_get_text(model: llama_model_p, token: Union[llama_token, int]) -> bytes: - return _lib.llama_token_get_text(model, token) +def llama_token_get_text(model: llama_model_p, token: Union[llama_token, int], /) -> bytes: + ... -_lib.llama_token_get_text.argtypes = [llama_model_p, llama_token] -_lib.llama_token_get_text.restype = c_char_p +llama_token_get_text = _lib.llama_token_get_text +llama_token_get_text.argtypes = [llama_model_p_ctypes, llama_token] +llama_token_get_text.restype = ctypes.c_char_p # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); def llama_token_get_score( - model: llama_model_p, token: Union[llama_token, int] + model: llama_model_p, token: Union[llama_token, int], / ) -> float: - return _lib.llama_token_get_score(model, token) + ... -_lib.llama_token_get_score.argtypes = [llama_model_p, llama_token] -_lib.llama_token_get_score.restype = c_float +llama_token_get_score = _lib.llama_token_get_score +llama_token_get_score.argtypes = [llama_model_p_ctypes, llama_token] +llama_token_get_score.restype = ctypes.c_float # LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); -def llama_token_get_type(model: llama_model_p, token: Union[llama_token, int]) -> int: - return _lib.llama_token_get_type(model, token) +def llama_token_get_type(model: llama_model_p, token: Union[llama_token, int], /) -> int: + ... -_lib.llama_token_get_type.argtypes = [llama_model_p, llama_token] -_lib.llama_token_get_type.restype = ctypes.c_int +llama_token_get_type = _lib.llama_token_get_type +llama_token_get_type.argtypes = [llama_model_p_ctypes, llama_token] +llama_token_get_type.restype = ctypes.c_int # // Special tokens # LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence -def llama_token_bos(model: llama_model_p) -> int: +def llama_token_bos(model: llama_model_p, /) -> int: """beginning-of-sentence""" - return _lib.llama_token_bos(model) + ... -_lib.llama_token_bos.argtypes = [llama_model_p] -_lib.llama_token_bos.restype = llama_token +llama_token_bos = _lib.llama_token_bos +llama_token_bos.argtypes = [llama_model_p_ctypes] +llama_token_bos.restype = llama_token # LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence -def llama_token_eos(model: llama_model_p) -> int: +def llama_token_eos(model: llama_model_p, /) -> int: """end-of-sentence""" - return _lib.llama_token_eos(model) + ... -_lib.llama_token_eos.argtypes = [llama_model_p] -_lib.llama_token_eos.restype = llama_token +llama_token_eos = _lib.llama_token_eos +llama_token_eos.argtypes = [llama_model_p_ctypes] +llama_token_eos.restype = llama_token # LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line -def llama_token_nl(model: llama_model_p) -> int: +def llama_token_nl(model: llama_model_p, /) -> int: """next-line""" - return _lib.llama_token_nl(model) + ... -_lib.llama_token_nl.argtypes = [llama_model_p] -_lib.llama_token_nl.restype = llama_token +llama_token_nl = _lib.llama_token_nl +llama_token_nl.argtypes = [llama_model_p_ctypes] +llama_token_nl.restype = llama_token # // Returns -1 if unknown, 1 for true or 0 for false. # LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); -def llama_add_bos_token(model: llama_model_p) -> int: +def llama_add_bos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" - return _lib.llama_add_bos_token(model) + ... -_lib.llama_add_bos_token.argtypes = [llama_model_p] -_lib.llama_add_bos_token.restype = c_int32 +llama_add_bos_token = _lib.llama_add_bos_token +llama_add_bos_token.argtypes = [llama_model_p_ctypes] +llama_add_bos_token.restype = ctypes.c_int32 # // Returns -1 if unknown, 1 for true or 0 for false. # LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); -def llama_add_eos_token(model: llama_model_p) -> int: +def llama_add_eos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" - return _lib.llama_add_eos_token(model) + ... -_lib.llama_add_eos_token.argtypes = [llama_model_p] -_lib.llama_add_eos_token.restype = c_int32 +llama_add_eos_token = _lib.llama_add_eos_token +llama_add_eos_token.argtypes = [llama_model_p_ctypes] +llama_add_eos_token.restype = ctypes.c_int32 # // codellama infill tokens # LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix def llama_token_prefix(model: llama_model_p) -> int: """codellama infill tokens""" - return _lib.llama_token_prefix(model) + ... -_lib.llama_token_prefix.argtypes = [llama_model_p] -_lib.llama_token_prefix.restype = llama_token +llama_token_prefix = _lib.llama_token_prefix +llama_token_prefix.argtypes = [llama_model_p_ctypes] +llama_token_prefix.restype = llama_token # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle -def llama_token_middle(model: llama_model_p) -> int: - return _lib.llama_token_middle(model) +def llama_token_middle(model: llama_model_p, /) -> int: + ... -_lib.llama_token_middle.argtypes = [llama_model_p] -_lib.llama_token_middle.restype = llama_token +llama_token_middle = _lib.llama_token_middle +llama_token_middle.argtypes = [llama_model_p_ctypes] +llama_token_middle.restype = llama_token # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix -def llama_token_suffix(model: llama_model_p) -> int: - return _lib.llama_token_suffix(model) +def llama_token_suffix(model: llama_model_p, /) -> int: + ... -_lib.llama_token_suffix.argtypes = [llama_model_p] -_lib.llama_token_suffix.restype = llama_token +llama_token_suffix = _lib.llama_token_suffix +llama_token_suffix.argtypes = [llama_model_p_ctypes] +llama_token_suffix.restype = llama_token # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle -def llama_token_eot(model: llama_model_p) -> int: - return _lib.llama_token_eot(model) +def llama_token_eot(model: llama_model_p, /) -> int: + ... -_lib.llama_token_eot.argtypes = [llama_model_p] -_lib.llama_token_eot.restype = llama_token +llama_token_eot = _lib.llama_token_eot +llama_token_eot.argtypes = [llama_model_p_ctypes] +llama_token_eot.restype = llama_token # // @@ -1919,28 +1994,28 @@ def llama_token_eot(model: llama_model_p) -> int: def llama_tokenize( model: llama_model_p, text: bytes, - text_len: Union[c_int, int], + text_len: Union[ctypes.c_int, int], tokens, # type: Array[llama_token] - n_max_tokens: Union[c_int, int], - add_bos: Union[c_bool, bool], - special: Union[c_bool, bool], + n_max_tokens: Union[ctypes.c_int, int], + add_bos: Union[ctypes.c_bool, bool], + special: Union[ctypes.c_bool, bool], + / ) -> int: """Convert the provided text into tokens.""" - return _lib.llama_tokenize( - model, text, text_len, tokens, n_max_tokens, add_bos, special - ) + ... -_lib.llama_tokenize.argtypes = [ - llama_model_p, - c_char_p, - c_int32, +llama_tokenize = _lib.llama_tokenize +llama_tokenize.argtypes = [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_int32, llama_token_p, - c_int32, - c_bool, - c_bool, + ctypes.c_int32, + ctypes.c_bool, + ctypes.c_bool, ] -_lib.llama_tokenize.restype = c_int32 +llama_tokenize.restype = ctypes.c_int32 # // Token Id -> Piece. @@ -1955,19 +2030,21 @@ def llama_tokenize( def llama_token_to_piece( model: llama_model_p, token: Union[llama_token, int], - buf: Union[c_char_p, bytes], - length: Union[c_int, int], + buf: Union[ctypes.c_char_p, bytes], + length: Union[ctypes.c_int, int], + / ) -> int: """Token Id -> Piece. Uses the vocabulary in the provided context. Does not write null terminator to the buffer. User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. """ - return _lib.llama_token_to_piece(model, token, buf, length) + ... -_lib.llama_token_to_piece.argtypes = [llama_model_p, llama_token, c_char_p, c_int32] -_lib.llama_token_to_piece.restype = c_int32 +llama_token_to_piece = _lib.llama_token_to_piece +llama_token_to_piece.argtypes = [llama_model_p_ctypes, llama_token, ctypes.c_char_p, ctypes.c_int32] +llama_token_to_piece.restype = ctypes.c_int32 # /// Apply chat template. Inspired by hf apply_chat_template() on python. @@ -1991,23 +2068,20 @@ def llama_token_to_piece( def llama_chat_apply_template( model: llama_model_p, tmpl: bytes, - chat: "ctypes._Pointer[llama_chat_message]", + chat: "ctypes._Pointer[llama_chat_message]", # type: ignore n_msg: int, + / ) -> int: - return _lib.llama_chat_apply_template( - model, - tmpl, - chat, - n_msg - ) + ... -_lib.llama_chat_apply_template.argtypes = [ +llama_chat_apply_template = _lib.llama_chat_apply_template +llama_chat_apply_template.argtypes = [ ctypes.c_void_p, ctypes.c_char_p, ctypes.POINTER(llama_chat_message), ctypes.c_size_t ] -_lib.llama_chat_apply_template.restype = ctypes.c_int32 +llama_chat_apply_template.restype = ctypes.c_int32 @@ -2022,39 +2096,43 @@ def llama_chat_apply_template( # size_t start_rule_index); def llama_grammar_init( rules, # type: Array[llama_grammar_element_p] # type: ignore - n_rules: Union[c_size_t, int], - start_rule_index: Union[c_size_t, int], + n_rules: Union[ctypes.c_size_t, int], + start_rule_index: Union[ctypes.c_size_t, int], + / ) -> llama_grammar_p: """Initialize a grammar from a set of rules.""" - return _lib.llama_grammar_init(rules, n_rules, start_rule_index) + ... -_lib.llama_grammar_init.argtypes = [ - POINTER(llama_grammar_element_p), - c_size_t, - c_size_t, +llama_grammar_init = _lib.llama_grammar_init +llama_grammar_init.argtypes = [ + ctypes.POINTER(llama_grammar_element_p), + ctypes.c_size_t, + ctypes.c_size_t, ] -_lib.llama_grammar_init.restype = llama_grammar_p +llama_grammar_init.restype = llama_grammar_p # LLAMA_API void llama_grammar_free(struct llama_grammar * grammar); -def llama_grammar_free(grammar: llama_grammar_p): +def llama_grammar_free(grammar: llama_grammar_p, /): """Free a grammar.""" - return _lib.llama_grammar_free(grammar) + ... -_lib.llama_grammar_free.argtypes = [llama_grammar_p] -_lib.llama_grammar_free.restype = None +llama_grammar_free = _lib.llama_grammar_free +llama_grammar_free.argtypes = [llama_grammar_p] +llama_grammar_free.restype = None # LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar); -def llama_grammar_copy(grammar: llama_grammar_p) -> llama_grammar_p: +def llama_grammar_copy(grammar: llama_grammar_p, /) -> llama_grammar_p: """Copy a grammar.""" - return _lib.llama_grammar_copy(grammar) + ... -_lib.llama_grammar_copy.argtypes = [llama_grammar_p] -_lib.llama_grammar_copy.restype = llama_grammar_p +llama_grammar_copy = _lib.llama_grammar_copy +llama_grammar_copy.argtypes = [llama_grammar_p] +llama_grammar_copy.restype = llama_grammar_p # // # // Sampling functions @@ -2063,13 +2141,14 @@ def llama_grammar_copy(grammar: llama_grammar_p) -> llama_grammar_p: # // Sets the current rng seed. # LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); -def llama_set_rng_seed(ctx: llama_context_p, seed: Union[c_uint32, int]): +def llama_set_rng_seed(ctx: llama_context_p, seed: Union[ctypes.c_uint32, int], /): """Sets the current rng seed.""" - return _lib.llama_set_rng_seed(ctx, seed) + ... -_lib.llama_set_rng_seed.argtypes = [llama_context_p, c_uint32] -_lib.llama_set_rng_seed.restype = None +llama_set_rng_seed = _lib.llama_set_rng_seed +llama_set_rng_seed.argtypes = [llama_context_p_ctypes, ctypes.c_uint32] +llama_set_rng_seed.restype = None # /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. @@ -2086,35 +2165,29 @@ def llama_sample_repetition_penalties( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] last_tokens_data, # type: Array[llama_token] - penalty_last_n: Union[c_size_t, int], - penalty_repeat: Union[c_float, float], - penalty_freq: Union[c_float, float], - penalty_present: Union[c_float, float], + penalty_last_n: Union[ctypes.c_size_t, int], + penalty_repeat: Union[ctypes.c_float, float], + penalty_freq: Union[ctypes.c_float, float], + penalty_present: Union[ctypes.c_float, float], + / ): """Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. """ - return _lib.llama_sample_repetition_penalties( - ctx, - candidates, - last_tokens_data, - penalty_last_n, - penalty_repeat, - penalty_freq, - penalty_present, - ) + ... -_lib.llama_sample_repetition_penalties.argtypes = [ - llama_context_p, +llama_sample_repetition_penalties = _lib.llama_sample_repetition_penalties +llama_sample_repetition_penalties.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, llama_token_p, - c_size_t, - c_float, - c_float, - c_float, + ctypes.c_size_t, + ctypes.c_float, + ctypes.c_float, + ctypes.c_float, ] -_lib.llama_sample_repetition_penalties.restype = None +llama_sample_repetition_penalties.restype = None # /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 @@ -2128,21 +2201,23 @@ def llama_sample_repetition_penalties( # float scale); def llama_sample_apply_guidance( ctx: llama_context_p, - logits, # type: _Pointer[c_float] - logits_guidance, # type: _Pointer[c_float] - scale: Union[c_float, float], + logits, # type: _Pointer[ctypes.c_float] + logits_guidance, # type: _Pointer[ctypes.c_float] + scale: Union[ctypes.c_float, float], + / ): """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" - return _lib.llama_sample_apply_guidance(ctx, logits, logits_guidance, scale) + ... -_lib.llama_sample_apply_guidance.argtypes = [ - llama_context_p, +llama_sample_apply_guidance = _lib.llama_sample_apply_guidance +llama_sample_apply_guidance.argtypes = [ + llama_context_p_ctypes, c_float_p, c_float_p, - c_float, + ctypes.c_float, ] -_lib.llama_sample_apply_guidance.restype = None +llama_sample_apply_guidance.restype = None # LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance( @@ -2155,21 +2230,21 @@ def llama_sample_classifier_free_guidance( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] guidance_ctx: llama_context_p, - scale: Union[c_float, float], + scale: Union[ctypes.c_float, float], + / ): """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" - return _lib.llama_sample_classifier_free_guidance( - ctx, candidates, guidance_ctx, scale - ) + ... -_lib.llama_sample_classifier_free_guidance.argtypes = [ - llama_context_p, +llama_sample_classifier_free_guidance = _lib.llama_sample_classifier_free_guidance +llama_sample_classifier_free_guidance.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - llama_context_p, - c_float, + llama_context_p_ctypes, + ctypes.c_float, ] -_lib.llama_sample_classifier_free_guidance.restype = None +llama_sample_classifier_free_guidance.restype = None # /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. @@ -2177,17 +2252,19 @@ def llama_sample_classifier_free_guidance( # struct llama_context * ctx, # llama_token_data_array * candidates); def llama_sample_softmax( - ctx: llama_context_p, candidates # type: _Pointer[llama_token_data] + ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data] + / ): """Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.""" - return _lib.llama_sample_softmax(ctx, candidates) + ... -_lib.llama_sample_softmax.argtypes = [ - llama_context_p, +llama_sample_softmax = _lib.llama_sample_softmax +llama_sample_softmax.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, ] -_lib.llama_sample_softmax.restype = None +llama_sample_softmax.restype = None # /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 @@ -2199,20 +2276,22 @@ def llama_sample_softmax( def llama_sample_top_k( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - k: Union[c_int, int], - min_keep: Union[c_size_t, int], + k: Union[ctypes.c_int, int], + min_keep: Union[ctypes.c_size_t, int], + / ): """Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" - return _lib.llama_sample_top_k(ctx, candidates, k, min_keep) + ... -_lib.llama_sample_top_k.argtypes = [ - llama_context_p, +llama_sample_top_k = _lib.llama_sample_top_k +llama_sample_top_k.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_int32, - c_size_t, + ctypes.c_int32, + ctypes.c_size_t, ] -_lib.llama_sample_top_k.restype = None +llama_sample_top_k.restype = None # /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 @@ -2224,20 +2303,22 @@ def llama_sample_top_k( def llama_sample_top_p( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - p: Union[c_float, float], - min_keep: Union[c_size_t, int], + p: Union[ctypes.c_float, float], + min_keep: Union[ctypes.c_size_t, int], + / ): """Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" - return _lib.llama_sample_top_p(ctx, candidates, p, min_keep) + ... -_lib.llama_sample_top_p.argtypes = [ - llama_context_p, +llama_sample_top_p = _lib.llama_sample_top_p +llama_sample_top_p.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_size_t, + ctypes.c_float, + ctypes.c_size_t, ] -_lib.llama_sample_top_p.restype = None +llama_sample_top_p.restype = None # /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 @@ -2249,20 +2330,22 @@ def llama_sample_top_p( def llama_sample_min_p( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - p: Union[c_float, float], - min_keep: Union[c_size_t, int], + p: Union[ctypes.c_float, float], + min_keep: Union[ctypes.c_size_t, int], + / ): """Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841""" - return _lib.llama_sample_min_p(ctx, candidates, p, min_keep) + ... -_lib.llama_sample_min_p.argtypes = [ - llama_context_p, +llama_sample_min_p = _lib.llama_sample_min_p +llama_sample_min_p.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_size_t, + ctypes.c_float, + ctypes.c_size_t, ] -_lib.llama_sample_min_p.restype = None +llama_sample_min_p.restype = None # /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. @@ -2274,20 +2357,22 @@ def llama_sample_min_p( def llama_sample_tail_free( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - z: Union[c_float, float], - min_keep: Union[c_size_t, int], + z: Union[ctypes.c_float, float], + min_keep: Union[ctypes.c_size_t, int], + / ): """Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.""" - return _lib.llama_sample_tail_free(ctx, candidates, z, min_keep) + ... -_lib.llama_sample_tail_free.argtypes = [ - llama_context_p, +llama_sample_tail_free = _lib.llama_sample_tail_free +llama_sample_tail_free.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_size_t, + ctypes.c_float, + ctypes.c_size_t, ] -_lib.llama_sample_tail_free.restype = None +llama_sample_tail_free.restype = None # /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. @@ -2299,20 +2384,22 @@ def llama_sample_tail_free( def llama_sample_typical( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - p: Union[c_float, float], - min_keep: Union[c_size_t, int], + p: Union[ctypes.c_float, float], + min_keep: Union[ctypes.c_size_t, int], + / ): """Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.""" - return _lib.llama_sample_typical(ctx, candidates, p, min_keep) + ... -_lib.llama_sample_typical.argtypes = [ - llama_context_p, +llama_sample_typical = _lib.llama_sample_typical +llama_sample_typical.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_size_t, + ctypes.c_float, + ctypes.c_size_t, ] -_lib.llama_sample_typical.restype = None +llama_sample_typical.restype = None # /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772. @@ -2325,22 +2412,24 @@ def llama_sample_typical( def llama_sample_entropy( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - min_temp: Union[c_float, float], - max_temp: Union[c_float, float], - exponent_val: Union[c_float, float], + min_temp: Union[ctypes.c_float, float], + max_temp: Union[ctypes.c_float, float], + exponent_val: Union[ctypes.c_float, float], + / ): """Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.""" - return _lib.llama_sample_entropy(ctx, candidates, min_temp, max_temp, exponent_val) + ... -_lib.llama_sample_entropy.argtypes = [ - llama_context_p, +llama_sample_entropy = _lib.llama_sample_entropy +llama_sample_entropy.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_float, - c_float, + ctypes.c_float, + ctypes.c_float, + ctypes.c_float, ] -_lib.llama_sample_entropy.restype = None +llama_sample_entropy.restype = None # LLAMA_API void llama_sample_temp( @@ -2350,7 +2439,8 @@ def llama_sample_entropy( def llama_sample_temp( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - temp: Union[c_float, float], + temp: Union[ctypes.c_float, float], + / ): """Temperature sampling described in academic paper "Generating Long Sequences with Sparse Transformers" https://arxiv.org/abs/1904.10509 @@ -2358,15 +2448,16 @@ def llama_sample_temp( candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. temp: The temperature value to use for the sampling. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. """ - return _lib.llama_sample_temp(ctx, candidates, temp) + ... -_lib.llama_sample_temp.argtypes = [ - llama_context_p, +llama_sample_temp = _lib.llama_sample_temp +llama_sample_temp.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, + ctypes.c_float, ] -_lib.llama_sample_temp.restype = None +llama_sample_temp.restype = None # LLAMA_API DEPRECATED(void llama_sample_temperature( @@ -2377,18 +2468,20 @@ def llama_sample_temp( def llama_sample_temperature( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - temp: Union[c_float, float], + temp: Union[ctypes.c_float, float], + / ): """use llama_sample_temp instead""" - return _lib.llama_sample_temperature(ctx, candidates, temp) + ... -_lib.llama_sample_temperature.argtypes = [ - llama_context_p, +llama_sample_temperature = _lib.llama_sample_temperature +llama_sample_temperature.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, + ctypes.c_float, ] -_lib.llama_sample_temperature.restype = None +llama_sample_temperature.restype = None # /// @details Apply constraints from grammar @@ -2400,6 +2493,7 @@ def llama_sample_grammar( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] grammar, # type: llama_grammar_p + / ): """Apply constraints from grammar @@ -2407,15 +2501,16 @@ def llama_sample_grammar( candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. grammar: A grammar object containing the rules and constraints to apply to the generated text. """ - return _lib.llama_sample_grammar(ctx, candidates, grammar) + ... -_lib.llama_sample_grammar.argtypes = [ - llama_context_p, +llama_sample_grammar = _lib.llama_sample_grammar +llama_sample_grammar.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, llama_grammar_p, ] -_lib.llama_sample_grammar.restype = None +llama_sample_grammar.restype = None # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2434,10 +2529,11 @@ def llama_sample_grammar( def llama_sample_token_mirostat( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - tau: Union[c_float, float], - eta: Union[c_float, float], - m: Union[c_int, int], - mu, # type: _Pointer[c_float] + tau: Union[ctypes.c_float, float], + eta: Union[ctypes.c_float, float], + m: Union[ctypes.c_int, int], + mu, # type: _Pointer[ctypes.c_float] + / ) -> int: """Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2448,18 +2544,19 @@ def llama_sample_token_mirostat( m: The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. """ - return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu) + ... -_lib.llama_sample_token_mirostat.argtypes = [ - llama_context_p, +llama_sample_token_mirostat = _lib.llama_sample_token_mirostat +llama_sample_token_mirostat.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_float, - c_int32, + ctypes.c_float, + ctypes.c_float, + ctypes.c_int32, c_float_p, ] -_lib.llama_sample_token_mirostat.restype = llama_token +llama_sample_token_mirostat.restype = llama_token # /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2476,9 +2573,10 @@ def llama_sample_token_mirostat( def llama_sample_token_mirostat_v2( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] - tau: Union[c_float, float], - eta: Union[c_float, float], - mu, # type: _Pointer[c_float] + tau: Union[ctypes.c_float, float], + eta: Union[ctypes.c_float, float], + mu, # type: _Pointer[ctypes.c_float] + / ) -> int: """Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2488,17 +2586,18 @@ def llama_sample_token_mirostat_v2( eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. """ - return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu) + ... -_lib.llama_sample_token_mirostat_v2.argtypes = [ - llama_context_p, +llama_sample_token_mirostat_v2 = _lib.llama_sample_token_mirostat_v2 +llama_sample_token_mirostat_v2.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, - c_float, - c_float, + ctypes.c_float, + ctypes.c_float, c_float_p, ] -_lib.llama_sample_token_mirostat_v2.restype = llama_token +llama_sample_token_mirostat_v2.restype = llama_token # /// @details Selects the token with the highest probability. @@ -2509,16 +2608,18 @@ def llama_sample_token_mirostat_v2( def llama_sample_token_greedy( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] + / ) -> int: """Selects the token with the highest probability.""" - return _lib.llama_sample_token_greedy(ctx, candidates) + ... -_lib.llama_sample_token_greedy.argtypes = [ - llama_context_p, +llama_sample_token_greedy = _lib.llama_sample_token_greedy +llama_sample_token_greedy.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, ] -_lib.llama_sample_token_greedy.restype = llama_token +llama_sample_token_greedy.restype = llama_token # /// @details Randomly selects a token from the candidates based on their probabilities. @@ -2528,16 +2629,18 @@ def llama_sample_token_greedy( def llama_sample_token( ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data_array] + / ) -> int: """Randomly selects a token from the candidates based on their probabilities.""" - return _lib.llama_sample_token(ctx, candidates) + ... -_lib.llama_sample_token.argtypes = [ - llama_context_p, +llama_sample_token = _lib.llama_sample_token +llama_sample_token.argtypes = [ + llama_context_p_ctypes, llama_token_data_array_p, ] -_lib.llama_sample_token.restype = llama_token +llama_sample_token.restype = llama_token # /// @details Accepts the sampled token into the grammar @@ -2549,17 +2652,19 @@ def llama_grammar_accept_token( ctx: llama_context_p, grammar: llama_grammar_p, token: Union[llama_token, int], + / ) -> None: """Accepts the sampled token into the grammar""" - _lib.llama_grammar_accept_token(ctx, grammar, token) + ... -_lib.llama_grammar_accept_token.argtypes = [ - llama_context_p, +llama_grammar_accept_token = _lib.llama_grammar_accept_token +llama_grammar_accept_token.argtypes = [ + llama_context_p_ctypes, llama_grammar_p, llama_token, ] -_lib.llama_grammar_accept_token.restype = None +llama_grammar_accept_token.restype = None # // @@ -2577,9 +2682,9 @@ def llama_grammar_accept_token( class llama_beam_view(ctypes.Structure): _fields_ = [ ("tokens", llama_token_p), - ("n_tokens", c_size_t), - ("p", c_float), - ("eob", c_bool), + ("n_tokens", ctypes.c_size_t), + ("p", ctypes.c_float), + ("eob", ctypes.c_bool), ] @@ -2595,10 +2700,10 @@ class llama_beam_view(ctypes.Structure): # }; class llama_beams_state(ctypes.Structure): _fields_ = [ - ("beam_views", POINTER(llama_beam_view)), - ("n_beams", c_size_t), - ("common_prefix_length", c_size_t), - ("last_call", c_bool), + ("beam_views", ctypes.POINTER(llama_beam_view)), + ("n_beams", ctypes.c_size_t), + ("common_prefix_length", ctypes.c_size_t), + ("last_call", ctypes.c_bool), ] @@ -2606,7 +2711,7 @@ class llama_beams_state(ctypes.Structure): # // void* callback_data is any custom data passed to llama_beam_search, that is subsequently # // passed back to beam_search_callback. This avoids having to use global variables in the callback. # typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); -llama_beam_search_callback_fn_t = ctypes.CFUNCTYPE(None, c_void_p, llama_beams_state) +llama_beam_search_callback_fn_t = ctypes.CFUNCTYPE(None, ctypes.c_void_p, llama_beams_state) # /// @details Deterministically returns entire sentence constructed by a beam search. @@ -2626,70 +2731,74 @@ class llama_beams_state(ctypes.Structure): # int32_t n_predict); def llama_beam_search( ctx: llama_context_p, - callback: "ctypes._CFuncPtr[None, c_void_p, llama_beams_state]", # type: ignore - callback_data: c_void_p, - n_beams: Union[c_size_t, int], - n_past: Union[c_int, int], - n_predict: Union[c_int, int], + callback: "ctypes._CFuncPtr[None, ctypes.c_void_p, llama_beams_state]", # type: ignore + callback_data: ctypes.c_void_p, + n_beams: Union[ctypes.c_size_t, int], + n_past: Union[ctypes.c_int, int], + n_predict: Union[ctypes.c_int, int], + / ): - return _lib.llama_beam_search( - ctx, callback, callback_data, n_beams, n_past, n_predict - ) + ... -_lib.llama_beam_search.argtypes = [ - llama_context_p, +llama_beam_search = _lib.llama_beam_search +llama_beam_search.argtypes = [ + llama_context_p_ctypes, llama_beam_search_callback_fn_t, - c_void_p, - c_size_t, - c_int32, - c_int32, + ctypes.c_void_p, + ctypes.c_size_t, + ctypes.c_int32, + ctypes.c_int32, ] -_lib.llama_beam_search.restype = None +llama_beam_search.restype = None # Performance information # LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); -def llama_get_timings(ctx: llama_context_p) -> llama_timings: +def llama_get_timings(ctx: llama_context_p, /) -> llama_timings: """Get performance information""" - return _lib.llama_get_timings(ctx) + ... -_lib.llama_get_timings.argtypes = [llama_context_p] -_lib.llama_get_timings.restype = llama_timings +llama_get_timings = _lib.llama_get_timings +llama_get_timings.argtypes = [llama_context_p_ctypes] +llama_get_timings.restype = llama_timings # LLAMA_API void llama_print_timings(struct llama_context * ctx); -def llama_print_timings(ctx: llama_context_p): +def llama_print_timings(ctx: llama_context_p, /): """Print performance information""" - _lib.llama_print_timings(ctx) + ... -_lib.llama_print_timings.argtypes = [llama_context_p] -_lib.llama_print_timings.restype = None +llama_print_timings = _lib.llama_print_timings +llama_print_timings.argtypes = [llama_context_p_ctypes] +llama_print_timings.restype = None # LLAMA_API void llama_reset_timings(struct llama_context * ctx); -def llama_reset_timings(ctx: llama_context_p): +def llama_reset_timings(ctx: llama_context_p, /): """Reset performance information""" - _lib.llama_reset_timings(ctx) + ... -_lib.llama_reset_timings.argtypes = [llama_context_p] -_lib.llama_reset_timings.restype = None +llama_reset_timings = _lib.llama_reset_timings +llama_reset_timings.argtypes = [llama_context_p_ctypes] +llama_reset_timings.restype = None # Print system information # LLAMA_API const char * llama_print_system_info(void); def llama_print_system_info() -> bytes: """Print system information""" - return _lib.llama_print_system_info() + ... -_lib.llama_print_system_info.argtypes = [] -_lib.llama_print_system_info.restype = c_char_p +llama_print_system_info = _lib.llama_print_system_info +llama_print_system_info.argtypes = [] +llama_print_system_info.restype = ctypes.c_char_p # NOTE: THIS IS CURRENTLY BROKEN AS ggml_log_callback IS NOT EXPOSED IN LLAMA.H @@ -2697,22 +2806,25 @@ def llama_print_system_info() -> bytes: # // If this is not called, or NULL is supplied, everything is output on stderr. # LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); def llama_log_set( - log_callback: Union["ctypes._FuncPointer", c_void_p], user_data: c_void_p # type: ignore + log_callback: Union["ctypes._FuncPointer", ctypes.c_void_p], user_data: ctypes.c_void_p, # type: ignore + / ): """Set callback for all future logging events. If this is not called, or NULL is supplied, everything is output on stderr.""" - return _lib.llama_log_set(log_callback, user_data) + ... -_lib.llama_log_set.argtypes = [ctypes.c_void_p, c_void_p] -_lib.llama_log_set.restype = None +llama_log_set = _lib.llama_log_set +llama_log_set.argtypes = [ctypes.c_void_p, ctypes.c_void_p] +llama_log_set.restype = None # LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p): - return _lib.llama_dump_timing_info_yaml(stream, ctx) +def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): + ... -_lib.llama_dump_timing_info_yaml.argtypes = [ctypes.c_void_p, llama_context_p] -_lib.llama_dump_timing_info_yaml.restype = None +llama_dump_timing_info_yaml = _lib.llama_dump_timing_info_yaml +llama_dump_timing_info_yaml.argtypes = [ctypes.c_void_p, llama_context_p_ctypes] +llama_dump_timing_info_yaml.restype = None diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index 8195bd40d..4eaa9e553 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -5,21 +5,15 @@ c_bool, c_char_p, c_int, - c_int8, - c_int32, c_uint8, - c_uint32, - c_size_t, c_float, - c_double, c_void_p, POINTER, _Pointer, # type: ignore Structure, - Array, ) import pathlib -from typing import List, Union +from typing import List, Union, NewType, Optional import llama_cpp.llama_cpp as llama_cpp @@ -67,7 +61,7 @@ def _load_shared_library(lib_base_name: str): for _lib_path in _lib_paths: if _lib_path.exists(): try: - return ctypes.CDLL(str(_lib_path), **cdll_args) + return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore except Exception as e: raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") @@ -88,7 +82,8 @@ def _load_shared_library(lib_base_name: str): ################################################ # struct clip_ctx; -clip_ctx_p = c_void_p +clip_ctx_p = NewType("clip_ctx_p", int) +clip_ctx_p_ctypes = c_void_p # struct llava_image_embed { # float * embed; @@ -102,43 +97,48 @@ class llava_image_embed(Structure): # /** sanity check for clip <-> llava embed size match */ # LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip); -def llava_validate_embed_size(ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p) -> bool: - return _libllava.llava_validate_embed_size(ctx_llama, ctx_clip) +def llava_validate_embed_size(ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p, /) -> bool: + ... -_libllava.llava_validate_embed_size.argtypes = [llama_cpp.llama_context_p, clip_ctx_p] -_libllava.llava_validate_embed_size.restype = c_bool +llava_validate_embed_size = _libllava.llava_validate_embed_size +llava_validate_embed_size.argtypes = [llama_cpp.llama_context_p_ctypes, clip_ctx_p_ctypes] +llava_validate_embed_size.restype = c_bool # /** build an image embed from image file bytes */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); -def llava_image_embed_make_with_bytes(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_bytes: bytes, image_bytes_length: Union[c_int, int]) -> "_Pointer[llava_image_embed]": - return _libllava.llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length) +def llava_image_embed_make_with_bytes(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_bytes: bytes, image_bytes_length: Union[c_int, int], /) -> "_Pointer[llava_image_embed]": + ... -_libllava.llava_image_embed_make_with_bytes.argtypes = [clip_ctx_p, c_int, POINTER(c_uint8), c_int] -_libllava.llava_image_embed_make_with_bytes.restype = POINTER(llava_image_embed) +llava_image_embed_make_with_bytes = _libllava.llava_image_embed_make_with_bytes +llava_image_embed_make_with_bytes.argtypes = [clip_ctx_p_ctypes, c_int, POINTER(c_uint8), c_int] +llava_image_embed_make_with_bytes.restype = POINTER(llava_image_embed) # /** build an image embed from a path to an image filename */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); -def llava_image_embed_make_with_filename(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes) -> "_Pointer[llava_image_embed]": - return _libllava.llava_image_embed_make_with_filename(ctx_clip, n_threads, image_path) +def llava_image_embed_make_with_filename(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes, /) -> "_Pointer[llava_image_embed]": + ... -_libllava.llava_image_embed_make_with_filename.argtypes = [clip_ctx_p, c_int, c_char_p] -_libllava.llava_image_embed_make_with_filename.restype = POINTER(llava_image_embed) +llava_image_embed_make_with_filename = _libllava.llava_image_embed_make_with_filename +llava_image_embed_make_with_filename.argtypes = [clip_ctx_p_ctypes, c_int, c_char_p] +llava_image_embed_make_with_filename.restype = POINTER(llava_image_embed) # LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); # /** free an embedding made with llava_image_embed_make_* */ -def llava_image_embed_free(embed: "_Pointer[llava_image_embed]"): - return _libllava.llava_image_embed_free(embed) +def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): + ... -_libllava.llava_image_embed_free.argtypes = [POINTER(llava_image_embed)] -_libllava.llava_image_embed_free.restype = None +llava_image_embed_free = _libllava.llava_image_embed_free +llava_image_embed_free.argtypes = [POINTER(llava_image_embed)] +llava_image_embed_free.restype = None # /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ # LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); -def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointer[llava_image_embed]", n_batch: Union[c_int, int], n_past: "_Pointer[c_int]") -> bool: - return _libllava.llava_eval_image_embed(ctx_llama, embed, n_batch, n_past) +def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointer[llava_image_embed]", n_batch: Union[c_int, int], n_past: "_Pointer[c_int]", /) -> bool: + ... -_libllava.llava_eval_image_embed.argtypes = [llama_cpp.llama_context_p, POINTER(llava_image_embed), c_int, POINTER(c_int)] -_libllava.llava_eval_image_embed.restype = c_bool +llava_eval_image_embed = _libllava.llava_eval_image_embed +llava_eval_image_embed.argtypes = [llama_cpp.llama_context_p_ctypes, POINTER(llava_image_embed), c_int, POINTER(c_int)] +llava_eval_image_embed.restype = c_bool ################################################ @@ -148,16 +148,18 @@ def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointe # /** load mmproj model */ # CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); -def clip_model_load(fname: bytes, verbosity: Union[c_int, int]) -> clip_ctx_p: - return _libllava.clip_model_load(fname, verbosity) +def clip_model_load(fname: bytes, verbosity: Union[c_int, int], /) -> Optional[clip_ctx_p]: + ... -_libllava.clip_model_load.argtypes = [c_char_p, c_int] -_libllava.clip_model_load.restype = clip_ctx_p +clip_model_load = _libllava.clip_model_load +clip_model_load.argtypes = [c_char_p, c_int] +clip_model_load.restype = clip_ctx_p_ctypes # /** free mmproj model */ # CLIP_API void clip_free(struct clip_ctx * ctx); -def clip_free(ctx: clip_ctx_p): - return _libllava.clip_free(ctx) +def clip_free(ctx: clip_ctx_p, /): + ... -_libllava.clip_free.argtypes = [clip_ctx_p] -_libllava.clip_free.restype = None +clip_free = _libllava.clip_free +clip_free.argtypes = [clip_ctx_p_ctypes] +clip_free.restype = None diff --git a/tests/test_llama.py b/tests/test_llama.py index dac33b74d..5cf421b56 100644 --- a/tests/test_llama.py +++ b/tests/test_llama.py @@ -54,7 +54,7 @@ def setup_mock(llama: llama_cpp.Llama, output_text: str): output_tokens = llama.tokenize( output_text.encode("utf-8"), add_bos=True, special=True ) - logits = (llama_cpp.c_float * (n_vocab * n_ctx))(-100.0) + logits = (ctypes.c_float * (n_vocab * n_ctx))(-100.0) for i in range(n_ctx): output_idx = i + 1 # logits for first tokens predict second token if output_idx < len(output_tokens): @@ -90,9 +90,9 @@ def mock_get_logits(ctx: llama_cpp.llama_context_p): assert n > 0, "mock_llama_decode not called" assert last_n_tokens > 0, "mock_llama_decode not called" # Return view of logits for last_n_tokens - return (llama_cpp.c_float * (last_n_tokens * n_vocab)).from_address( + return (ctypes.c_float * (last_n_tokens * n_vocab)).from_address( ctypes.addressof(logits) - + (n - last_n_tokens) * n_vocab * ctypes.sizeof(llama_cpp.c_float) + + (n - last_n_tokens) * n_vocab * ctypes.sizeof(ctypes.c_float) ) monkeypatch.setattr("llama_cpp.llama_cpp.llama_decode", mock_decode) From fe5626cd4077037f9f0f1ecfc598f57b6a52afa8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 16:26:30 -0500 Subject: [PATCH 158/624] misc: add .local pattern to gitignore --- .gitignore | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.gitignore b/.gitignore index 51f357200..9d68dbcd9 100644 --- a/.gitignore +++ b/.gitignore @@ -1,3 +1,5 @@ +*.local + .python-version .vscode/ From 14191e9036c5d22dcc2499a44abd146709b701a2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 16:26:49 -0500 Subject: [PATCH 159/624] docs: Add create_chat_completion_openai_v1 to api reference --- docs/api-reference.md | 1 + 1 file changed, 1 insertion(+) diff --git a/docs/api-reference.md b/docs/api-reference.md index c635e1178..ab51ef754 100644 --- a/docs/api-reference.md +++ b/docs/api-reference.md @@ -21,6 +21,7 @@ High-level Python bindings for llama.cpp. - create_completion - __call__ - create_chat_completion + - create_chat_completion_openai_v1 - set_cache - save_state - load_state From 7f3962e11c739879cf3755acf9d97d19128c3c9e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 16:27:56 -0500 Subject: [PATCH 160/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 89febfed9..ba2135cca 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 89febfed9322c8849520dc63c93ee4f5fd72556e +Subproject commit ba2135ccae7462470b3865c6e41d2e1d734eac05 From 7981e9ce1efe97efb44a21e4be7b8d7bbac7660a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 16:30:59 -0500 Subject: [PATCH 161/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 43b62c89b..bf54a6405 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.46] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba2135ccae7462470b3865c6e41d2e1d734eac05 +- feat: Pull models directly from huggingface by @abetlen in #1206 +- feat(low-level-api): Improve API static type-safety and performance. Low level api functions are positional args only now. by @abetlen in #1205 + ## [0.2.45] - feat: Update llama.cpp to ggerganov/llama.cpp@89febfed9322c8849520dc63c93ee4f5fd72556e diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 197fc68b7..e44adc796 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.45" \ No newline at end of file +__version__ = "0.2.46" \ No newline at end of file From 0653e15c206b5ef1f8cb064d2cd18beb662e6cd1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 23:04:52 -0500 Subject: [PATCH 162/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 +- vendor/llama.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 69dbe097f..2ab326daf 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2049,7 +2049,7 @@ def llama_token_to_piece( # /// Apply chat template. Inspired by hf apply_chat_template() on python. # /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" -# /// NOTE: This function only support some known jinja templates. It is not a jinja parser. +# /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template # /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. # /// @param chat Pointer to a list of multiple llama_chat_message # /// @param n_msg Number of llama_chat_message in this chat diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ba2135cca..973053d8b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ba2135ccae7462470b3865c6e41d2e1d734eac05 +Subproject commit 973053d8b0d04809836b3339a50f68d9c842de90 From 3632241e986fe466bd67d66e10f60a5433c85c69 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Feb 2024 23:09:13 -0500 Subject: [PATCH 163/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index bf54a6405..aa6b4b4ea 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.47] + +- feat: Update llama.cpp to ggerganov/llama.cpp@973053d8b0d04809836b3339a50f68d9c842de90 + ## [0.2.46] - feat: Update llama.cpp to ggerganov/llama.cpp@ba2135ccae7462470b3865c6e41d2e1d734eac05 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index e44adc796..bd701e383 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.46" \ No newline at end of file +__version__ = "0.2.47" \ No newline at end of file From dd22010e85265ae840c76ec835d67a29ed852722 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 00:09:45 -0500 Subject: [PATCH 164/624] fix: Raise exceptions when llama model or context fails to load --- llama_cpp/_internals.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 10c1a6fbb..8690843cd 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -51,6 +51,9 @@ def __init__( self.path_model.encode("utf-8"), self.params ) + if self.model is None: + raise ValueError(f"Failed to load model from file: {path_model}") + def __del__(self): if self.model is not None and self._llama_free_model is not None: self._llama_free_model(self.model) @@ -258,6 +261,9 @@ def __init__( self.model.model, self.params ) + if self.ctx is None: + raise ValueError("Failed to create llama_context") + def __del__(self): if self.ctx is not None and self._llama_free is not None: self._llama_free(self.ctx) From e6d6260a91b7831733f7d1f73c7af46a3e8185ed Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 00:10:23 -0500 Subject: [PATCH 165/624] fix: Update from_pretrained defaults to match hf_hub_download --- llama_cpp/llama.py | 22 +++++++++++++++++----- 1 file changed, 17 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 9fc4ec2d7..122654514 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1885,8 +1885,9 @@ def from_pretrained( cls, repo_id: str, filename: Optional[str], - local_dir: Optional[Union[str, os.PathLike[str]]] = ".", + local_dir: Optional[Union[str, os.PathLike[str]]] = None, local_dir_use_symlinks: Union[bool, Literal["auto"]] = "auto", + cache_dir: Optional[Union[str, os.PathLike[str]]] = None, **kwargs: Any, ) -> "Llama": """Create a Llama model from a pretrained model name or path. @@ -1945,18 +1946,29 @@ def from_pretrained( subfolder = str(Path(matching_file).parent) filename = Path(matching_file).name - local_dir = "." - # download the file hf_hub_download( repo_id=repo_id, - local_dir=local_dir, filename=filename, subfolder=subfolder, + local_dir=local_dir, local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cache_dir, ) - model_path = os.path.join(local_dir, filename) + if local_dir is None: + model_path = hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cache_dir, + local_files_only=True, + + ) + else: + model_path = os.path.join(local_dir, filename) return cls( model_path=model_path, From 3921e10770996d95a9eb22c8248bacef39f69365 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 00:17:06 -0500 Subject: [PATCH 166/624] feat: support minItems/maxItems in JSON grammar converter (by @nopperl) --- llama_cpp/llama_grammar.py | 18 +++++++++++++++--- 1 file changed, 15 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 3eb3b965b..6a37857b9 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1498,9 +1498,21 @@ def visit(self, schema: Dict[str, Any], name: str) -> str: item_rule_name = self.visit( schema["items"], f'{name}{"-" if name else ""}item' ) - rule = ( - f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space' - ) + list_item_operator = f'("," space {item_rule_name})' + successive_items = "" + min_items = schema.get("minItems", 0) + if min_items > 0: + first_item = f"({item_rule_name})" + successive_items = list_item_operator * (min_items - 1) + min_items -= 1 + else: + first_item = f"({item_rule_name})?" + max_items = schema.get("maxItems") + if max_items is not None and max_items > min_items: + successive_items += (list_item_operator + "?") * (max_items - min_items - 1) + else: + successive_items += list_item_operator + "*" + rule = f'"[" space {first_item} {successive_items} "]" space' return self._add_rule(rule_name, rule) else: From aefcb8f71abe3ff611bcd4c877ccb8d69edeec56 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 02:00:09 -0500 Subject: [PATCH 167/624] misc: additional type annotations for low level api --- llama_cpp/llama_cpp.py | 352 ++++++++++++++++++++++++++--------------- 1 file changed, 227 insertions(+), 125 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 2ab326daf..193919024 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import sys import os import ctypes @@ -6,7 +8,16 @@ Array, ) import pathlib -from typing import List, Union, NewType, Optional +from typing import ( + List, + Union, + NewType, + Optional, + TYPE_CHECKING, + TypeVar, + TypeAlias, + Generic, +) # Load the library @@ -56,7 +67,7 @@ def _load_shared_library(lib_base_name: str): for _lib_path in _lib_paths: if _lib_path.exists(): try: - return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore + return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore except Exception as e: raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") @@ -71,14 +82,39 @@ def _load_shared_library(lib_base_name: str): # Load the library _lib = _load_shared_library(_lib_base_name) -# Misc -c_float_p = ctypes.POINTER(ctypes.c_float) -c_uint8_p = ctypes.POINTER(ctypes.c_uint8) -c_size_t_p = ctypes.POINTER(ctypes.c_size_t) + +# ctypes sane type hint helpers +# +# - Generic Pointer and Array types +# - PointerOrRef type with a type hinted byref function +# +# NOTE: Only use these for static type checking not for runtime checks +# no good will come of that + +if TYPE_CHECKING: + CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore + + CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore + + CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore + + CtypesVoidPointer: TypeAlias = ctypes.c_void_p + + class CtypesRef(Generic[CtypesCData]): + pass + + CtypesPointerOrRef: TypeAlias = Union[ + CtypesPointer[CtypesCData], CtypesRef[CtypesCData] + ] + + CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore + # from ggml-backend.h # typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); -ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p, ctypes.c_bool, ctypes.c_void_p) +ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE( + ctypes.c_bool, ctypes.c_void_p, ctypes.c_bool, ctypes.c_void_p +) # llama.h bindings @@ -286,7 +322,9 @@ class llama_token_data_array(ctypes.Structure): llama_token_data_array_p = ctypes.POINTER(llama_token_data_array) # typedef bool (*llama_progress_callback)(float progress, void *ctx); -llama_progress_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_float, ctypes.c_void_p) +llama_progress_callback = ctypes.CFUNCTYPE( + ctypes.c_bool, ctypes.c_float, ctypes.c_void_p +) # // Input data for llama_decode @@ -336,7 +374,7 @@ class llama_batch(ctypes.Structure): _fields_ = [ ("n_tokens", ctypes.c_int32), ("token", ctypes.POINTER(llama_token)), - ("embd", c_float_p), + ("embd", ctypes.POINTER(ctypes.c_float)), ("pos", ctypes.POINTER(llama_pos)), ("n_seq_id", ctypes.POINTER(ctypes.c_int32)), ("seq_id", ctypes.POINTER(ctypes.POINTER(llama_seq_id))), @@ -431,7 +469,7 @@ class llama_model_params(ctypes.Structure): ("n_gpu_layers", ctypes.c_int32), ("split_mode", ctypes.c_int), ("main_gpu", ctypes.c_int32), - ("tensor_split", c_float_p), + ("tensor_split", ctypes.POINTER(ctypes.c_float)), ("progress_callback", llama_progress_callback), ("progress_callback_user_data", ctypes.c_void_p), ("kv_overrides", ctypes.POINTER(llama_model_kv_override)), @@ -532,7 +570,9 @@ class llama_context_params(ctypes.Structure): # // if it exists. # // It might not exist for progress report where '.' is output repeatedly. # typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data); -llama_log_callback = ctypes.CFUNCTYPE(None, ctypes.c_int, ctypes.c_char_p, ctypes.c_void_p) +llama_log_callback = ctypes.CFUNCTYPE( + None, ctypes.c_int, ctypes.c_char_p, ctypes.c_void_p +) """Signature for logging events Note that text includes the new line character at the end for most events. If your logging mechanism cannot handle that, check if the last character is '\n' and strip it @@ -966,14 +1006,23 @@ def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: # // Get metadata value as a string by key name # LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); def llama_model_meta_val_str( - model: llama_model_p, key: Union[ctypes.c_char_p, bytes], buf: bytes, buf_size: int, / + model: llama_model_p, + key: Union[ctypes.c_char_p, bytes], + buf: bytes, + buf_size: int, + /, ) -> int: """Get metadata value as a string by key name""" ... llama_model_meta_val_str = _lib.llama_model_meta_val_str -llama_model_meta_val_str.argtypes = [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_size_t] +llama_model_meta_val_str.argtypes = [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.c_size_t, +] llama_model_meta_val_str.restype = ctypes.c_int32 @@ -1087,8 +1136,8 @@ def llama_get_model_tensor( def llama_model_quantize( fname_inp: bytes, fname_out: bytes, - params, # type: ctypes.POINTER(llama_model_quantize_params) # type: ignore - / + params: CtypesPointerOrRef[llama_model_quantize_params], + /, ) -> int: """Returns 0 on success""" ... @@ -1121,8 +1170,8 @@ def llama_apply_lora_from_file( path_lora: Union[ctypes.c_char_p, bytes], scale: Union[ctypes.c_float, float], path_base_model: Union[ctypes.c_char_p, bytes], - n_threads: Union[ctypes.c_int, int], - / + n_threads: Union[ctypes.c_int32, int], + /, ) -> int: """Apply a LoRA adapter to a loaded model path_base_model is the path to a higher quality model to use as a base for @@ -1155,8 +1204,8 @@ def llama_model_apply_lora_from_file( path_lora: Union[ctypes.c_char_p, bytes], scale: Union[ctypes.c_float, float], path_base_model: Union[ctypes.c_char_p, bytes], - n_threads: Union[ctypes.c_int, int], - / + n_threads: Union[ctypes.c_int32, int], + /, ) -> int: ... @@ -1262,7 +1311,7 @@ def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # # // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); -def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore +def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[llama_kv_cache_view], /): # type: ignore """Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)""" ... @@ -1326,7 +1375,7 @@ def llama_kv_cache_seq_rm( seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], - / + /, ): """Removes all tokens that belong to the specified sequence and have positions in [p0, p1) seq_id < 0 : match any sequence @@ -1361,7 +1410,7 @@ def llama_kv_cache_seq_cp( seq_id_dst: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], - / + /, ): """Copy all tokens that belong to the specified sequence to another sequence Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence @@ -1385,11 +1434,7 @@ def llama_kv_cache_seq_cp( # LLAMA_API void llama_kv_cache_seq_keep( # struct llama_context * ctx, # llama_seq_id seq_id); -def llama_kv_cache_seq_keep( - ctx: llama_context_p, - seq_id: Union[llama_seq_id, int], - / -): +def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): """Removes all tokens that do not belong to the specified sequence""" ... @@ -1415,7 +1460,7 @@ def llama_kv_cache_seq_shift( p0: Union[llama_pos, int], p1: Union[llama_pos, int], delta: Union[llama_pos, int], - / + /, ): """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) If the KV cache is RoPEd, the KV data is updated accordingly @@ -1451,7 +1496,7 @@ def llama_kv_cache_seq_div( p0: Union[llama_pos, int], p1: Union[llama_pos, int], d: Union[ctypes.c_int, int], - / + /, ): """Integer division of the positions by factor of `d > 1` If the KV cache is RoPEd, the KV data is updated accordingly @@ -1496,8 +1541,7 @@ def llama_get_state_size(ctx: llama_context_p, /) -> int: # struct llama_context * ctx, # uint8_t * dst); def llama_copy_state_data( - ctx: llama_context_p, dst, # type: Array[ctypes.c_uint8] - / + ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], / ) -> int: """Copies the state to the specified destination address. Destination needs to have allocated enough memory. @@ -1506,7 +1550,10 @@ def llama_copy_state_data( llama_copy_state_data = _lib.llama_copy_state_data -llama_copy_state_data.argtypes = [llama_context_p_ctypes, c_uint8_p] +llama_copy_state_data.argtypes = [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), +] llama_copy_state_data.restype = ctypes.c_size_t @@ -1516,15 +1563,14 @@ def llama_copy_state_data( # struct llama_context * ctx, # uint8_t * src); def llama_set_state_data( - ctx: llama_context_p, src, # type: Array[ctypes.c_uint8] - / + ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], / ) -> int: """Set the state reading from the specified address""" ... llama_set_state_data = _lib.llama_set_state_data -llama_set_state_data.argtypes = [llama_context_p_ctypes, c_uint8_p] +llama_set_state_data.argtypes = [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)] llama_set_state_data.restype = ctypes.c_size_t @@ -1538,10 +1584,10 @@ def llama_set_state_data( def llama_load_session_file( ctx: llama_context_p, path_session: bytes, - tokens_out, # type: Array[llama_token] + tokens_out: CtypesArray[llama_token], n_token_capacity: Union[ctypes.c_size_t, int], - n_token_count_out, # type: _Pointer[ctypes.c_size_t] - / + n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], + /, ) -> int: ... @@ -1552,7 +1598,7 @@ def llama_load_session_file( ctypes.c_char_p, llama_token_p, ctypes.c_size_t, - c_size_t_p, + ctypes.POINTER(ctypes.c_size_t), ] llama_load_session_file.restype = ctypes.c_size_t @@ -1565,9 +1611,9 @@ def llama_load_session_file( def llama_save_session_file( ctx: llama_context_p, path_session: bytes, - tokens, # type: Array[llama_token] + tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], - / + /, ) -> int: ... @@ -1599,10 +1645,10 @@ def llama_save_session_file( # "use llama_decode() instead"); def llama_eval( ctx: llama_context_p, - tokens, # type: Array[llama_token] + tokens: CtypesArray[llama_token], n_tokens: Union[ctypes.c_int, int], n_past: Union[ctypes.c_int, int], - / + /, ) -> int: """Run the llama inference to obtain the logits and probabilities for the next token(s). tokens + n_tokens is the provided batch of new tokens to process @@ -1613,7 +1659,12 @@ def llama_eval( llama_eval = _lib.llama_eval -llama_eval.argtypes = [llama_context_p_ctypes, llama_token_p, ctypes.c_int32, ctypes.c_int32] +llama_eval.argtypes = [ + llama_context_p_ctypes, + llama_token_p, + ctypes.c_int32, + ctypes.c_int32, +] llama_eval.restype = ctypes.c_int @@ -1627,10 +1678,10 @@ def llama_eval( # "use llama_decode() instead"); def llama_eval_embd( ctx: llama_context_p, - embd, # type: Array[ctypes.c_float] + embd: CtypesArray[ctypes.c_float], n_tokens: Union[ctypes.c_int, int], n_past: Union[ctypes.c_int, int], - / + /, ) -> int: """Same as llama_eval, but use float matrix input directly. DEPRECATED: use llama_decode() instead""" @@ -1638,7 +1689,12 @@ def llama_eval_embd( llama_eval_embd = _lib.llama_eval_embd -llama_eval_embd.argtypes = [llama_context_p_ctypes, c_float_p, ctypes.c_int32, ctypes.c_int32] +llama_eval_embd.argtypes = [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_float), + ctypes.c_int32, + ctypes.c_int32, +] llama_eval_embd.restype = ctypes.c_int @@ -1652,11 +1708,11 @@ def llama_eval_embd( # llama_pos pos_0, # llama_seq_id seq_id); def llama_batch_get_one( - tokens, # type: Array[llama_token] + tokens: CtypesArray[llama_token], n_tokens: Union[ctypes.c_int, int], pos_0: Union[llama_pos, int], seq_id: llama_seq_id, - / + /, ) -> llama_batch: """Return batch for single sequence of tokens starting at pos_0 @@ -1690,7 +1746,7 @@ def llama_batch_init( n_tokens: Union[ctypes.c_int32, int], embd: Union[ctypes.c_int32, int], n_seq_max: Union[ctypes.c_int32, int], - / + /, ) -> llama_batch: """Allocates a batch of tokens on the heap that can hold a maximum of n_tokens Each token can be assigned up to n_seq_max sequence ids @@ -1747,7 +1803,7 @@ def llama_set_n_threads( ctx: llama_context_p, n_threads: Union[ctypes.c_uint32, int], n_threads_batch: Union[ctypes.c_uint32, int], - / + /, ): """Set the number of threads used for decoding n_threads is the number of threads used for generation (single token) @@ -1757,7 +1813,11 @@ def llama_set_n_threads( llama_set_n_threads = _lib.llama_set_n_threads -llama_set_n_threads.argtypes = [llama_context_p_ctypes, ctypes.c_uint32, ctypes.c_uint32] +llama_set_n_threads.argtypes = [ + llama_context_p_ctypes, + ctypes.c_uint32, + ctypes.c_uint32, +] llama_set_n_threads.restype = None @@ -1768,8 +1828,7 @@ def llama_set_n_threads( # // Cols: n_vocab # LLAMA_API float * llama_get_logits(struct llama_context * ctx); def llama_get_logits( - ctx: llama_context_p, - / + ctx: llama_context_p, / ): # type: (...) -> Array[float] # type: ignore """Token logits obtained from the last call to llama_eval() The logits for the last token are stored in the last row @@ -1781,15 +1840,14 @@ def llama_get_logits( llama_get_logits = _lib.llama_get_logits llama_get_logits.argtypes = [llama_context_p_ctypes] -llama_get_logits.restype = c_float_p +llama_get_logits.restype = ctypes.POINTER(ctypes.c_float) # // Logits for the ith token. Equivalent to: # // llama_get_logits(ctx) + i*n_vocab # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); def llama_get_logits_ith( - ctx: llama_context_p, i: Union[ctypes.c_int32, int] - , / + ctx: llama_context_p, i: Union[ctypes.c_int32, int], / ): # type: (...) -> Array[float] # type: ignore """Logits for the ith token. Equivalent to: llama_get_logits(ctx) + i*n_vocab""" @@ -1798,7 +1856,7 @@ def llama_get_logits_ith( llama_get_logits_ith = _lib.llama_get_logits_ith llama_get_logits_ith.argtypes = [llama_context_p_ctypes, ctypes.c_int32] -llama_get_logits_ith.restype = c_float_p +llama_get_logits_ith.restype = ctypes.POINTER(ctypes.c_float) # Get the embeddings for the input @@ -1814,7 +1872,7 @@ def llama_get_embeddings( llama_get_embeddings = _lib.llama_get_embeddings llama_get_embeddings.argtypes = [llama_context_p_ctypes] -llama_get_embeddings.restype = c_float_p +llama_get_embeddings.restype = ctypes.POINTER(ctypes.c_float) # // Get the embeddings for the ith sequence @@ -1830,7 +1888,7 @@ def llama_get_embeddings_ith( llama_get_embeddings_ith = _lib.llama_get_embeddings_ith llama_get_embeddings_ith.argtypes = [llama_context_p_ctypes, ctypes.c_int32] -llama_get_embeddings_ith.restype = c_float_p +llama_get_embeddings_ith.restype = ctypes.POINTER(ctypes.c_float) # // @@ -1839,7 +1897,9 @@ def llama_get_embeddings_ith( # LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); -def llama_token_get_text(model: llama_model_p, token: Union[llama_token, int], /) -> bytes: +def llama_token_get_text( + model: llama_model_p, token: Union[llama_token, int], / +) -> bytes: ... @@ -1861,7 +1921,9 @@ def llama_token_get_score( # LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); -def llama_token_get_type(model: llama_model_p, token: Union[llama_token, int], /) -> int: +def llama_token_get_type( + model: llama_model_p, token: Union[llama_token, int], / +) -> int: ... @@ -1995,11 +2057,11 @@ def llama_tokenize( model: llama_model_p, text: bytes, text_len: Union[ctypes.c_int, int], - tokens, # type: Array[llama_token] + tokens: CtypesArray[llama_token], n_max_tokens: Union[ctypes.c_int, int], add_bos: Union[ctypes.c_bool, bool], special: Union[ctypes.c_bool, bool], - / + /, ) -> int: """Convert the provided text into tokens.""" ... @@ -2032,7 +2094,7 @@ def llama_token_to_piece( token: Union[llama_token, int], buf: Union[ctypes.c_char_p, bytes], length: Union[ctypes.c_int, int], - / + /, ) -> int: """Token Id -> Piece. Uses the vocabulary in the provided context. @@ -2043,7 +2105,12 @@ def llama_token_to_piece( llama_token_to_piece = _lib.llama_token_to_piece -llama_token_to_piece.argtypes = [llama_model_p_ctypes, llama_token, ctypes.c_char_p, ctypes.c_int32] +llama_token_to_piece.argtypes = [ + llama_model_p_ctypes, + llama_token, + ctypes.c_char_p, + ctypes.c_int32, +] llama_token_to_piece.restype = ctypes.c_int32 @@ -2066,25 +2133,25 @@ def llama_token_to_piece( # char * buf, # int32_t length); def llama_chat_apply_template( - model: llama_model_p, - tmpl: bytes, - chat: "ctypes._Pointer[llama_chat_message]", # type: ignore - n_msg: int, - / + model: llama_model_p, + tmpl: bytes, + chat: CtypesArray[llama_chat_message], + n_msg: int, + /, ) -> int: ... + llama_chat_apply_template = _lib.llama_chat_apply_template llama_chat_apply_template.argtypes = [ ctypes.c_void_p, ctypes.c_char_p, ctypes.POINTER(llama_chat_message), - ctypes.c_size_t + ctypes.c_size_t, ] llama_chat_apply_template.restype = ctypes.c_int32 - # // # // Grammar # // @@ -2095,10 +2162,12 @@ def llama_chat_apply_template( # size_t n_rules, # size_t start_rule_index); def llama_grammar_init( - rules, # type: Array[llama_grammar_element_p] # type: ignore + rules: CtypesArray[ + CtypesPointer[llama_grammar_element] + ], # NOTE: This might be wrong type sig n_rules: Union[ctypes.c_size_t, int], start_rule_index: Union[ctypes.c_size_t, int], - / + /, ) -> llama_grammar_p: """Initialize a grammar from a set of rules.""" ... @@ -2163,13 +2232,15 @@ def llama_set_rng_seed(ctx: llama_context_p, seed: Union[ctypes.c_uint32, int], # float penalty_present); def llama_sample_repetition_penalties( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - last_tokens_data, # type: Array[llama_token] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], + last_tokens_data: CtypesArray[llama_token], penalty_last_n: Union[ctypes.c_size_t, int], penalty_repeat: Union[ctypes.c_float, float], penalty_freq: Union[ctypes.c_float, float], penalty_present: Union[ctypes.c_float, float], - / + /, ): """Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. @@ -2201,10 +2272,10 @@ def llama_sample_repetition_penalties( # float scale); def llama_sample_apply_guidance( ctx: llama_context_p, - logits, # type: _Pointer[ctypes.c_float] - logits_guidance, # type: _Pointer[ctypes.c_float] + logits: CtypesArray[ctypes.c_float], + logits_guidance: CtypesArray[ctypes.c_float], scale: Union[ctypes.c_float, float], - / + /, ): """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" ... @@ -2213,8 +2284,8 @@ def llama_sample_apply_guidance( llama_sample_apply_guidance = _lib.llama_sample_apply_guidance llama_sample_apply_guidance.argtypes = [ llama_context_p_ctypes, - c_float_p, - c_float_p, + ctypes.POINTER(ctypes.c_float), + ctypes.POINTER(ctypes.c_float), ctypes.c_float, ] llama_sample_apply_guidance.restype = None @@ -2228,10 +2299,12 @@ def llama_sample_apply_guidance( # "use llama_sample_apply_guidance() instead"); def llama_sample_classifier_free_guidance( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], guidance_ctx: llama_context_p, scale: Union[ctypes.c_float, float], - / + /, ): """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" ... @@ -2252,8 +2325,11 @@ def llama_sample_classifier_free_guidance( # struct llama_context * ctx, # llama_token_data_array * candidates); def llama_sample_softmax( - ctx: llama_context_p, candidates, # type: _Pointer[llama_token_data] - / + ctx: llama_context_p, + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], + /, ): """Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.""" ... @@ -2275,10 +2351,12 @@ def llama_sample_softmax( # size_t min_keep); def llama_sample_top_k( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], k: Union[ctypes.c_int, int], min_keep: Union[ctypes.c_size_t, int], - / + /, ): """Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" ... @@ -2302,10 +2380,12 @@ def llama_sample_top_k( # size_t min_keep); def llama_sample_top_p( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], p: Union[ctypes.c_float, float], min_keep: Union[ctypes.c_size_t, int], - / + /, ): """Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" ... @@ -2329,10 +2409,12 @@ def llama_sample_top_p( # size_t min_keep); def llama_sample_min_p( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], p: Union[ctypes.c_float, float], min_keep: Union[ctypes.c_size_t, int], - / + /, ): """Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841""" ... @@ -2356,10 +2438,12 @@ def llama_sample_min_p( # size_t min_keep); def llama_sample_tail_free( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], z: Union[ctypes.c_float, float], min_keep: Union[ctypes.c_size_t, int], - / + /, ): """Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.""" ... @@ -2383,10 +2467,12 @@ def llama_sample_tail_free( # size_t min_keep); def llama_sample_typical( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], p: Union[ctypes.c_float, float], min_keep: Union[ctypes.c_size_t, int], - / + /, ): """Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.""" ... @@ -2411,11 +2497,13 @@ def llama_sample_typical( # float exponent_val); def llama_sample_entropy( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], min_temp: Union[ctypes.c_float, float], max_temp: Union[ctypes.c_float, float], exponent_val: Union[ctypes.c_float, float], - / + /, ): """Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.""" ... @@ -2438,9 +2526,11 @@ def llama_sample_entropy( # float temp); def llama_sample_temp( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], temp: Union[ctypes.c_float, float], - / + /, ): """Temperature sampling described in academic paper "Generating Long Sequences with Sparse Transformers" https://arxiv.org/abs/1904.10509 @@ -2467,9 +2557,11 @@ def llama_sample_temp( # "use llama_sample_temp instead"); def llama_sample_temperature( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], temp: Union[ctypes.c_float, float], - / + /, ): """use llama_sample_temp instead""" ... @@ -2491,9 +2583,11 @@ def llama_sample_temperature( # const struct llama_grammar * grammar); def llama_sample_grammar( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], grammar, # type: llama_grammar_p - / + /, ): """Apply constraints from grammar @@ -2528,12 +2622,14 @@ def llama_sample_grammar( # float * mu); def llama_sample_token_mirostat( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], tau: Union[ctypes.c_float, float], eta: Union[ctypes.c_float, float], m: Union[ctypes.c_int, int], - mu, # type: _Pointer[ctypes.c_float] - / + mu: CtypesPointerOrRef[ctypes.c_float], + /, ) -> int: """Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2554,7 +2650,7 @@ def llama_sample_token_mirostat( ctypes.c_float, ctypes.c_float, ctypes.c_int32, - c_float_p, + ctypes.POINTER(ctypes.c_float), ] llama_sample_token_mirostat.restype = llama_token @@ -2572,11 +2668,13 @@ def llama_sample_token_mirostat( # float * mu); def llama_sample_token_mirostat_v2( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], tau: Union[ctypes.c_float, float], eta: Union[ctypes.c_float, float], mu, # type: _Pointer[ctypes.c_float] - / + /, ) -> int: """Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2595,7 +2693,7 @@ def llama_sample_token_mirostat_v2( llama_token_data_array_p, ctypes.c_float, ctypes.c_float, - c_float_p, + ctypes.POINTER(ctypes.c_float), ] llama_sample_token_mirostat_v2.restype = llama_token @@ -2607,8 +2705,10 @@ def llama_sample_token_mirostat_v2( # llama_token_data_array * candidates); def llama_sample_token_greedy( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - / + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], + /, ) -> int: """Selects the token with the highest probability.""" ... @@ -2628,8 +2728,10 @@ def llama_sample_token_greedy( # llama_token_data_array * candidates); def llama_sample_token( ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - / + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], + /, ) -> int: """Randomly selects a token from the candidates based on their probabilities.""" ... @@ -2649,10 +2751,7 @@ def llama_sample_token( # struct llama_grammar * grammar, # llama_token token); def llama_grammar_accept_token( - ctx: llama_context_p, - grammar: llama_grammar_p, - token: Union[llama_token, int], - / + ctx: llama_context_p, grammar: llama_grammar_p, token: Union[llama_token, int], / ) -> None: """Accepts the sampled token into the grammar""" ... @@ -2711,7 +2810,9 @@ class llama_beams_state(ctypes.Structure): # // void* callback_data is any custom data passed to llama_beam_search, that is subsequently # // passed back to beam_search_callback. This avoids having to use global variables in the callback. # typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); -llama_beam_search_callback_fn_t = ctypes.CFUNCTYPE(None, ctypes.c_void_p, llama_beams_state) +llama_beam_search_callback_fn_t = ctypes.CFUNCTYPE( + None, ctypes.c_void_p, llama_beams_state +) # /// @details Deterministically returns entire sentence constructed by a beam search. @@ -2731,12 +2832,12 @@ class llama_beams_state(ctypes.Structure): # int32_t n_predict); def llama_beam_search( ctx: llama_context_p, - callback: "ctypes._CFuncPtr[None, ctypes.c_void_p, llama_beams_state]", # type: ignore + callback: CtypesFuncPointer, callback_data: ctypes.c_void_p, n_beams: Union[ctypes.c_size_t, int], n_past: Union[ctypes.c_int, int], n_predict: Union[ctypes.c_int, int], - / + /, ): ... @@ -2806,8 +2907,9 @@ def llama_print_system_info() -> bytes: # // If this is not called, or NULL is supplied, everything is output on stderr. # LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); def llama_log_set( - log_callback: Union["ctypes._FuncPointer", ctypes.c_void_p], user_data: ctypes.c_void_p, # type: ignore - / + log_callback: Optional[CtypesFuncPointer], + user_data: ctypes.c_void_p, # type: ignore + /, ): """Set callback for all future logging events. From d80c5cf29dcf355286ce15baa2dde44f4d72b439 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 02:30:24 -0500 Subject: [PATCH 168/624] docs: fix indentation for mkdocs-material --- README.md | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index 03c95e84d..05706dc21 100644 --- a/README.md +++ b/README.md @@ -12,14 +12,14 @@ This package provides: - Low-level access to C API via `ctypes` interface. - High-level Python API for text completion - - OpenAI-like API - - [LangChain compatibility](https://python.langchain.com/docs/integrations/llms/llamacpp) - - [LlamaIndex compatibility](https://docs.llamaindex.ai/en/stable/examples/llm/llama_2_llama_cpp.html) + - OpenAI-like API + - [LangChain compatibility](https://python.langchain.com/docs/integrations/llms/llamacpp) + - [LlamaIndex compatibility](https://docs.llamaindex.ai/en/stable/examples/llm/llama_2_llama_cpp.html) - OpenAI compatible web server - - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) - - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) - - [Vision API support](https://llama-cpp-python.readthedocs.io/en/latest/server/#multimodal-models) - - [Multiple Models](https://llama-cpp-python.readthedocs.io/en/latest/server/#configuration-and-multi-model-support) + - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) + - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) + - [Vision API support](https://llama-cpp-python.readthedocs.io/en/latest/server/#multimodal-models) + - [Multiple Models](https://llama-cpp-python.readthedocs.io/en/latest/server/#configuration-and-multi-model-support) Documentation is available at [https://llama-cpp-python.readthedocs.io/en/latest](https://llama-cpp-python.readthedocs.io/en/latest). From 32efed7b07bcf4d451485283cee4e21bc557e686 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 03:25:11 -0500 Subject: [PATCH 169/624] docs: Update README --- README.md | 124 +++++++++++++++++++++++++++++++++++++++--------------- 1 file changed, 91 insertions(+), 33 deletions(-) diff --git a/README.md b/README.md index 05706dc21..5d8bbc520 100644 --- a/README.md +++ b/README.md @@ -25,47 +25,82 @@ Documentation is available at [https://llama-cpp-python.readthedocs.io/en/latest ## Installation -`llama-cpp-python` can be installed directly from PyPI as a source distribution by running: +Requirements: + + - Python 3.8+ + - C compiler + - Linux: gcc or clang + - Windows: Visual Studio or MinGW + - MacOS: Xcode + +To install the package, run: ```bash pip install llama-cpp-python ``` -This will build `llama.cpp` from source using cmake and your system's c compiler (required) and install the library alongside this python package. +This will also build `llama.cpp` from source and install it alongside this python package. -If you run into issues during installation add the `--verbose` flag to the `pip install` command to see the full cmake build log. +If this fails, add `--verbose` to the `pip install` see the full cmake build log. -### Installation with Specific Hardware Acceleration (BLAS, CUDA, Metal, etc) +### Installation Configuration -The default pip install behaviour is to build `llama.cpp` for CPU only on Linux and Windows and use Metal on MacOS. +`llama.cpp` supports a number of hardware acceleration backends to speed up inference as well as backend specific options. See the [llama.cpp README](https://github.com/ggerganov/llama.cpp#build) for a full list. -`llama.cpp` supports a number of hardware acceleration backends depending including OpenBLAS, cuBLAS, CLBlast, HIPBLAS, and Metal. -See the [llama.cpp README](https://github.com/ggerganov/llama.cpp#build) for a full list of supported backends. +All `llama.cpp` cmake build options can be set via the `CMAKE_ARGS` environment variable or via the `--config-settings / -C` cli flag during installation. -All of these backends are supported by `llama-cpp-python` and can be enabled by setting the `CMAKE_ARGS` environment variable before installing. - -On Linux and Mac you set the `CMAKE_ARGS` like this: +
+Environment Variables ```bash -CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python +# Linux and Mac +CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" \ + pip install llama-cpp-python ``` -On Windows you can set the `CMAKE_ARGS` like this: - -```ps +```powershell +# Windows $env:CMAKE_ARGS = "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python ``` +
+ +
+CLI / requirements.txt + +They can also be set via `pip install -C / --config-settings` command and saved to a `requirements.txt` file: + +```bash +pip install --upgrade pip # ensure pip is up to date +pip install llama-cpp-python \ + -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS" +``` + +```txt +# requirements.txt + +llama-cpp-python -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS" +``` + +
-#### OpenBLAS -To install with OpenBLAS, set the `LLAMA_BLAS and LLAMA_BLAS_VENDOR` environment variables before installing: +### Supported Backends + +Below are some common backends, their build commands and any additional environment variables required. + +
+OpenBLAS (CPU) + +To install with OpenBLAS, set the `LLAMA_BLAS` and `LLAMA_BLAS_VENDOR` environment variables before installing: ```bash CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python ``` +
-#### cuBLAS +
+cuBLAS (CUDA) To install with cuBLAS, set the `LLAMA_CUBLAS=on` environment variable before installing: @@ -73,7 +108,10 @@ To install with cuBLAS, set the `LLAMA_CUBLAS=on` environment variable before in CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python ``` -#### Metal +
+ +
+Metal To install with Metal (MPS), set the `LLAMA_METAL=on` environment variable before installing: @@ -81,7 +119,10 @@ To install with Metal (MPS), set the `LLAMA_METAL=on` environment variable befor CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python ``` -#### CLBlast +
+
+ +CLBlast (OpenCL) To install with CLBlast, set the `LLAMA_CLBLAST=on` environment variable before installing: @@ -89,7 +130,10 @@ To install with CLBlast, set the `LLAMA_CLBLAST=on` environment variable before CMAKE_ARGS="-DLLAMA_CLBLAST=on" pip install llama-cpp-python ``` -#### hipBLAS +
+ +
+hipBLAS (ROCm) To install with hipBLAS / ROCm support for AMD cards, set the `LLAMA_HIPBLAS=on` environment variable before installing: @@ -97,7 +141,10 @@ To install with hipBLAS / ROCm support for AMD cards, set the `LLAMA_HIPBLAS=on` CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python ``` -#### Vulkan +
+ +
+Vulkan To install with Vulkan support, set the `LLAMA_VULKAN=on` environment variable before installing: @@ -105,15 +152,20 @@ To install with Vulkan support, set the `LLAMA_VULKAN=on` environment variable b CMAKE_ARGS="-DLLAMA_VULKAN=on" pip install llama-cpp-python ``` -#### Kompute +
+ +
+Kompute To install with Kompute support, set the `LLAMA_KOMPUTE=on` environment variable before installing: ```bash CMAKE_ARGS="-DLLAMA_KOMPUTE=on" pip install llama-cpp-python ``` +
-#### SYCL +
+SYCL To install with SYCL support, set the `LLAMA_SYCL=on` environment variable before installing: @@ -121,9 +173,14 @@ To install with SYCL support, set the `LLAMA_SYCL=on` environment variable befor source /opt/intel/oneapi/setvars.sh CMAKE_ARGS="-DLLAMA_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install llama-cpp-python ``` +
+ ### Windows Notes +
+Error: Can't find 'nmake' or 'CMAKE_C_COMPILER' + If you run into issues where it complains it can't find `'nmake'` `'?'` or CMAKE_C_COMPILER, you can extract w64devkit as [mentioned in llama.cpp repo](https://github.com/ggerganov/llama.cpp#openblas) and add those manually to CMAKE_ARGS before running `pip` install: ```ps @@ -132,12 +189,14 @@ $env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.e ``` See the above instructions and set `CMAKE_ARGS` to the BLAS backend you want to use. +
### MacOS Notes Detailed MacOS Metal GPU install documentation is available at [docs/install/macos.md](https://llama-cpp-python.readthedocs.io/en/latest/install/macos/) -#### M1 Mac Performance Issue +
+M1 Mac Performance Issue Note: If you are using Apple Silicon (M1) Mac, make sure you have installed a version of Python that supports arm64 architecture. For example: @@ -147,24 +206,21 @@ bash Miniforge3-MacOSX-arm64.sh ``` Otherwise, while installing it will build the llama.cpp x86 version which will be 10x slower on Apple Silicon (M1) Mac. +
-#### M Series Mac Error: `(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64'))` +
+M Series Mac Error: `(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64'))` Try installing with ```bash CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on" pip install --upgrade --verbose --force-reinstall --no-cache-dir llama-cpp-python ``` +
### Upgrading and Reinstalling -To upgrade or rebuild `llama-cpp-python` add the following flags to ensure that the package is rebuilt correctly: - -```bash -pip install llama-cpp-python --upgrade --force-reinstall --no-cache-dir -``` - -This will ensure that all source files are re-built with the most recently set `CMAKE_ARGS` flags. +To upgrade and rebuild `llama-cpp-python` add `--upgrade --force-reinstall --no-cache-dir` flags to the `pip install` command to ensure the package is rebuilt from source. ## High-level API @@ -218,13 +274,15 @@ You can pull `Llama` models from Hugging Face using the `from_pretrained` method You'll need to install the `huggingface-hub` package to use this feature (`pip install huggingface-hub`). ```python -llama = Llama.from_pretrained( +llm = Llama.from_pretrained( repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", filename="*q8_0.gguf", verbose=False ) ``` +By default the `from_pretrained` method will download the model to the huggingface cache directory so you can manage installed model files with the `huggingface-cli` tool. + ### Chat Completion The high-level API also provides a simple interface for chat completion. From e10af30cf1daec7f0717cc87b1cc09fa3cfd9f12 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 03:27:28 -0500 Subject: [PATCH 170/624] fix: TypeAlias import error --- llama_cpp/llama_cpp.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 193919024..3da131e1a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -15,9 +15,9 @@ Optional, TYPE_CHECKING, TypeVar, - TypeAlias, Generic, ) +from typing_extensions import TypeAlias # Load the library From 045cc1267059e3a13732662c964167305e478011 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 03:53:52 -0500 Subject: [PATCH 171/624] docs: Update README --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 5d8bbc520..1b7634fde 100644 --- a/README.md +++ b/README.md @@ -49,7 +49,7 @@ If this fails, add `--verbose` to the `pip install` see the full cmake build log All `llama.cpp` cmake build options can be set via the `CMAKE_ARGS` environment variable or via the `--config-settings / -C` cli flag during installation. -
+
Environment Variables ```bash @@ -89,7 +89,7 @@ llama-cpp-python -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS" Below are some common backends, their build commands and any additional environment variables required. -
+
OpenBLAS (CPU) To install with OpenBLAS, set the `LLAMA_BLAS` and `LLAMA_BLAS_VENDOR` environment variables before installing: From 0f8cad6cb7c43c6dca7198bcab741d53c623af86 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 11:31:44 -0500 Subject: [PATCH 172/624] docs: Update README --- README.md | 20 ++++++++++++-------- 1 file changed, 12 insertions(+), 8 deletions(-) diff --git a/README.md b/README.md index 1b7634fde..b208d6f5d 100644 --- a/README.md +++ b/README.md @@ -84,6 +84,7 @@ llama-cpp-python -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS"
+https://github.com/abetlen/llama-cpp-python/releases/download/${VERSION}/ ### Supported Backends @@ -268,9 +269,9 @@ Below is a short example demonstrating how to use the high-level API to for basi Text completion is available through the [`__call__`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.__call__) and [`create_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_completion) methods of the [`Llama`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama) class. -## Pulling models from Hugging Face +### Pulling models from Hugging Face Hub -You can pull `Llama` models from Hugging Face using the `from_pretrained` method. +You can download `Llama` models in `gguf` format directly from Hugging Face using the [`from_pretrained`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.from_pretrained) method. You'll need to install the `huggingface-hub` package to use this feature (`pip install huggingface-hub`). ```python @@ -281,7 +282,7 @@ llm = Llama.from_pretrained( ) ``` -By default the `from_pretrained` method will download the model to the huggingface cache directory so you can manage installed model files with the `huggingface-cli` tool. +By default [`from_pretrained`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.from_pretrained) will download the model to the huggingface cache directory, you can then manage installed model files with the [`huggingface-cli`](https://huggingface.co/docs/huggingface_hub/en/guides/cli) tool. ### Chat Completion @@ -308,13 +309,16 @@ Note that `chat_format` option must be set for the particular model you are usin Chat completion is available through the [`create_chat_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion) method of the [`Llama`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama) class. +For OpenAI API v1 compatibility, you use the [`create_chat_completion_openai_v1`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion_openai_v1) method which will return pydantic models instead of dicts. + + ### JSON and JSON Schema Mode -If you want to constrain chat responses to only valid JSON or a specific JSON Schema you can use the `response_format` argument to the `create_chat_completion` method. +To constrain chat responses to only valid JSON or a specific JSON Schema use the `response_format` argument in [`create_chat_completion`](http://localhost:8000/api-reference/#llama_cpp.Llama.create_chat_completion). #### JSON Mode -The following example will constrain the response to be valid JSON. +The following example will constrain the response to valid JSON strings only. ```python >>> from llama_cpp import Llama @@ -336,7 +340,7 @@ The following example will constrain the response to be valid JSON. #### JSON Schema Mode -To constrain the response to a specific JSON Schema, you can use the `schema` property of the `response_format` argument. +To constrain the response further to a specific JSON Schema add the schema to the `schema` property of the `response_format` argument. ```python >>> from llama_cpp import Llama @@ -471,7 +475,7 @@ llama = Llama( ### Embeddings -`llama-cpp-python` supports generating embeddings from the text. +To generate text embeddings use [`create_embedding`](http://localhost:8000/api-reference/#llama_cpp.Llama.create_embedding). ```python import llama_cpp @@ -480,7 +484,7 @@ llm = llama_cpp.Llama(model_path="path/to/model.gguf", embeddings=True) embeddings = llm.create_embedding("Hello, world!") -# or batched +# or create multiple embeddings at once embeddings = llm.create_embedding(["Hello, world!", "Goodbye, world!"]) ``` From eb56ce2e2ab2007616905c64d93aa456b8149065 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Feb 2024 11:33:05 -0500 Subject: [PATCH 173/624] docs: fix low-level api example --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index b208d6f5d..505f4d259 100644 --- a/README.md +++ b/README.md @@ -561,7 +561,7 @@ Below is a short example demonstrating how to use the low-level API to tokenize ```python >>> import llama_cpp >>> import ctypes ->>> llama_cpp.llama_backend_init(numa=False) # Must be called once at the start of each program +>>> llama_cpp.llama_backend_init(False) # Must be called once at the start of each program >>> params = llama_cpp.llama_context_default_params() # use bytes for char * params >>> model = llama_cpp.llama_load_model_from_file(b"./models/7b/llama-model.gguf", params) @@ -569,7 +569,7 @@ Below is a short example demonstrating how to use the low-level API to tokenize >>> max_tokens = params.n_ctx # use ctypes arrays for array params >>> tokens = (llama_cpp.llama_token * int(max_tokens))() ->>> n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, add_bos=llama_cpp.c_bool(True)) +>>> n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, llama_cpp.c_bool(True)) >>> llama_cpp.llama_free(ctx) ``` From 410e02da515bd494fe53979c5578be9de9653753 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 00:43:31 -0500 Subject: [PATCH 174/624] docs: Fix typo --- README.md | 2 -- 1 file changed, 2 deletions(-) diff --git a/README.md b/README.md index 505f4d259..6dbd79dd4 100644 --- a/README.md +++ b/README.md @@ -84,8 +84,6 @@ llama-cpp-python -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS"
-https://github.com/abetlen/llama-cpp-python/releases/download/${VERSION}/ - ### Supported Backends Below are some common backends, their build commands and any additional environment variables required. From a0ce429dc09476b7575a92e00c5f8af87f80d109 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 03:39:38 -0500 Subject: [PATCH 175/624] misc: use decorator to bind low level api functions, fixes docs --- llama_cpp/llama_cpp.py | 1444 ++++++++++++++++++++-------------------- 1 file changed, 713 insertions(+), 731 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 3da131e1a..f4d523be0 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -3,12 +3,12 @@ import sys import os import ctypes -from ctypes import ( - _Pointer, # type: ignore - Array, -) +import functools import pathlib + from typing import ( + Any, + Callable, List, Union, NewType, @@ -110,6 +110,36 @@ class CtypesRef(Generic[CtypesCData]): CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore +def ctypes_function_for_shared_library(lib: ctypes.CDLL): + def ctypes_function( + name: str, argtypes: List[Any], restype: Any, enabled: bool = True + ): + def decorator(f: Callable[..., Any]): + if enabled: + func = getattr(lib, name) + func.argtypes = argtypes + func.restype = restype + functools.wraps(f)(func) + return func + else: + return f + + return decorator + + return ctypes_function + + +ctypes_function = ctypes_function_for_shared_library(_lib) + + +def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCData]: + """Type-annotated version of ctypes.byref""" + ... + + +byref = ctypes.byref # type: ignore + + # from ggml-backend.h # typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE( @@ -707,43 +737,48 @@ class llama_chat_message(ctypes.Structure): # // Helpers for getting default parameters # LLAMA_API struct llama_model_params llama_model_default_params(void); +@ctypes_function( + "llama_model_default_params", + [], + llama_model_params, +) def llama_model_default_params() -> llama_model_params: """Get default parameters for llama_model""" ... -llama_model_default_params = _lib.llama_model_default_params -llama_model_default_params.argtypes = [] -llama_model_default_params.restype = llama_model_params - - # LLAMA_API struct llama_context_params llama_context_default_params(void); +@ctypes_function( + "llama_context_default_params", + [], + llama_context_params, +) def llama_context_default_params() -> llama_context_params: """Get default parameters for llama_context""" ... -llama_context_default_params = _lib.llama_context_default_params -llama_context_default_params.argtypes = [] -llama_context_default_params.restype = llama_context_params - - # LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void); +@ctypes_function( + "llama_model_quantize_default_params", + [], + llama_model_quantize_params, +) def llama_model_quantize_default_params() -> llama_model_quantize_params: """Get default parameters for llama_model_quantize""" ... -llama_model_quantize_default_params = _lib.llama_model_quantize_default_params -llama_model_quantize_default_params.argtypes = [] -llama_model_quantize_default_params.restype = llama_model_quantize_params - - # // Initialize the llama + ggml backend # // If numa is true, use NUMA optimizations # // Call once at the start of the program # LLAMA_API void llama_backend_init(bool numa); # LLAMA_API void llama_backend_init(void); +@ctypes_function( + "llama_backend_init", + [], + None, +) def llama_backend_init(): """Initialize the llama + ggml backend If numa is true, use NUMA optimizations @@ -751,11 +786,6 @@ def llama_backend_init(): ... -llama_backend_init = _lib.llama_backend_init -llama_backend_init.argtypes = [] -llama_backend_init.restype = None - - # // numa strategies # enum ggml_numa_strategy { # GGML_NUMA_STRATEGY_DISABLED = 0, @@ -775,228 +805,201 @@ def llama_backend_init(): # //optional: # LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); +@ctypes_function( + "llama_numa_init", + [ctypes.c_int], + None, +) def llama_numa_init(numa: int, /): ... -llama_numa_init = _lib.llama_numa_init -llama_numa_init.argtypes = [ctypes.c_int] -llama_numa_init.restype = None - - # // Call once at the end of the program - currently only used for MPI # LLAMA_API void llama_backend_free(void); +@ctypes_function( + "llama_backend_free", + [], + None, +) def llama_backend_free(): """Call once at the end of the program - currently only used for MPI""" ... -llama_backend_free = _lib.llama_backend_free -llama_backend_free.argtypes = [] -llama_backend_free.restype = None - - # LLAMA_API struct llama_model * llama_load_model_from_file( # const char * path_model, # struct llama_model_params params); +@ctypes_function( + "llama_load_model_from_file", + [ctypes.c_char_p, llama_model_params], + llama_model_p_ctypes, +) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / ) -> Optional[llama_model_p]: ... -llama_load_model_from_file = _lib.llama_load_model_from_file -llama_load_model_from_file.argtypes = [ctypes.c_char_p, llama_model_params] -llama_load_model_from_file.restype = llama_model_p_ctypes - - # LLAMA_API void llama_free_model(struct llama_model * model); +@ctypes_function( + "llama_free_model", + [llama_model_p_ctypes], + None, +) def llama_free_model(model: llama_model_p, /): ... -llama_free_model = _lib.llama_free_model -llama_free_model.argtypes = [llama_model_p_ctypes] -llama_free_model.restype = None - - # LLAMA_API struct llama_context * llama_new_context_with_model( # struct llama_model * model, # struct llama_context_params params); +@ctypes_function( + "llama_new_context_with_model", + [llama_model_p_ctypes, llama_context_params], + llama_context_p_ctypes, +) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / ) -> Optional[llama_context_p]: ... -llama_new_context_with_model = _lib.llama_new_context_with_model -llama_new_context_with_model.argtypes = [llama_model_p_ctypes, llama_context_params] -llama_new_context_with_model.restype = llama_context_p_ctypes - - # // Frees all allocated memory # LLAMA_API void llama_free(struct llama_context * ctx); +@ctypes_function( + "llama_free", + [llama_context_p_ctypes], + None, +) def llama_free(ctx: llama_context_p, /): """Frees all allocated memory""" ... -llama_free = _lib.llama_free -llama_free.argtypes = [llama_context_p_ctypes] -llama_free.restype = None - - # LLAMA_API int64_t llama_time_us(void); +@ctypes_function( + "llama_time_us", + [], + ctypes.c_int64, +) def llama_time_us() -> int: ... -llama_time_us = _lib.llama_time_us -llama_time_us.argtypes = [] -llama_time_us.restype = ctypes.c_int64 +# LLAMA_API size_t llama_max_devices(void); -# LLAMA_API size_t llama_max_devices(void); +@ctypes_function("llama_max_devices", [], ctypes.c_size_t) def llama_max_devices() -> int: ... -llama_max_devices = _lib.llama_max_devices -llama_max_devices.argtypes = [] -llama_max_devices.restype = ctypes.c_size_t +# LLAMA_API bool llama_supports_mmap (void); -# LLAMA_API bool llama_supports_mmap (void); +@ctypes_function("llama_supports_mmap", [], ctypes.c_bool) def llama_supports_mmap() -> bool: ... -llama_supports_mmap = _lib.llama_supports_mmap -llama_supports_mmap.argtypes = [] -llama_supports_mmap.restype = ctypes.c_bool +# LLAMA_API bool llama_supports_mlock (void); -# LLAMA_API bool llama_supports_mlock (void); +@ctypes_function("llama_supports_mlock", [], ctypes.c_bool) def llama_supports_mlock() -> bool: ... -llama_supports_mlock = _lib.llama_supports_mlock -llama_supports_mlock.argtypes = [] -llama_supports_mlock.restype = ctypes.c_bool +# LLAMA_API bool llama_supports_gpu_offload(void); -# LLAMA_API bool llama_supports_gpu_offload(void); +@ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) def llama_supports_gpu_offload() -> bool: ... -llama_supports_gpu_offload = _lib.llama_supports_gpu_offload -llama_supports_gpu_offload.argtypes = [] -llama_supports_gpu_offload.restype = ctypes.c_bool +# LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); -# LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); +@ctypes_function("llama_mmap_supported", [], ctypes.c_bool) def llama_mmap_supported() -> bool: ... -llama_mmap_supported = _lib.llama_mmap_supported -llama_mmap_supported.argtypes = [] -llama_mmap_supported.restype = ctypes.c_bool +# LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); -# LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); +@ctypes_function("llama_mlock_supported", [], ctypes.c_bool) def llama_mlock_supported() -> bool: ... -llama_mlock_supported = _lib.llama_mlock_supported -llama_mlock_supported.argtypes = [] -llama_mlock_supported.restype = ctypes.c_bool +# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); -# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); +@ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... -llama_get_model = _lib.llama_get_model -llama_get_model.argtypes = [llama_context_p_ctypes] -llama_get_model.restype = llama_model_p_ctypes +# LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); -# LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); +@ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) def llama_n_ctx(ctx: llama_context_p, /) -> int: ... -llama_n_ctx = _lib.llama_n_ctx -llama_n_ctx.argtypes = [llama_context_p_ctypes] -llama_n_ctx.restype = ctypes.c_uint32 +# LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); -# LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); +@ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) def llama_n_batch(ctx: llama_context_p, /) -> int: ... -llama_n_batch = _lib.llama_n_batch -llama_n_batch.argtypes = [llama_context_p_ctypes] -llama_n_batch.restype = ctypes.c_uint32 +# LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); -# LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); +@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) def llama_vocab_type(model: llama_model_p, /) -> int: ... -llama_vocab_type = _lib.llama_vocab_type -llama_vocab_type.argtypes = [llama_model_p_ctypes] -llama_vocab_type.restype = ctypes.c_int +# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); -# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); +@ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_vocab(model: llama_model_p, /) -> int: ... -llama_n_vocab = _lib.llama_n_vocab -llama_n_vocab.argtypes = [llama_model_p_ctypes] -llama_n_vocab.restype = ctypes.c_int32 +# LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); -# LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); +@ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_ctx_train(model: llama_model_p, /) -> int: ... -llama_n_ctx_train = _lib.llama_n_ctx_train -llama_n_ctx_train.argtypes = [llama_model_p_ctypes] -llama_n_ctx_train.restype = ctypes.c_int32 +# LLAMA_API int32_t llama_n_embd (const struct llama_model * model); -# LLAMA_API int32_t llama_n_embd (const struct llama_model * model); +@ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_embd(model: llama_model_p, /) -> int: ... -llama_n_embd = _lib.llama_n_embd -llama_n_embd.argtypes = [llama_model_p_ctypes] -llama_n_embd.restype = ctypes.c_int32 - - # // Get the model's RoPE frequency scaling factor # LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); + + +@ctypes_function("llama_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: """Get the model's RoPE frequency scaling factor""" ... -llama_rope_freq_scale_train = _lib.llama_rope_freq_scale_train -llama_rope_freq_scale_train.argtypes = [llama_model_p_ctypes] -llama_rope_freq_scale_train.restype = ctypes.c_float - # // Functions to access the model's GGUF metadata scalar values # // - The functions return the length of the string on success, or -1 on failure # // - The output string is always null-terminated and cleared on failure @@ -1005,6 +1008,18 @@ def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: # // Get metadata value as a string by key name # LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); + + +@ctypes_function( + "llama_model_meta_val_str", + [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_model_meta_val_str( model: llama_model_p, key: Union[ctypes.c_char_p, bytes], @@ -1016,106 +1031,112 @@ def llama_model_meta_val_str( ... -llama_model_meta_val_str = _lib.llama_model_meta_val_str -llama_model_meta_val_str.argtypes = [ - llama_model_p_ctypes, - ctypes.c_char_p, - ctypes.c_char_p, - ctypes.c_size_t, -] -llama_model_meta_val_str.restype = ctypes.c_int32 - - # // Get the number of metadata key/value pairs # LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model); + + +@ctypes_function("llama_model_meta_count", [llama_model_p_ctypes], ctypes.c_int32) def llama_model_meta_count(model: llama_model_p, /) -> int: """Get the number of metadata key/value pairs""" ... -llama_model_meta_count = _lib.llama_model_meta_count -llama_model_meta_count.argtypes = [llama_model_p_ctypes] -llama_model_meta_count.restype = ctypes.c_int32 - - # // Get metadata key name by index # LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); + + +@ctypes_function( + "llama_model_meta_key_by_index", + [ + llama_model_p_ctypes, + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_model_meta_key_by_index( - model: llama_model_p, i: Union[ctypes.c_int, int], buf: bytes, buf_size: int, / + model: llama_model_p, + i: Union[ctypes.c_int, int], + buf: Union[bytes, CtypesArray[ctypes.c_char]], + buf_size: int, + /, ) -> int: """Get metadata key name by index""" ... -llama_model_meta_key_by_index = _lib.llama_model_meta_key_by_index -llama_model_meta_key_by_index.argtypes = [ - llama_model_p_ctypes, - ctypes.c_int32, - ctypes.c_char_p, - ctypes.c_size_t, -] -llama_model_meta_key_by_index.restype = ctypes.c_int32 - - # // Get metadata value as a string by index # LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); + + +@ctypes_function( + "llama_model_meta_val_str_by_index", + [ + llama_model_p_ctypes, + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_model_meta_val_str_by_index( - model: llama_model_p, i: Union[ctypes.c_int, int], buf: bytes, buf_size: int, / + model: llama_model_p, + i: Union[ctypes.c_int, int], + buf: Union[bytes, CtypesArray[ctypes.c_char]], + buf_size: int, + /, ) -> int: """Get metadata value as a string by index""" ... -llama_model_meta_val_str_by_index = _lib.llama_model_meta_val_str_by_index -llama_model_meta_val_str_by_index.argtypes = [ - llama_model_p_ctypes, - ctypes.c_int32, - ctypes.c_char_p, - ctypes.c_size_t, -] -llama_model_meta_val_str_by_index.restype = ctypes.c_int32 - - # // Get a string describing the model type # LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); + + +@ctypes_function( + "llama_model_desc", + [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_size_t], + ctypes.c_int32, +) def llama_model_desc( - model: llama_model_p, buf: bytes, buf_size: Union[ctypes.c_size_t, int], / + model: llama_model_p, + buf: Union[bytes, CtypesArray[ctypes.c_char]], + buf_size: Union[ctypes.c_size_t, int], + /, ) -> int: """Get a string describing the model type""" ... -llama_model_desc = _lib.llama_model_desc -llama_model_desc.argtypes = [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_size_t] -llama_model_desc.restype = ctypes.c_int32 - - # // Returns the total size of all the tensors in the model in bytes # LLAMA_API uint64_t llama_model_size(const struct llama_model * model); + + +@ctypes_function("llama_model_size", [llama_model_p_ctypes], ctypes.c_uint64) def llama_model_size(model: llama_model_p, /) -> int: """Returns the total size of all the tensors in the model in bytes""" ... -llama_model_size = _lib.llama_model_size -llama_model_size.argtypes = [llama_model_p_ctypes] -llama_model_size.restype = ctypes.c_uint64 - - # // Returns the total number of parameters in the model # LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); + + +@ctypes_function("llama_model_n_params", [llama_model_p_ctypes], ctypes.c_uint64) def llama_model_n_params(model: llama_model_p, /) -> int: """Returns the total number of parameters in the model""" ... -llama_model_n_params = _lib.llama_model_n_params -llama_model_n_params.argtypes = [llama_model_p_ctypes] -llama_model_n_params.restype = ctypes.c_uint64 - - # // Get a llama model tensor # LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); + + +@ctypes_function( + "llama_get_model_tensor", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_void_p +) def llama_get_model_tensor( model: llama_model_p, name: Union[ctypes.c_char_p, bytes], / ) -> ctypes.c_void_p: @@ -1123,16 +1144,22 @@ def llama_get_model_tensor( ... -llama_get_model_tensor = _lib.llama_get_model_tensor -llama_get_model_tensor.argtypes = [llama_model_p_ctypes, ctypes.c_char_p] -llama_get_model_tensor.restype = ctypes.c_void_p - - # // Returns 0 on success # LLAMA_API uint32_t llama_model_quantize( # const char * fname_inp, # const char * fname_out, # const llama_model_quantize_params * params); + + +@ctypes_function( + "llama_model_quantize", + [ + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(llama_model_quantize_params), + ], + ctypes.c_uint32, +) def llama_model_quantize( fname_inp: bytes, fname_out: bytes, @@ -1143,15 +1170,6 @@ def llama_model_quantize( ... -llama_model_quantize = _lib.llama_model_quantize -llama_model_quantize.argtypes = [ - ctypes.c_char_p, - ctypes.c_char_p, - ctypes.POINTER(llama_model_quantize_params), -] -llama_model_quantize.restype = ctypes.c_uint32 - - # // Apply a LoRA adapter to a loaded model # // path_base_model is the path to a higher quality model to use as a base for # // the layers modified by the adapter. Can be NULL to use the current loaded model. @@ -1165,6 +1183,19 @@ def llama_model_quantize( # const char * path_base_model, # int32_t n_threads), # "use llama_model_apply_lora_from_file instead"); + + +@ctypes_function( + "llama_apply_lora_from_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + ctypes.c_float, + ctypes.c_char_p, + ctypes.c_int32, + ], + ctypes.c_int32, +) def llama_apply_lora_from_file( ctx: llama_context_p, path_lora: Union[ctypes.c_char_p, bytes], @@ -1182,23 +1213,25 @@ def llama_apply_lora_from_file( ... -llama_apply_lora_from_file = _lib.llama_apply_lora_from_file -llama_apply_lora_from_file.argtypes = [ - llama_context_p_ctypes, - ctypes.c_char_p, - ctypes.c_float, - ctypes.c_char_p, - ctypes.c_int32, -] -llama_apply_lora_from_file.restype = ctypes.c_int32 - - # LLAMA_API int32_t llama_model_apply_lora_from_file( # const struct llama_model * model, # const char * path_lora, # float scale, # const char * path_base_model, # int32_t n_threads); + + +@ctypes_function( + "llama_model_apply_lora_from_file", + [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_float, + ctypes.c_char_p, + ctypes.c_int32, + ], + ctypes.c_int32, +) def llama_model_apply_lora_from_file( model: llama_model_p, path_lora: Union[ctypes.c_char_p, bytes], @@ -1210,16 +1243,6 @@ def llama_model_apply_lora_from_file( ... -llama_model_apply_lora_from_file = _lib.llama_model_apply_lora_from_file -llama_model_apply_lora_from_file.argtypes = [ - llama_model_p_ctypes, - ctypes.c_char_p, - ctypes.c_float, - ctypes.c_char_p, - ctypes.c_int32, -] -llama_model_apply_lora_from_file.restype = ctypes.c_int32 - # // # // KV cache # // @@ -1285,6 +1308,13 @@ class llama_kv_cache_view(ctypes.Structure): # // Create an empty KV cache view. (use only for debugging purposes) # LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq); + + +@ctypes_function( + "llama_kv_cache_view_init", + [llama_context_p_ctypes, ctypes.c_int32], + llama_kv_cache_view, +) def llama_kv_cache_view_init( ctx: llama_context_p, n_max_seq: Union[ctypes.c_int32, int], / ) -> llama_kv_cache_view: @@ -1292,38 +1322,36 @@ def llama_kv_cache_view_init( ... -llama_kv_cache_view_init = _lib.llama_kv_cache_view_init -llama_kv_cache_view_init.argtypes = [llama_context_p_ctypes, ctypes.c_int32] -llama_kv_cache_view_init.restype = llama_kv_cache_view - - # // Free a KV cache view. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); + + +@ctypes_function("llama_kv_cache_view_free", [llama_kv_cache_view_p], None) def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore """Free a KV cache view. (use only for debugging purposes)""" ... -llama_kv_cache_view_free = _lib.llama_kv_cache_view_free -llama_kv_cache_view_free.argtypes = [llama_kv_cache_view_p] -llama_kv_cache_view_free.restype = None - - # // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); + + +@ctypes_function( + "llama_kv_cache_view_update", [llama_context_p_ctypes, llama_kv_cache_view_p], None +) def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[llama_kv_cache_view], /): # type: ignore """Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)""" ... -llama_kv_cache_view_update = _lib.llama_kv_cache_view_update -llama_kv_cache_view_update.argtypes = [llama_context_p_ctypes, llama_kv_cache_view_p] -llama_kv_cache_view_update.restype = None - - # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times # LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); + + +@ctypes_function( + "llama_get_kv_cache_token_count", [llama_context_p_ctypes], ctypes.c_int32 +) def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: """Returns the number of tokens in the KV cache (slow, use only for debug) If a KV cell has multiple sequences assigned to it, it will be counted multiple times @@ -1331,36 +1359,29 @@ def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: ... -llama_get_kv_cache_token_count = _lib.llama_get_kv_cache_token_count -llama_get_kv_cache_token_count.argtypes = [llama_context_p_ctypes] -llama_get_kv_cache_token_count.restype = ctypes.c_int32 - - # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) # LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); + + +@ctypes_function( + "llama_get_kv_cache_used_cells", [llama_context_p_ctypes], ctypes.c_int32 +) def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" ... -llama_get_kv_cache_used_cells = _lib.llama_get_kv_cache_used_cells -llama_get_kv_cache_used_cells.argtypes = [llama_context_p_ctypes] -llama_get_kv_cache_used_cells.restype = ctypes.c_int32 - - # // Clear the KV cache # LLAMA_API void llama_kv_cache_clear( # struct llama_context * ctx); + + +@ctypes_function("llama_kv_cache_clear", [llama_context_p_ctypes], None) def llama_kv_cache_clear(ctx: llama_context_p, /): """Clear the KV cache""" ... -llama_kv_cache_clear = _lib.llama_kv_cache_clear -llama_kv_cache_clear.argtypes = [llama_context_p_ctypes] -llama_kv_cache_clear.restype = None - - # // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) # // seq_id < 0 : match any sequence # // p0 < 0 : [0, p1] @@ -1370,6 +1391,18 @@ def llama_kv_cache_clear(ctx: llama_context_p, /): # llama_seq_id seq_id, # llama_pos p0, # llama_pos p1); + + +@ctypes_function( + "llama_kv_cache_seq_rm", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) def llama_kv_cache_seq_rm( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], @@ -1384,16 +1417,6 @@ def llama_kv_cache_seq_rm( ... -llama_kv_cache_seq_rm = _lib.llama_kv_cache_seq_rm -llama_kv_cache_seq_rm.argtypes = [ - llama_context_p_ctypes, - llama_seq_id, - llama_pos, - llama_pos, -] -llama_kv_cache_seq_rm.restype = None - - # // Copy all tokens that belong to the specified sequence to another sequence # // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence # // p0 < 0 : [0, p1] @@ -1404,6 +1427,19 @@ def llama_kv_cache_seq_rm( # llama_seq_id seq_id_dst, # llama_pos p0, # llama_pos p1); + + +@ctypes_function( + "llama_kv_cache_seq_cp", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) def llama_kv_cache_seq_cp( ctx: llama_context_p, seq_id_src: Union[llama_seq_id, int], @@ -1419,31 +1455,20 @@ def llama_kv_cache_seq_cp( ... -llama_kv_cache_seq_cp = _lib.llama_kv_cache_seq_cp -llama_kv_cache_seq_cp.argtypes = [ - llama_context_p_ctypes, - llama_seq_id, - llama_seq_id, - llama_pos, - llama_pos, -] -llama_kv_cache_seq_cp.restype = None - - # // Removes all tokens that do not belong to the specified sequence # LLAMA_API void llama_kv_cache_seq_keep( # struct llama_context * ctx, # llama_seq_id seq_id); + + +@ctypes_function( + "llama_kv_cache_seq_keep", [llama_context_p_ctypes, llama_seq_id], None +) def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): """Removes all tokens that do not belong to the specified sequence""" ... -llama_kv_cache_seq_keep = _lib.llama_kv_cache_seq_keep -llama_kv_cache_seq_keep.argtypes = [llama_context_p_ctypes, llama_seq_id] -llama_kv_cache_seq_keep.restype = None - - # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) # // If the KV cache is RoPEd, the KV data is updated accordingly # // p0 < 0 : [0, p1] @@ -1454,6 +1479,19 @@ def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, in # llama_pos p0, # llama_pos p1, # llama_pos delta); + + +@ctypes_function( + "llama_kv_cache_seq_shift", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + llama_pos, + ], + None, +) def llama_kv_cache_seq_shift( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], @@ -1469,17 +1507,6 @@ def llama_kv_cache_seq_shift( ... -llama_kv_cache_seq_shift = _lib.llama_kv_cache_seq_shift -llama_kv_cache_seq_shift.argtypes = [ - llama_context_p_ctypes, - llama_seq_id, - llama_pos, - llama_pos, - llama_pos, -] -llama_kv_cache_seq_shift.restype = None - - # // Integer division of the positions by factor of `d > 1` # // If the KV cache is RoPEd, the KV data is updated accordingly # // p0 < 0 : [0, p1] @@ -1490,6 +1517,19 @@ def llama_kv_cache_seq_shift( # llama_pos p0, # llama_pos p1, # int d); + + +@ctypes_function( + "llama_kv_cache_seq_div", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ctypes.c_int, + ], + None, +) def llama_kv_cache_seq_div( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], @@ -1505,16 +1545,6 @@ def llama_kv_cache_seq_div( ... -llama_kv_cache_seq_div = _lib.llama_kv_cache_seq_div -llama_kv_cache_seq_div.argtypes = [ - llama_context_p_ctypes, - llama_seq_id, - llama_pos, - llama_pos, - ctypes.c_int, -] -llama_kv_cache_seq_div.restype = None - # // # // State / sessions # // @@ -1523,23 +1553,31 @@ def llama_kv_cache_seq_div( # Returns the maximum size in bytes of the state (rng, logits, embedding # and kv_cache) - will often be smaller after compacting tokens # LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); + + +@ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_get_state_size(ctx: llama_context_p, /) -> int: """Returns the maximum size in bytes of the state (rng, logits, embedding and kv_cache) - will often be smaller after compacting tokens""" ... -llama_get_state_size = _lib.llama_get_state_size -llama_get_state_size.argtypes = [llama_context_p_ctypes] -llama_get_state_size.restype = ctypes.c_size_t - - # Copies the state to the specified destination address. # Destination needs to have allocated enough memory. # Returns the number of bytes copied # LLAMA_API size_t llama_copy_state_data( # struct llama_context * ctx, # uint8_t * dst); + + +@ctypes_function( + "llama_copy_state_data", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ], + ctypes.c_size_t, +) def llama_copy_state_data( ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], / ) -> int: @@ -1549,19 +1587,18 @@ def llama_copy_state_data( ... -llama_copy_state_data = _lib.llama_copy_state_data -llama_copy_state_data.argtypes = [ - llama_context_p_ctypes, - ctypes.POINTER(ctypes.c_uint8), -] -llama_copy_state_data.restype = ctypes.c_size_t - - # Set the state reading from the specified address # Returns the number of bytes read # LLAMA_API size_t llama_set_state_data( # struct llama_context * ctx, # uint8_t * src); + + +@ctypes_function( + "llama_set_state_data", + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], + ctypes.c_size_t, +) def llama_set_state_data( ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], / ) -> int: @@ -1569,11 +1606,6 @@ def llama_set_state_data( ... -llama_set_state_data = _lib.llama_set_state_data -llama_set_state_data.argtypes = [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)] -llama_set_state_data.restype = ctypes.c_size_t - - # Save/load session file # LLAMA_API bool llama_load_session_file( # struct llama_context * ctx, @@ -1581,6 +1613,19 @@ def llama_set_state_data( # llama_token * tokens_out, # size_t n_token_capacity, # size_t * n_token_count_out); + + +@ctypes_function( + "llama_load_session_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ctypes.POINTER(ctypes.c_size_t), + ], + ctypes.c_size_t, +) def llama_load_session_file( ctx: llama_context_p, path_session: bytes, @@ -1592,22 +1637,23 @@ def llama_load_session_file( ... -llama_load_session_file = _lib.llama_load_session_file -llama_load_session_file.argtypes = [ - llama_context_p_ctypes, - ctypes.c_char_p, - llama_token_p, - ctypes.c_size_t, - ctypes.POINTER(ctypes.c_size_t), -] -llama_load_session_file.restype = ctypes.c_size_t - - # LLAMA_API bool llama_save_session_file( # struct llama_context * ctx, # const char * path_session, # const llama_token * tokens, # size_t n_token_count); + + +@ctypes_function( + "llama_save_session_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ], + ctypes.c_size_t, +) def llama_save_session_file( ctx: llama_context_p, path_session: bytes, @@ -1618,15 +1664,6 @@ def llama_save_session_file( ... -llama_save_session_file = _lib.llama_save_session_file -llama_save_session_file.argtypes = [ - llama_context_p_ctypes, - ctypes.c_char_p, - llama_token_p, - ctypes.c_size_t, -] -llama_save_session_file.restype = ctypes.c_size_t - # // # // Decoding # // @@ -1643,6 +1680,18 @@ def llama_save_session_file( # int32_t n_tokens, # int32_t n_past), # "use llama_decode() instead"); + + +@ctypes_function( + "llama_eval", + [ + llama_context_p_ctypes, + llama_token_p, + ctypes.c_int32, + ctypes.c_int32, + ], + ctypes.c_int, +) def llama_eval( ctx: llama_context_p, tokens: CtypesArray[llama_token], @@ -1658,16 +1707,6 @@ def llama_eval( ... -llama_eval = _lib.llama_eval -llama_eval.argtypes = [ - llama_context_p_ctypes, - llama_token_p, - ctypes.c_int32, - ctypes.c_int32, -] -llama_eval.restype = ctypes.c_int - - # // Same as llama_eval, but use float matrix input directly. # // DEPRECATED: use llama_decode() instead # LLAMA_API DEPRECATED(int llama_eval_embd( @@ -1676,6 +1715,18 @@ def llama_eval( # int32_t n_tokens, # int32_t n_past), # "use llama_decode() instead"); + + +@ctypes_function( + "llama_eval_embd", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_float), + ctypes.c_int32, + ctypes.c_int32, + ], + ctypes.c_int, +) def llama_eval_embd( ctx: llama_context_p, embd: CtypesArray[ctypes.c_float], @@ -1688,16 +1739,6 @@ def llama_eval_embd( ... -llama_eval_embd = _lib.llama_eval_embd -llama_eval_embd.argtypes = [ - llama_context_p_ctypes, - ctypes.POINTER(ctypes.c_float), - ctypes.c_int32, - ctypes.c_int32, -] -llama_eval_embd.restype = ctypes.c_int - - # // Return batch for single sequence of tokens starting at pos_0 # // # // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it @@ -1707,6 +1748,18 @@ def llama_eval_embd( # int32_t n_tokens, # llama_pos pos_0, # llama_seq_id seq_id); + + +@ctypes_function( + "llama_batch_get_one", + [ + llama_token_p, + ctypes.c_int, + llama_pos, + llama_seq_id, + ], + llama_batch, +) def llama_batch_get_one( tokens: CtypesArray[llama_token], n_tokens: Union[ctypes.c_int, int], @@ -1721,16 +1774,6 @@ def llama_batch_get_one( ... -llama_batch_get_one = _lib.llama_batch_get_one -llama_batch_get_one.argtypes = [ - llama_token_p, - ctypes.c_int, - llama_pos, - llama_seq_id, -] -llama_batch_get_one.restype = llama_batch - - # // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens # // Each token can be assigned up to n_seq_max sequence ids # // The batch has to be freed with llama_batch_free() @@ -1742,6 +1785,11 @@ def llama_batch_get_one( # int32_t n_tokens, # int32_t embd, # int32_t n_seq_max); + + +@ctypes_function( + "llama_batch_init", [ctypes.c_int32, ctypes.c_int32, ctypes.c_int32], llama_batch +) def llama_batch_init( n_tokens: Union[ctypes.c_int32, int], embd: Union[ctypes.c_int32, int], @@ -1758,23 +1806,16 @@ def llama_batch_init( ... -llama_batch_init = _lib.llama_batch_init -llama_batch_init.argtypes = [ctypes.c_int32, ctypes.c_int32, ctypes.c_int32] -llama_batch_init.restype = llama_batch - - # // Frees a batch of tokens allocated with llama_batch_init() # LLAMA_API void llama_batch_free(struct llama_batch batch); + + +@ctypes_function("llama_batch_free", [llama_batch], None) def llama_batch_free(batch: llama_batch, /): """Frees a batch of tokens allocated with llama_batch_init()""" ... -llama_batch_free = _lib.llama_batch_free -llama_batch_free.argtypes = [llama_batch] -llama_batch_free.restype = None - - # // Positive return values does not mean a fatal error, but rather a warning. # // 0 - success # // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) @@ -1782,6 +1823,9 @@ def llama_batch_free(batch: llama_batch, /): # LLAMA_API int32_t llama_decode( # struct llama_context * ctx, # struct llama_batch batch); + + +@ctypes_function("llama_decode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: """Positive return values does not mean a fatal error, but rather a warning. 0 - success @@ -1790,15 +1834,21 @@ def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: ... -llama_decode = _lib.llama_decode -llama_decode.argtypes = [llama_context_p_ctypes, llama_batch] -llama_decode.restype = ctypes.c_int32 - - # // Set the number of threads used for decoding # // n_threads is the number of threads used for generation (single token) # // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) # LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); + + +@ctypes_function( + "llama_set_n_threads", + [ + llama_context_p_ctypes, + ctypes.c_uint32, + ctypes.c_uint32, + ], + None, +) def llama_set_n_threads( ctx: llama_context_p, n_threads: Union[ctypes.c_uint32, int], @@ -1812,24 +1862,18 @@ def llama_set_n_threads( ... -llama_set_n_threads = _lib.llama_set_n_threads -llama_set_n_threads.argtypes = [ - llama_context_p_ctypes, - ctypes.c_uint32, - ctypes.c_uint32, -] -llama_set_n_threads.restype = None - - # // Token logits obtained from the last call to llama_eval() # // The logits for the last token are stored in the last row # // Logits for which llama_batch.logits[i] == 0 are undefined # // Rows: n_tokens provided with llama_batch # // Cols: n_vocab # LLAMA_API float * llama_get_logits(struct llama_context * ctx); -def llama_get_logits( - ctx: llama_context_p, / -): # type: (...) -> Array[float] # type: ignore + + +@ctypes_function( + "llama_get_logits", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) +) +def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: """Token logits obtained from the last call to llama_eval() The logits for the last token are stored in the last row Logits for which llama_batch.logits[i] == 0 are undefined @@ -1838,202 +1882,181 @@ def llama_get_logits( ... -llama_get_logits = _lib.llama_get_logits -llama_get_logits.argtypes = [llama_context_p_ctypes] -llama_get_logits.restype = ctypes.POINTER(ctypes.c_float) - - # // Logits for the ith token. Equivalent to: # // llama_get_logits(ctx) + i*n_vocab # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); + + +@ctypes_function( + "llama_get_logits_ith", + [llama_context_p_ctypes, ctypes.c_int32], + ctypes.POINTER(ctypes.c_float), +) def llama_get_logits_ith( ctx: llama_context_p, i: Union[ctypes.c_int32, int], / -): # type: (...) -> Array[float] # type: ignore +) -> CtypesArray[ctypes.c_float]: """Logits for the ith token. Equivalent to: llama_get_logits(ctx) + i*n_vocab""" ... -llama_get_logits_ith = _lib.llama_get_logits_ith -llama_get_logits_ith.argtypes = [llama_context_p_ctypes, ctypes.c_int32] -llama_get_logits_ith.restype = ctypes.POINTER(ctypes.c_float) - - # Get the embeddings for the input # shape: [n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); -def llama_get_embeddings( - ctx: llama_context_p, / -): # type: (...) -> Array[float] # type: ignore + + +@ctypes_function( + "llama_get_embeddings", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) +) +def llama_get_embeddings(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: """Get the embeddings for the input shape: [n_embd] (1-dimensional)""" ... -llama_get_embeddings = _lib.llama_get_embeddings -llama_get_embeddings.argtypes = [llama_context_p_ctypes] -llama_get_embeddings.restype = ctypes.POINTER(ctypes.c_float) - - # // Get the embeddings for the ith sequence # // llama_get_embeddings(ctx) + i*n_embd # LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); + + +@ctypes_function( + "llama_get_embeddings_ith", + [llama_context_p_ctypes, ctypes.c_int32], + ctypes.POINTER(ctypes.c_float), +) def llama_get_embeddings_ith( ctx: llama_context_p, i: Union[ctypes.c_int32, int], / -): # type: (...) -> Array[float] # type: ignore +) -> CtypesArray[ctypes.c_float]: """Get the embeddings for the ith sequence llama_get_embeddings(ctx) + i*n_embd""" ... -llama_get_embeddings_ith = _lib.llama_get_embeddings_ith -llama_get_embeddings_ith.argtypes = [llama_context_p_ctypes, ctypes.c_int32] -llama_get_embeddings_ith.restype = ctypes.POINTER(ctypes.c_float) - - # // # // Vocab # // # LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); + + +@ctypes_function( + "llama_token_get_text", [llama_model_p_ctypes, llama_token], ctypes.c_char_p +) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / ) -> bytes: ... -llama_token_get_text = _lib.llama_token_get_text -llama_token_get_text.argtypes = [llama_model_p_ctypes, llama_token] -llama_token_get_text.restype = ctypes.c_char_p +# LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); -# LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); +@ctypes_function( + "llama_token_get_score", [llama_model_p_ctypes, llama_token], ctypes.c_float +) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / ) -> float: ... -llama_token_get_score = _lib.llama_token_get_score -llama_token_get_score.argtypes = [llama_model_p_ctypes, llama_token] -llama_token_get_score.restype = ctypes.c_float +# LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); -# LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); +@ctypes_function( + "llama_token_get_type", [llama_model_p_ctypes, llama_token], ctypes.c_int +) def llama_token_get_type( model: llama_model_p, token: Union[llama_token, int], / ) -> int: ... -llama_token_get_type = _lib.llama_token_get_type -llama_token_get_type.argtypes = [llama_model_p_ctypes, llama_token] -llama_token_get_type.restype = ctypes.c_int - - # // Special tokens # LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence + + +@ctypes_function("llama_token_bos", [llama_model_p_ctypes], llama_token) def llama_token_bos(model: llama_model_p, /) -> int: """beginning-of-sentence""" ... -llama_token_bos = _lib.llama_token_bos -llama_token_bos.argtypes = [llama_model_p_ctypes] -llama_token_bos.restype = llama_token +# LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence -# LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence +@ctypes_function("llama_token_eos", [llama_model_p_ctypes], llama_token) def llama_token_eos(model: llama_model_p, /) -> int: """end-of-sentence""" ... -llama_token_eos = _lib.llama_token_eos -llama_token_eos.argtypes = [llama_model_p_ctypes] -llama_token_eos.restype = llama_token +# LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line -# LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line +@ctypes_function("llama_token_nl", [llama_model_p_ctypes], llama_token) def llama_token_nl(model: llama_model_p, /) -> int: """next-line""" ... -llama_token_nl = _lib.llama_token_nl -llama_token_nl.argtypes = [llama_model_p_ctypes] -llama_token_nl.restype = llama_token - - # // Returns -1 if unknown, 1 for true or 0 for false. # LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); + + +@ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_int32) def llama_add_bos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" ... -llama_add_bos_token = _lib.llama_add_bos_token -llama_add_bos_token.argtypes = [llama_model_p_ctypes] -llama_add_bos_token.restype = ctypes.c_int32 - - # // Returns -1 if unknown, 1 for true or 0 for false. # LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); + + +@ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_int32) def llama_add_eos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" ... -llama_add_eos_token = _lib.llama_add_eos_token -llama_add_eos_token.argtypes = [llama_model_p_ctypes] -llama_add_eos_token.restype = ctypes.c_int32 - - # // codellama infill tokens # LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix + + +@ctypes_function("llama_token_prefix", [llama_model_p_ctypes], llama_token) def llama_token_prefix(model: llama_model_p) -> int: """codellama infill tokens""" ... -llama_token_prefix = _lib.llama_token_prefix -llama_token_prefix.argtypes = [llama_model_p_ctypes] -llama_token_prefix.restype = llama_token +# LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle -# LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle +@ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) def llama_token_middle(model: llama_model_p, /) -> int: ... -llama_token_middle = _lib.llama_token_middle -llama_token_middle.argtypes = [llama_model_p_ctypes] -llama_token_middle.restype = llama_token +# LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix -# LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix +@ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) def llama_token_suffix(model: llama_model_p, /) -> int: ... -llama_token_suffix = _lib.llama_token_suffix -llama_token_suffix.argtypes = [llama_model_p_ctypes] -llama_token_suffix.restype = llama_token +# LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle -# LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle +@ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) def llama_token_eot(model: llama_model_p, /) -> int: ... -llama_token_eot = _lib.llama_token_eot -llama_token_eot.argtypes = [llama_model_p_ctypes] -llama_token_eot.restype = llama_token - - # // # // Tokenization # // @@ -2053,6 +2076,21 @@ def llama_token_eot(model: llama_model_p, /) -> int: # int32_t n_max_tokens, # bool add_bos, # bool special); + + +@ctypes_function( + "llama_tokenize", + [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_int32, + llama_token_p, + ctypes.c_int32, + ctypes.c_bool, + ctypes.c_bool, + ], + ctypes.c_int32, +) def llama_tokenize( model: llama_model_p, text: bytes, @@ -2067,19 +2105,6 @@ def llama_tokenize( ... -llama_tokenize = _lib.llama_tokenize -llama_tokenize.argtypes = [ - llama_model_p_ctypes, - ctypes.c_char_p, - ctypes.c_int32, - llama_token_p, - ctypes.c_int32, - ctypes.c_bool, - ctypes.c_bool, -] -llama_tokenize.restype = ctypes.c_int32 - - # // Token Id -> Piece. # // Uses the vocabulary in the provided context. # // Does not write null terminator to the buffer. @@ -2089,10 +2114,22 @@ def llama_tokenize( # llama_token token, # char * buf, # int32_t length); + + +@ctypes_function( + "llama_token_to_piece", + [ + llama_model_p_ctypes, + llama_token, + ctypes.c_char_p, + ctypes.c_int32, + ], + ctypes.c_int32, +) def llama_token_to_piece( model: llama_model_p, token: Union[llama_token, int], - buf: Union[ctypes.c_char_p, bytes], + buf: Union[ctypes.c_char_p, bytes, CtypesArray[ctypes.c_char]], length: Union[ctypes.c_int, int], /, ) -> int: @@ -2104,16 +2141,6 @@ def llama_token_to_piece( ... -llama_token_to_piece = _lib.llama_token_to_piece -llama_token_to_piece.argtypes = [ - llama_model_p_ctypes, - llama_token, - ctypes.c_char_p, - ctypes.c_int32, -] -llama_token_to_piece.restype = ctypes.c_int32 - - # /// Apply chat template. Inspired by hf apply_chat_template() on python. # /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" # /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template @@ -2132,6 +2159,18 @@ def llama_token_to_piece( # bool add_ass, # char * buf, # int32_t length); + + +@ctypes_function( + "llama_chat_apply_template", + [ + ctypes.c_void_p, + ctypes.c_char_p, + ctypes.POINTER(llama_chat_message), + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_chat_apply_template( model: llama_model_p, tmpl: bytes, @@ -2142,16 +2181,6 @@ def llama_chat_apply_template( ... -llama_chat_apply_template = _lib.llama_chat_apply_template -llama_chat_apply_template.argtypes = [ - ctypes.c_void_p, - ctypes.c_char_p, - ctypes.POINTER(llama_chat_message), - ctypes.c_size_t, -] -llama_chat_apply_template.restype = ctypes.c_int32 - - # // # // Grammar # // @@ -2161,6 +2190,17 @@ def llama_chat_apply_template( # const llama_grammar_element ** rules, # size_t n_rules, # size_t start_rule_index); + + +@ctypes_function( + "llama_grammar_init", + [ + ctypes.POINTER(llama_grammar_element_p), + ctypes.c_size_t, + ctypes.c_size_t, + ], + llama_grammar_p, +) def llama_grammar_init( rules: CtypesArray[ CtypesPointer[llama_grammar_element] @@ -2173,36 +2213,28 @@ def llama_grammar_init( ... -llama_grammar_init = _lib.llama_grammar_init -llama_grammar_init.argtypes = [ - ctypes.POINTER(llama_grammar_element_p), - ctypes.c_size_t, - ctypes.c_size_t, -] -llama_grammar_init.restype = llama_grammar_p - - # LLAMA_API void llama_grammar_free(struct llama_grammar * grammar); +@ctypes_function( + "llama_grammar_free", + [llama_grammar_p], + None, +) def llama_grammar_free(grammar: llama_grammar_p, /): """Free a grammar.""" ... -llama_grammar_free = _lib.llama_grammar_free -llama_grammar_free.argtypes = [llama_grammar_p] -llama_grammar_free.restype = None - - # LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar); +@ctypes_function( + "llama_grammar_copy", + [llama_grammar_p], + llama_grammar_p, +) def llama_grammar_copy(grammar: llama_grammar_p, /) -> llama_grammar_p: """Copy a grammar.""" ... -llama_grammar_copy = _lib.llama_grammar_copy -llama_grammar_copy.argtypes = [llama_grammar_p] -llama_grammar_copy.restype = llama_grammar_p - # // # // Sampling functions # // @@ -2210,16 +2242,16 @@ def llama_grammar_copy(grammar: llama_grammar_p, /) -> llama_grammar_p: # // Sets the current rng seed. # LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); +@ctypes_function( + "llama_set_rng_seed", + [llama_context_p_ctypes, ctypes.c_uint32], + None, +) def llama_set_rng_seed(ctx: llama_context_p, seed: Union[ctypes.c_uint32, int], /): """Sets the current rng seed.""" ... -llama_set_rng_seed = _lib.llama_set_rng_seed -llama_set_rng_seed.argtypes = [llama_context_p_ctypes, ctypes.c_uint32] -llama_set_rng_seed.restype = None - - # /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. # /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. # LLAMA_API void llama_sample_repetition_penalties( @@ -2230,6 +2262,19 @@ def llama_set_rng_seed(ctx: llama_context_p, seed: Union[ctypes.c_uint32, int], # float penalty_repeat, # float penalty_freq, # float penalty_present); +@ctypes_function( + "llama_sample_repetition_penalties", + [ + llama_context_p_ctypes, + llama_token_data_array_p, + llama_token_p, + ctypes.c_size_t, + ctypes.c_float, + ctypes.c_float, + ctypes.c_float, + ], + None, +) def llama_sample_repetition_penalties( ctx: llama_context_p, candidates: Union[ @@ -2248,19 +2293,6 @@ def llama_sample_repetition_penalties( ... -llama_sample_repetition_penalties = _lib.llama_sample_repetition_penalties -llama_sample_repetition_penalties.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - llama_token_p, - ctypes.c_size_t, - ctypes.c_float, - ctypes.c_float, - ctypes.c_float, -] -llama_sample_repetition_penalties.restype = None - - # /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 # /// @param logits Logits extracted from the original generation context. # /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. @@ -2270,6 +2302,16 @@ def llama_sample_repetition_penalties( # float * logits, # float * logits_guidance, # float scale); +@ctypes_function( + "llama_sample_apply_guidance", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_float), + ctypes.POINTER(ctypes.c_float), + ctypes.c_float, + ], + None, +) def llama_sample_apply_guidance( ctx: llama_context_p, logits: CtypesArray[ctypes.c_float], @@ -2281,22 +2323,22 @@ def llama_sample_apply_guidance( ... -llama_sample_apply_guidance = _lib.llama_sample_apply_guidance -llama_sample_apply_guidance.argtypes = [ - llama_context_p_ctypes, - ctypes.POINTER(ctypes.c_float), - ctypes.POINTER(ctypes.c_float), - ctypes.c_float, -] -llama_sample_apply_guidance.restype = None - - # LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance( # struct llama_context * ctx, # llama_token_data_array * candidates, # struct llama_context * guidance_ctx, # float scale), # "use llama_sample_apply_guidance() instead"); +@ctypes_function( + "llama_sample_classifier_free_guidance", + [ + llama_context_p_ctypes, + llama_token_data_array_p, + llama_context_p_ctypes, + ctypes.c_float, + ], + None, +) def llama_sample_classifier_free_guidance( ctx: llama_context_p, candidates: Union[ @@ -2310,20 +2352,15 @@ def llama_sample_classifier_free_guidance( ... -llama_sample_classifier_free_guidance = _lib.llama_sample_classifier_free_guidance -llama_sample_classifier_free_guidance.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - llama_context_p_ctypes, - ctypes.c_float, -] -llama_sample_classifier_free_guidance.restype = None - - # /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. # LLAMA_API void llama_sample_softmax( # struct llama_context * ctx, # llama_token_data_array * candidates); +@ctypes_function( + "llama_sample_softmax", + [llama_context_p_ctypes, llama_token_data_array_p], + None, +) def llama_sample_softmax( ctx: llama_context_p, candidates: Union[ @@ -2335,20 +2372,17 @@ def llama_sample_softmax( ... -llama_sample_softmax = _lib.llama_sample_softmax -llama_sample_softmax.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, -] -llama_sample_softmax.restype = None - - # /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 # LLAMA_API void llama_sample_top_k( # struct llama_context * ctx, # llama_token_data_array * candidates, # int32_t k, # size_t min_keep); +@ctypes_function( + "llama_sample_top_k", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_int32, ctypes.c_size_t], + None, +) def llama_sample_top_k( ctx: llama_context_p, candidates: Union[ @@ -2362,22 +2396,17 @@ def llama_sample_top_k( ... -llama_sample_top_k = _lib.llama_sample_top_k -llama_sample_top_k.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_int32, - ctypes.c_size_t, -] -llama_sample_top_k.restype = None - - # /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 # LLAMA_API void llama_sample_top_p( # struct llama_context * ctx, # llama_token_data_array * candidates, # float p, # size_t min_keep); +@ctypes_function( + "llama_sample_top_p", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], + None, +) def llama_sample_top_p( ctx: llama_context_p, candidates: Union[ @@ -2391,22 +2420,17 @@ def llama_sample_top_p( ... -llama_sample_top_p = _lib.llama_sample_top_p -llama_sample_top_p.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_size_t, -] -llama_sample_top_p.restype = None - - # /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 # LLAMA_API void llama_sample_min_p( # struct llama_context * ctx, # llama_token_data_array * candidates, # float p, # size_t min_keep); +@ctypes_function( + "llama_sample_min_p", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], + None, +) def llama_sample_min_p( ctx: llama_context_p, candidates: Union[ @@ -2420,22 +2444,17 @@ def llama_sample_min_p( ... -llama_sample_min_p = _lib.llama_sample_min_p -llama_sample_min_p.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_size_t, -] -llama_sample_min_p.restype = None - - # /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. # LLAMA_API void llama_sample_tail_free( # struct llama_context * ctx, # llama_token_data_array * candidates, # float z, # size_t min_keep); +@ctypes_function( + "llama_sample_tail_free", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], + None, +) def llama_sample_tail_free( ctx: llama_context_p, candidates: Union[ @@ -2449,22 +2468,17 @@ def llama_sample_tail_free( ... -llama_sample_tail_free = _lib.llama_sample_tail_free -llama_sample_tail_free.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_size_t, -] -llama_sample_tail_free.restype = None - - # /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. # LLAMA_API void llama_sample_typical( # struct llama_context * ctx, # llama_token_data_array * candidates, # float p, # size_t min_keep); +@ctypes_function( + "llama_sample_typical", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], + None, +) def llama_sample_typical( ctx: llama_context_p, candidates: Union[ @@ -2478,16 +2492,6 @@ def llama_sample_typical( ... -llama_sample_typical = _lib.llama_sample_typical -llama_sample_typical.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_size_t, -] -llama_sample_typical.restype = None - - # /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772. # LLAMA_API void llama_sample_entropy( # struct llama_context * ctx, @@ -2495,6 +2499,17 @@ def llama_sample_typical( # float min_temp, # float max_temp, # float exponent_val); +@ctypes_function( + "llama_sample_entropy", + [ + llama_context_p_ctypes, + llama_token_data_array_p, + ctypes.c_float, + ctypes.c_float, + ctypes.c_float, + ], + None, +) def llama_sample_entropy( ctx: llama_context_p, candidates: Union[ @@ -2509,21 +2524,15 @@ def llama_sample_entropy( ... -llama_sample_entropy = _lib.llama_sample_entropy -llama_sample_entropy.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_float, - ctypes.c_float, -] -llama_sample_entropy.restype = None - - # LLAMA_API void llama_sample_temp( # struct llama_context * ctx, # llama_token_data_array * candidates, # float temp); +@ctypes_function( + "llama_sample_temp", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float], + None, +) def llama_sample_temp( ctx: llama_context_p, candidates: Union[ @@ -2541,20 +2550,16 @@ def llama_sample_temp( ... -llama_sample_temp = _lib.llama_sample_temp -llama_sample_temp.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, -] -llama_sample_temp.restype = None - - # LLAMA_API DEPRECATED(void llama_sample_temperature( # struct llama_context * ctx, # llama_token_data_array * candidates, # float temp), # "use llama_sample_temp instead"); +@ctypes_function( + "llama_sample_temperature", + [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float], + None, +) def llama_sample_temperature( ctx: llama_context_p, candidates: Union[ @@ -2567,20 +2572,16 @@ def llama_sample_temperature( ... -llama_sample_temperature = _lib.llama_sample_temperature -llama_sample_temperature.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, -] -llama_sample_temperature.restype = None - - # /// @details Apply constraints from grammar # LLAMA_API void llama_sample_grammar( # struct llama_context * ctx, # llama_token_data_array * candidates, # const struct llama_grammar * grammar); +@ctypes_function( + "llama_sample_grammar", + [llama_context_p_ctypes, llama_token_data_array_p, llama_grammar_p], + None, +) def llama_sample_grammar( ctx: llama_context_p, candidates: Union[ @@ -2598,15 +2599,6 @@ def llama_sample_grammar( ... -llama_sample_grammar = _lib.llama_sample_grammar -llama_sample_grammar.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - llama_grammar_p, -] -llama_sample_grammar.restype = None - - # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. # /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -2620,6 +2612,18 @@ def llama_sample_grammar( # float eta, # int32_t m, # float * mu); +@ctypes_function( + "llama_sample_token_mirostat", + [ + llama_context_p_ctypes, + llama_token_data_array_p, + ctypes.c_float, + ctypes.c_float, + ctypes.c_int32, + ctypes.POINTER(ctypes.c_float), + ], + llama_token, +) def llama_sample_token_mirostat( ctx: llama_context_p, candidates: Union[ @@ -2643,18 +2647,6 @@ def llama_sample_token_mirostat( ... -llama_sample_token_mirostat = _lib.llama_sample_token_mirostat -llama_sample_token_mirostat.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_float, - ctypes.c_int32, - ctypes.POINTER(ctypes.c_float), -] -llama_sample_token_mirostat.restype = llama_token - - # /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. # /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -2666,6 +2658,17 @@ def llama_sample_token_mirostat( # float tau, # float eta, # float * mu); +@ctypes_function( + "llama_sample_token_mirostat_v2", + [ + llama_context_p_ctypes, + llama_token_data_array_p, + ctypes.c_float, + ctypes.c_float, + ctypes.POINTER(ctypes.c_float), + ], + llama_token, +) def llama_sample_token_mirostat_v2( ctx: llama_context_p, candidates: Union[ @@ -2673,7 +2676,7 @@ def llama_sample_token_mirostat_v2( ], tau: Union[ctypes.c_float, float], eta: Union[ctypes.c_float, float], - mu, # type: _Pointer[ctypes.c_float] + mu: CtypesPointerOrRef[ctypes.c_float], /, ) -> int: """Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2687,22 +2690,16 @@ def llama_sample_token_mirostat_v2( ... -llama_sample_token_mirostat_v2 = _lib.llama_sample_token_mirostat_v2 -llama_sample_token_mirostat_v2.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_float, - ctypes.POINTER(ctypes.c_float), -] -llama_sample_token_mirostat_v2.restype = llama_token - - # /// @details Selects the token with the highest probability. # /// Does not compute the token probabilities. Use llama_sample_softmax() instead. # LLAMA_API llama_token llama_sample_token_greedy( # struct llama_context * ctx, # llama_token_data_array * candidates); +@ctypes_function( + "llama_sample_token_greedy", + [llama_context_p_ctypes, llama_token_data_array_p], + llama_token, +) def llama_sample_token_greedy( ctx: llama_context_p, candidates: Union[ @@ -2714,18 +2711,15 @@ def llama_sample_token_greedy( ... -llama_sample_token_greedy = _lib.llama_sample_token_greedy -llama_sample_token_greedy.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, -] -llama_sample_token_greedy.restype = llama_token - - # /// @details Randomly selects a token from the candidates based on their probabilities. # LLAMA_API llama_token llama_sample_token( # struct llama_context * ctx, # llama_token_data_array * candidates); +@ctypes_function( + "llama_sample_token", + [llama_context_p_ctypes, llama_token_data_array_p], + llama_token, +) def llama_sample_token( ctx: llama_context_p, candidates: Union[ @@ -2737,19 +2731,16 @@ def llama_sample_token( ... -llama_sample_token = _lib.llama_sample_token -llama_sample_token.argtypes = [ - llama_context_p_ctypes, - llama_token_data_array_p, -] -llama_sample_token.restype = llama_token - - # /// @details Accepts the sampled token into the grammar # LLAMA_API void llama_grammar_accept_token( # struct llama_context * ctx, # struct llama_grammar * grammar, # llama_token token); +@ctypes_function( + "llama_grammar_accept_token", + [llama_context_p_ctypes, llama_grammar_p, llama_token], + None, +) def llama_grammar_accept_token( ctx: llama_context_p, grammar: llama_grammar_p, token: Union[llama_token, int], / ) -> None: @@ -2757,15 +2748,6 @@ def llama_grammar_accept_token( ... -llama_grammar_accept_token = _lib.llama_grammar_accept_token -llama_grammar_accept_token.argtypes = [ - llama_context_p_ctypes, - llama_grammar_p, - llama_token, -] -llama_grammar_accept_token.restype = None - - # // # // Beam search # // @@ -2830,6 +2812,18 @@ class llama_beams_state(ctypes.Structure): # size_t n_beams, # int32_t n_past, # int32_t n_predict); +@ctypes_function( + "llama_beam_search", + [ + llama_context_p_ctypes, + llama_beam_search_callback_fn_t, + ctypes.c_void_p, + ctypes.c_size_t, + ctypes.c_int32, + ctypes.c_int32, + ], + None, +) def llama_beam_search( ctx: llama_context_p, callback: CtypesFuncPointer, @@ -2842,73 +2836,66 @@ def llama_beam_search( ... -llama_beam_search = _lib.llama_beam_search -llama_beam_search.argtypes = [ - llama_context_p_ctypes, - llama_beam_search_callback_fn_t, - ctypes.c_void_p, - ctypes.c_size_t, - ctypes.c_int32, - ctypes.c_int32, -] -llama_beam_search.restype = None - - # Performance information # LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); +@ctypes_function( + "llama_get_timings", + [llama_context_p_ctypes], + llama_timings, +) def llama_get_timings(ctx: llama_context_p, /) -> llama_timings: """Get performance information""" ... -llama_get_timings = _lib.llama_get_timings -llama_get_timings.argtypes = [llama_context_p_ctypes] -llama_get_timings.restype = llama_timings - - # LLAMA_API void llama_print_timings(struct llama_context * ctx); +@ctypes_function( + "llama_print_timings", + [llama_context_p_ctypes], + None, +) def llama_print_timings(ctx: llama_context_p, /): """Print performance information""" ... -llama_print_timings = _lib.llama_print_timings -llama_print_timings.argtypes = [llama_context_p_ctypes] -llama_print_timings.restype = None - - # LLAMA_API void llama_reset_timings(struct llama_context * ctx); +@ctypes_function( + "llama_reset_timings", + [llama_context_p_ctypes], + None, +) def llama_reset_timings(ctx: llama_context_p, /): """Reset performance information""" ... -llama_reset_timings = _lib.llama_reset_timings -llama_reset_timings.argtypes = [llama_context_p_ctypes] -llama_reset_timings.restype = None - - # Print system information # LLAMA_API const char * llama_print_system_info(void); +@ctypes_function( + "llama_print_system_info", + [], + ctypes.c_char_p, +) def llama_print_system_info() -> bytes: """Print system information""" ... -llama_print_system_info = _lib.llama_print_system_info -llama_print_system_info.argtypes = [] -llama_print_system_info.restype = ctypes.c_char_p - - # NOTE: THIS IS CURRENTLY BROKEN AS ggml_log_callback IS NOT EXPOSED IN LLAMA.H # // Set callback for all future logging events. # // If this is not called, or NULL is supplied, everything is output on stderr. # LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); +@ctypes_function( + "llama_log_set", + [ctypes.c_void_p, ctypes.c_void_p], + None, +) def llama_log_set( log_callback: Optional[CtypesFuncPointer], - user_data: ctypes.c_void_p, # type: ignore + user_data: ctypes.c_void_p, /, ): """Set callback for all future logging events. @@ -2917,16 +2904,11 @@ def llama_log_set( ... -llama_log_set = _lib.llama_log_set -llama_log_set.argtypes = [ctypes.c_void_p, ctypes.c_void_p] -llama_log_set.restype = None - - # LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); +@ctypes_function( + "llama_dump_timing_info_yaml", + [ctypes.c_void_p, llama_context_p_ctypes], + None, +) def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): ... - - -llama_dump_timing_info_yaml = _lib.llama_dump_timing_info_yaml -llama_dump_timing_info_yaml.argtypes = [ctypes.c_void_p, llama_context_p_ctypes] -llama_dump_timing_info_yaml.restype = None From b9aca612af63a102a54b6b5069ad600305081dd3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 03:40:07 -0500 Subject: [PATCH 176/624] misc: use typesafe byref for internal classes --- llama_cpp/_internals.py | 40 ++++++++++++++++++++-------------------- 1 file changed, 20 insertions(+), 20 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 8690843cd..98ad6d718 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -82,7 +82,7 @@ def rope_freq_scale_train(self) -> float: def desc(self) -> str: assert self.model is not None buf = ctypes.create_string_buffer(1024) - llama_cpp.llama_model_desc(self.model, buf, 1024) # type: ignore + llama_cpp.llama_model_desc(self.model, buf, 1024) return buf.value.decode("utf-8") def size(self) -> int: @@ -184,7 +184,7 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): def token_to_piece(self, token: int) -> bytes: assert self.model is not None buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32) # type: ignore + llama_cpp.llama_token_to_piece(self.model, token, buf, 32) return bytes(buf) def detokenize(self, tokens: List[int]) -> bytes: @@ -349,7 +349,7 @@ def sample_repetition_penalties( assert self.ctx is not None llama_cpp.llama_sample_repetition_penalties( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), last_tokens_data, penalty_last_n, penalty_repeat, @@ -367,7 +367,7 @@ def sample_classifier_free_guidance( assert guidance_ctx.ctx is not None llama_cpp.llama_sample_classifier_free_guidance( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), guidance_ctx.ctx, scale, ) @@ -376,25 +376,25 @@ def sample_softmax(self, candidates: "_LlamaTokenDataArray"): assert self.ctx is not None llama_cpp.llama_sample_softmax( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), ) def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): assert self.ctx is not None llama_cpp.llama_sample_top_k( - self.ctx, ctypes.byref(candidates.candidates), k, min_keep # type: ignore + self.ctx, llama_cpp.byref(candidates.candidates), k, min_keep ) def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): assert self.ctx is not None llama_cpp.llama_sample_top_p( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore + self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): assert self.ctx is not None llama_cpp.llama_sample_min_p( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore + self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) def sample_tail_free( @@ -402,7 +402,7 @@ def sample_tail_free( ): assert self.ctx is not None llama_cpp.llama_sample_tail_free( - self.ctx, ctypes.byref(candidates.candidates), z, min_keep # type: ignore + self.ctx, llama_cpp.byref(candidates.candidates), z, min_keep ) def sample_typical( @@ -410,13 +410,13 @@ def sample_typical( ): assert self.ctx is not None llama_cpp.llama_sample_typical( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore + self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): assert self.ctx is not None llama_cpp.llama_sample_temp( - self.ctx, ctypes.byref(candidates.candidates), temp # type: ignore + self.ctx, llama_cpp.byref(candidates.candidates), temp ) def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): @@ -424,7 +424,7 @@ def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGramm assert grammar.grammar is not None llama_cpp.llama_sample_grammar( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), grammar.grammar, ) @@ -434,12 +434,12 @@ def sample_token_mirostat( tau: float, eta: float, m: int, - mu: ctypes._Pointer[ctypes.c_float], # type: ignore + mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: assert self.ctx is not None return llama_cpp.llama_sample_token_mirostat( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), tau, eta, m, @@ -447,12 +447,12 @@ def sample_token_mirostat( ) def sample_token_mirostat_v2( - self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: ctypes._Pointer[ctypes.c_float] # type: ignore + self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float] ) -> int: assert self.ctx is not None return llama_cpp.llama_sample_token_mirostat_v2( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), tau, eta, mu, @@ -462,14 +462,14 @@ def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: assert self.ctx is not None return llama_cpp.llama_sample_token_greedy( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), ) def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: assert self.ctx is not None return llama_cpp.llama_sample_token( self.ctx, - ctypes.byref(candidates.candidates), # type: ignore + llama_cpp.byref(candidates.candidates), ) # Grammar @@ -566,7 +566,7 @@ def __init__(self, *, n_vocab: int): size=self.n_vocab, sorted=False, ) - self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) + self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) # type: ignore self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) def copy_logits(self, logits: npt.NDArray[np.single]): @@ -754,7 +754,7 @@ def sample( ctx_main.sample_repetition_penalties( token_data_array, # TODO: Only create this once - (llama_cpp.llama_token * len(self.prev))(*self.prev), # type: ignore + (llama_cpp.llama_token * len(self.prev))(*self.prev), self.params.penalty_last_n, self.params.penalty_repeat, self.params.penalty_freq, From 5f96621e92a0a5408ab479aaad1ccffc0a10d4d1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 03:40:25 -0500 Subject: [PATCH 177/624] misc: only search tests folder for tests --- pyproject.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/pyproject.toml b/pyproject.toml index 413097201..2f3d3ced0 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -72,4 +72,4 @@ Documentation = "https://llama-cpp-python.readthedocs.io/en/latest/" Changelog = "https://llama-cpp-python.readthedocs.io/en/latest/changelog/" [tool.pytest.ini_options] -addopts = "--ignore=vendor" +testpaths = "tests" From eebb102df76299751c722a54eb6f4964de4e6650 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 03:42:08 -0500 Subject: [PATCH 178/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 973053d8b..15499eb94 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 973053d8b0d04809836b3339a50f68d9c842de90 +Subproject commit 15499eb94227401bdc8875da6eb85c15d37068f7 From 251a8a2cadb4c0df4671062144d168a7874086a2 Mon Sep 17 00:00:00 2001 From: Alvaro Bartolome Date: Fri, 23 Feb 2024 18:40:52 +0900 Subject: [PATCH 179/624] feat: Add Google's Gemma formatting via `chat_format="gemma"` (#1210) * Add Google's Gemma formatting via `chat_format="gemma"` * Replace `raise ValueError` with `logger.debug` Co-authored-by: Andrei --------- Co-authored-by: Andrei --- llama_cpp/llama_chat_format.py | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 8dd0ddfd2..16bccb942 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -14,6 +14,7 @@ import llama_cpp.llama_types as llama_types import llama_cpp.llama_grammar as llama_grammar +from ._logger import logger from ._utils import suppress_stdout_stderr, Singleton ### Common Chat Templates and Special Tokens ### @@ -993,6 +994,26 @@ def format_saiga( return ChatFormatterResponse(prompt=_prompt.strip()) +# Chat format for Google's Gemma models, see more details and available models: +# https://huggingface.co/collections/google/gemma-release-65d5efbccdbb8c4202ec078b +@register_chat_format("gemma") +def format_gemma( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + system_message = _get_system_message(messages) + if system_message is not None and system_message != "": + logger.debug( + "`role='system'` messages are not allowed on Google's Gemma models." + ) + _roles = dict(user="user\n", assistant="model\n") + _sep = "\n" + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_no_colon_single(system_message="", messages=_messages, sep=_sep) + return ChatFormatterResponse(prompt=_prompt, stop=_sep) + + # Tricky chat formats that require custom chat handlers From 52d9d70076e28cc93a8a6a63bbf8285bc7ed8cb6 Mon Sep 17 00:00:00 2001 From: Aditya Purandare Date: Fri, 23 Feb 2024 15:11:22 +0530 Subject: [PATCH 180/624] docs: Update README.md to fix pip install llama cpp server (#1187) Without the single quotes, when running the command, an error is printed saying no matching packages found on pypi. Adding the quotes fixes it ```bash $ pip install llama-cpp-python[server] zsh: no matches found: llama-cpp-python[server] ``` Co-authored-by: Andrei --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 6dbd79dd4..ccd66ea82 100644 --- a/README.md +++ b/README.md @@ -505,14 +505,14 @@ This allows you to use llama.cpp compatible models with any OpenAI compatible cl To install the server package and get started: ```bash -pip install llama-cpp-python[server] +pip install 'llama-cpp-python[server]' python3 -m llama_cpp.server --model models/7B/llama-model.gguf ``` Similar to Hardware Acceleration section above, you can also install with GPU (cuBLAS) support like this: ```bash -CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python[server] +CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install 'llama-cpp-python[server]' python3 -m llama_cpp.server --model models/7B/llama-model.gguf --n_gpu_layers 35 ``` From 427d816ebf01c9b2212852e77ce8b9498f176fb1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 04:54:08 -0500 Subject: [PATCH 181/624] chore: Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index aa6b4b4ea..c539ade24 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.48] + +- feat: Update llama.cpp to ggerganov/llama.cpp@15499eb94227401bdc8875da6eb85c15d37068f7 +- feat: Add Google's Gemma formatting via chat_format="gemma" by @alvarobartt in #1210 +- feat: support minItems/maxItems in JSON grammar converter by @nopperl in 3921e10770996d95a9eb22c8248bacef39f69365 +- fix: Update from_pretrained defaults to match hf_hub_download and pull to local cache folder by @abetlen in e6d6260a91b7831733f7d1f73c7af46a3e8185ed +- fix: Raise exceptions when llama model or context fails to load by @abetlen in dd22010e85265ae840c76ec835d67a29ed852722 +- docs: Update README.md to fix pip install llama cpp server by @audip in #1187 + ## [0.2.47] - feat: Update llama.cpp to ggerganov/llama.cpp@973053d8b0d04809836b3339a50f68d9c842de90 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index bd701e383..8c4f4002e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.47" \ No newline at end of file +__version__ = "0.2.48" \ No newline at end of file From db776a885cd4c20811f22f8bd1a27ecc71dba927 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 11:24:53 -0500 Subject: [PATCH 182/624] fix: module 'llama_cpp.llama_cpp' has no attribute 'c_uint8' --- llama_cpp/llama.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 122654514..4b3d36a16 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -5,8 +5,10 @@ import uuid import time import json +import ctypes import fnmatch import multiprocessing + from typing import ( List, Optional, @@ -20,7 +22,6 @@ from collections import deque from pathlib import Path -import ctypes from llama_cpp.llama_types import List @@ -1789,7 +1790,7 @@ def save_state(self) -> LlamaState: state_size = llama_cpp.llama_get_state_size(self._ctx.ctx) if self.verbose: print(f"Llama.save_state: got state size: {state_size}", file=sys.stderr) - llama_state = (llama_cpp.c_uint8 * int(state_size))() + llama_state = (ctypes.c_uint8 * int(state_size))() if self.verbose: print("Llama.save_state: allocated state", file=sys.stderr) n_bytes = llama_cpp.llama_copy_state_data(self._ctx.ctx, llama_state) @@ -1797,7 +1798,7 @@ def save_state(self) -> LlamaState: print(f"Llama.save_state: copied llama state: {n_bytes}", file=sys.stderr) if int(n_bytes) > int(state_size): raise RuntimeError("Failed to copy llama state data") - llama_state_compact = (llama_cpp.c_uint8 * int(n_bytes))() + llama_state_compact = (ctypes.c_uint8 * int(n_bytes))() llama_cpp.ctypes.memmove(llama_state_compact, llama_state, int(n_bytes)) if self.verbose: print( From 858496224ea6eb4e22f681dc60a9afe064c33f1d Mon Sep 17 00:00:00 2001 From: Luke Stanley <306671+lukestanley@users.noreply.github.com> Date: Fri, 23 Feb 2024 16:27:38 +0000 Subject: [PATCH 183/624] feat: Auto detect Mixtral's slightly different format (#1214) --- llama_cpp/llama_chat_format.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 16bccb942..eec06b899 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -29,6 +29,8 @@ MISTRAL_INSTRUCT_BOS_TOKEN = "" MISTRAL_INSTRUCT_EOS_TOKEN = "" +# Source: https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1/blob/main/tokenizer_config.json +MIXTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" ### Chat Completion Handler ### @@ -470,7 +472,8 @@ def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[s if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE: return "chatml" - if metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE: + if (metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE or + metadata["tokenizer.chat_template"] == MIXTRAL_INSTRUCT_CHAT_TEMPLATE): return "mistral-instruct" return None From ded5d627a5bf1a7e130135b1624d092ad3997edc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 11:32:43 -0500 Subject: [PATCH 184/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index c539ade24..c1ba40fa9 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.49] + +- fix: module 'llama_cpp.llama_cpp' has no attribute 'c_uint8' in Llama.save_state by @abetlen in db776a885cd4c20811f22f8bd1a27ecc71dba927 +- feat: Auto detect Mixtral's slightly different format by @lukestanley in #1214 + + ## [0.2.48] - feat: Update llama.cpp to ggerganov/llama.cpp@15499eb94227401bdc8875da6eb85c15d37068f7 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 8c4f4002e..637f52087 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.48" \ No newline at end of file +__version__ = "0.2.49" \ No newline at end of file From 47bad30dd716443652275099fa3851811168ff4a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 12:23:24 -0500 Subject: [PATCH 185/624] fix: LlamaHFTokenizer now receives pre_tokens --- llama_cpp/llama.py | 38 +++++++++++++++++++----------------- llama_cpp/llama_tokenizer.py | 18 ++++++++++++----- 2 files changed, 33 insertions(+), 23 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 4b3d36a16..01a2e706e 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -480,7 +480,7 @@ def detokenize( Returns: The detokenized string. """ - return self.tokenizer_.detokenize(tokens, prev_tokens) + return self.tokenizer_.detokenize(tokens, prev_tokens=prev_tokens) def set_cache(self, cache: Optional[BaseLlamaCache]): """Set the cache. @@ -1016,13 +1016,13 @@ def logit_bias_processor( grammar=grammar, ): if token == self._token_eos: - text = self.detokenize(completion_tokens) + text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" break completion_tokens.append(token) - all_text = self.detokenize(completion_tokens) + all_text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) # Contains multi-byte UTF8 for k, char in enumerate(all_text[-3:]): @@ -1046,7 +1046,7 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - remaining_text = self.detokenize(remaining_tokens) + remaining_text = self.detokenize(remaining_tokens, prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) remaining_length = len(remaining_text) # We want to avoid yielding any characters from @@ -1068,17 +1068,17 @@ def logit_bias_processor( for token in remaining_tokens: if token == self.token_bos(): continue - token_end_position += len(self.detokenize([token])) + token_end_position += len(self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens])) # Check if stop sequence is in the token if token_end_position > ( remaining_length - first_stop_position ): break - token_str = self.detokenize([token]).decode( + token_str = self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( "utf-8", errors="ignore" ) text_offset = len(prompt) + len( - self.detokenize(completion_tokens[:returned_tokens]).decode( + self.detokenize(completion_tokens[:returned_tokens], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( "utf-8", errors="ignore" ) ) @@ -1100,7 +1100,7 @@ def logit_bias_processor( top_logprob.update({token_str: current_logprobs[int(token)]}) logprobs_or_none = { "tokens": [ - self.detokenize([token]).decode( + self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( "utf-8", errors="ignore" ) ], @@ -1116,7 +1116,7 @@ def logit_bias_processor( "model": model_name, "choices": [ { - "text": self.detokenize([token]).decode( + "text": self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( "utf-8", errors="ignore" ), "index": 0, @@ -1130,7 +1130,7 @@ def logit_bias_processor( decode_success = False for i in range(1, len(remaining_tokens) + 1): try: - bs = self.detokenize(remaining_tokens[:i]) + bs = self.detokenize(remaining_tokens[:i], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) ts = bs.decode("utf-8") decode_success = True break @@ -1165,14 +1165,14 @@ def logit_bias_processor( } if len(completion_tokens) >= max_tokens: - text = self.detokenize(completion_tokens) + text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "length" break if stopping_criteria is not None and stopping_criteria( self._input_ids, self._scores[-1, :] ): - text = self.detokenize(completion_tokens) + text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" if self.verbose: @@ -1180,7 +1180,7 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - all_text = self.detokenize(remaining_tokens) + all_text = self.detokenize(remaining_tokens, prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) any_stop = [s for s in stop_sequences if s in all_text] if len(any_stop) > 0: end = min(all_text.index(stop) for stop in any_stop) @@ -1189,7 +1189,7 @@ def logit_bias_processor( token_end_position = 0 for token in remaining_tokens: - token_end_position += len(self.detokenize([token])) + token_end_position += len(self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens])) logprobs_or_none: Optional[CompletionLogprobs] = None if logprobs is not None: @@ -1199,7 +1199,7 @@ def logit_bias_processor( "utf-8", errors="ignore" ) text_offset = len(prompt) + len( - self.detokenize(completion_tokens[:returned_tokens]) + self.detokenize(completion_tokens[:returned_tokens], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) ) token_offset = len(prompt_tokens) + returned_tokens - 1 logits = self._scores[token_offset, :] @@ -1313,8 +1313,8 @@ def logit_bias_processor( all_tokens = completion_tokens all_token_strs = [ - self.detokenize([token]).decode("utf-8", errors="ignore") - for token in all_tokens + self.detokenize([token], prev_tokens=all_tokens[:i]).decode("utf-8", errors="ignore") + for i, token in enumerate(all_tokens) ] all_logprobs = Llama.logits_to_logprobs(self._scores)[token_offset:] # TODO: may be able to change this loop to use np.take_along_dim @@ -1339,7 +1339,7 @@ def logit_bias_processor( ) token_logprobs.append(logprobs_token[int(token)]) top_logprob: Optional[Dict[str, float]] = { - self.detokenize([i]).decode("utf-8", errors="ignore"): logprob + self.detokenize([i], prev_tokens=all_tokens[:idx]).decode("utf-8", errors="ignore"): logprob for logprob, i in sorted_logprobs[:logprobs] } top_logprob.update({token_str: logprobs_token[int(token)]}) @@ -1594,6 +1594,8 @@ def create_chat_completion( logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, logit_bias: Optional[Dict[str, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, ) -> Union[ CreateChatCompletionResponse, Iterator[CreateChatCompletionStreamResponse] ]: diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py index c2aad4744..5a8b13b7a 100644 --- a/llama_cpp/llama_tokenizer.py +++ b/llama_cpp/llama_tokenizer.py @@ -16,12 +16,23 @@ class BaseLlamaTokenizer(abc.ABC): def tokenize( self, text: bytes, add_bos: bool = True, special: bool = True ) -> List[int]: + """Tokenize the text into tokens. + + Args: + text: The text to tokenize. + add_bos: Whether to add a beginning of sequence token. + special: Whether to tokenize text literally or as special tokens.""" raise NotImplementedError @abc.abstractmethod def detokenize( self, tokens: List[int], prev_tokens: Optional[List[int]] = None ) -> bytes: + """Detokenize the tokens into text. + + Args: + tokens: The tokens to detokenize. + prev_tokens: If tokens is a continuation of a previous sequence, the previous tokens.""" raise NotImplementedError @@ -37,10 +48,7 @@ def tokenize( def detokenize( self, tokens: List[int], prev_tokens: Optional[List[int]] = None ) -> bytes: - if prev_tokens is not None: - return self._model.detokenize(tokens[len(prev_tokens) :]) - else: - return self._model.detokenize(tokens) + return self._model.detokenize(tokens) def encode( self, text: str, add_bos: bool = True, special: bool = True @@ -72,7 +80,7 @@ def detokenize( self, tokens: List[int], prev_tokens: Optional[List[int]] = None ) -> bytes: if prev_tokens is not None: - text = self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") + text = self.hf_tokenizer.decode(prev_tokens + tokens).encode("utf-8", errors="ignore") prev_text = self.hf_tokenizer.decode(prev_tokens).encode( "utf-8", errors="ignore" ) From bce6dc0ac2f0570a32ced3cd14408237cf08f8ad Mon Sep 17 00:00:00 2001 From: Jeffrey Fong Date: Sat, 24 Feb 2024 01:24:10 +0800 Subject: [PATCH 186/624] docs: Update Functionary OpenAI Server Readme (#1193) * update functionary parts in server readme * add write-up about hf tokenizer --- docs/server.md | 8 +++++--- examples/notebooks/Functions.ipynb | 2 +- 2 files changed, 6 insertions(+), 4 deletions(-) diff --git a/docs/server.md b/docs/server.md index cd351ba33..c66c9cce0 100644 --- a/docs/server.md +++ b/docs/server.md @@ -76,12 +76,14 @@ Function calling is completely compatible with the OpenAI function calling API a You'll first need to download one of the available function calling models in GGUF format: -- [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) +- [functionary](https://huggingface.co/meetkai) -Then when you run the server you'll need to also specify the `functionary` chat_format +Then when you run the server you'll need to also specify either `functionary-v1` or `functionary-v2` chat_format. + +Note that since functionary requires a HF Tokenizer due to discrepancies between llama.cpp and HuggingFace's tokenizers as mentioned [here](https://github.com/abetlen/llama-cpp-python/blob/main?tab=readme-ov-file#function-calling), you will need to pass in the path to the tokenizer too. The tokenizer files are already included in the respective HF repositories hosting the gguf files. ```bash -python3 -m llama_cpp.server --model --chat_format functionary +python3 -m llama_cpp.server --model --chat_format functionary-v2 --hf_pretrained_model_name_or_path ``` Check out this [example notebook](https://github.com/abetlen/llama-cpp-python/blob/main/examples/notebooks/Functions.ipynb) for a walkthrough of some interesting use cases for function calling. diff --git a/examples/notebooks/Functions.ipynb b/examples/notebooks/Functions.ipynb index 7a7b899ec..f1e5e9a1d 100644 --- a/examples/notebooks/Functions.ipynb +++ b/examples/notebooks/Functions.ipynb @@ -9,7 +9,7 @@ "The OpenAI compatbile web server in `llama-cpp-python` supports function calling.\n", "\n", "Function calling allows API clients to specify a schema that gives the model a format it should respond in.\n", - "Function calling in `llama-cpp-python` works by combining models pretrained for function calling such as [`functionary`](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) with constrained sampling to produce a response that is compatible with the schema.\n", + "Function calling in `llama-cpp-python` works by combining models pretrained for function calling such as [`functionary`](https://huggingface.co/meetkai) with constrained sampling to produce a response that is compatible with the schema.\n", "\n", "Note however that this improves but does not guarantee that the response will be compatible with the schema.\n", "\n", From 702306b381d47040e147d1a7377b82fb40666475 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 12:34:02 -0500 Subject: [PATCH 187/624] docs: Restore functionary docs in README --- README.md | 25 ++++++++++++++++++++----- 1 file changed, 20 insertions(+), 5 deletions(-) diff --git a/README.md b/README.md index ccd66ea82..4ce8030e1 100644 --- a/README.md +++ b/README.md @@ -365,14 +365,10 @@ To constrain the response further to a specific JSON Schema add the schema to th ### Function Calling -The high-level API also provides a simple interface for function calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat format. - -The gguf-converted files for functionary can be found here: [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) +The high-level API supports OpenAI compatible function and tool calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat format. ```python >>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", chat_format="functionary") ->>> # or >>> llm = Llama(model_path="path/to/chatml/llama-model.gguf", chat_format="chatml-function-calling") >>> llm.create_chat_completion( messages = [ @@ -416,6 +412,25 @@ The gguf-converted files for functionary can be found here: [functionary-7b-v1]( ) ``` +
+Functionary v2 + +The various gguf-converted files for this set of models can be found [here](https://huggingface.co/meetkai). Functionary is able to intelligently call functions and also analyze any provided function outputs to generate coherent responses. All v2 models of functionary supports **parallel function calling**. You can provide either `functionary-v1` or `functionary-v2` for the `chat_format` when initializing the Llama class. + +Due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is required to provide HF Tokenizer for functionary. The `LlamaHFTokenizer` class can be initialized and passed into the Llama class. This will override the default llama.cpp tokenizer used in Llama class. The tokenizer files are already included in the respective HF repositories hosting the gguf files. + +```python +>>> from llama_cpp import Llama +>>> from llama_cpp.llama_tokenizer import LlamaHFTokenizer +>>> llm = Llama.from_pretrained( + repo_id="meetkai/functionary-7b-v1-GGUF", + filename="functionary-small-v2.2.q4_0.gguf", + chat_format="functionary-v2", + tokenizer=LlamaHFTokenizer.from_pretrained("meetkai/functionary-7b-v1-GGUF") +) +``` +
+ ### Multi-modal Models `llama-cpp-python` supports the llava1.5 family of multi-modal models which allow the language model to From b681674bf29471fb8447f122356203a07fc543b1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 12:36:13 -0500 Subject: [PATCH 188/624] docs: Fix functionary repo_id --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 4ce8030e1..a3f82e43e 100644 --- a/README.md +++ b/README.md @@ -423,10 +423,10 @@ Due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is requi >>> from llama_cpp import Llama >>> from llama_cpp.llama_tokenizer import LlamaHFTokenizer >>> llm = Llama.from_pretrained( - repo_id="meetkai/functionary-7b-v1-GGUF", + repo_id="meetkai/functionary-small-v2.2-GGUF", filename="functionary-small-v2.2.q4_0.gguf", chat_format="functionary-v2", - tokenizer=LlamaHFTokenizer.from_pretrained("meetkai/functionary-7b-v1-GGUF") + tokenizer=LlamaHFTokenizer.from_pretrained("meetkai/functionary-small-v2.2-GGUF") ) ```
From 20ea6fd7d614245d60f1b729cf524b211f991095 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 23 Feb 2024 12:38:36 -0500 Subject: [PATCH 189/624] chore: Bump version --- CHANGELOG.md | 14 +++++++++----- llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index c1ba40fa9..bb42f61ee 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,12 +7,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.50] + +- docs: Update Functionary OpenAI Server Readme by @jeffrey-fong in #1193 +- fix: LlamaHFTokenizer now receives pre_tokens by @abetlen in 47bad30dd716443652275099fa3851811168ff4a + ## [0.2.49] - fix: module 'llama_cpp.llama_cpp' has no attribute 'c_uint8' in Llama.save_state by @abetlen in db776a885cd4c20811f22f8bd1a27ecc71dba927 - feat: Auto detect Mixtral's slightly different format by @lukestanley in #1214 - ## [0.2.48] - feat: Update llama.cpp to ggerganov/llama.cpp@15499eb94227401bdc8875da6eb85c15d37068f7 @@ -151,7 +155,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 - feat: Update llama.cpp to ggerganov/llama.cpp@b3a7c20b5c035250257d2b62851c379b159c899a - feat: Add `saiga` chat format by @femoiseev in #1050 - feat: Added `chatglm3` chat format by @xaviviro in #1059 -- fix: Correct typo in README.md by @qeleb in (#1058) +- fix: Correct typo in README.md by @qeleb in (#1058) ## [0.2.26] @@ -284,7 +288,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [0.2.11] -- Fix bug in `llama_model_params` object has no attribute `logits_all` by @abetlen in d696251fbe40015e8616ea7a7d7ad5257fd1b896 +- Fix bug in `llama_model_params` object has no attribute `logits_all` by @abetlen in d696251fbe40015e8616ea7a7d7ad5257fd1b896 ## [0.2.10] @@ -472,7 +476,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [0.1.60] -NOTE: This release was deleted due to a bug with the packaging system that caused pip installations to fail. +NOTE: This release was deleted due to a bug with the packaging system that caused pip installations to fail. - Truncate max_tokens in create_completion so requested tokens doesn't exceed context size. - Temporarily disable cache for completion requests @@ -496,4 +500,4 @@ NOTE: This release was deleted due to a bug with the packaging system that caus - (misc) Added first version of the changelog - (server) Use async routes - (python-api) Use numpy for internal buffers to reduce memory usage and improve performance. -- (python-api) Performance bug in stop sequence check slowing down streaming. \ No newline at end of file +- (python-api) Performance bug in stop sequence check slowing down streaming. diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 637f52087..fc2efdceb 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.49" \ No newline at end of file +__version__ = "0.2.50" \ No newline at end of file From 221edb9ef1aa2155404d0a121921f7e0e418feda Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 24 Feb 2024 23:47:29 -0500 Subject: [PATCH 190/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++++ vendor/llama.cpp | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index f4d523be0..b8e74d7a3 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -245,6 +245,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_S = 26, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -272,6 +274,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 LLAMA_FTYPE_MOSTLY_IQ1_S = 24 LLAMA_FTYPE_MOSTLY_IQ4_NL = 25 +LLAMA_FTYPE_MOSTLY_IQ3_S = 26 +LLAMA_FTYPE_MOSTLY_IQ3_M = 27 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 15499eb94..9e359a4f4 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 15499eb94227401bdc8875da6eb85c15d37068f7 +Subproject commit 9e359a4f47c1b2dceb99e29706c9f7403d32ab5e From 2292af5796fb0ce2092289246eabe246abd09c0f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 16:53:58 -0500 Subject: [PATCH 191/624] feat: Update llama.cpp --- llama_cpp/llama.py | 12 ++++---- llama_cpp/llama_cpp.py | 56 ++++++++++++++++++------------------ llama_cpp/server/settings.py | 8 ++++-- vendor/llama.cpp | 2 +- 4 files changed, 41 insertions(+), 37 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 01a2e706e..81bfce40d 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -65,7 +65,7 @@ def __init__( *, # Model Params n_gpu_layers: int = 0, - split_mode: int = llama_cpp.LLAMA_SPLIT_LAYER, + split_mode: int = llama_cpp.LLAMA_SPLIT_MODE_LAYER, main_gpu: int = 0, tensor_split: Optional[List[float]] = None, vocab_only: bool = False, @@ -78,7 +78,7 @@ def __init__( n_batch: int = 512, n_threads: Optional[int] = None, n_threads_batch: Optional[int] = None, - rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED, + rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, rope_freq_base: float = 0.0, rope_freq_scale: float = 0.0, yarn_ext_factor: float = -1.0, @@ -238,13 +238,13 @@ def __init__( for i, (k, v) in enumerate(kv_overrides.items()): self._kv_overrides_array[i].key = k.encode("utf-8") if isinstance(v, bool): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL self._kv_overrides_array[i].value.bool_value = v elif isinstance(v, int): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT self._kv_overrides_array[i].value.int_value = v elif isinstance(v, float): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_FLOAT + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT self._kv_overrides_array[i].value.float_value = v else: raise ValueError(f"Unknown value type for {k}: {v}") @@ -270,7 +270,7 @@ def __init__( self.context_params.rope_scaling_type = ( rope_scaling_type if rope_scaling_type is not None - else llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED + else llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED ) self.context_params.rope_freq_base = ( rope_freq_base if rope_freq_base != 0.0 else 0 diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index b8e74d7a3..37d46373c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -279,35 +279,35 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { -# LLAMA_ROPE_SCALING_UNSPECIFIED = -1, -# LLAMA_ROPE_SCALING_NONE = 0, -# LLAMA_ROPE_SCALING_LINEAR = 1, -# LLAMA_ROPE_SCALING_YARN = 2, -# LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN, +# LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1, +# LLAMA_ROPE_SCALING_TYPE_NONE = 0, +# LLAMA_ROPE_SCALING_TYPE_LINEAR = 1, +# LLAMA_ROPE_SCALING_TYPE_YARN = 2, +# LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN, # }; -LLAMA_ROPE_SCALING_UNSPECIFIED = -1 -LLAMA_ROPE_SCALING_NONE = 0 -LLAMA_ROPE_SCALING_LINEAR = 1 -LLAMA_ROPE_SCALING_YARN = 2 -LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN +LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1 +LLAMA_ROPE_SCALING_TYPE_NONE = 0 +LLAMA_ROPE_SCALING_TYPE_LINEAR = 1 +LLAMA_ROPE_SCALING_TYPE_YARN = 2 +LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN # enum llama_pooling_type { -# LLAMA_POOLING_NONE = 0, -# LLAMA_POOLING_MEAN = 1, -# LLAMA_POOLING_CLS = 2, +# LLAMA_POOLING_TYPE_NONE = 0, +# LLAMA_POOLING_TYPE_MEAN = 1, +# LLAMA_POOLING_TYPE_CLS = 2, # }; -LLAMA_POOLING_NONE = 0 -LLAMA_POOLING_MEAN = 1 -LLAMA_POOLING_CLS = 2 +LLAMA_POOLING_TYPE_NONE = 0 +LLAMA_POOLING_TYPE_MEAN = 1 +LLAMA_POOLING_TYPE_CLS = 2 # enum llama_split_mode { -# LLAMA_SPLIT_NONE = 0, // single GPU -# LLAMA_SPLIT_LAYER = 1, // split layers and KV across GPUs -# LLAMA_SPLIT_ROW = 2, // split rows across GPUs +# LLAMA_SPLIT_MODE_NONE = 0, // single GPU +# LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs +# LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs # }; -LLAMA_SPLIT_NONE = 0 -LLAMA_SPLIT_LAYER = 1 -LLAMA_SPLIT_ROW = 2 +LLAMA_SPLIT_MODE_NONE = 0 +LLAMA_SPLIT_MODE_LAYER = 1 +LLAMA_SPLIT_MODE_ROW = 2 # typedef struct llama_token_data { @@ -420,13 +420,13 @@ class llama_batch(ctypes.Structure): # enum llama_model_kv_override_type { -# LLAMA_KV_OVERRIDE_INT, -# LLAMA_KV_OVERRIDE_FLOAT, -# LLAMA_KV_OVERRIDE_BOOL, +# LLAMA_KV_OVERRIDE_TYPE_INT, +# LLAMA_KV_OVERRIDE_TYPE_FLOAT, +# LLAMA_KV_OVERRIDE_TYPE_BOOL, # }; -LLAMA_KV_OVERRIDE_INT = 0 -LLAMA_KV_OVERRIDE_FLOAT = 1 -LLAMA_KV_OVERRIDE_BOOL = 2 +LLAMA_KV_OVERRIDE_TYPE_INT = 0 +LLAMA_KV_OVERRIDE_TYPE_FLOAT = 1 +LLAMA_KV_OVERRIDE_TYPE_BOOL = 2 # struct llama_model_kv_override { diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 790c6b129..8989ffac8 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -29,7 +29,7 @@ class ModelSettings(BaseSettings): description="The number of layers to put on the GPU. The rest will be on the CPU. Set -1 to move all to GPU.", ) split_mode: int = Field( - default=llama_cpp.LLAMA_SPLIT_LAYER, + default=llama_cpp.LLAMA_SPLIT_MODE_LAYER, description="The split mode to use.", ) main_gpu: int = Field( @@ -74,7 +74,7 @@ class ModelSettings(BaseSettings): ge=0, description="The number of threads to use when batch processing.", ) - rope_scaling_type: int = Field(default=llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED) + rope_scaling_type: int = Field(default=llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) rope_freq_base: float = Field(default=0.0, description="RoPE base frequency") rope_freq_scale: float = Field( default=0.0, description="RoPE frequency scaling factor" @@ -143,6 +143,10 @@ class ModelSettings(BaseSettings): default=None, description="The model name or path to a pretrained HuggingFace tokenizer model. Same as you would pass to AutoTokenizer.from_pretrained().", ) + hf_model_repo_id: Optional[str] = Field( + default=None, + description="The HuggingFace repo_id to use to load model files from", + ) # Speculative Decoding draft_model: Optional[str] = Field( default=None, diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 9e359a4f4..f7625019c 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 9e359a4f47c1b2dceb99e29706c9f7403d32ab5e +Subproject commit f7625019c51ca437a5840576d92362cfa710e4a2 From 19234aa0dbd0c3c87656e65dd2b064665371925b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 16:54:37 -0500 Subject: [PATCH 192/624] fix: Restore type hints for low-level api --- llama_cpp/llama_cpp.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 37d46373c..e5ae4f7d7 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -109,12 +109,13 @@ class CtypesRef(Generic[CtypesCData]): CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore +F = TypeVar("F", bound=Callable[..., Any]) def ctypes_function_for_shared_library(lib: ctypes.CDLL): def ctypes_function( name: str, argtypes: List[Any], restype: Any, enabled: bool = True ): - def decorator(f: Callable[..., Any]): + def decorator(f: F) -> F: if enabled: func = getattr(lib, name) func.argtypes = argtypes From cbbcd888afd016762ae1adfdcbeda3d083f360f9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 20:52:14 -0500 Subject: [PATCH 193/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 2 +- llama_cpp/llama_cpp.py | 190 +++++++++++----------------------------- tests/test_llama.py | 4 +- vendor/llama.cpp | 2 +- 4 files changed, 54 insertions(+), 144 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 98ad6d718..13bf1fb06 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -291,7 +291,7 @@ def kv_cache_seq_keep(self, seq_id: int): def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_shift(self.ctx, seq_id, p0, p1, shift) + llama_cpp.llama_kv_cache_seq_add(self.ctx, seq_id, p0, p1, shift) def get_state_size(self) -> int: assert self.ctx is not None diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index e5ae4f7d7..5a934dc26 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -200,6 +200,20 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_TYPE_WPM = 2 +# // note: these values should be synchronized with ggml_rope +# // TODO: maybe move this enum to ggml.h (ggml_rope_type) +# enum llama_rope_type { +# LLAMA_ROPE_TYPE_NONE = -1, +# LLAMA_ROPE_TYPE_NORM = 0, +# LLAMA_ROPE_TYPE_NEOX = 2, +# LLAMA_ROPE_TYPE_GLM = 4, +# }; +LLAMA_ROPE_TYPE_NONE = -1 +LLAMA_ROPE_TYPE_NORM = 0 +LLAMA_ROPE_TYPE_NEOX = 2 +LLAMA_ROPE_TYPE_GLM = 4 + + # enum llama_token_type { # LLAMA_TOKEN_TYPE_UNDEFINED = 0, # LLAMA_TOKEN_TYPE_NORMAL = 1, @@ -892,104 +906,84 @@ def llama_time_us() -> int: # LLAMA_API size_t llama_max_devices(void); - - @ctypes_function("llama_max_devices", [], ctypes.c_size_t) def llama_max_devices() -> int: ... # LLAMA_API bool llama_supports_mmap (void); - - @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) def llama_supports_mmap() -> bool: ... # LLAMA_API bool llama_supports_mlock (void); - - @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) def llama_supports_mlock() -> bool: ... # LLAMA_API bool llama_supports_gpu_offload(void); - - @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) def llama_supports_gpu_offload() -> bool: ... # LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); - - @ctypes_function("llama_mmap_supported", [], ctypes.c_bool) def llama_mmap_supported() -> bool: ... # LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); - - @ctypes_function("llama_mlock_supported", [], ctypes.c_bool) def llama_mlock_supported() -> bool: ... # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); - - @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); - - @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) def llama_n_ctx(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); - - @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) def llama_n_batch(ctx: llama_context_p, /) -> int: ... # LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); - - @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) def llama_vocab_type(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); +# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); +@ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_rope_type(model: llama_model_p, /) -> int: + ... +# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_vocab(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); - - @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_ctx_train(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); - - @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_embd(model: llama_model_p, /) -> int: ... @@ -997,8 +991,6 @@ def llama_n_embd(model: llama_model_p, /) -> int: # // Get the model's RoPE frequency scaling factor # LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); - - @ctypes_function("llama_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: """Get the model's RoPE frequency scaling factor""" @@ -1013,8 +1005,6 @@ def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: # // Get metadata value as a string by key name # LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); - - @ctypes_function( "llama_model_meta_val_str", [ @@ -1038,8 +1028,6 @@ def llama_model_meta_val_str( # // Get the number of metadata key/value pairs # LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model); - - @ctypes_function("llama_model_meta_count", [llama_model_p_ctypes], ctypes.c_int32) def llama_model_meta_count(model: llama_model_p, /) -> int: """Get the number of metadata key/value pairs""" @@ -1048,8 +1036,6 @@ def llama_model_meta_count(model: llama_model_p, /) -> int: # // Get metadata key name by index # LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); - - @ctypes_function( "llama_model_meta_key_by_index", [ @@ -1073,8 +1059,6 @@ def llama_model_meta_key_by_index( # // Get metadata value as a string by index # LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); - - @ctypes_function( "llama_model_meta_val_str_by_index", [ @@ -1098,8 +1082,6 @@ def llama_model_meta_val_str_by_index( # // Get a string describing the model type # LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); - - @ctypes_function( "llama_model_desc", [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_size_t], @@ -1117,8 +1099,6 @@ def llama_model_desc( # // Returns the total size of all the tensors in the model in bytes # LLAMA_API uint64_t llama_model_size(const struct llama_model * model); - - @ctypes_function("llama_model_size", [llama_model_p_ctypes], ctypes.c_uint64) def llama_model_size(model: llama_model_p, /) -> int: """Returns the total size of all the tensors in the model in bytes""" @@ -1127,8 +1107,6 @@ def llama_model_size(model: llama_model_p, /) -> int: # // Returns the total number of parameters in the model # LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); - - @ctypes_function("llama_model_n_params", [llama_model_p_ctypes], ctypes.c_uint64) def llama_model_n_params(model: llama_model_p, /) -> int: """Returns the total number of parameters in the model""" @@ -1137,8 +1115,6 @@ def llama_model_n_params(model: llama_model_p, /) -> int: # // Get a llama model tensor # LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); - - @ctypes_function( "llama_get_model_tensor", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_void_p ) @@ -1154,8 +1130,6 @@ def llama_get_model_tensor( # const char * fname_inp, # const char * fname_out, # const llama_model_quantize_params * params); - - @ctypes_function( "llama_model_quantize", [ @@ -1188,8 +1162,6 @@ def llama_model_quantize( # const char * path_base_model, # int32_t n_threads), # "use llama_model_apply_lora_from_file instead"); - - @ctypes_function( "llama_apply_lora_from_file", [ @@ -1224,8 +1196,6 @@ def llama_apply_lora_from_file( # float scale, # const char * path_base_model, # int32_t n_threads); - - @ctypes_function( "llama_model_apply_lora_from_file", [ @@ -1313,8 +1283,6 @@ class llama_kv_cache_view(ctypes.Structure): # // Create an empty KV cache view. (use only for debugging purposes) # LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq); - - @ctypes_function( "llama_kv_cache_view_init", [llama_context_p_ctypes, ctypes.c_int32], @@ -1329,8 +1297,6 @@ def llama_kv_cache_view_init( # // Free a KV cache view. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); - - @ctypes_function("llama_kv_cache_view_free", [llama_kv_cache_view_p], None) def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore """Free a KV cache view. (use only for debugging purposes)""" @@ -1339,8 +1305,6 @@ def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # # // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); - - @ctypes_function( "llama_kv_cache_view_update", [llama_context_p_ctypes, llama_kv_cache_view_p], None ) @@ -1352,8 +1316,6 @@ def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[ll # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times # LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); - - @ctypes_function( "llama_get_kv_cache_token_count", [llama_context_p_ctypes], ctypes.c_int32 ) @@ -1366,8 +1328,6 @@ def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) # LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); - - @ctypes_function( "llama_get_kv_cache_used_cells", [llama_context_p_ctypes], ctypes.c_int32 ) @@ -1379,8 +1339,6 @@ def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: # // Clear the KV cache # LLAMA_API void llama_kv_cache_clear( # struct llama_context * ctx); - - @ctypes_function("llama_kv_cache_clear", [llama_context_p_ctypes], None) def llama_kv_cache_clear(ctx: llama_context_p, /): """Clear the KV cache""" @@ -1396,8 +1354,6 @@ def llama_kv_cache_clear(ctx: llama_context_p, /): # llama_seq_id seq_id, # llama_pos p0, # llama_pos p1); - - @ctypes_function( "llama_kv_cache_seq_rm", [ @@ -1432,8 +1388,6 @@ def llama_kv_cache_seq_rm( # llama_seq_id seq_id_dst, # llama_pos p0, # llama_pos p1); - - @ctypes_function( "llama_kv_cache_seq_cp", [ @@ -1464,8 +1418,6 @@ def llama_kv_cache_seq_cp( # LLAMA_API void llama_kv_cache_seq_keep( # struct llama_context * ctx, # llama_seq_id seq_id); - - @ctypes_function( "llama_kv_cache_seq_keep", [llama_context_p_ctypes, llama_seq_id], None ) @@ -1475,19 +1427,19 @@ def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, in # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) -# // If the KV cache is RoPEd, the KV data is updated accordingly +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_cache_update() # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_shift( +# LLAMA_API void llama_kv_cache_seq_add( # struct llama_context * ctx, # llama_seq_id seq_id, # llama_pos p0, # llama_pos p1, # llama_pos delta); - - @ctypes_function( - "llama_kv_cache_seq_shift", + "llama_kv_cache_seq_add", [ llama_context_p_ctypes, llama_seq_id, @@ -1497,7 +1449,7 @@ def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, in ], None, ) -def llama_kv_cache_seq_shift( +def llama_kv_cache_seq_add( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], @@ -1506,7 +1458,9 @@ def llama_kv_cache_seq_shift( /, ): """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) - If the KV cache is RoPEd, the KV data is updated accordingly + If the KV cache is RoPEd, the KV data is updated accordingly: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update() p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" ... @@ -1522,8 +1476,6 @@ def llama_kv_cache_seq_shift( # llama_pos p0, # llama_pos p1, # int d); - - @ctypes_function( "llama_kv_cache_seq_div", [ @@ -1550,6 +1502,28 @@ def llama_kv_cache_seq_div( ... +# // Defragment the KV cache +# // This will be applied: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_cache_update() +# LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); +@ctypes_function("llama_kv_cache_defrag", [llama_context_p_ctypes], None) +def llama_kv_cache_defrag(ctx: llama_context_p, /): + """Defragment the KV cache + This will be applied: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update()""" + ... + + +# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) +# LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); +@ctypes_function("llama_kv_cache_update", [llama_context_p_ctypes], None) +def llama_kv_cache_update(ctx: llama_context_p, /): + """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" + ... + + # // # // State / sessions # // @@ -1558,8 +1532,6 @@ def llama_kv_cache_seq_div( # Returns the maximum size in bytes of the state (rng, logits, embedding # and kv_cache) - will often be smaller after compacting tokens # LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); - - @ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_get_state_size(ctx: llama_context_p, /) -> int: """Returns the maximum size in bytes of the state (rng, logits, embedding @@ -1573,8 +1545,6 @@ def llama_get_state_size(ctx: llama_context_p, /) -> int: # LLAMA_API size_t llama_copy_state_data( # struct llama_context * ctx, # uint8_t * dst); - - @ctypes_function( "llama_copy_state_data", [ @@ -1597,8 +1567,6 @@ def llama_copy_state_data( # LLAMA_API size_t llama_set_state_data( # struct llama_context * ctx, # uint8_t * src); - - @ctypes_function( "llama_set_state_data", [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], @@ -1618,8 +1586,6 @@ def llama_set_state_data( # llama_token * tokens_out, # size_t n_token_capacity, # size_t * n_token_count_out); - - @ctypes_function( "llama_load_session_file", [ @@ -1647,8 +1613,6 @@ def llama_load_session_file( # const char * path_session, # const llama_token * tokens, # size_t n_token_count); - - @ctypes_function( "llama_save_session_file", [ @@ -1685,8 +1649,6 @@ def llama_save_session_file( # int32_t n_tokens, # int32_t n_past), # "use llama_decode() instead"); - - @ctypes_function( "llama_eval", [ @@ -1720,8 +1682,6 @@ def llama_eval( # int32_t n_tokens, # int32_t n_past), # "use llama_decode() instead"); - - @ctypes_function( "llama_eval_embd", [ @@ -1753,8 +1713,6 @@ def llama_eval_embd( # int32_t n_tokens, # llama_pos pos_0, # llama_seq_id seq_id); - - @ctypes_function( "llama_batch_get_one", [ @@ -1790,8 +1748,6 @@ def llama_batch_get_one( # int32_t n_tokens, # int32_t embd, # int32_t n_seq_max); - - @ctypes_function( "llama_batch_init", [ctypes.c_int32, ctypes.c_int32, ctypes.c_int32], llama_batch ) @@ -1813,8 +1769,6 @@ def llama_batch_init( # // Frees a batch of tokens allocated with llama_batch_init() # LLAMA_API void llama_batch_free(struct llama_batch batch); - - @ctypes_function("llama_batch_free", [llama_batch], None) def llama_batch_free(batch: llama_batch, /): """Frees a batch of tokens allocated with llama_batch_init()""" @@ -1828,8 +1782,6 @@ def llama_batch_free(batch: llama_batch, /): # LLAMA_API int32_t llama_decode( # struct llama_context * ctx, # struct llama_batch batch); - - @ctypes_function("llama_decode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: """Positive return values does not mean a fatal error, but rather a warning. @@ -1843,8 +1795,6 @@ def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: # // n_threads is the number of threads used for generation (single token) # // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) # LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); - - @ctypes_function( "llama_set_n_threads", [ @@ -1873,8 +1823,6 @@ def llama_set_n_threads( # // Rows: n_tokens provided with llama_batch # // Cols: n_vocab # LLAMA_API float * llama_get_logits(struct llama_context * ctx); - - @ctypes_function( "llama_get_logits", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) ) @@ -1890,8 +1838,6 @@ def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: # // Logits for the ith token. Equivalent to: # // llama_get_logits(ctx) + i*n_vocab # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); - - @ctypes_function( "llama_get_logits_ith", [llama_context_p_ctypes, ctypes.c_int32], @@ -1908,8 +1854,6 @@ def llama_get_logits_ith( # Get the embeddings for the input # shape: [n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); - - @ctypes_function( "llama_get_embeddings", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) ) @@ -1922,8 +1866,6 @@ def llama_get_embeddings(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float] # // Get the embeddings for the ith sequence # // llama_get_embeddings(ctx) + i*n_embd # LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); - - @ctypes_function( "llama_get_embeddings_ith", [llama_context_p_ctypes, ctypes.c_int32], @@ -1943,8 +1885,6 @@ def llama_get_embeddings_ith( # LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); - - @ctypes_function( "llama_token_get_text", [llama_model_p_ctypes, llama_token], ctypes.c_char_p ) @@ -1955,8 +1895,6 @@ def llama_token_get_text( # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); - - @ctypes_function( "llama_token_get_score", [llama_model_p_ctypes, llama_token], ctypes.c_float ) @@ -1967,8 +1905,6 @@ def llama_token_get_score( # LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); - - @ctypes_function( "llama_token_get_type", [llama_model_p_ctypes, llama_token], ctypes.c_int ) @@ -1982,8 +1918,6 @@ def llama_token_get_type( # LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence - - @ctypes_function("llama_token_bos", [llama_model_p_ctypes], llama_token) def llama_token_bos(model: llama_model_p, /) -> int: """beginning-of-sentence""" @@ -1991,8 +1925,6 @@ def llama_token_bos(model: llama_model_p, /) -> int: # LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence - - @ctypes_function("llama_token_eos", [llama_model_p_ctypes], llama_token) def llama_token_eos(model: llama_model_p, /) -> int: """end-of-sentence""" @@ -2000,8 +1932,6 @@ def llama_token_eos(model: llama_model_p, /) -> int: # LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line - - @ctypes_function("llama_token_nl", [llama_model_p_ctypes], llama_token) def llama_token_nl(model: llama_model_p, /) -> int: """next-line""" @@ -2010,8 +1940,6 @@ def llama_token_nl(model: llama_model_p, /) -> int: # // Returns -1 if unknown, 1 for true or 0 for false. # LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); - - @ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_int32) def llama_add_bos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" @@ -2020,8 +1948,6 @@ def llama_add_bos_token(model: llama_model_p, /) -> int: # // Returns -1 if unknown, 1 for true or 0 for false. # LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); - - @ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_int32) def llama_add_eos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" @@ -2030,8 +1956,6 @@ def llama_add_eos_token(model: llama_model_p, /) -> int: # // codellama infill tokens # LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix - - @ctypes_function("llama_token_prefix", [llama_model_p_ctypes], llama_token) def llama_token_prefix(model: llama_model_p) -> int: """codellama infill tokens""" @@ -2039,24 +1963,18 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle - - @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) def llama_token_middle(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix - - @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) def llama_token_suffix(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle - - @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) def llama_token_eot(model: llama_model_p, /) -> int: ... @@ -2081,8 +1999,6 @@ def llama_token_eot(model: llama_model_p, /) -> int: # int32_t n_max_tokens, # bool add_bos, # bool special); - - @ctypes_function( "llama_tokenize", [ @@ -2119,8 +2035,6 @@ def llama_tokenize( # llama_token token, # char * buf, # int32_t length); - - @ctypes_function( "llama_token_to_piece", [ @@ -2164,8 +2078,6 @@ def llama_token_to_piece( # bool add_ass, # char * buf, # int32_t length); - - @ctypes_function( "llama_chat_apply_template", [ @@ -2195,8 +2107,6 @@ def llama_chat_apply_template( # const llama_grammar_element ** rules, # size_t n_rules, # size_t start_rule_index); - - @ctypes_function( "llama_grammar_init", [ diff --git a/tests/test_llama.py b/tests/test_llama.py index 5cf421b56..fa2f6dfc1 100644 --- a/tests/test_llama.py +++ b/tests/test_llama.py @@ -132,7 +132,7 @@ def mock_kv_cache_seq_keep( assert ctx == llama._ctx.ctx, "context does not match mock_llama" return - def mock_kv_cache_seq_shift( + def mock_kv_cache_seq_add( ctx: llama_cpp.llama_context_p, seq_id: llama_cpp.llama_seq_id, pos0: llama_cpp.llama_pos, @@ -146,7 +146,7 @@ def mock_kv_cache_seq_shift( monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_rm", mock_kv_cache_seq_rm) monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_cp", mock_kv_cache_seq_cp) monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_keep", mock_kv_cache_seq_keep) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_shift", mock_kv_cache_seq_shift) + monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_add", mock_kv_cache_seq_add) return setup_mock diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f7625019c..c39373398 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f7625019c51ca437a5840576d92362cfa710e4a2 +Subproject commit c39373398803c669056304090050fe3f44b41bf9 From dcf38f61411fbf51860978d5b02c4cb9d573dcf0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 21:00:37 -0500 Subject: [PATCH 194/624] fix: remove prematurely commited change --- llama_cpp/server/settings.py | 4 ---- 1 file changed, 4 deletions(-) diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 8989ffac8..8ac8df881 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -143,10 +143,6 @@ class ModelSettings(BaseSettings): default=None, description="The model name or path to a pretrained HuggingFace tokenizer model. Same as you would pass to AutoTokenizer.from_pretrained().", ) - hf_model_repo_id: Optional[str] = Field( - default=None, - description="The HuggingFace repo_id to use to load model files from", - ) # Speculative Decoding draft_model: Optional[str] = Field( default=None, From bd4ec2e612cdba200a5e38544ae2e1b0eda132f6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 21:09:13 -0500 Subject: [PATCH 195/624] docs(examples): Add gradio chat example --- examples/gradio_chat/local.py | 59 ++++++++++++++++++++++++++++++++++ examples/gradio_chat/server.py | 56 ++++++++++++++++++++++++++++++++ 2 files changed, 115 insertions(+) create mode 100644 examples/gradio_chat/local.py create mode 100644 examples/gradio_chat/server.py diff --git a/examples/gradio_chat/local.py b/examples/gradio_chat/local.py new file mode 100644 index 000000000..a7de8e842 --- /dev/null +++ b/examples/gradio_chat/local.py @@ -0,0 +1,59 @@ +import llama_cpp +import llama_cpp.llama_tokenizer + +import gradio as gr + +llama = llama_cpp.Llama.from_pretrained( + repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", + filename="*q8_0.gguf", + tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B"), + verbose=False +) + +model = "gpt-3.5-turbo" + +def predict(message, history): + messages = [] + + for user_message, assistant_message in history: + messages.append({"role": "user", "content": user_message}) + messages.append({"role": "assistant", "content": assistant_message}) + + messages.append({"role": "user", "content": message}) + + response = llama.create_chat_completion_openai_v1( + model=model, + messages=messages, + stream=True + ) + + text = "" + for chunk in response: + content = chunk.choices[0].delta.content + if content: + text += content + yield text + + +js = """function () { + gradioURL = window.location.href + if (!gradioURL.endsWith('?__theme=dark')) { + window.location.replace(gradioURL + '?__theme=dark'); + } +}""" + +css = """ +footer { + visibility: hidden; +} +full-height { + height: 100%; +} +""" + +with gr.Blocks(theme=gr.themes.Soft(), js=js, css=css, fill_height=True) as demo: + gr.ChatInterface(predict, fill_height=True, examples=["What is the capital of France?", "Who was the first person on the moon?"]) + + +if __name__ == "__main__": + demo.launch() diff --git a/examples/gradio_chat/server.py b/examples/gradio_chat/server.py new file mode 100644 index 000000000..36fa43fbd --- /dev/null +++ b/examples/gradio_chat/server.py @@ -0,0 +1,56 @@ +import gradio as gr + +from openai import OpenAI + +client = OpenAI( + base_url="http://localhost:8000/v1", + api_key="llama.cpp" +) + +model = "gpt-3.5-turbo" + +def predict(message, history): + messages = [] + + for user_message, assistant_message in history: + messages.append({"role": "user", "content": user_message}) + messages.append({"role": "assistant", "content": assistant_message}) + + messages.append({"role": "user", "content": message}) + + response = client.chat.completions.create( + model=model, + messages=messages, + stream=True + ) + + text = "" + for chunk in response: + content = chunk.choices[0].delta.content + if content: + text += content + yield text + + +js = """function () { + gradioURL = window.location.href + if (!gradioURL.endsWith('?__theme=dark')) { + window.location.replace(gradioURL + '?__theme=dark'); + } +}""" + +css = """ +footer { + visibility: hidden; +} +full-height { + height: 100%; +} +""" + +with gr.Blocks(theme=gr.themes.Soft(), js=js, css=css, fill_height=True) as demo: + gr.ChatInterface(predict, fill_height=True, examples=["What is the capital of France?", "Who was the first person on the moon?"]) + + +if __name__ == "__main__": + demo.launch() From 252e1ff2b4b2d92f8aaa267e3828a74da3084f43 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 21:09:41 -0500 Subject: [PATCH 196/624] docs(examples): Add huggingface pull example --- examples/hf_pull/main.py | 39 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 39 insertions(+) create mode 100644 examples/hf_pull/main.py diff --git a/examples/hf_pull/main.py b/examples/hf_pull/main.py new file mode 100644 index 000000000..d3eb11c39 --- /dev/null +++ b/examples/hf_pull/main.py @@ -0,0 +1,39 @@ +import llama_cpp +import llama_cpp.llama_tokenizer + + +llama = llama_cpp.Llama.from_pretrained( + repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", + filename="*q8_0.gguf", + tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B"), + verbose=False +) + +response = llama.create_chat_completion( + messages=[ + { + "role": "user", + "content": "What is the capital of France?" + } + ], + response_format={ + "type": "json_object", + "schema": { + "type": "object", + "properties": { + "country": {"type": "string"}, + "capital": {"type": "string"} + }, + "required": ["country", "capital"], + } + }, + stream=True +) + +for chunk in response: + delta = chunk["choices"][0]["delta"] + if "content" not in delta: + continue + print(delta["content"], end="", flush=True) + +print() \ No newline at end of file From e857c133fbbb88cbc0374497550dc78fd927a00b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 21:14:01 -0500 Subject: [PATCH 197/624] feat: Update llama.cpp --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index bb42f61ee..036dbe039 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.51] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c39373398803c669056304090050fe3f44b41bf9 +- fix: Restore type hints for low-level api by @abetlen in 19234aa0dbd0c3c87656e65dd2b064665371925b + ## [0.2.50] - docs: Update Functionary OpenAI Server Readme by @jeffrey-fong in #1193 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index fc2efdceb..83a70cf76 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.50" \ No newline at end of file +__version__ = "0.2.51" \ No newline at end of file From 8e03fd995788b73d434d0b461bc07acc5b5636d2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 25 Feb 2024 21:14:01 -0500 Subject: [PATCH 198/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index bb42f61ee..036dbe039 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.51] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c39373398803c669056304090050fe3f44b41bf9 +- fix: Restore type hints for low-level api by @abetlen in 19234aa0dbd0c3c87656e65dd2b064665371925b + ## [0.2.50] - docs: Update Functionary OpenAI Server Readme by @jeffrey-fong in #1193 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index fc2efdceb..83a70cf76 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.50" \ No newline at end of file +__version__ = "0.2.51" \ No newline at end of file From 34111788feb56172021f5804d96021b0a8432e65 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 10:58:41 -0500 Subject: [PATCH 199/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c39373398..47bb7b48c 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c39373398803c669056304090050fe3f44b41bf9 +Subproject commit 47bb7b48c7cec9d8f57d56812ce811ec130b89a3 From 8383a9e5620f5df5a88f62da16813eac200dd706 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:03:20 -0500 Subject: [PATCH 200/624] fix: llava this function takes at least 4 arguments (0 given) --- llama_cpp/llama_chat_format.py | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index eec06b899..875581761 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1840,7 +1840,7 @@ def create_completion(stop): class Llava15ChatHandler: _clip_free = None - def __init__(self, clip_model_path: str, verbose: bool = False): + def __init__(self, clip_model_path: str, verbose: bool = False): import llama_cpp.llava_cpp as llava_cpp self._llava_cpp = llava_cpp @@ -1957,10 +1957,10 @@ def __call__( with suppress_stdout_stderr(disable=self.verbose): embed = ( self._llava_cpp.llava_image_embed_make_with_bytes( - ctx_clip=self.clip_ctx, - n_threads=llama.context_params.n_threads, - image_bytes=c_ubyte_ptr, - image_bytes_length=len(image_bytes), + self.clip_ctx, + llama.context_params.n_threads, + c_ubyte_ptr, + length=len(image_bytes), ) ) try: @@ -1968,10 +1968,10 @@ def __call__( n_past_p = ctypes.pointer(n_past) with suppress_stdout_stderr(disable=self.verbose): self._llava_cpp.llava_eval_image_embed( - ctx_llama=llama.ctx, - embed=embed, - n_batch=llama.n_batch, - n_past=n_past_p, + llama.ctx, + embed, + llama.n_batch, + n_past_p, ) assert llama.n_ctx() >= n_past.value llama.n_tokens = n_past.value From 44558cbd7ae9267559b3c4cf2224ac8ffdb60b78 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:07:33 -0500 Subject: [PATCH 201/624] misc: llava_cpp use ctypes function decorator for binding --- llama_cpp/llava_cpp.py | 62 +++++++++++++++++++++++------------------- 1 file changed, 34 insertions(+), 28 deletions(-) diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index 4eaa9e553..543c87d0b 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -1,6 +1,7 @@ import sys import os import ctypes +import functools from ctypes import ( c_bool, c_char_p, @@ -13,7 +14,7 @@ Structure, ) import pathlib -from typing import List, Union, NewType, Optional +from typing import List, Union, NewType, Optional, TypeVar, Callable, Any import llama_cpp.llama_cpp as llama_cpp @@ -76,6 +77,31 @@ def _load_shared_library(lib_base_name: str): # Load the library _libllava = _load_shared_library(_libllava_base_name) +# ctypes helper + +F = TypeVar("F", bound=Callable[..., Any]) + +def ctypes_function_for_shared_library(lib: ctypes.CDLL): + def ctypes_function( + name: str, argtypes: List[Any], restype: Any, enabled: bool = True + ): + def decorator(f: F) -> F: + if enabled: + func = getattr(lib, name) + func.argtypes = argtypes + func.restype = restype + functools.wraps(f)(func) + return func + else: + return f + + return decorator + + return ctypes_function + + +ctypes_function = ctypes_function_for_shared_library(_libllava) + ################################################ # llava.h @@ -97,49 +123,35 @@ class llava_image_embed(Structure): # /** sanity check for clip <-> llava embed size match */ # LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip); +@ctypes_function("llava_validate_embed_size", [llama_cpp.llama_context_p_ctypes, clip_ctx_p_ctypes], c_bool) def llava_validate_embed_size(ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p, /) -> bool: ... -llava_validate_embed_size = _libllava.llava_validate_embed_size -llava_validate_embed_size.argtypes = [llama_cpp.llama_context_p_ctypes, clip_ctx_p_ctypes] -llava_validate_embed_size.restype = c_bool # /** build an image embed from image file bytes */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); +@ctypes_function("llava_image_embed_make_with_bytes", [clip_ctx_p_ctypes, c_int, POINTER(c_uint8), c_int], POINTER(llava_image_embed)) def llava_image_embed_make_with_bytes(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_bytes: bytes, image_bytes_length: Union[c_int, int], /) -> "_Pointer[llava_image_embed]": ... -llava_image_embed_make_with_bytes = _libllava.llava_image_embed_make_with_bytes -llava_image_embed_make_with_bytes.argtypes = [clip_ctx_p_ctypes, c_int, POINTER(c_uint8), c_int] -llava_image_embed_make_with_bytes.restype = POINTER(llava_image_embed) - # /** build an image embed from a path to an image filename */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); +@ctypes_function("llava_image_embed_make_with_filename", [clip_ctx_p_ctypes, c_int, c_char_p], POINTER(llava_image_embed)) def llava_image_embed_make_with_filename(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes, /) -> "_Pointer[llava_image_embed]": ... -llava_image_embed_make_with_filename = _libllava.llava_image_embed_make_with_filename -llava_image_embed_make_with_filename.argtypes = [clip_ctx_p_ctypes, c_int, c_char_p] -llava_image_embed_make_with_filename.restype = POINTER(llava_image_embed) - # LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); # /** free an embedding made with llava_image_embed_make_* */ +@ctypes_function("llava_image_embed_free", [POINTER(llava_image_embed)], None) def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): ... -llava_image_embed_free = _libllava.llava_image_embed_free -llava_image_embed_free.argtypes = [POINTER(llava_image_embed)] -llava_image_embed_free.restype = None - # /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ # LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); +@ctypes_function("llava_eval_image_embed", [llama_cpp.llama_context_p_ctypes, POINTER(llava_image_embed), c_int, POINTER(c_int)], c_bool) def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointer[llava_image_embed]", n_batch: Union[c_int, int], n_past: "_Pointer[c_int]", /) -> bool: ... -llava_eval_image_embed = _libllava.llava_eval_image_embed -llava_eval_image_embed.argtypes = [llama_cpp.llama_context_p_ctypes, POINTER(llava_image_embed), c_int, POINTER(c_int)] -llava_eval_image_embed.restype = c_bool - ################################################ # clip.h @@ -148,18 +160,12 @@ def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointe # /** load mmproj model */ # CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); +@ctypes_function("clip_model_load", [c_char_p, c_int], clip_ctx_p_ctypes) def clip_model_load(fname: bytes, verbosity: Union[c_int, int], /) -> Optional[clip_ctx_p]: ... -clip_model_load = _libllava.clip_model_load -clip_model_load.argtypes = [c_char_p, c_int] -clip_model_load.restype = clip_ctx_p_ctypes - # /** free mmproj model */ # CLIP_API void clip_free(struct clip_ctx * ctx); +@ctypes_function("clip_free", [clip_ctx_p_ctypes], None) def clip_free(ctx: clip_ctx_p, /): ... - -clip_free = _libllava.clip_free -clip_free.argtypes = [clip_ctx_p_ctypes] -clip_free.restype = None From 78e536dcfe28bde33fb2db414e836369448a2e78 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:14:26 -0500 Subject: [PATCH 202/624] fix: typo --- llama_cpp/llama_chat_format.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 875581761..4fcc7dbf5 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1840,7 +1840,7 @@ def create_completion(stop): class Llava15ChatHandler: _clip_free = None - def __init__(self, clip_model_path: str, verbose: bool = False): + def __init__(self, clip_model_path: str, verbose: bool = False): import llama_cpp.llava_cpp as llava_cpp self._llava_cpp = llava_cpp From dbaba3059d3d854035a675d371559a1c2538820f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:31:11 -0500 Subject: [PATCH 203/624] fix: positional arguments only for low-level api --- llama_cpp/llama_chat_format.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 4fcc7dbf5..69ed6019a 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1960,7 +1960,7 @@ def __call__( self.clip_ctx, llama.context_params.n_threads, c_ubyte_ptr, - length=len(image_bytes), + len(image_bytes), ) ) try: From bf315ee7a9807026b9a13a416e81854a74709052 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:32:11 -0500 Subject: [PATCH 204/624] docs: Update multimodal example --- examples/notebooks/Multimodal.ipynb | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/examples/notebooks/Multimodal.ipynb b/examples/notebooks/Multimodal.ipynb index f1b8e9d15..63e332367 100644 --- a/examples/notebooks/Multimodal.ipynb +++ b/examples/notebooks/Multimodal.ipynb @@ -11,21 +11,21 @@ }, { "cell_type": "code", - "execution_count": 1, + "execution_count": 10, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ - "{'text': 'Llama C++'}\n" + "{'text': 'Llama +'}\n" ] } ], "source": [ "from openai import OpenAI\n", "\n", - "client = OpenAI(base_url=\"http://100.64.159.73:8000/v1\", api_key=\"sk-1234\")\n", + "client = OpenAI(base_url=\"http://localhost:8000/v1\", api_key=\"llama.cpp\")\n", "response = client.chat.completions.create(\n", " model=\"gpt-4-vision-preview\",\n", " messages=[\n", From 79c649c2d13ffaccd8cbd43788668ceb9294bbaa Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:34:45 -0500 Subject: [PATCH 205/624] docs: Update multimodal example --- examples/notebooks/Multimodal.ipynb | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) diff --git a/examples/notebooks/Multimodal.ipynb b/examples/notebooks/Multimodal.ipynb index 63e332367..def68df00 100644 --- a/examples/notebooks/Multimodal.ipynb +++ b/examples/notebooks/Multimodal.ipynb @@ -11,14 +11,14 @@ }, { "cell_type": "code", - "execution_count": 10, + "execution_count": 13, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ - "{'text': 'Llama +'}\n" + "{'text': 'Llama C++'}\n" ] } ], @@ -42,7 +42,17 @@ " ],\n", " }\n", " ],\n", - " response_format={ \"type\": \"json_object\" }\n", + " response_format={ \n", + " \"type\": \"json_object\",\n", + " \"schema\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"text\": {\n", + " \"type\": \"string\"\n", + " }\n", + " }\n", + " }\n", + " }\n", ")\n", "import json\n", "print(json.loads(response.choices[0].message.content))" From a57d5dff867ae86f86c1f06f6a85a590f7c88cf2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:37:43 -0500 Subject: [PATCH 206/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 47bb7b48c..a33e6a0d2 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 47bb7b48c7cec9d8f57d56812ce811ec130b89a3 +Subproject commit a33e6a0d2a66104ea9a906bdbf8a94d050189d91 From 9558ce7878bebde52025cf985f3975c8c1d0cee6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:40:58 -0500 Subject: [PATCH 207/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 5a934dc26..fe20dba28 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -256,12 +256,14 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors -# LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_XS = 22, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ3_S = 26, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -285,12 +287,14 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19 LLAMA_FTYPE_MOSTLY_IQ2_XS = 20 LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 -LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22 +LLAMA_FTYPE_MOSTLY_IQ3_XS = 22 LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 LLAMA_FTYPE_MOSTLY_IQ1_S = 24 LLAMA_FTYPE_MOSTLY_IQ4_NL = 25 LLAMA_FTYPE_MOSTLY_IQ3_S = 26 LLAMA_FTYPE_MOSTLY_IQ3_M = 27 +LLAMA_FTYPE_MOSTLY_IQ2_S = 28 +LLAMA_FTYPE_MOSTLY_IQ2_M = 29 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { From afe1e445c9eb8d4cf53fdee370f3758b8c1734ba Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:43:24 -0500 Subject: [PATCH 208/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 036dbe039..b6d51a664 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.52] + +- feat: Update llama.cpp to ggerganov/llama.cpp@a33e6a0d2a66104ea9a906bdbf8a94d050189d91 +- fix: Llava15ChatHandler (this function takes at least 4 arguments) by @abetlen in 8383a9e5620f5df5a88f62da16813eac200dd706 + ## [0.2.51] - feat: Update llama.cpp to ggerganov/llama.cpp@c39373398803c669056304090050fe3f44b41bf9 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 83a70cf76..949aed0bf 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.51" \ No newline at end of file +__version__ = "0.2.52" \ No newline at end of file From b3e358dee41b1bbeec856354ab58f2bc8c7003d8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 26 Feb 2024 11:58:33 -0500 Subject: [PATCH 209/624] docs: Add example of local image loading to README --- README.md | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 insertions(+) diff --git a/README.md b/README.md index a3f82e43e..e3d76436a 100644 --- a/README.md +++ b/README.md @@ -468,6 +468,38 @@ Then you'll need to use a custom chat handler to load the clip model and process ) ``` +
+Loading a Local Image + +Images can be passed as base64 encoded data URIs. The following example demonstrates how to do this. + +```python +import base64 + +def image_to_base64_data_uri(file_path): + with open(file_path, "rb") as img_file: + base64_data = base64.b64encode(img_file.read()).decode('utf-8') + return f"data:image/png;base64,{base64_data}" + +# Replace 'file_path.png' with the actual path to your PNG file +file_path = 'file_path.png' +data_uri = image_to_base64_data_uri(file_path) + +messages = [ + {"role": "system", "content": "You are an assistant who perfectly describes images."}, + { + "role": "user", + "content": [ + {"type": "image_url", "image_url": {"url": data_uri }}, + {"type" : "text", "text": "Describe this image in detail please."} + ] + } +] + +``` + +
+ ### Speculative Decoding `llama-cpp-python` supports speculative decoding which allows the model to generate completions based on a draft model. From 4d574bd765912534304b196b838a380a88e602a3 Mon Sep 17 00:00:00 2001 From: Andrei Date: Mon, 26 Feb 2024 14:35:08 -0500 Subject: [PATCH 210/624] feat(server): Add support for pulling models from Huggingface Hub (#1222) * Basic support for hf pull on server * Add hf_model_repo_id setting * Update README --- README.md | 6 ++++++ llama_cpp/server/model.py | 15 +++++++++++++-- llama_cpp/server/settings.py | 5 +++++ 3 files changed, 24 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index e3d76436a..8d0980e0d 100644 --- a/README.md +++ b/README.md @@ -577,6 +577,12 @@ python3 -m llama_cpp.server --model models/7B/llama-model.gguf --chat_format cha That will format the prompt according to how model expects it. You can find the prompt format in the model card. For possible options, see [llama_cpp/llama_chat_format.py](llama_cpp/llama_chat_format.py) and look for lines starting with "@register_chat_format". +If you have `huggingface-hub` installed, you can also use the `--hf_model_repo_id` flag to load a model from the Hugging Face Hub. + +```bash +python3 -m llama_cpp.server --hf_model_repo_id Qwen/Qwen1.5-0.5B-Chat-GGUF --model '*q8_0.gguf' +``` + ### Web Server Features - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 5308dc2a8..816d089f3 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -120,9 +120,20 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: kv_overrides[key] = float(value) else: raise ValueError(f"Unknown value type {value_type}") + + import functools - _model = llama_cpp.Llama( - model_path=settings.model, + kwargs = {} + + if settings.hf_model_repo_id is not None: + create_fn = functools.partial(llama_cpp.Llama.from_pretrained, repo_id=settings.hf_model_repo_id, filename=settings.model) + else: + create_fn = llama_cpp.Llama + kwargs["model_path"] = settings.model + + + _model = create_fn( + **kwargs, # Model Params n_gpu_layers=settings.n_gpu_layers, main_gpu=settings.main_gpu, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 8ac8df881..bd806748c 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -143,6 +143,11 @@ class ModelSettings(BaseSettings): default=None, description="The model name or path to a pretrained HuggingFace tokenizer model. Same as you would pass to AutoTokenizer.from_pretrained().", ) + # Loading from HuggingFace Model Hub + hf_model_repo_id: Optional[str] = Field( + default=None, + description="The model repo id to use for the HuggingFace tokenizer model.", + ) # Speculative Decoding draft_model: Optional[str] = Field( default=None, From fea33c9b94b0ee3b1f8787e28226d8829c2cff36 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 27 Feb 2024 12:22:17 -0500 Subject: [PATCH 211/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 5 +++++ vendor/llama.cpp | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index fe20dba28..8151709b2 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -264,6 +264,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -295,6 +296,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_MOSTLY_IQ3_M = 27 LLAMA_FTYPE_MOSTLY_IQ2_S = 28 LLAMA_FTYPE_MOSTLY_IQ2_M = 29 +LLAMA_FTYPE_MOSTLY_IQ4_XS = 30 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { @@ -548,6 +550,7 @@ class llama_model_params(ctypes.Structure): # float yarn_beta_fast; // YaRN low correction dim # float yarn_beta_slow; // YaRN high correction dim # uint32_t yarn_orig_ctx; // YaRN original context size +# float defrag_thold; // defragment the KV cache if holes/size > thold, < 0 disabled (default) # ggml_backend_sched_eval_callback cb_eval; # void * cb_eval_user_data; @@ -580,6 +583,7 @@ class llama_context_params(ctypes.Structure): yarn_beta_fast (float): YaRN low correction dim yarn_beta_slow (float): YaRN high correction dim yarn_orig_ctx (int): YaRN original context size + defrag_thold (float): defragment the KV cache if holes/size > thold, < 0 disabled (default) cb_eval (ggml_backend_sched_eval_callback): callback for scheduling eval cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache @@ -605,6 +609,7 @@ class llama_context_params(ctypes.Structure): ("yarn_beta_fast", ctypes.c_float), ("yarn_beta_slow", ctypes.c_float), ("yarn_orig_ctx", ctypes.c_uint32), + ("defrag_thold", ctypes.c_float), ("cb_eval", ggml_backend_sched_eval_callback), ("cb_eval_user_data", ctypes.c_void_p), ("type_k", ctypes.c_int), diff --git a/vendor/llama.cpp b/vendor/llama.cpp index a33e6a0d2..cb49e0f8c 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit a33e6a0d2a66104ea9a906bdbf8a94d050189d91 +Subproject commit cb49e0f8c906e5da49e9f6d64a57742a9a241c6a From c36ab15e68b9686c6031552b4944f6d4cb09aebe Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Wed, 28 Feb 2024 07:30:31 +0100 Subject: [PATCH 212/624] fix: eos/bos_token set correctly for Jinja2ChatFormatter and automatic chat formatter (#1230) The token strings were not correctly retrieved (empty). --- llama_cpp/llama.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 81bfce40d..d1bac9b8f 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -408,8 +408,8 @@ def __init__( except: bos_token_id = self.token_bos() - eos_token = self.detokenize([eos_token_id]).decode("utf-8") - bos_token = self.detokenize([bos_token_id]).decode("utf-8") + eos_token = self._model.token_get_text(eos_token_id) + bos_token = self._model.token_get_text(bos_token_id) if self.verbose: print(f"Using chat template: {template}", file=sys.stderr) From ffcd4b2636c56b7ae383a6dfb1058b3af01b4769 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 28 Feb 2024 01:38:32 -0500 Subject: [PATCH 213/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index b6d51a664..d7a96a9e2 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.53] + +- feat: Update llama.cpp to ggerganov/llama.cpp@cb49e0f8c906e5da49e9f6d64a57742a9a241c6a +- fix: eos/bos_token set correctly for Jinja2ChatFormatter and automatic chat formatter by @CISC in #1230 + ## [0.2.52] - feat: Update llama.cpp to ggerganov/llama.cpp@a33e6a0d2a66104ea9a906bdbf8a94d050189d91 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 949aed0bf..aa0536c8b 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.52" \ No newline at end of file +__version__ = "0.2.53" \ No newline at end of file From 0d37ce52b1ec9da5a6c61159b0c8de78ed84efcb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 28 Feb 2024 14:27:16 -0500 Subject: [PATCH 214/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 144 +---------------------------------- llama_cpp/server/settings.py | 4 +- vendor/llama.cpp | 2 +- 3 files changed, 5 insertions(+), 145 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 8151709b2..ce65449ab 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -111,6 +111,7 @@ class CtypesRef(Generic[CtypesCData]): F = TypeVar("F", bound=Callable[..., Any]) + def ctypes_function_for_shared_library(lib: ctypes.CDLL): def ctypes_function( name: str, argtypes: List[Any], restype: Any, enabled: bool = True @@ -938,18 +939,6 @@ def llama_supports_gpu_offload() -> bool: ... -# LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); -@ctypes_function("llama_mmap_supported", [], ctypes.c_bool) -def llama_mmap_supported() -> bool: - ... - - -# LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); -@ctypes_function("llama_mlock_supported", [], ctypes.c_bool) -def llama_mlock_supported() -> bool: - ... - - # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: @@ -1158,47 +1147,6 @@ def llama_model_quantize( ... -# // Apply a LoRA adapter to a loaded model -# // path_base_model is the path to a higher quality model to use as a base for -# // the layers modified by the adapter. Can be NULL to use the current loaded model. -# // The model needs to be reloaded before applying a new adapter, otherwise the adapter -# // will be applied on top of the previous one -# // Returns 0 on success -# LLAMA_API DEPRECATED(int32_t llama_apply_lora_from_file( -# struct llama_context * ctx, -# const char * path_lora, -# float scale, -# const char * path_base_model, -# int32_t n_threads), -# "use llama_model_apply_lora_from_file instead"); -@ctypes_function( - "llama_apply_lora_from_file", - [ - llama_context_p_ctypes, - ctypes.c_char_p, - ctypes.c_float, - ctypes.c_char_p, - ctypes.c_int32, - ], - ctypes.c_int32, -) -def llama_apply_lora_from_file( - ctx: llama_context_p, - path_lora: Union[ctypes.c_char_p, bytes], - scale: Union[ctypes.c_float, float], - path_base_model: Union[ctypes.c_char_p, bytes], - n_threads: Union[ctypes.c_int32, int], - /, -) -> int: - """Apply a LoRA adapter to a loaded model - path_base_model is the path to a higher quality model to use as a base for - the layers modified by the adapter. Can be NULL to use the current loaded model. - The model needs to be reloaded before applying a new adapter, otherwise the adapter - will be applied on top of the previous one - Returns 0 on success""" - ... - - # LLAMA_API int32_t llama_model_apply_lora_from_file( # const struct llama_model * model, # const char * path_lora, @@ -1220,7 +1168,7 @@ def llama_model_apply_lora_from_file( model: llama_model_p, path_lora: Union[ctypes.c_char_p, bytes], scale: Union[ctypes.c_float, float], - path_base_model: Union[ctypes.c_char_p, bytes], + path_base_model: Union[ctypes.c_char_p, bytes, None], n_threads: Union[ctypes.c_int32, int], /, ) -> int: @@ -1647,72 +1595,6 @@ def llama_save_session_file( # // -# // Run the llama inference to obtain the logits and probabilities for the next token(s). -# // tokens + n_tokens is the provided batch of new tokens to process -# // n_past is the number of tokens to use from previous eval calls -# // Returns 0 on success -# // DEPRECATED: use llama_decode() instead -# LLAMA_API DEPRECATED(int llama_eval( -# struct llama_context * ctx, -# llama_token * tokens, -# int32_t n_tokens, -# int32_t n_past), -# "use llama_decode() instead"); -@ctypes_function( - "llama_eval", - [ - llama_context_p_ctypes, - llama_token_p, - ctypes.c_int32, - ctypes.c_int32, - ], - ctypes.c_int, -) -def llama_eval( - ctx: llama_context_p, - tokens: CtypesArray[llama_token], - n_tokens: Union[ctypes.c_int, int], - n_past: Union[ctypes.c_int, int], - /, -) -> int: - """Run the llama inference to obtain the logits and probabilities for the next token(s). - tokens + n_tokens is the provided batch of new tokens to process - n_past is the number of tokens to use from previous eval calls - Returns 0 on success - DEPRECATED: use llama_decode() instead""" - ... - - -# // Same as llama_eval, but use float matrix input directly. -# // DEPRECATED: use llama_decode() instead -# LLAMA_API DEPRECATED(int llama_eval_embd( -# struct llama_context * ctx, -# float * embd, -# int32_t n_tokens, -# int32_t n_past), -# "use llama_decode() instead"); -@ctypes_function( - "llama_eval_embd", - [ - llama_context_p_ctypes, - ctypes.POINTER(ctypes.c_float), - ctypes.c_int32, - ctypes.c_int32, - ], - ctypes.c_int, -) -def llama_eval_embd( - ctx: llama_context_p, - embd: CtypesArray[ctypes.c_float], - n_tokens: Union[ctypes.c_int, int], - n_past: Union[ctypes.c_int, int], - /, -) -> int: - """Same as llama_eval, but use float matrix input directly. - DEPRECATED: use llama_decode() instead""" - ... - - # // Return batch for single sequence of tokens starting at pos_0 # // # // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it @@ -2474,28 +2356,6 @@ def llama_sample_temp( ... -# LLAMA_API DEPRECATED(void llama_sample_temperature( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float temp), -# "use llama_sample_temp instead"); -@ctypes_function( - "llama_sample_temperature", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float], - None, -) -def llama_sample_temperature( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - temp: Union[ctypes.c_float, float], - /, -): - """use llama_sample_temp instead""" - ... - - # /// @details Apply constraints from grammar # LLAMA_API void llama_sample_grammar( # struct llama_context * ctx, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index bd806748c..292d7ebf0 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -45,11 +45,11 @@ class ModelSettings(BaseSettings): default=False, description="Whether to only return the vocabulary." ) use_mmap: bool = Field( - default=llama_cpp.llama_mmap_supported(), + default=llama_cpp.llama_supports_mmap(), description="Use mmap.", ) use_mlock: bool = Field( - default=llama_cpp.llama_mlock_supported(), + default=llama_cpp.llama_supports_mlock(), description="Use mlock.", ) kv_overrides: Optional[List[str]] = Field( diff --git a/vendor/llama.cpp b/vendor/llama.cpp index cb49e0f8c..08c5ee87e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit cb49e0f8c906e5da49e9f6d64a57742a9a241c6a +Subproject commit 08c5ee87e4cceb603ecceac90734fcdade57311b From 727d60c28a76312c3cbed412f290af8cad2f92ec Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 28 Feb 2024 14:27:40 -0500 Subject: [PATCH 215/624] misc: Format --- llama_cpp/server/app.py | 50 +++++++++++++++++++----------------- llama_cpp/server/errors.py | 4 +-- llama_cpp/server/model.py | 19 ++++++++------ llama_cpp/server/settings.py | 8 +++--- llama_cpp/server/types.py | 2 +- 5 files changed, 44 insertions(+), 39 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 7a1391df9..ec9280986 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -199,8 +199,8 @@ async def authenticate( @router.post( "/v1/completions", summary="Completion", - dependencies=[Depends(authenticate)], - response_model= Union[ + dependencies=[Depends(authenticate)], + response_model=Union[ llama_cpp.CreateCompletionResponse, str, ], @@ -211,19 +211,19 @@ async def authenticate( "application/json": { "schema": { "anyOf": [ - {"$ref": "#/components/schemas/CreateCompletionResponse"} + {"$ref": "#/components/schemas/CreateCompletionResponse"} ], "title": "Completion response, when stream=False", } }, - "text/event-stream":{ - "schema": { - "type": "string", - "title": "Server Side Streaming response, when stream=True. " + - "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 - "example": """data: {... see CreateCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""" + "text/event-stream": { + "schema": { + "type": "string", + "title": "Server Side Streaming response, when stream=True. " + + "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 + "example": """data: {... see CreateCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""", } - } + }, }, } }, @@ -290,7 +290,7 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: inner_send_chan=send_chan, iterator=iterator(), ), - sep='\n', + sep="\n", ) else: return iterator_or_completion @@ -310,10 +310,10 @@ async def create_embedding( @router.post( - "/v1/chat/completions", summary="Chat", dependencies=[Depends(authenticate)], - response_model= Union[ - llama_cpp.ChatCompletion, str - ], + "/v1/chat/completions", + summary="Chat", + dependencies=[Depends(authenticate)], + response_model=Union[llama_cpp.ChatCompletion, str], responses={ "200": { "description": "Successful Response", @@ -321,19 +321,21 @@ async def create_embedding( "application/json": { "schema": { "anyOf": [ - {"$ref": "#/components/schemas/CreateChatCompletionResponse"} + { + "$ref": "#/components/schemas/CreateChatCompletionResponse" + } ], "title": "Completion response, when stream=False", } }, - "text/event-stream":{ - "schema": { - "type": "string", - "title": "Server Side Streaming response, when stream=True" + - "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 - "example": """data: {... see CreateChatCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""" + "text/event-stream": { + "schema": { + "type": "string", + "title": "Server Side Streaming response, when stream=True" + + "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 + "example": """data: {... see CreateChatCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""", } - } + }, }, } }, @@ -383,7 +385,7 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: inner_send_chan=send_chan, iterator=iterator(), ), - sep='\n', + sep="\n", ) else: return iterator_or_completion diff --git a/llama_cpp/server/errors.py b/llama_cpp/server/errors.py index 9d3d35598..fbf9fd80d 100644 --- a/llama_cpp/server/errors.py +++ b/llama_cpp/server/errors.py @@ -22,6 +22,7 @@ CreateChatCompletionRequest, ) + class ErrorResponse(TypedDict): """OpenAI style error response""" @@ -75,7 +76,7 @@ def context_length_exceeded( (completion_tokens or 0) + prompt_tokens, prompt_tokens, completion_tokens, - ), # type: ignore + ), # type: ignore type="invalid_request_error", param="messages", code="context_length_exceeded", @@ -207,4 +208,3 @@ async def custom_route_handler(request: Request) -> Response: ) return custom_route_handler - diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 816d089f3..dace8d547 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -88,15 +88,15 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: assert ( settings.hf_tokenizer_config_path is not None ), "hf_tokenizer_config_path must be set for hf-tokenizer-config" - chat_handler = ( - llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_completion_handler( - json.load(open(settings.hf_tokenizer_config_path)) - ) + chat_handler = llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_completion_handler( + json.load(open(settings.hf_tokenizer_config_path)) ) tokenizer: Optional[llama_cpp.BaseLlamaTokenizer] = None if settings.hf_pretrained_model_name_or_path is not None: - tokenizer = llama_tokenizer.LlamaHFTokenizer.from_pretrained(settings.hf_pretrained_model_name_or_path) + tokenizer = llama_tokenizer.LlamaHFTokenizer.from_pretrained( + settings.hf_pretrained_model_name_or_path + ) draft_model = None if settings.draft_model is not None: @@ -120,17 +120,20 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: kv_overrides[key] = float(value) else: raise ValueError(f"Unknown value type {value_type}") - + import functools kwargs = {} if settings.hf_model_repo_id is not None: - create_fn = functools.partial(llama_cpp.Llama.from_pretrained, repo_id=settings.hf_model_repo_id, filename=settings.model) + create_fn = functools.partial( + llama_cpp.Llama.from_pretrained, + repo_id=settings.hf_model_repo_id, + filename=settings.model, + ) else: create_fn = llama_cpp.Llama kwargs["model_path"] = settings.model - _model = create_fn( **kwargs, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 292d7ebf0..daa913fac 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -74,7 +74,9 @@ class ModelSettings(BaseSettings): ge=0, description="The number of threads to use when batch processing.", ) - rope_scaling_type: int = Field(default=llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) + rope_scaling_type: int = Field( + default=llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED + ) rope_freq_base: float = Field(default=0.0, description="RoPE base frequency") rope_freq_scale: float = Field( default=0.0, description="RoPE frequency scaling factor" @@ -193,6 +195,4 @@ class Settings(ServerSettings, ModelSettings): class ConfigFileSettings(ServerSettings): """Configuration file format settings.""" - models: List[ModelSettings] = Field( - default=[], description="Model configs" - ) + models: List[ModelSettings] = Field(default=[], description="Model configs") diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index f0827d762..9a4b81e09 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -110,7 +110,7 @@ class CreateCompletionRequest(BaseModel): default=None, description="A suffix to append to the generated text. If None, no suffix is appended. Useful for chatbots.", ) - max_tokens: Optional[int] = Field( + max_tokens: Optional[int] = Field( default=16, ge=0, description="The maximum number of tokens to generate." ) temperature: float = temperature_field From 8c71725d5345b2ba325e6d28209875057c5873e6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 28 Feb 2024 14:37:07 -0500 Subject: [PATCH 216/624] fix: Remove deprecated cfg sampling functions --- llama_cpp/_internals.py | 22 +--------------------- llama_cpp/llama_cpp.py | 29 ----------------------------- 2 files changed, 1 insertion(+), 50 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 13bf1fb06..22d0bef8a 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -357,21 +357,6 @@ def sample_repetition_penalties( penalty_present, ) - def sample_classifier_free_guidance( - self, - candidates: "_LlamaTokenDataArray", - guidance_ctx: "_LlamaContext", - scale: float, - ): - assert self.ctx is not None - assert guidance_ctx.ctx is not None - llama_cpp.llama_sample_classifier_free_guidance( - self.ctx, - llama_cpp.byref(candidates.candidates), - guidance_ctx.ctx, - scale, - ) - def sample_softmax(self, candidates: "_LlamaTokenDataArray"): assert self.ctx is not None llama_cpp.llama_sample_softmax( @@ -720,7 +705,7 @@ def prev_str(self, ctx_main: _LlamaContext, n: int) -> str: return ctx_main.model.detokenize(self.prev[-n:]).decode("utf-8") def sample( - self, ctx_main: _LlamaContext, ctx_cfg: Optional[_LlamaContext] = None, idx: int = 0, logits_array: Optional[npt.NDArray[np.single]] = None + self, ctx_main: _LlamaContext, idx: int = 0, logits_array: Optional[npt.NDArray[np.single]] = None ): n_vocab = ctx_main.model.n_vocab() id: int = 0 @@ -741,11 +726,6 @@ def sample( ) # TODO: Only create this once token_data_array.copy_logits(logits_array) - if ctx_cfg is not None: - ctx_main.sample_classifier_free_guidance( - token_data_array, ctx_cfg, self.params.cfg_scale - ) - # apply penalties if len(self.prev) > 0: nl_token = ctx_main.model.token_nl() diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index ce65449ab..038a6f8b2 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2129,35 +2129,6 @@ def llama_sample_apply_guidance( ... -# LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# struct llama_context * guidance_ctx, -# float scale), -# "use llama_sample_apply_guidance() instead"); -@ctypes_function( - "llama_sample_classifier_free_guidance", - [ - llama_context_p_ctypes, - llama_token_data_array_p, - llama_context_p_ctypes, - ctypes.c_float, - ], - None, -) -def llama_sample_classifier_free_guidance( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - guidance_ctx: llama_context_p, - scale: Union[ctypes.c_float, float], - /, -): - """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" - ... - - # /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. # LLAMA_API void llama_sample_softmax( # struct llama_context * ctx, From cf1fdd8a9a9f461c095b45c2797fb9f19576ae9c Mon Sep 17 00:00:00 2001 From: Douglas Hanley Date: Thu, 29 Feb 2024 12:55:50 -0600 Subject: [PATCH 217/624] docs: fix typo in README.md embeddings example. (#1232) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 8d0980e0d..32c624f49 100644 --- a/README.md +++ b/README.md @@ -525,7 +525,7 @@ To generate text embeddings use [`create_embedding`](http://localhost:8000/api-r ```python import llama_cpp -llm = llama_cpp.Llama(model_path="path/to/model.gguf", embeddings=True) +llm = llama_cpp.Llama(model_path="path/to/model.gguf", embedding=True) embeddings = llm.create_embedding("Hello, world!") From f062a7f51d9826df36b605ec8664df6b84a70a1b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 1 Mar 2024 12:57:16 -0500 Subject: [PATCH 218/624] feat: Update llama.cpp --- llama_cpp/llama.py | 4 ---- llama_cpp/llama_cpp.py | 10 +++------- vendor/llama.cpp | 2 +- 3 files changed, 4 insertions(+), 12 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index d1bac9b8f..70498f39c 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -86,7 +86,6 @@ def __init__( yarn_beta_fast: float = 32.0, yarn_beta_slow: float = 1.0, yarn_orig_ctx: int = 0, - mul_mat_q: bool = True, logits_all: bool = False, embedding: bool = False, offload_kqv: bool = True, @@ -291,7 +290,6 @@ def __init__( yarn_beta_slow if yarn_beta_slow != 0.0 else 0 ) self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0 - self.context_params.mul_mat_q = mul_mat_q self.context_params.logits_all = ( logits_all if draft_model is None else True ) # Must be set to True for speculative decoding @@ -1724,7 +1722,6 @@ def __getstate__(self): yarn_beta_fast=self.context_params.yarn_beta_fast, yarn_beta_slow=self.context_params.yarn_beta_slow, yarn_orig_ctx=self.context_params.yarn_orig_ctx, - mul_mat_q=self.context_params.mul_mat_q, logits_all=self.context_params.logits_all, embedding=self.context_params.embedding, # Sampling Params @@ -1768,7 +1765,6 @@ def __setstate__(self, state): yarn_beta_fast=state["yarn_beta_fast"], yarn_beta_slow=state["yarn_beta_slow"], yarn_orig_ctx=state["yarn_orig_ctx"], - mul_mat_q=state["mul_mat_q"], logits_all=state["logits_all"], embedding=state["embedding"], # Sampling Params diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 038a6f8b2..15932561c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -559,9 +559,7 @@ class llama_model_params(ctypes.Structure): # enum ggml_type type_k; // data type for K cache # enum ggml_type type_v; // data type for V cache - # // Keep the booleans together to avoid misalignment during copy-by-value. -# bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) # bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) # bool embedding; // embedding mode only # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU @@ -589,7 +587,6 @@ class llama_context_params(ctypes.Structure): cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache type_v (int): data type for V cache - mul_mat_q (bool): if true, use experimental mul_mat_q kernels (DEPRECATED - always true) logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) embedding (bool): embedding mode only offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU @@ -615,7 +612,6 @@ class llama_context_params(ctypes.Structure): ("cb_eval_user_data", ctypes.c_void_p), ("type_k", ctypes.c_int), ("type_v", ctypes.c_int), - ("mul_mat_q", ctypes.c_bool), ("logits_all", ctypes.c_bool), ("embedding", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), @@ -1519,11 +1515,11 @@ def llama_copy_state_data( ... -# Set the state reading from the specified address -# Returns the number of bytes read +# // Set the state reading from the specified address +# // Returns the number of bytes read # LLAMA_API size_t llama_set_state_data( # struct llama_context * ctx, -# uint8_t * src); +# const uint8_t * src); @ctypes_function( "llama_set_state_data", [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 08c5ee87e..c2224f003 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 08c5ee87e4cceb603ecceac90734fcdade57311b +Subproject commit c2224f003bf9cf558b1a3c57033563e11a4de9a5 From 97aa3a153debe25df874055a6f96db0ac943091c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 1 Mar 2024 13:10:25 -0500 Subject: [PATCH 219/624] docs: Add information re: auto chat formats. Closes #1236 --- README.md | 11 ++++++++++- llama_cpp/llama.py | 4 +++- 2 files changed, 13 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 32c624f49..1d296e905 100644 --- a/README.md +++ b/README.md @@ -286,7 +286,16 @@ By default [`from_pretrained`](https://llama-cpp-python.readthedocs.io/en/latest The high-level API also provides a simple interface for chat completion. -Note that `chat_format` option must be set for the particular model you are using. +Chat completion requires that the model know how to format the messages into a single prompt. +The `Llama` class does this using pre-registered chat formats (ie. `chatml`, `llama-2`, `gemma`, etc) or by providing a custom chat handler object. + +The model will will format the messages into a single prompt using the following order of precedence: + - Use the `chat_handler` if provided + - Use the `chat_format` if provided + - Use the `tokenizer.chat_template` from the `gguf` model's metadata (should work for most new models, older models may not have this) + - else, fallback to the `llama-2` chat format + +Set `verbose=True` to see the selected chat format. ```python >>> from llama_cpp import Llama diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 70498f39c..108a4cfa9 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -410,7 +410,7 @@ def __init__( bos_token = self._model.token_get_text(bos_token_id) if self.verbose: - print(f"Using chat template: {template}", file=sys.stderr) + print(f"Using gguf chat template: {template}", file=sys.stderr) print(f"Using chat eos_token: {eos_token}", file=sys.stderr) print(f"Using chat bos_token: {bos_token}", file=sys.stderr) @@ -420,6 +420,8 @@ def __init__( if self.chat_format is None and self.chat_handler is None: self.chat_format = "llama-2" + if self.verbose: + print(f"Using fallback chat format: {chat_format}", file=sys.stderr) @property def ctx(self) -> llama_cpp.llama_context_p: From d5df431278433b580e52222dbf4174f5102585b1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 1 Mar 2024 13:15:16 -0500 Subject: [PATCH 220/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index d7a96a9e2..375c6efa7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.54] + +- feat: Update llama.cpp to ggerganov/llama.cpp@cb49e0f8c906e5da49e9f6d64a57742a9a241c6a +- docs: fix typo in README.md embeddings example by @iamlemec in #1232 + ## [0.2.53] - feat: Update llama.cpp to ggerganov/llama.cpp@cb49e0f8c906e5da49e9f6d64a57742a9a241c6a diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index aa0536c8b..a9a82220d 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.53" \ No newline at end of file +__version__ = "0.2.54" \ No newline at end of file From 0e70984fb69d621c191913bf870b7d9201bcc3d5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 2 Mar 2024 22:20:04 -0500 Subject: [PATCH 221/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 36 ++++++++++++++++++++++++++++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 35 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 15932561c..88ba41ca8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -148,6 +148,12 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa ctypes.c_bool, ctypes.c_void_p, ctypes.c_bool, ctypes.c_void_p ) +# // Abort callback +# // If not NULL, called before ggml computation +# // If it returns true, the computation is aborted +# typedef bool (*ggml_abort_callback)(void * data); +ggml_abort_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p) + # llama.h bindings _lib.llama_max_devices.argtypes = [] @@ -560,10 +566,16 @@ class llama_model_params(ctypes.Structure): # enum ggml_type type_v; // data type for V cache # // Keep the booleans together to avoid misalignment during copy-by-value. -# bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) +# bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) # bool embedding; // embedding mode only # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU # bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) + +# // Abort callback +# // if it returns true, execution of llama_decode() will be aborted +# // currently works only with CPU execution +# ggml_abort_callback abort_callback; +# void * abort_callback_data; # }; class llama_context_params(ctypes.Structure): """Parameters for llama_context @@ -591,6 +603,8 @@ class llama_context_params(ctypes.Structure): embedding (bool): embedding mode only offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU do_pooling (bool): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) + abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted + abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback """ _fields_ = [ @@ -616,6 +630,8 @@ class llama_context_params(ctypes.Structure): ("embedding", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), ("do_pooling", ctypes.c_bool), + ("abort_callback", ggml_abort_callback), + ("abort_callback_data", ctypes.c_void_p), ] @@ -1703,8 +1719,24 @@ def llama_set_n_threads( """ ... +# // Set abort callback +# LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); +@ctypes_function( + "llama_set_abort_callback", + [llama_context_p_ctypes, ggml_abort_callback, ctypes.c_void_p], + None, +) +def llama_set_abort_callback( + ctx: llama_context_p, + abort_callback: Callable[[ctypes.c_void_p], None], + abort_callback_data: ctypes.c_void_p, + /, +): + """Set abort callback""" + ... + -# // Token logits obtained from the last call to llama_eval() +# // Token logits obtained from the last call to llama_decode() # // The logits for the last token are stored in the last row # // Logits for which llama_batch.logits[i] == 0 are undefined # // Rows: n_tokens provided with llama_batch diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c2224f003..973113429 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c2224f003bf9cf558b1a3c57033563e11a4de9a5 +Subproject commit 9731134296af3a6839cd682e51d9c2109a871de5 From 663659f7301963e0a3e98662e14668a6632c6295 Mon Sep 17 00:00:00 2001 From: Kenneth Hoste Date: Sun, 3 Mar 2024 04:20:41 +0100 Subject: [PATCH 222/624] docs: fix small typo in README: 'model know how' -> 'model knows how' (#1244) Co-authored-by: Andrei --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 1d296e905..3323f3899 100644 --- a/README.md +++ b/README.md @@ -286,7 +286,7 @@ By default [`from_pretrained`](https://llama-cpp-python.readthedocs.io/en/latest The high-level API also provides a simple interface for chat completion. -Chat completion requires that the model know how to format the messages into a single prompt. +Chat completion requires that the model knows how to format the messages into a single prompt. The `Llama` class does this using pre-registered chat formats (ie. `chatml`, `llama-2`, `gemma`, etc) or by providing a custom chat handler object. The model will will format the messages into a single prompt using the following order of precedence: From 13177aae0f674100f7a7d23c54fc9f14012bf6a2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 2 Mar 2024 22:46:40 -0500 Subject: [PATCH 223/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 375c6efa7..e16a6dfcc 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.55] + +- feat: Update llama.cpp to ggerganov/9731134296af3a6839cd682e51d9c2109a871de5 +- docs: fix small typo in README: 'model know how' -> 'model knows how' by @boegel in #1244 + ## [0.2.54] - feat: Update llama.cpp to ggerganov/llama.cpp@cb49e0f8c906e5da49e9f6d64a57742a9a241c6a diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index a9a82220d..519ab51b7 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.54" \ No newline at end of file +__version__ = "0.2.55" \ No newline at end of file From 87a6e5797eb7b0cd63ad27c528fb950c80c84ad8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 3 Mar 2024 11:27:04 -0500 Subject: [PATCH 224/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 14 +++++++++----- vendor/llama.cpp | 2 +- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 88ba41ca8..08adfe205 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -320,10 +320,12 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN # enum llama_pooling_type { +# LLAMA_POOLING_TYPE_UNSPECIFIED = -1, # LLAMA_POOLING_TYPE_NONE = 0, # LLAMA_POOLING_TYPE_MEAN = 1, # LLAMA_POOLING_TYPE_CLS = 2, # }; +LLAMA_POOLING_TYPE_UNSPECIFIED = -1 LLAMA_POOLING_TYPE_NONE = 0 LLAMA_POOLING_TYPE_MEAN = 1 LLAMA_POOLING_TYPE_CLS = 2 @@ -547,7 +549,10 @@ class llama_model_params(ctypes.Structure): # uint32_t n_batch; // prompt processing maximum batch size # uint32_t n_threads; // number of threads to use for generation # uint32_t n_threads_batch; // number of threads to use for batch processing -# int32_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` + +# enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` +# enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id +# // (ignored if no pooling layer) # // ref: https://github.com/ggerganov/llama.cpp/pull/2054 # float rope_freq_base; // RoPE base frequency, 0 = from model @@ -569,7 +574,6 @@ class llama_model_params(ctypes.Structure): # bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) # bool embedding; // embedding mode only # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU -# bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) # // Abort callback # // if it returns true, execution of llama_decode() will be aborted @@ -587,6 +591,7 @@ class llama_context_params(ctypes.Structure): n_threads (int): number of threads to use for generation n_threads_batch (int): number of threads to use for batch processing rope_scaling_type (int): RoPE scaling type, from `enum llama_rope_scaling_type` + pooling_type (int): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) rope_freq_base (float): RoPE base frequency, 0 = from model rope_freq_scale (float): RoPE frequency scaling factor, 0 = from model yarn_ext_factor (float): YaRN extrapolation mix factor, negative = from model @@ -602,7 +607,6 @@ class llama_context_params(ctypes.Structure): logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) embedding (bool): embedding mode only offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU - do_pooling (bool): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback """ @@ -613,7 +617,8 @@ class llama_context_params(ctypes.Structure): ("n_batch", ctypes.c_uint32), ("n_threads", ctypes.c_uint32), ("n_threads_batch", ctypes.c_uint32), - ("rope_scaling_type", ctypes.c_int32), + ("rope_scaling_type", ctypes.c_int), + ("pooling_type", ctypes.c_int), ("rope_freq_base", ctypes.c_float), ("rope_freq_scale", ctypes.c_float), ("yarn_ext_factor", ctypes.c_float), @@ -629,7 +634,6 @@ class llama_context_params(ctypes.Structure): ("logits_all", ctypes.c_bool), ("embedding", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), - ("do_pooling", ctypes.c_bool), ("abort_callback", ggml_abort_callback), ("abort_callback_data", ctypes.c_void_p), ] diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 973113429..67be2ce10 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 9731134296af3a6839cd682e51d9c2109a871de5 +Subproject commit 67be2ce1015d070b3b2cd488bcb041eefb61de72 From 93dc56ace8e3de97f6f39a7071ff63aaf29d376f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 6 Mar 2024 01:32:00 -0500 Subject: [PATCH 225/624] Update llama.cpp --- llama_cpp/llama.py | 6 +++--- llama_cpp/llama_cpp.py | 34 ++++++++++++++++++++++++++-------- vendor/llama.cpp | 2 +- 3 files changed, 30 insertions(+), 12 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 108a4cfa9..7187b4a17 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -293,7 +293,7 @@ def __init__( self.context_params.logits_all = ( logits_all if draft_model is None else True ) # Must be set to True for speculative decoding - self.context_params.embedding = embedding + self.context_params.embeddings = embedding # TODO: Rename to embeddings self.context_params.offload_kqv = offload_kqv # Sampling Params @@ -787,7 +787,7 @@ def embed( n_embd = self.n_embd() n_batch = self.n_batch - if self.context_params.embedding == False: + if self.context_params.embeddings == False: raise RuntimeError( "Llama model must be created with embedding=True to call this method" ) @@ -1725,7 +1725,7 @@ def __getstate__(self): yarn_beta_slow=self.context_params.yarn_beta_slow, yarn_orig_ctx=self.context_params.yarn_orig_ctx, logits_all=self.context_params.logits_all, - embedding=self.context_params.embedding, + embedding=self.context_params.embeddings, # Sampling Params last_n_tokens_size=self.last_n_tokens_size, # LoRA Params diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 08adfe205..92b96766c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -399,7 +399,7 @@ class llama_token_data_array(ctypes.Structure): # // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) # // - pos : the positions of the respective token in the sequence # // - seq_id : the sequence to which the respective token belongs -# // - logits : if zero, the logits for the respective token will not be output +# // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output # // # typedef struct llama_batch { # int32_t n_tokens; @@ -409,7 +409,7 @@ class llama_token_data_array(ctypes.Structure): # llama_pos * pos; # int32_t * n_seq_id; # llama_seq_id ** seq_id; -# int8_t * logits; +# int8_t * logits; // TODO: rename this to "output" # // NOTE: helpers for smooth API transition - can be deprecated in the future @@ -572,7 +572,7 @@ class llama_model_params(ctypes.Structure): # // Keep the booleans together to avoid misalignment during copy-by-value. # bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) -# bool embedding; // embedding mode only +# bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU # // Abort callback @@ -605,7 +605,7 @@ class llama_context_params(ctypes.Structure): type_k (int): data type for K cache type_v (int): data type for V cache logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) - embedding (bool): embedding mode only + embeddings (bool): if true, extract embeddings (together with logits) offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback @@ -632,7 +632,7 @@ class llama_context_params(ctypes.Structure): ("type_k", ctypes.c_int), ("type_v", ctypes.c_int), ("logits_all", ctypes.c_bool), - ("embedding", ctypes.c_bool), + ("embeddings", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), ("abort_callback", ggml_abort_callback), ("abort_callback_data", ctypes.c_void_p), @@ -1774,8 +1774,8 @@ def llama_get_logits_ith( ... -# Get the embeddings for the input -# shape: [n_embd] (1-dimensional) +# // Get all output token embeddings +# // shape: [n_tokens*n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); @ctypes_function( "llama_get_embeddings", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) @@ -1786,8 +1786,9 @@ def llama_get_embeddings(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float] ... -# // Get the embeddings for the ith sequence +# // Get the embeddings for the ith token # // llama_get_embeddings(ctx) + i*n_embd +# // shape: [n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); @ctypes_function( "llama_get_embeddings_ith", @@ -1802,6 +1803,23 @@ def llama_get_embeddings_ith( ... +# // Get the embeddings for a sequence id +# // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE +# // shape: [n_embd] (1-dimensional) +# LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id); +@ctypes_function( + "llama_get_embeddings_seq", + [llama_context_p_ctypes, llama_seq_id], + ctypes.POINTER(ctypes.c_float), +) +def llama_get_embeddings_seq( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> CtypesArray[ctypes.c_float]: + """Get the embeddings for a sequence id + Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE + shape: [n_embd] (1-dimensional)""" + ... + # // # // Vocab # // diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 67be2ce10..8ced9f7e3 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 67be2ce1015d070b3b2cd488bcb041eefb61de72 +Subproject commit 8ced9f7e3225adb8501e9821ed1bbd92e3a5c7ae From 40c6b54f6880e1cbb8f6393d9097328ffd422e13 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 8 Mar 2024 20:58:50 -0500 Subject: [PATCH 226/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 16 ++++++++++++---- vendor/llama.cpp | 2 +- 2 files changed, 13 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 92b96766c..0176e49a8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -429,10 +429,12 @@ class llama_batch(ctypes.Structure): The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens Attributes: + n_tokens (int): number of tokens token (ctypes.Array[llama_token]): the token ids of the input (used when embd is NULL) embd (ctypes.Array[ctypes.ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL) pos (ctypes.Array[ctypes.Array[llama_pos]]): the positions of the respective token in the sequence seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs + logits (ctypes.Array[ctypes.ctypes.c_int8]): if zero, the logits for the respective token will not be output """ _fields_ = [ @@ -547,6 +549,7 @@ class llama_model_params(ctypes.Structure): # uint32_t seed; // RNG seed, -1 for random # uint32_t n_ctx; // text context, 0 = from model # uint32_t n_batch; // prompt processing maximum batch size +# uint32_t n_parallel; // number of parallel sequences (i.e. distinct states for recurrent models) # uint32_t n_threads; // number of threads to use for generation # uint32_t n_threads_batch; // number of threads to use for batch processing @@ -588,6 +591,7 @@ class llama_context_params(ctypes.Structure): seed (int): RNG seed, -1 for random n_ctx (int): text context, 0 = from model n_batch (int): prompt processing maximum batch size + n_parallel (int): number of parallel sequences (i.e. distinct states for recurrent models) n_threads (int): number of threads to use for generation n_threads_batch (int): number of threads to use for batch processing rope_scaling_type (int): RoPE scaling type, from `enum llama_rope_scaling_type` @@ -615,6 +619,7 @@ class llama_context_params(ctypes.Structure): ("seed", ctypes.c_uint32), ("n_ctx", ctypes.c_uint32), ("n_batch", ctypes.c_uint32), + ("n_parallel", ctypes.c_uint32), ("n_threads", ctypes.c_uint32), ("n_threads_batch", ctypes.c_uint32), ("rope_scaling_type", ctypes.c_int), @@ -1322,7 +1327,7 @@ def llama_kv_cache_clear(ctx: llama_context_p, /): # // seq_id < 0 : match any sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_rm( +# LLAMA_API bool llama_kv_cache_seq_rm( # struct llama_context * ctx, # llama_seq_id seq_id, # llama_pos p0, @@ -1335,7 +1340,7 @@ def llama_kv_cache_clear(ctx: llama_context_p, /): llama_pos, llama_pos, ], - None, + ctypes.c_bool, ) def llama_kv_cache_seq_rm( ctx: llama_context_p, @@ -1343,7 +1348,7 @@ def llama_kv_cache_seq_rm( p0: Union[llama_pos, int], p1: Union[llama_pos, int], /, -): +) -> bool: """Removes all tokens that belong to the specified sequence and have positions in [p0, p1) seq_id < 0 : match any sequence p0 < 0 : [0, p1] @@ -1754,7 +1759,10 @@ def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: The logits for the last token are stored in the last row Logits for which llama_batch.logits[i] == 0 are undefined Rows: n_tokens provided with llama_batch - Cols: n_vocab""" + Cols: n_vocab + + Returns: + Pointer to the logits buffer of shape (n_tokens, n_vocab)""" ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8ced9f7e3..c2101a2e9 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8ced9f7e3225adb8501e9821ed1bbd92e3a5c7ae +Subproject commit c2101a2e909ac7c08976d414e64e96c90ee5fa9e From 2811014bae356401856a9c0796f42e719f2e8c3c Mon Sep 17 00:00:00 2001 From: Douglas Hanley Date: Fri, 8 Mar 2024 19:59:35 -0600 Subject: [PATCH 227/624] feat: Switch embed to llama_get_embeddings_seq (#1263) * switch to llama_get_embeddings_seq * Remove duplicate definition of llama_get_embeddings_seq Co-authored-by: Andrei --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 7187b4a17..aabbb7e71 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -814,7 +814,7 @@ def decode_batch(n_seq: int): # store embeddings for i in range(n_seq): - embedding: List[float] = llama_cpp.llama_get_embeddings_ith( + embedding: List[float] = llama_cpp.llama_get_embeddings_seq( self._ctx.ctx, i )[:n_embd] if normalize: From 1f3156d4f2d8c5439dbb2ad72b8c1de84703eb09 Mon Sep 17 00:00:00 2001 From: Kevin Cao Date: Fri, 8 Mar 2024 21:00:10 -0500 Subject: [PATCH 228/624] fix: Check for existence of clip model path (#1264) --- llama_cpp/llama_chat_format.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 69ed6019a..4eb2b0291 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1848,6 +1848,9 @@ def __init__(self, clip_model_path: str, verbose: bool = False): self.verbose = verbose self._clip_free = self._llava_cpp._libllava.clip_free # type: ignore + if not os.path.exists(clip_model_path): + raise ValueError(f"Clip model path does not exist: {clip_model_path}") + with suppress_stdout_stderr(disable=self.verbose): self.clip_ctx = self._llava_cpp.clip_model_load( self.clip_model_path.encode(), 0 From c139f8b5d50f6f416a24c0ba65983a3fb84bf2f3 Mon Sep 17 00:00:00 2001 From: Felipe Lorenz Date: Fri, 8 Mar 2024 21:09:00 -0500 Subject: [PATCH 229/624] feat: Add endpoints for tokenize, detokenize and count tokens (#1136) * Add endpoint to count tokens * Add tokenize and detokenize endpoints * Change response key to tokens for tokenize endpoint * Fix dependency bug * Cleanup * Remove example added by mistake * Move tokenize, detokenize, and count to Extras namespace. Tag existing endpoints --------- Co-authored-by: Andrei Betlen --- llama_cpp/server/app.py | 71 +++++++++++++++++++++++++++++++++++++-- llama_cpp/server/types.py | 36 ++++++++++++++++++++ 2 files changed, 105 insertions(+), 2 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index ec9280986..aa6afc112 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -41,6 +41,11 @@ CreateEmbeddingRequest, CreateChatCompletionRequest, ModelList, + TokenizeInputRequest, + TokenizeInputResponse, + TokenizeInputCountResponse, + DetokenizeInputRequest, + DetokenizeInputResponse, ) from llama_cpp.server.errors import RouteErrorHandler @@ -196,6 +201,9 @@ async def authenticate( ) +openai_v1_tag = "OpenAI V1" + + @router.post( "/v1/completions", summary="Completion", @@ -227,11 +235,13 @@ async def authenticate( }, } }, + tags=[openai_v1_tag], ) @router.post( "/v1/engines/copilot-codex/completions", include_in_schema=False, dependencies=[Depends(authenticate)], + tags=[openai_v1_tag], ) async def create_completion( request: Request, @@ -297,7 +307,10 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: @router.post( - "/v1/embeddings", summary="Embedding", dependencies=[Depends(authenticate)] + "/v1/embeddings", + summary="Embedding", + dependencies=[Depends(authenticate)], + tags=[openai_v1_tag], ) async def create_embedding( request: CreateEmbeddingRequest, @@ -339,6 +352,7 @@ async def create_embedding( }, } }, + tags=[openai_v1_tag], ) async def create_chat_completion( request: Request, @@ -391,7 +405,12 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: return iterator_or_completion -@router.get("/v1/models", summary="Models", dependencies=[Depends(authenticate)]) +@router.get( + "/v1/models", + summary="Models", + dependencies=[Depends(authenticate)], + tags=[openai_v1_tag], +) async def get_models( llama_proxy: LlamaProxy = Depends(get_llama_proxy), ) -> ModelList: @@ -407,3 +426,51 @@ async def get_models( for model_alias in llama_proxy ], } + + +extras_tag = "Extras" + + +@router.post( + "/extras/tokenize", + summary="Tokenize", + dependencies=[Depends(authenticate)], + tags=[extras_tag], +) +async def tokenize( + body: TokenizeInputRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), +) -> TokenizeInputResponse: + tokens = llama_proxy(body.model).tokenize(body.input.encode("utf-8"), special=True) + + return {"tokens": tokens} + + +@router.post( + "/extras/tokenize/count", + summary="Tokenize Count", + dependencies=[Depends(authenticate)], + tags=[extras_tag], +) +async def count_query_tokens( + body: TokenizeInputRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), +) -> TokenizeInputCountResponse: + tokens = llama_proxy(body.model).tokenize(body.input.encode("utf-8"), special=True) + + return {"count": len(tokens)} + + +@router.post( + "/extras/detokenize", + summary="Detokenize", + dependencies=[Depends(authenticate)], + tags=[extras_tag], +) +async def detokenize( + body: DetokenizeInputRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), +) -> DetokenizeInputResponse: + text = llama_proxy(body.model).detokenize(body.tokens).decode("utf-8") + + return {"text": text} diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index 9a4b81e09..c8b2ebc6c 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -264,3 +264,39 @@ class ModelData(TypedDict): class ModelList(TypedDict): object: Literal["list"] data: List[ModelData] + + +class TokenizeInputRequest(BaseModel): + model: Optional[str] = model_field + input: Optional[str] = Field(description="The input to tokenize.") + + model_config = { + "json_schema_extra": {"examples": [{"input": "How many tokens in this query?"}]} + } + + +class TokenizeInputResponse(BaseModel): + tokens: List[int] = Field(description="A list of tokens.") + + model_config = {"json_schema_extra": {"example": {"tokens": [123, 321, 222]}}} + + +class TokenizeInputCountResponse(BaseModel): + count: int = Field(description="The number of tokens in the input.") + + model_config = {"json_schema_extra": {"example": {"count": 5}}} + + +class DetokenizeInputRequest(BaseModel): + model: Optional[str] = model_field + tokens: List[int] = Field(description="A list of toekns to detokenize.") + + model_config = {"json_schema_extra": {"example": [{"tokens": [123, 321, 222]}]}} + + +class DetokenizeInputResponse(BaseModel): + text: str = Field(description="The detokenized text.") + + model_config = { + "json_schema_extra": {"example": {"text": "How many tokens in this query?"}} + } From d02a9cf16ff88ad011e2eb1ce29f4d9400f13cd1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 8 Mar 2024 21:10:53 -0500 Subject: [PATCH 230/624] Fixed json strings grammar by blacklisting character control set. Closes #1259 --- llama_cpp/llama_grammar.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 6a37857b9..9cc48a93b 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1337,7 +1337,7 @@ def print_grammar(file: TextIO, state: parse_state) -> None: string ::= "\"" ( - [^"\\] | + [^"\\\x7F\x00-\x1F] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes )* "\"" ws @@ -1366,7 +1366,7 @@ def print_grammar(file: TextIO, state: parse_state) -> None: string ::= "\"" ( - [^"\\] | + [^"\\\x7F\x00-\x1F] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes )* "\"" ws From a7281994d87927e42d8e636295c786057e98d8fe Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 8 Mar 2024 21:14:44 -0500 Subject: [PATCH 231/624] chore: Bump version --- CHANGELOG.md | 8 ++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index e16a6dfcc..90dd1e690 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.56] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c2101a2e909ac7c08976d414e64e96c90ee5fa9e +- feat(server): Add endpoints for tokenize, detokenize and count tokens by @felipelo in #1136 +- feat: Switch embed to llama_get_embeddings_seq by @iamlemec in #1263 +- fix: Fixed json strings grammar by blacklisting character control set by @ExtReMLapin in d02a9cf16ff88ad011e2eb1ce29f4d9400f13cd1 +- fix: Check for existence of clip model path by @kejcao in #1264 + ## [0.2.55] - feat: Update llama.cpp to ggerganov/9731134296af3a6839cd682e51d9c2109a871de5 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 519ab51b7..fcbc7150c 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.55" \ No newline at end of file +__version__ = "0.2.56" \ No newline at end of file From 08e910f7a7e3657cf210ab8633a6994e1bde7164 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 10 Mar 2024 23:45:05 -0400 Subject: [PATCH 232/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 11 +++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 12 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 0176e49a8..8f0ba546c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1728,6 +1728,17 @@ def llama_set_n_threads( """ ... + +# // Set whether to use causal attention or not +# // If set to true, the model will only attend to the past tokens +# LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn); +@ctypes_function("llama_set_causal_attn", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_causal_attn(ctx: llama_context_p, causal_attn: bool, /): + """Set whether to use causal attention or not + If set to true, the model will only attend to the past tokens""" + ... + + # // Set abort callback # LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); @ctypes_function( diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c2101a2e9..3814a0739 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c2101a2e909ac7c08976d414e64e96c90ee5fa9e +Subproject commit 3814a07392d2bdc22911652bc7c2f9bdb0ce042e From dd0ee56217c60a20a192dc7f1523dba9a006bbc9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 13 Mar 2024 15:57:35 -0400 Subject: [PATCH 233/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 160 ++++++++++++++++++++++------------------- vendor/llama.cpp | 2 +- 2 files changed, 87 insertions(+), 75 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 8f0ba546c..884a06384 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -548,8 +548,9 @@ class llama_model_params(ctypes.Structure): # struct llama_context_params { # uint32_t seed; // RNG seed, -1 for random # uint32_t n_ctx; // text context, 0 = from model -# uint32_t n_batch; // prompt processing maximum batch size -# uint32_t n_parallel; // number of parallel sequences (i.e. distinct states for recurrent models) +# uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode +# uint32_t n_ubatch; // physical maximum batch size +# uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models) # uint32_t n_threads; // number of threads to use for generation # uint32_t n_threads_batch; // number of threads to use for batch processing @@ -578,6 +579,7 @@ class llama_model_params(ctypes.Structure): # bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU + # // Abort callback # // if it returns true, execution of llama_decode() will be aborted # // currently works only with CPU execution @@ -590,8 +592,9 @@ class llama_context_params(ctypes.Structure): Attributes: seed (int): RNG seed, -1 for random n_ctx (int): text context, 0 = from model - n_batch (int): prompt processing maximum batch size - n_parallel (int): number of parallel sequences (i.e. distinct states for recurrent models) + n_batch (int): logical maximum batch size that can be submitted to llama_decode + n_ubatch (int): physical maximum batch size + n_seq_max (int): max number of sequences (i.e. distinct states for recurrent models) n_threads (int): number of threads to use for generation n_threads_batch (int): number of threads to use for batch processing rope_scaling_type (int): RoPE scaling type, from `enum llama_rope_scaling_type` @@ -619,7 +622,8 @@ class llama_context_params(ctypes.Structure): ("seed", ctypes.c_uint32), ("n_ctx", ctypes.c_uint32), ("n_batch", ctypes.c_uint32), - ("n_parallel", ctypes.c_uint32), + ("n_ubatch", ctypes.c_uint32), + ("n_seq_max", ctypes.c_uint32), ("n_threads", ctypes.c_uint32), ("n_threads_batch", ctypes.c_uint32), ("rope_scaling_type", ctypes.c_int), @@ -667,7 +671,7 @@ class llama_context_params(ctypes.Structure): # bool allow_requantize; // allow quantizing non-f32/f16 tensors # bool quantize_output_tensor; // quantize output.weight # bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored -# bool pure; // disable k-quant mixtures and quantize all tensors to the same type +# bool pure; // quantize all tensors to the default type # void * imatrix; // pointer to importance matrix data # } llama_model_quantize_params; class llama_model_quantize_params(ctypes.Structure): @@ -679,7 +683,7 @@ class llama_model_quantize_params(ctypes.Structure): allow_requantize (bool): allow quantizing non-f32/f16 tensors quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored - pure (bool): disable k-quant mixtures and quantize all tensors to the same type + pure (bool): quantize all tensors to the default type imatrix (ctypes.ctypes.c_void_p): pointer to importance matrix data """ @@ -860,8 +864,7 @@ def llama_backend_init(): [ctypes.c_int], None, ) -def llama_numa_init(numa: int, /): - ... +def llama_numa_init(numa: int, /): ... # // Call once at the end of the program - currently only used for MPI @@ -886,8 +889,7 @@ def llama_backend_free(): ) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / -) -> Optional[llama_model_p]: - ... +) -> Optional[llama_model_p]: ... # LLAMA_API void llama_free_model(struct llama_model * model); @@ -896,8 +898,7 @@ def llama_load_model_from_file( [llama_model_p_ctypes], None, ) -def llama_free_model(model: llama_model_p, /): - ... +def llama_free_model(model: llama_model_p, /): ... # LLAMA_API struct llama_context * llama_new_context_with_model( @@ -910,8 +911,7 @@ def llama_free_model(model: llama_model_p, /): ) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / -) -> Optional[llama_context_p]: - ... +) -> Optional[llama_context_p]: ... # // Frees all allocated memory @@ -932,80 +932,77 @@ def llama_free(ctx: llama_context_p, /): [], ctypes.c_int64, ) -def llama_time_us() -> int: - ... +def llama_time_us() -> int: ... # LLAMA_API size_t llama_max_devices(void); @ctypes_function("llama_max_devices", [], ctypes.c_size_t) -def llama_max_devices() -> int: - ... +def llama_max_devices() -> int: ... # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) -def llama_supports_mmap() -> bool: - ... +def llama_supports_mmap() -> bool: ... # LLAMA_API bool llama_supports_mlock (void); @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) -def llama_supports_mlock() -> bool: - ... +def llama_supports_mlock() -> bool: ... # LLAMA_API bool llama_supports_gpu_offload(void); @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) -def llama_supports_gpu_offload() -> bool: - ... +def llama_supports_gpu_offload() -> bool: ... # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) -def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: - ... +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ctx(ctx: llama_context_p, /) -> int: - ... +def llama_n_ctx(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_batch(ctx: llama_context_p, /) -> int: - ... +def llama_n_batch(ctx: llama_context_p, /) -> int: ... + + +# LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); +@ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_ubatch(ctx: llama_context_p, /) -> int: ... + + +# LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); +@ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... # LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: - ... +def llama_vocab_type(model: llama_model_p, /) -> int: ... # LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: - ... +def llama_rope_type(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: - ... +def llama_n_vocab(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_ctx_train(model: llama_model_p, /) -> int: - ... +def llama_n_ctx_train(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_embd(model: llama_model_p, /) -> int: - ... +def llama_n_embd(model: llama_model_p, /) -> int: ... # // Get the model's RoPE frequency scaling factor @@ -1192,8 +1189,7 @@ def llama_model_apply_lora_from_file( path_base_model: Union[ctypes.c_char_p, bytes, None], n_threads: Union[ctypes.c_int32, int], /, -) -> int: - ... +) -> int: ... # // @@ -1219,7 +1215,7 @@ class llama_kv_cache_view_cell(ctypes.Structure): # // Maximum number of sequences that can exist in a cell. It's not an error # // if there are more sequences in a cell than this value, however they will # // not be visible in the view cells_sequences. -# int32_t n_max_seq; +# int32_t n_seq_max; # // Number of tokens in the cache. For example, if there are two populated # // cells, the first with 1 sequence id in it and the second with 2 sequence @@ -1240,7 +1236,7 @@ class llama_kv_cache_view_cell(ctypes.Structure): # struct llama_kv_cache_view_cell * cells; -# // The sequences for each cell. There will be n_max_seq items per cell. +# // The sequences for each cell. There will be n_seq_max items per cell. # llama_seq_id * cells_sequences; # }; class llama_kv_cache_view(ctypes.Structure): @@ -1260,14 +1256,14 @@ class llama_kv_cache_view(ctypes.Structure): # // Create an empty KV cache view. (use only for debugging purposes) -# LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq); +# LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max); @ctypes_function( "llama_kv_cache_view_init", [llama_context_p_ctypes, ctypes.c_int32], llama_kv_cache_view, ) def llama_kv_cache_view_init( - ctx: llama_context_p, n_max_seq: Union[ctypes.c_int32, int], / + ctx: llama_context_p, n_seq_max: Union[ctypes.c_int32, int], / ) -> llama_kv_cache_view: """Create an empty KV cache view. (use only for debugging purposes)""" ... @@ -1582,8 +1578,7 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: - ... +) -> int: ... # LLAMA_API bool llama_save_session_file( @@ -1607,8 +1602,7 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: - ... +) -> int: ... # // @@ -1756,6 +1750,18 @@ def llama_set_abort_callback( ... +# // Wait until all computations are finished +# // This is automatically done when using one of the functions below to obtain the computation results +# // and is not necessary to call it explicitly in most cases +# LLAMA_API void llama_synchronize(struct llama_context * ctx); +@ctypes_function("llama_synchronize", [llama_context_p_ctypes], None) +def llama_synchronize(ctx: llama_context_p, /): + """Wait until all computations are finished + This is automatically done when using one of the functions below to obtain the computation results + and is not necessary to call it explicitly in most cases""" + ... + + # // Token logits obtained from the last call to llama_decode() # // The logits for the last token are stored in the last row # // Logits for which llama_batch.logits[i] == 0 are undefined @@ -1771,7 +1777,7 @@ def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: Logits for which llama_batch.logits[i] == 0 are undefined Rows: n_tokens provided with llama_batch Cols: n_vocab - + Returns: Pointer to the logits buffer of shape (n_tokens, n_vocab)""" ... @@ -1839,6 +1845,7 @@ def llama_get_embeddings_seq( shape: [n_embd] (1-dimensional)""" ... + # // # // Vocab # // @@ -1850,8 +1857,7 @@ def llama_get_embeddings_seq( ) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / -) -> bytes: - ... +) -> bytes: ... # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); @@ -1860,8 +1866,7 @@ def llama_token_get_text( ) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / -) -> float: - ... +) -> float: ... # LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); @@ -1870,8 +1875,7 @@ def llama_token_get_score( ) def llama_token_get_type( model: llama_model_p, token: Union[llama_token, int], / -) -> int: - ... +) -> int: ... # // Special tokens @@ -1924,20 +1928,17 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: - ... +def llama_token_middle(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: - ... +def llama_token_suffix(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: - ... +def llama_token_eot(model: llama_model_p, /) -> int: ... # // @@ -1947,7 +1948,7 @@ def llama_token_eot(model: llama_model_p, /) -> int: # /// @details Convert the provided text into tokens. # /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. -# /// @return Returns the number of tokens on success, no more than n_max_tokens +# /// @return Returns the number of tokens on success, no more than n_tokens_max # /// @return Returns a negative number on failure - the number of tokens that would have been returned # /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. # /// Does not insert a leading space. @@ -1956,7 +1957,7 @@ def llama_token_eot(model: llama_model_p, /) -> int: # const char * text, # int32_t text_len, # llama_token * tokens, -# int32_t n_max_tokens, +# int32_t n_tokens_max, # bool add_bos, # bool special); @ctypes_function( @@ -1977,12 +1978,26 @@ def llama_tokenize( text: bytes, text_len: Union[ctypes.c_int, int], tokens: CtypesArray[llama_token], - n_max_tokens: Union[ctypes.c_int, int], + n_tokens_max: Union[ctypes.c_int, int], add_bos: Union[ctypes.c_bool, bool], special: Union[ctypes.c_bool, bool], /, ) -> int: - """Convert the provided text into tokens.""" + """Convert the provided text into tokens. + + Args: + model: The model to use for tokenization. + text: The text to tokenize. + text_len: The length of the text. + tokens: The tokens pointer must be large enough to hold the resulting tokens. + n_max_tokens: The maximum number of tokens to return. + add_bos: Whether to add a beginning-of-sentence token. + special: Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. + Does not insert a leading space. + + Returns: + Returns the number of tokens on success, no more than n_tokens_max + Returns a negative number on failure - the number of tokens that would have been returned""" ... @@ -2054,8 +2069,7 @@ def llama_chat_apply_template( chat: CtypesArray[llama_chat_message], n_msg: int, /, -) -> int: - ... +) -> int: ... # // @@ -2656,8 +2670,7 @@ def llama_beam_search( n_past: Union[ctypes.c_int, int], n_predict: Union[ctypes.c_int, int], /, -): - ... +): ... # Performance information @@ -2734,5 +2747,4 @@ def llama_log_set( [ctypes.c_void_p, llama_context_p_ctypes], None, ) -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): - ... +def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 3814a0739..19885d205 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 3814a07392d2bdc22911652bc7c2f9bdb0ce042e +Subproject commit 19885d205e768579ab090d1e99281cae58c21b54 From d318cc8b83981833fbab2314d97af21104e52d99 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 14 Mar 2024 09:17:41 -0400 Subject: [PATCH 234/624] fix: Set default pooling_type to mean, check for null pointer. --- llama_cpp/llama.py | 10 ++++++++-- llama_cpp/llama_cpp.py | 1 - 2 files changed, 8 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index aabbb7e71..18db1a0b6 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -79,6 +79,7 @@ def __init__( n_threads: Optional[int] = None, n_threads_batch: Optional[int] = None, rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, + pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_MEAN, rope_freq_base: float = 0.0, rope_freq_scale: float = 0.0, yarn_ext_factor: float = -1.0, @@ -151,6 +152,7 @@ def __init__( n_threads: Number of threads to use for generation n_threads_batch: Number of threads to use for batch processing rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054 + pooling_type: Pooling type, from `enum llama_pooling_type`. rope_freq_base: RoPE base frequency, 0 = from model rope_freq_scale: RoPE frequency scaling factor, 0 = from model yarn_ext_factor: YaRN extrapolation mix factor, negative = from model @@ -271,6 +273,7 @@ def __init__( if rope_scaling_type is not None else llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED ) + self.context_params.pooling_type = pooling_type self.context_params.rope_freq_base = ( rope_freq_base if rope_freq_base != 0.0 else 0 ) @@ -814,9 +817,12 @@ def decode_batch(n_seq: int): # store embeddings for i in range(n_seq): - embedding: List[float] = llama_cpp.llama_get_embeddings_seq( + ptr = llama_cpp.llama_get_embeddings_seq( self._ctx.ctx, i - )[:n_embd] + ) + if not ptr: + raise RuntimeError("Failed to get embeddings from sequence pooling type is not set") + embedding: List[float] = ptr[:n_embd] if normalize: norm = float(np.linalg.norm(embedding)) embedding = [v / norm for v in embedding] diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 884a06384..dc45fd933 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -579,7 +579,6 @@ class llama_model_params(ctypes.Structure): # bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU - # // Abort callback # // if it returns true, execution of llama_decode() will be aborted # // currently works only with CPU execution From 4084aabe867b8ec2aba1b22659e59c9318b0d1f3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 14 Mar 2024 10:04:57 -0400 Subject: [PATCH 235/624] fix: set default pooling type to unspecified --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 18db1a0b6..fed84d579 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -79,7 +79,7 @@ def __init__( n_threads: Optional[int] = None, n_threads_batch: Optional[int] = None, rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, - pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_MEAN, + pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_UNSPECIFIED, rope_freq_base: float = 0.0, rope_freq_scale: float = 0.0, yarn_ext_factor: float = -1.0, From 1a9b8af2dd39ed1c9050df0df2714e61d2740ee2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 14 Mar 2024 11:46:48 -0400 Subject: [PATCH 236/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 19885d205..044ec4b2a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 19885d205e768579ab090d1e99281cae58c21b54 +Subproject commit 044ec4b2a567f649459ccd20af2f387c784faa51 From 20e6815252d0efd9f015f7adbf108faaf36e3f3c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 15 Mar 2024 12:58:34 -0400 Subject: [PATCH 237/624] fix: json mode --- llama_cpp/llama_chat_format.py | 73 ++++++++++++++++------------------ 1 file changed, 34 insertions(+), 39 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 4eb2b0291..81ca5520b 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -339,16 +339,7 @@ def chat_completion_handler( stop = stop + rstop if response_format is not None and response_format["type"] == "json_object": - try: - # create grammar from json schema - if "schema" in response_format: - grammar = llama_grammar.LlamaGrammar.from_json_schema( - json.dumps(response_format["schema"]), verbose=llama.verbose - ) - except Exception as e: - grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF, verbose=llama.verbose - ) + grammar = _grammar_for_response_format(response_format, verbose=llama.verbose) completion_or_chunks = llama.create_completion( prompt=prompt, @@ -606,6 +597,35 @@ def _format_chatglm3( ret += role return ret +def _grammar_for_json(verbose:bool=False): + return llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF, verbose=verbose) + +def _grammar_for_json_schema( + schema: str, + verbose: bool = False, + fallback_to_json: bool = True +): + try: + return llama_grammar.LlamaGrammar.from_json_schema(schema, verbose=verbose) + except Exception as e: + if fallback_to_json: + return _grammar_for_json(verbose=verbose) + else: + raise e + +def _grammar_for_response_format( + response_format: llama_types.ChatCompletionRequestResponseFormat, + verbose: bool = False +): + if response_format["type"] != "json_object": + return None + + if "schema" in response_format: + return _grammar_for_json_schema( + json.dumps(response_format["schema"]), verbose=verbose + ) + else: + return _grammar_for_json(verbose=verbose) ### Chat Formats ### @@ -1994,16 +2014,7 @@ def __call__( prompt = llama.input_ids[: llama.n_tokens].tolist() if response_format is not None and response_format["type"] == "json_object": - try: - # create grammar from json schema - if "schema" in response_format: - grammar = llama_grammar.LlamaGrammar.from_json_schema( - json.dumps(response_format["schema"]) - ) - except Exception as e: - grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF - ) + grammar = _grammar_for_response_format(response_format) return _convert_completion_to_chat( llama.create_completion( @@ -2159,26 +2170,10 @@ def chatml_function_calling( tool_calls=None, add_generation_prompt=True, ) + if response_format is not None and response_format["type"] == "json_object": - try: - grammar = ( - llama_grammar.LlamaGrammar.from_json_schema( - json.dumps(response_format["schema"]) - ) - if "schema" in response_format - else None - ) - except Exception as e: - if llama.verbose: - print( - "Failed to parse response format as JSON schema, falling back to default grammar" - ) - print(e) - grammar = ( - llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) - if grammar is None - else grammar - ) + grammar = _grammar_for_response_format(response_format) + return _convert_completion_to_chat( llama.create_completion( prompt=prompt, From 6eb25231e4dafeb792ffc9597c27330344c970b1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 15 Mar 2024 12:58:45 -0400 Subject: [PATCH 238/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 14 ++++++++------ vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 7 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index dc45fd933..b9593cf7a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -198,13 +198,15 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # enum llama_vocab_type { -# LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece -# LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding -# LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece +# LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab +# LLAMA_VOCAB_TYPE_SPM = 1, // SentencePiece +# LLAMA_VOCAB_TYPE_BPE = 2, // Byte Pair Encoding +# LLAMA_VOCAB_TYPE_WPM = 3, // WordPiece # }; -LLAMA_VOCAB_TYPE_SPM = 0 -LLAMA_VOCAB_TYPE_BPE = 1 -LLAMA_VOCAB_TYPE_WPM = 2 +LLAMA_VOCAB_TYPE_NONE = 0 +LLAMA_VOCAB_TYPE_SPM = 1 +LLAMA_VOCAB_TYPE_BPE = 2 +LLAMA_VOCAB_TYPE_WPM = 3 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 044ec4b2a..4e9a7f7f7 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 044ec4b2a567f649459ccd20af2f387c784faa51 +Subproject commit 4e9a7f7f7fb6acbddd1462909c8d696e38edbfcc From 8d298b47507c9d82945f0add20d1ddb3e8ea0aa4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 18 Mar 2024 10:26:36 -0400 Subject: [PATCH 239/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 85 ++++++++++++++++++++++++++++++++++++++---- vendor/llama.cpp | 2 +- 2 files changed, 78 insertions(+), 9 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index b9593cf7a..6b5c1bc15 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -581,6 +581,7 @@ class llama_model_params(ctypes.Structure): # bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU + # // Abort callback # // if it returns true, execution of llama_decode() will be aborted # // currently works only with CPU execution @@ -1006,6 +1007,11 @@ def llama_n_ctx_train(model: llama_model_p, /) -> int: ... def llama_n_embd(model: llama_model_p, /) -> int: ... +# LLAMA_API int32_t llama_n_layer (const struct llama_model * model); +@ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) +def llama_n_layer(model: llama_model_p, /) -> int: ... + + # // Get the model's RoPE frequency scaling factor # LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); @ctypes_function("llama_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) @@ -1166,12 +1172,18 @@ def llama_model_quantize( ... +# // Apply a LoRA adapter to a loaded model +# // path_base_model is the path to a higher quality model to use as a base for +# // the layers modified by the adapter. Can be NULL to use the current loaded model. +# // The model needs to be reloaded before applying a new adapter, otherwise the adapter +# // will be applied on top of the previous one +# // Returns 0 on success # LLAMA_API int32_t llama_model_apply_lora_from_file( # const struct llama_model * model, -# const char * path_lora, -# float scale, -# const char * path_base_model, -# int32_t n_threads); +# const char * path_lora, +# float scale, +# const char * path_base_model, +# int32_t n_threads); @ctypes_function( "llama_model_apply_lora_from_file", [ @@ -1190,7 +1202,57 @@ def llama_model_apply_lora_from_file( path_base_model: Union[ctypes.c_char_p, bytes, None], n_threads: Union[ctypes.c_int32, int], /, -) -> int: ... +) -> int: + """Apply a LoRA adapter to a loaded model + path_base_model is the path to a higher quality model to use as a base for + the layers modified by the adapter. Can be NULL to use the current loaded model. + The model needs to be reloaded before applying a new adapter, otherwise the adapter + will be applied on top of the previous one + Returns 0 on success""" + ... + + +# // Apply a loaded control vector to a llama_context, or if data is NULL, clear +# // the currently loaded vector. +# // n_embd should be the size of a single layer's control, and data should point +# // to an n_embd x n_layers buffer starting from layer 1. +# // il_start and il_end are the layer range the vector should apply to (both inclusive) +# // See llama_control_vector_load in common to load a control vector. +# LLAMA_API int32_t llama_control_vector_apply( +# struct llama_context * lctx, +# const float * data, +# size_t len, +# int32_t n_embd, +# int32_t il_start, +# int32_t il_end); +@ctypes_function( + "llama_control_vector_apply", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_float), + ctypes.c_size_t, + ctypes.c_int32, + ctypes.c_int32, + ctypes.c_int32, + ], + ctypes.c_int32, +) +def llama_control_vector_apply( + lctx: llama_context_p, + data: CtypesPointerOrRef[ctypes.c_float], + len: int, + n_embd: int, + il_start: int, + il_end: int, + /, +) -> int: + """Apply a loaded control vector to a llama_context, or if data is NULL, clear + the currently loaded vector. + n_embd should be the size of a single layer's control, and data should point + to an n_embd x n_layers buffer starting from layer 1. + il_start and il_end are the layer range the vector should apply to (both inclusive) + See llama_control_vector_load in common to load a control vector.""" + ... # // @@ -1205,6 +1267,12 @@ def llama_model_apply_lora_from_file( # llama_pos pos; # }; class llama_kv_cache_view_cell(ctypes.Structure): + """Information associated with an individual cell in the KV cache view. + + Attributes: + pos (llama_pos): The position for this cell. Takes KV cache shifts into account. + May be negative if the cell is not populated.""" + _fields_ = [("pos", llama_pos)] @@ -1985,7 +2053,7 @@ def llama_tokenize( /, ) -> int: """Convert the provided text into tokens. - + Args: model: The model to use for tokenization. text: The text to tokenize. @@ -1995,10 +2063,11 @@ def llama_tokenize( add_bos: Whether to add a beginning-of-sentence token. special: Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. Does not insert a leading space. - + Returns: Returns the number of tokens on success, no more than n_tokens_max - Returns a negative number on failure - the number of tokens that would have been returned""" + Returns a negative number on failure - the number of tokens that would have been returned + """ ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4e9a7f7f7..ac9ee6a4a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4e9a7f7f7fb6acbddd1462909c8d696e38edbfcc +Subproject commit ac9ee6a4ad740bc1ee484ede43e9f92b5af244c1 From 8a60c7bc8cae7aa9770eeac0f482d39350763a6f Mon Sep 17 00:00:00 2001 From: Jeffrey Fong Date: Mon, 18 Mar 2024 22:40:57 +0800 Subject: [PATCH 240/624] fix: Fix and optimize functionary chat handler (#1282) * fix functionary chat logic * further fixes --------- Co-authored-by: Andrei --- llama_cpp/llama_chat_format.py | 131 ++++++++++++++++----------------- 1 file changed, 65 insertions(+), 66 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 81ca5520b..c89cce879 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1596,13 +1596,15 @@ def prepare_messages_for_inference( function_call = ( tool_choice if isinstance(tool_choice, str) else tool_choice["function"] ) + else: + function_call = "auto" prompt = prepare_messages_for_inference( messages, tokenizer, version, functions, tools ) # If no tools/functions are provided - if function_call is None and (functions is None or len(functions) == 0): + if function_call == "none" or functions is None or len(functions) == 0: if version == "v1": stop = END_ASSISTANT_TOKEN else: @@ -1630,6 +1632,7 @@ def prepare_messages_for_inference( logits_processor=logits_processor, grammar=grammar, ) + completion_or_completion_chunks["choices"][0]["text"] = completion_or_completion_chunks["choices"][0]["text"].lstrip() return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore assert stream is False # TODO: support stream mode @@ -1692,13 +1695,12 @@ def create_completion(stop): return completion + content = "" function_calls, function_bodies = [], [] if version == "v1": # If no or "auto" tool_choice/function_call - if function_call is None or ( - isinstance(function_call, str) and function_call == "auto" - ): + if isinstance(function_call, str) and function_call == "auto": stops = ["\n", END_ASSISTANT_TOKEN] # If tool_choice/function_call is "none" elif isinstance(function_call, str) and function_call == "none": @@ -1747,70 +1749,67 @@ def create_completion(stop): else: function_bodies.append(completion_text.strip()) else: - # Loop until all parallel function calls are generated - while True: - # If no or "auto" tool_choice/function_call - if function_call is None or ( - isinstance(function_call, str) and function_call == "auto" - ): - grammar = None - stops = CONTENT_TOKEN - # If tool_choice/function_call is "none" - elif isinstance(function_call, str) and function_call == "none": - prompt = ( - prepare_messages_for_inference(messages, tokenizer, version, [], []) - + "all\n<|content|>" - ) - stops = STOP_TOKEN - # If tool_choice/function_call is provided - elif isinstance(function_call, dict): - prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" - stops = STOP_TOKEN - function_call = function_call["name"] - function_calls.append(function_call) - grammar = get_grammar(function_call) - else: - prompt = prompt - stops = STOP_TOKEN - + # If tool_choice/function_call is "none" + if isinstance(function_call, str) and function_call == "none": + prompt = ( + prepare_messages_for_inference(messages, tokenizer, version, [], []) + + "all\n<|content|>" + ) + stops = [STOP_TOKEN, FROM_TOKEN] + completion = create_completion(stop=stops) + completion["choices"][0]["text"] = completion["choices"][0]["text"].strip() + return _convert_completion_to_chat(completion, stream=stream) # type: ignore + # If tool_choice/function_call is provided + elif isinstance(function_call, dict): + prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + stops = [STOP_TOKEN, FROM_TOKEN] completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] - - # If the generation does not involve a function call - if prompt.endswith("all\n<|content|>") and not completion_text.startswith( - "all" - ): - return _convert_completion_to_chat(completion, stream=stream) # type: ignore - # Generate model response if the model decides not to call any function - elif prompt.endswith(RECIPIENT_TOKEN) and completion_text.startswith("all"): - prompt += completion_text + CONTENT_TOKEN - completion = create_completion(stop=STOP_TOKEN) - return _convert_completion_to_chat(completion, stream=stream) # type: ignore - # Generate parameters if model decides to call a function - elif prompt.endswith(RECIPIENT_TOKEN): - function_calls.append(completion_text[:-1]) - grammar = get_grammar(function_calls[-1]) - completion = create_completion(stop=[STOP_TOKEN, "\n"]) - function_bodies.append(completion["choices"][0]["text"].strip()) - prompt += f"{function_calls[-1]}\n{CONTENT_TOKEN}{function_bodies[-1]}" + function_bodies.append(completion_text.strip()) + # If "auto" or no tool_choice/function_call + elif isinstance(function_call, str) and function_call == "auto": + while True: + # Generate function name first grammar = None - - # Try to generate the beginning of next turn - # If empty completion, break from loop - next_turn_completion_text = create_completion( - stop=[STOP_TOKEN, RECIPIENT_TOKEN] - )["choices"][0]["text"] - if len(next_turn_completion_text) > 0: - prompt += f"\n{FROM_TOKEN}assistant\n{RECIPIENT_TOKEN}" + stops = CONTENT_TOKEN + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + function_name = completion_text.strip() + if function_name == "all": + prompt += "all\n<|content|>" else: - break - # Break from loop if tool_choice/function_call is provided as a dict - else: - function_bodies.append(completion_text.strip()) - break + function_call = completion_text.strip() + prompt += f"{function_call}\n<|content|>" + function_calls.append(function_call) + grammar = get_grammar(function_call) + # Generate content + stops = [RECIPIENT_TOKEN, STOP_TOKEN] + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + if function_name == "all": + content += completion_text.removesuffix("\n<|from|>assistant\n").removesuffix("\n<|from|> assistant\n") + content = content.lstrip() + # Check whether the model wants to generate another turn + if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text: + cleaned_completion_text = completion_text.removesuffix("\n<|from|>assistant\n").removesuffix("\n<|from|> assistant\n").strip() + prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" + else: + break + else: + function_bodies.append(completion_text.strip()) + # Check whether the model wants to generate another turn + prompt += completion_text.strip() + grammar = None + completion = create_completion(stop=stops) + if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]: + prompt += "\n<|from|>assistant\n<|recipient|>" + else: + break assert "usage" in completion - assert len(function_calls) > 0 assert len(function_calls) == len(function_bodies) tool_calls = [] @@ -1843,14 +1842,14 @@ def create_completion(stop): "index": 0, "message": { "role": "assistant", - "content": None, + "content": None if content == "" else content, "function_call": { "name": tool_calls[0]["function"]["name"], "arguments": tool_calls[0]["function"]["arguments"], - }, - "tool_calls": tool_calls, + } if len(tool_calls) > 0 else None, + "tool_calls": tool_calls if len(tool_calls) > 0 else None, }, - "finish_reason": "tool_calls", + "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", } ], usage=completion["usage"], From bf64752535ac73032a25d6ba9ae0f064246964f2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 18 Mar 2024 11:37:30 -0400 Subject: [PATCH 241/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 90dd1e690..a85eaa436 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.57] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ac9ee6a4ad740bc1ee484ede43e9f92b5af244c1 +- fix: set default embedding pooling type to unspecified by @abetlen in 4084aabe867b8ec2aba1b22659e59c9318b0d1f3 +- fix: Fix and optimize functionary chat handler by @jeffrey-fong in #1282 +- fix: json mode for basic chat formats by @abetlen in 20e6815252d0efd9f015f7adbf108faaf36e3f3c + ## [0.2.56] - feat: Update llama.cpp to ggerganov/llama.cpp@c2101a2e909ac7c08976d414e64e96c90ee5fa9e diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index fcbc7150c..1e802fa7e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.56" \ No newline at end of file +__version__ = "0.2.57" \ No newline at end of file From 18d7ce918f45bc2bcf80102239142c49ebe29925 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 19 Mar 2024 04:40:24 -0400 Subject: [PATCH 242/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ac9ee6a4a..b80cf3b2d 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ac9ee6a4ad740bc1ee484ede43e9f92b5af244c1 +Subproject commit b80cf3b2d1dee0ad325f7a794fecc66befce7336 From 60d8498f212ca1eb4303d95610022a055718bbb8 Mon Sep 17 00:00:00 2001 From: Andrei Date: Tue, 19 Mar 2024 04:55:57 -0400 Subject: [PATCH 243/624] feat: Add tools/functions variables to Jinja2ChatFormatter, add function response formatting for all simple chat formats (#1273) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * Add tools/functions variables to Jinja2ChatFormatter Also fixed missing tools/tool_choices parameters in chat_formatter_to_chat_completion_handler(). * Set grammar when doing explicit function calling * Add function / tool response for all chat formats --------- Co-authored-by: Sigbjørn Skjæret --- llama_cpp/llama_chat_format.py | 408 +++++++++++++++++++-------------- 1 file changed, 233 insertions(+), 175 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index c89cce879..5bda163fd 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -188,6 +188,10 @@ def __call__( self, *, messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, **kwargs: Any, ) -> ChatFormatterResponse: def raise_exception(message: str): @@ -199,6 +203,10 @@ def raise_exception(message: str): bos_token=self.bos_token, raise_exception=raise_exception, add_generation_prompt=self.add_generation_prompt, + functions=functions, + function_call=function_call, + tools=tools, + tool_choice=tool_choice, ) return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token]) @@ -288,6 +296,183 @@ def _convert_completion_to_chat( return _convert_text_completion_to_chat(completion) +def _convert_completion_to_chat_function( + tool_name: str, + completion_or_chunks: Union[ + llama_types.CreateCompletionResponse, + Iterator[llama_types.CreateCompletionStreamResponse], + ], + stream: bool, +): + if not stream: + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + assert "usage" in completion + tool_id = "call_" + "_0_" + tool_name + "_" + completion["id"] + # TODO: Fix for legacy function calls + chat_completion: llama_types.CreateChatCompletionResponse = { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "index": 0, + "message": { + "role": "assistant", + "content": None, + "function_call": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + "tool_calls": [ + { + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + ], + }, + "finish_reason": "tool_calls", + } + ], + "usage": completion["usage"], + } + return chat_completion + else: + chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore + + def _stream_response_to_function_stream( + chunks: Iterator[llama_types.CreateCompletionStreamResponse], + ) -> Iterator[llama_types.CreateChatCompletionStreamResponse]: + # blank first message + first = True + id_ = None + created = None + model = None + tool_id = None + for chunk in chunks: + if first: + id_ = "chat" + chunk["id"] + created = chunk["created"] + model = chunk["model"] + tool_id = "call_" + "_0_" + tool_name + "_" + chunk["id"] + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": "assistant", + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": "", + }, + } + ], + }, + } + ], + } + first = False + continue + assert tool_id is not None + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": chunk["choices"][0][ + "text" + ], + }, + } + ], + }, + } + ], + } + + if id_ is not None and created is not None and model is not None: + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": "tool_calls", + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + + return _stream_response_to_function_stream(chunks) + + + def chat_formatter_to_chat_completion_handler( chat_formatter: ChatFormatter, ) -> LlamaChatCompletionHandler: @@ -331,6 +516,8 @@ def chat_completion_handler( messages=messages, functions=functions, function_call=function_call, + tools=tools, + tool_choice=tool_choice, ) prompt = result.prompt if result.stop is not None: @@ -341,6 +528,47 @@ def chat_completion_handler( if response_format is not None and response_format["type"] == "json_object": grammar = _grammar_for_response_format(response_format, verbose=llama.verbose) + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + tool = None + if tool_choice is not None and isinstance(tool_choice, dict) and tools is not None: + name = tool_choice["function"]["name"] + tool = next((t for t in tools if t["function"]["name"] == name), None) + if tool is None: + raise ValueError(f"Tool choice '{name}' not found in tools.") + schema = tool["function"]["parameters"] + try: + # create grammar from json schema + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(schema), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + completion_or_chunks = llama.create_completion( prompt=prompt, temperature=temperature, @@ -364,6 +592,11 @@ def chat_completion_handler( grammar=grammar, logit_bias=logit_bias, ) + if tool is not None: + tool_name = tool["function"]["name"] + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) return _convert_completion_to_chat(completion_or_chunks, stream=stream) return chat_completion_handler @@ -2198,181 +2431,6 @@ def chatml_function_calling( stream=stream, ) - def _convert_completion_to_chat_function( - tool_name: str, - completion_or_chunks: Union[ - llama_types.CreateCompletionResponse, - Iterator[llama_types.CreateCompletionStreamResponse], - ], - stream: bool, - ): - if not stream: - completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore - assert "usage" in completion - tool_id = "call_" + "_0_" + tool_name + "_" + completion["id"] - # TODO: Fix for legacy function calls - chat_completion: llama_types.CreateChatCompletionResponse = { - "id": "chat" + completion["id"], - "object": "chat.completion", - "created": completion["created"], - "model": completion["model"], - "choices": [ - { - "index": 0, - "message": { - "role": "assistant", - "content": None, - "function_call": { - "name": tool_name, - "arguments": completion["choices"][0]["text"], - }, - "tool_calls": [ - { - "id": tool_id, - "type": "function", - "function": { - "name": tool_name, - "arguments": completion["choices"][0]["text"], - }, - } - ], - }, - "finish_reason": "tool_calls", - } - ], - "usage": completion["usage"], - } - return chat_completion - else: - chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore - - def _stream_response_to_function_stream( - chunks: Iterator[llama_types.CreateCompletionStreamResponse], - ) -> Iterator[llama_types.CreateChatCompletionStreamResponse]: - # blank first message - first = True - id_ = None - created = None - model = None - tool_id = None - for chunk in chunks: - if first: - id_ = "chat" + chunk["id"] - created = chunk["created"] - model = chunk["model"] - tool_id = "call_" + "_0_" + tool_name + "_" + chunk["id"] - yield { - "id": id_, - "object": "chat.completion.chunk", - "created": created, - "model": model, - "choices": [ - { - "index": 0, - "finish_reason": None, - "logprobs": None, - "delta": { - "role": "assistant", - "content": None, - "function_call": None, - "tool_calls": None, - }, - } - ], - } - yield { - "id": "chat" + chunk["id"], - "object": "chat.completion.chunk", - "created": chunk["created"], - "model": chunk["model"], - "choices": [ - { - "index": 0, - "finish_reason": None, - "logprobs": None, - "delta": { - "role": None, - "content": None, - "function_call": { - "name": tool_name, - "arguments": chunk["choices"][0]["text"], - }, - "tool_calls": [ - { - "index": 0, - "id": tool_id, - "type": "function", - "function": { - "name": tool_name, - "arguments": "", - }, - } - ], - }, - } - ], - } - first = False - continue - assert tool_id is not None - yield { - "id": "chat" + chunk["id"], - "object": "chat.completion.chunk", - "created": chunk["created"], - "model": chunk["model"], - "choices": [ - { - "index": 0, - "finish_reason": None, - "logprobs": None, - "delta": { - "role": None, - "content": None, - "function_call": { - "name": tool_name, - "arguments": chunk["choices"][0]["text"], - }, - "tool_calls": [ - { - "index": 0, - "id": tool_id, - "type": "function", - "function": { - "name": tool_name, - "arguments": chunk["choices"][0][ - "text" - ], - }, - } - ], - }, - } - ], - } - - if id_ is not None and created is not None and model is not None: - yield { - "id": id_, - "object": "chat.completion.chunk", - "created": created, - "model": model, - "choices": [ - { - "index": 0, - "finish_reason": "tool_calls", - "logprobs": None, - "delta": { - "role": None, - "content": None, - "function_call": None, - "tool_calls": None, - }, - } - ], - } - - return _stream_response_to_function_stream(chunks) - # Case 2: Tool choice by user if isinstance(tool_choice, dict): tool_name = tool_choice["function"]["name"] From f7decc956207f181710a43e2795f478ad2a5fc5e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 19 Mar 2024 10:52:53 -0400 Subject: [PATCH 244/624] docs: Add chat examples to openapi ui --- llama_cpp/server/app.py | 68 +++++++++++++++++++++++++++++++++++------ 1 file changed, 59 insertions(+), 9 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index aa6afc112..2e1081e42 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -12,14 +12,7 @@ import anyio from anyio.streams.memory import MemoryObjectSendStream from starlette.concurrency import run_in_threadpool, iterate_in_threadpool -from fastapi import ( - Depends, - FastAPI, - APIRouter, - Request, - HTTPException, - status, -) +from fastapi import Depends, FastAPI, APIRouter, Request, HTTPException, status, Body from fastapi.middleware import Middleware from fastapi.middleware.cors import CORSMiddleware from fastapi.security import HTTPBearer @@ -356,7 +349,64 @@ async def create_embedding( ) async def create_chat_completion( request: Request, - body: CreateChatCompletionRequest, + body: CreateChatCompletionRequest = Body( + openapi_examples={ + "normal": { + "summary": "Chat Completion", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "What is the capital of France?"}, + ], + }, + }, + "json_mode": { + "summary": "JSON Mode", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "Who won the world series in 2020"}, + ], + "response_format": { "type": "json_object" } + }, + }, + "tool_calling": { + "summary": "Tool Calling", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "Extract Jason is 30 years old."}, + ], + "tools": [ + { + "type": "function", + "function": { + "name": "User", + "description": "User record", + "parameters": { + "type": "object", + "properties": { + "name": {"type": "string"}, + "age": {"type": "number"}, + }, + "required": ["name", "age"], + }, + } + } + ], + "tool_choice": { + "type": "function", + "function": { + "name": "User", + } + } + }, + }, + } + ), llama_proxy: LlamaProxy = Depends(get_llama_proxy), ) -> llama_cpp.ChatCompletion: exclude = { From 740f3f38125d1cc1bbe80f25944a292d0e966868 Mon Sep 17 00:00:00 2001 From: bretello Date: Wed, 20 Mar 2024 17:46:09 +0100 Subject: [PATCH 245/624] fix: set LLAMA_METAL_EMBED_LIBRARY=on on MacOS arm64 (#1289) --- CMakeLists.txt | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index b4df8ef45..741514907 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -17,6 +17,11 @@ if (LLAMA_BUILD) set(LLAMA_FMA "Off" CACHE BOOL "llama: enable FMA" FORCE) set(LLAMA_F16C "Off" CACHE BOOL "llama: enable F16C" FORCE) endif() + + if (APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") + set(LLAMA_METAL_EMBED_LIBRARY "On" CACHE BOOL "llama: embed metal library" FORCE) + endif() + add_subdirectory(vendor/llama.cpp) install( TARGETS llama From 3db03b73027036cf336fda2448894c36d3899cab Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 20 Mar 2024 13:27:43 -0400 Subject: [PATCH 246/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index b80cf3b2d..f9c7ba344 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit b80cf3b2d1dee0ad325f7a794fecc66befce7336 +Subproject commit f9c7ba34476ffc4f13ae2cdb1aec493a16eb8d47 From c89be28ef945017e8b64ea5f194ae7073fd872e8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 20 Mar 2024 20:50:47 -0400 Subject: [PATCH 247/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f9c7ba344..42e21c688 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f9c7ba34476ffc4f13ae2cdb1aec493a16eb8d47 +Subproject commit 42e21c68826f2e56b9592dccd9f3c43895b6890d From e325a831f015fdc807b436bc3c48e52cce658a18 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 22 Mar 2024 23:43:29 -0400 Subject: [PATCH 248/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 64 ++++++++++++++++++++++++++++++++++++------ vendor/llama.cpp | 2 +- 2 files changed, 57 insertions(+), 9 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 6b5c1bc15..1b8f6caa9 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -668,13 +668,15 @@ class llama_context_params(ctypes.Structure): # // model quantization parameters # typedef struct llama_model_quantize_params { -# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() -# enum llama_ftype ftype; // quantize to this llama_ftype -# bool allow_requantize; // allow quantizing non-f32/f16 tensors -# bool quantize_output_tensor; // quantize output.weight -# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored -# bool pure; // quantize all tensors to the default type -# void * imatrix; // pointer to importance matrix data +# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() +# enum llama_ftype ftype; // quantize to this llama_ftype +# enum ggml_type output_tensor_type; // output tensor type +# enum ggml_type token_embedding_type; // itoken embeddings tensor type +# bool allow_requantize; // allow quantizing non-f32/f16 tensors +# bool quantize_output_tensor; // quantize output.weight +# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored +# bool pure; // quantize all tensors to the default type +# void * imatrix; // pointer to importance matrix data # } llama_model_quantize_params; class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize @@ -682,16 +684,20 @@ class llama_model_quantize_params(ctypes.Structure): Attributes: nthread (int): number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() ftype (int): quantize to this llama_ftype + output_tensor_type (int): output tensor type + token_embedding_type (int): itoken embeddings tensor type allow_requantize (bool): allow quantizing non-f32/f16 tensors quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored pure (bool): quantize all tensors to the default type - imatrix (ctypes.ctypes.c_void_p): pointer to importance matrix data + imatrix (ctypes.c_void_p): pointer to importance matrix data """ _fields_ = [ ("nthread", ctypes.c_int32), ("ftype", ctypes.c_int), + ("output_tensor_type", ctypes.c_int), + ("token_embedding_type", ctypes.c_int), ("allow_requantize", ctypes.c_bool), ("quantize_output_tensor", ctypes.c_bool), ("only_copy", ctypes.c_bool), @@ -2743,6 +2749,48 @@ def llama_beam_search( ): ... +# /// @details Build a split GGUF final path for this chunk. +# /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" +# // Returns the split_path length. +# LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count); +@ctypes_function( + "llama_split_path", + [ctypes.c_char_p, ctypes.c_size_t, ctypes.c_char_p, ctypes.c_int, ctypes.c_int], + ctypes.c_int, +) +def llama_split_path( + split_path: bytes, + maxlen: Union[ctypes.c_size_t, int], + path_prefix: bytes, + split_no: Union[ctypes.c_int, int], + split_count: Union[ctypes.c_int, int], + /, +) -> int: + """Build a split GGUF final path for this chunk.""" + ... + + +# /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match. +# /// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0" +# // Returns the split_prefix length. +# LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count); +@ctypes_function( + "llama_split_prefix", + [ctypes.c_char_p, ctypes.c_size_t, ctypes.c_char_p, ctypes.c_int, ctypes.c_int], + ctypes.c_int, +) +def llama_split_prefix( + split_prefix: bytes, + maxlen: Union[ctypes.c_size_t, int], + split_path: bytes, + split_no: Union[ctypes.c_int, int], + split_count: Union[ctypes.c_int, int], + /, +) -> int: + """Extract the path prefix from the split_path if and only if the split_no and split_count match.""" + ... + + # Performance information diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 42e21c688..50ccaf5ea 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 42e21c68826f2e56b9592dccd9f3c43895b6890d +Subproject commit 50ccaf5eacb50a2ca378a4ef0dc7aeb45fead652 From c1325dcdfba7cb331b4c09d110001658b94ebb9f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 22 Mar 2024 23:44:04 -0400 Subject: [PATCH 249/624] fix: tool_call missing first token. --- llama_cpp/llama_chat_format.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 5bda163fd..ccf4fd0b7 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -402,7 +402,7 @@ def _stream_response_to_function_stream( "type": "function", "function": { "name": tool_name, - "arguments": "", + "arguments": chunk["choices"][0]["text"], }, } ], From d11ccc30361a947177e46d31c632988661f5af02 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 23 Mar 2024 17:14:15 -0400 Subject: [PATCH 250/624] fix(server): minor type fixes --- llama_cpp/server/app.py | 6 +++--- llama_cpp/server/types.py | 2 +- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 2e1081e42..d5e2f7ee8 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -493,7 +493,7 @@ async def tokenize( ) -> TokenizeInputResponse: tokens = llama_proxy(body.model).tokenize(body.input.encode("utf-8"), special=True) - return {"tokens": tokens} + return TokenizeInputResponse(tokens=tokens) @router.post( @@ -508,7 +508,7 @@ async def count_query_tokens( ) -> TokenizeInputCountResponse: tokens = llama_proxy(body.model).tokenize(body.input.encode("utf-8"), special=True) - return {"count": len(tokens)} + return TokenizeInputCountResponse(count=len(tokens)) @router.post( @@ -523,4 +523,4 @@ async def detokenize( ) -> DetokenizeInputResponse: text = llama_proxy(body.model).detokenize(body.tokens).decode("utf-8") - return {"text": text} + return DetokenizeInputResponse(text=text) diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index c8b2ebc6c..378f8d74c 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -268,7 +268,7 @@ class ModelList(TypedDict): class TokenizeInputRequest(BaseModel): model: Optional[str] = model_field - input: Optional[str] = Field(description="The input to tokenize.") + input: str = Field(description="The input to tokenize.") model_config = { "json_schema_extra": {"examples": [{"input": "How many tokens in this query?"}]} From 364678bde55ca75db2788563e2d75e6303720e6d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 24 Mar 2024 12:27:49 -0400 Subject: [PATCH 251/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 50ccaf5ea..a0e584def 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 50ccaf5eacb50a2ca378a4ef0dc7aeb45fead652 +Subproject commit a0e584defd8c16e7a51ab895f595df0448d710d0 From a93b9149f8ebab3219680e8f8cd46e53177aad74 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 25 Mar 2024 11:10:14 -0400 Subject: [PATCH 252/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index a0e584def..2f34b865b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit a0e584defd8c16e7a51ab895f595df0448d710d0 +Subproject commit 2f34b865b62b1d2b5eb8a27885e4de220deeacbd From b64fa4e2c05385d7cb36154dce42a29ca2b669ee Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 25 Mar 2024 23:09:07 -0400 Subject: [PATCH 253/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 2f34b865b..e190f1fca 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 2f34b865b62b1d2b5eb8a27885e4de220deeacbd +Subproject commit e190f1fca6f60d80944f9e8709d343a025c4d245 From 901fe02461fbad0aa817060b8a4062924b89f418 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 26 Mar 2024 22:58:53 -0400 Subject: [PATCH 254/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 30 ++++++++++++++++++++---------- vendor/llama.cpp | 2 +- 2 files changed, 21 insertions(+), 11 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1b8f6caa9..dc39e4c3b 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -175,8 +175,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 4 -LLAMA_SESSION_VERSION = 4 +# define LLAMA_SESSION_VERSION 5 +LLAMA_SESSION_VERSION = 5 # struct llama_model; @@ -274,6 +274,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -677,6 +678,7 @@ class llama_context_params(ctypes.Structure): # bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored # bool pure; // quantize all tensors to the default type # void * imatrix; // pointer to importance matrix data +# void * kv_overrides; // pointer to vector containing overrides # } llama_model_quantize_params; class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize @@ -691,6 +693,7 @@ class llama_model_quantize_params(ctypes.Structure): only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored pure (bool): quantize all tensors to the default type imatrix (ctypes.c_void_p): pointer to importance matrix data + kv_overrides (ctypes.c_void_p): pointer to vector containing overrides """ _fields_ = [ @@ -703,6 +706,7 @@ class llama_model_quantize_params(ctypes.Structure): ("only_copy", ctypes.c_bool), ("pure", ctypes.c_bool), ("imatrix", ctypes.c_void_p), + ("kv_overrides", ctypes.c_void_p), ] @@ -1838,9 +1842,9 @@ def llama_synchronize(ctx: llama_context_p, /): # // Token logits obtained from the last call to llama_decode() -# // The logits for the last token are stored in the last row -# // Logits for which llama_batch.logits[i] == 0 are undefined -# // Rows: n_tokens provided with llama_batch +# // The logits for which llama_batch.logits[i] != 0 are stored contiguously +# // in the order they have appeared in the batch. +# // Rows: number of tokens for which llama_batch.logits[i] != 0 # // Cols: n_vocab # LLAMA_API float * llama_get_logits(struct llama_context * ctx); @ctypes_function( @@ -1859,7 +1863,8 @@ def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: # // Logits for the ith token. Equivalent to: -# // llama_get_logits(ctx) + i*n_vocab +# // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab +# // returns NULL for invalid ids. # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); @ctypes_function( "llama_get_logits_ith", @@ -1874,8 +1879,12 @@ def llama_get_logits_ith( ... -# // Get all output token embeddings -# // shape: [n_tokens*n_embd] (1-dimensional) +# // Get all output token embeddings. +# // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model, +# // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously +# // in the order they have appeared in the batch. +# // shape: [n_outputs*n_embd] +# // Otherwise, returns NULL. # LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); @ctypes_function( "llama_get_embeddings", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) @@ -1886,9 +1895,10 @@ def llama_get_embeddings(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float] ... -# // Get the embeddings for the ith token -# // llama_get_embeddings(ctx) + i*n_embd +# // Get the embeddings for the ith token. Equivalent to: +# // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd # // shape: [n_embd] (1-dimensional) +# // returns NULL for invalid ids. # LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); @ctypes_function( "llama_get_embeddings_ith", diff --git a/vendor/llama.cpp b/vendor/llama.cpp index e190f1fca..32c8486e1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit e190f1fca6f60d80944f9e8709d343a025c4d245 +Subproject commit 32c8486e1f0297393cb22ac0a0d26a6b17ad4d54 From 125b2358c9ea92cbcb416e2bd2c8f93132e641ac Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 28 Mar 2024 12:06:46 -0400 Subject: [PATCH 255/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 10 +++++++--- vendor/llama.cpp | 2 +- 2 files changed, 8 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index dc39e4c3b..1db47becc 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -199,14 +199,18 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # enum llama_vocab_type { # LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab -# LLAMA_VOCAB_TYPE_SPM = 1, // SentencePiece -# LLAMA_VOCAB_TYPE_BPE = 2, // Byte Pair Encoding -# LLAMA_VOCAB_TYPE_WPM = 3, // WordPiece +# LLAMA_VOCAB_TYPE_SPM = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback +# LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE +# LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece # }; LLAMA_VOCAB_TYPE_NONE = 0 +"""For models without vocab""" LLAMA_VOCAB_TYPE_SPM = 1 +"""LLaMA tokenizer based on byte-level BPE with byte fallback""" LLAMA_VOCAB_TYPE_BPE = 2 +"""GPT-2 tokenizer based on byte-level BPE""" LLAMA_VOCAB_TYPE_WPM = 3 +"""BERT tokenizer based on WordPiece""" # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 32c8486e1..5106ef482 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 32c8486e1f0297393cb22ac0a0d26a6b17ad4d54 +Subproject commit 5106ef482c65ac60ac14da9a68c7b37bca4c6993 From dcbe57fcf8c4af829ef0085720b8730b0516f99a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 29 Mar 2024 12:45:27 -0400 Subject: [PATCH 256/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5106ef482..069574775 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5106ef482c65ac60ac14da9a68c7b37bca4c6993 +Subproject commit 069574775cea67d8a977904e4d534aff47a671f4 From 1e60dba082464b8828dca0a6d05a2fe46fcc4f7c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 29 Mar 2024 13:34:23 -0400 Subject: [PATCH 257/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 069574775..ba0c7c70a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 069574775cea67d8a977904e4d534aff47a671f4 +Subproject commit ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c From aa9f1ae011fbc22893750209af500fee3167f21c Mon Sep 17 00:00:00 2001 From: windspirit95 <32880949+windspirit95@users.noreply.github.com> Date: Mon, 1 Apr 2024 02:30:13 +0900 Subject: [PATCH 258/624] feat: Add logprobs support to chat completions (#1311) * Add logprobs return in ChatCompletionResponse * Fix duplicate field * Set default to false * Simplify check * Add server example --------- Co-authored-by: Andrei Betlen --- llama_cpp/llama.py | 1 + llama_cpp/llama_chat_format.py | 5 +++++ llama_cpp/llama_types.py | 1 + llama_cpp/server/app.py | 12 ++++++++++++ llama_cpp/server/types.py | 10 +++++++++- 5 files changed, 28 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index fed84d579..66caaa958 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1653,6 +1653,7 @@ def create_chat_completion( top_k=top_k, min_p=min_p, typical_p=typical_p, + logprobs=top_logprobs if logprobs else None, stream=stream, stop=stop, seed=seed, diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index ccf4fd0b7..06cf9ced7 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -231,6 +231,7 @@ def _convert_text_completion_to_chat( "role": "assistant", "content": completion["choices"][0]["text"], }, + "logprobs": completion["choices"][0]["logprobs"], "finish_reason": completion["choices"][0]["finish_reason"], } ], @@ -254,6 +255,7 @@ def _convert_text_completion_chunks_to_chat( "delta": { "role": "assistant", }, + "logprobs": None, "finish_reason": None, } ], @@ -273,6 +275,7 @@ def _convert_text_completion_chunks_to_chat( if chunk["choices"][0]["finish_reason"] is None else {} ), + "logprobs": chunk["choices"][0]["logprobs"], "finish_reason": chunk["choices"][0]["finish_reason"], } ], @@ -487,6 +490,7 @@ def chat_completion_handler( temperature: float = 0.2, top_p: float = 0.95, top_k: int = 40, + logprobs: int = 0, min_p: float = 0.05, typical_p: float = 1.0, stream: bool = False, @@ -576,6 +580,7 @@ def chat_completion_handler( top_k=top_k, min_p=min_p, typical_p=typical_p, + logprobs=logprobs, stream=stream, stop=stop, seed=seed, diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 1b1befebe..87e000f14 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -84,6 +84,7 @@ class ChatCompletionFunction(TypedDict): class ChatCompletionResponseChoice(TypedDict): index: int message: "ChatCompletionResponseMessage" + logprobs: Optional[CompletionLogprobs] finish_reason: Optional[str] diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index d5e2f7ee8..815ed3c5e 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -405,6 +405,18 @@ async def create_chat_completion( } }, }, + "logprobs": { + "summary": "Logprobs", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "What is the capital of France?"}, + ], + "logprobs": True, + "top_logprobs": 10 + }, + }, } ), llama_proxy: LlamaProxy = Depends(get_llama_proxy), diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index 378f8d74c..ce9c87a69 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -130,7 +130,6 @@ class CreateCompletionRequest(BaseModel): presence_penalty: Optional[float] = presence_penalty_field frequency_penalty: Optional[float] = frequency_penalty_field logit_bias: Optional[Dict[str, float]] = Field(None) - logprobs: Optional[int] = Field(None) seed: Optional[int] = Field(None) # ignored or currently unsupported @@ -209,6 +208,15 @@ class CreateChatCompletionRequest(BaseModel): default=None, description="The maximum number of tokens to generate. Defaults to inf", ) + logprobs: Optional[bool] = Field( + default=False, + description="Whether to output the logprobs or not. Default is True" + ) + top_logprobs: Optional[int] = Field( + default=None, + ge=0, + description="The number of logprobs to generate. If None, no logprobs are generated. logprobs need to set to True.", + ) temperature: float = temperature_field top_p: float = top_p_field min_p: float = min_p_field From f165048a69b2ad5b70d4fd2ec4adebfb411b5b47 Mon Sep 17 00:00:00 2001 From: Limour <93720049+Limour-dev@users.noreply.github.com> Date: Mon, 1 Apr 2024 22:19:28 +0800 Subject: [PATCH 259/624] feat: add support for KV cache quantization options (#1307) * add KV cache quantization options https://github.com/abetlen/llama-cpp-python/discussions/1220 https://github.com/abetlen/llama-cpp-python/issues/1305 * Add ggml_type * Use ggml_type instead of string for quantization * Add server support --------- Co-authored-by: Andrei Betlen --- llama_cpp/llama.py | 59 ++++++++++----------------------- llama_cpp/llama_cpp.py | 64 ++++++++++++++++++++++++++++++++++++ llama_cpp/server/model.py | 3 ++ llama_cpp/server/settings.py | 9 +++++ 4 files changed, 94 insertions(+), 41 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 66caaa958..dcc7be758 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -105,6 +105,9 @@ def __init__( draft_model: Optional[LlamaDraftModel] = None, # Tokenizer Override tokenizer: Optional[BaseLlamaTokenizer] = None, + # KV cache quantization + type_k: Optional[int] = None, + type_v: Optional[int] = None, # Misc verbose: bool = True, # Extra Params @@ -172,6 +175,8 @@ def __init__( draft_model: Optional draft model to use for speculative decoding. tokenizer: Optional tokenizer to override the default tokenizer from llama.cpp. verbose: Print verbose output to stderr. + type_k: KV cache data type for K (default: f16) + type_v: KV cache data type for V (default: f16) Raises: ValueError: If the model path does not exist. @@ -298,7 +303,11 @@ def __init__( ) # Must be set to True for speculative decoding self.context_params.embeddings = embedding # TODO: Rename to embeddings self.context_params.offload_kqv = offload_kqv - + # KV cache quantization + if type_k is not None: + self.context_params.type_k = type_k + if type_v is not None: + self.context_params.type_v = type_v # Sampling Params self.last_n_tokens_size = last_n_tokens_size @@ -1724,6 +1733,7 @@ def __getstate__(self): n_threads=self.context_params.n_threads, n_threads_batch=self.context_params.n_threads_batch, rope_scaling_type=self.context_params.rope_scaling_type, + pooling_type=self.context_params.pooling_type, rope_freq_base=self.context_params.rope_freq_base, rope_freq_scale=self.context_params.rope_freq_scale, yarn_ext_factor=self.context_params.yarn_ext_factor, @@ -1733,6 +1743,7 @@ def __getstate__(self): yarn_orig_ctx=self.context_params.yarn_orig_ctx, logits_all=self.context_params.logits_all, embedding=self.context_params.embeddings, + offload_kqv=self.context_params.offload_kqv, # Sampling Params last_n_tokens_size=self.last_n_tokens_size, # LoRA Params @@ -1744,51 +1755,17 @@ def __getstate__(self): # Chat Format Params chat_format=self.chat_format, chat_handler=self.chat_handler, + # Speculative Decidng + draft_model=self.draft_model, + # KV cache quantization + type_k=self.context_params.type_k, + type_v=self.context_params.type_v, # Misc verbose=self.verbose, ) def __setstate__(self, state): - self.__init__( - model_path=state["model_path"], - # Model Params - n_gpu_layers=state["n_gpu_layers"], - split_mode=state["split_mode"], - main_gpu=state["main_gpu"], - tensor_split=state["tensor_split"], - vocab_only=state["vocab_only"], - use_mmap=state["use_mmap"], - use_mlock=state["use_mlock"], - kv_overrides=state["kv_overrides"], - # Context Params - seed=state["seed"], - n_ctx=state["n_ctx"], - n_batch=state["n_batch"], - n_threads=state["n_threads"], - n_threads_batch=state["n_threads_batch"], - rope_freq_base=state["rope_freq_base"], - rope_freq_scale=state["rope_freq_scale"], - rope_scaling_type=state["rope_scaling_type"], - yarn_ext_factor=state["yarn_ext_factor"], - yarn_attn_factor=state["yarn_attn_factor"], - yarn_beta_fast=state["yarn_beta_fast"], - yarn_beta_slow=state["yarn_beta_slow"], - yarn_orig_ctx=state["yarn_orig_ctx"], - logits_all=state["logits_all"], - embedding=state["embedding"], - # Sampling Params - last_n_tokens_size=state["last_n_tokens_size"], - # LoRA Params - lora_base=state["lora_base"], - lora_path=state["lora_path"], - # Backend Params - numa=state["numa"], - # Chat Format Params - chat_format=state["chat_format"], - chat_handler=state["chat_handler"], - # Misc - verbose=state["verbose"], - ) + self.__init__(**state) def save_state(self) -> LlamaState: assert self._ctx.ctx is not None diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1db47becc..accc02cb8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -141,6 +141,70 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa byref = ctypes.byref # type: ignore +# from ggml.h +# // NOTE: always add types at the end of the enum to keep backward compatibility +# enum ggml_type { +# GGML_TYPE_F32 = 0, +# GGML_TYPE_F16 = 1, +# GGML_TYPE_Q4_0 = 2, +# GGML_TYPE_Q4_1 = 3, +# // GGML_TYPE_Q4_2 = 4, support has been removed +# // GGML_TYPE_Q4_3 = 5, support has been removed +# GGML_TYPE_Q5_0 = 6, +# GGML_TYPE_Q5_1 = 7, +# GGML_TYPE_Q8_0 = 8, +# GGML_TYPE_Q8_1 = 9, +# GGML_TYPE_Q2_K = 10, +# GGML_TYPE_Q3_K = 11, +# GGML_TYPE_Q4_K = 12, +# GGML_TYPE_Q5_K = 13, +# GGML_TYPE_Q6_K = 14, +# GGML_TYPE_Q8_K = 15, +# GGML_TYPE_IQ2_XXS = 16, +# GGML_TYPE_IQ2_XS = 17, +# GGML_TYPE_IQ3_XXS = 18, +# GGML_TYPE_IQ1_S = 19, +# GGML_TYPE_IQ4_NL = 20, +# GGML_TYPE_IQ3_S = 21, +# GGML_TYPE_IQ2_S = 22, +# GGML_TYPE_IQ4_XS = 23, +# GGML_TYPE_I8 = 24, +# GGML_TYPE_I16 = 25, +# GGML_TYPE_I32 = 26, +# GGML_TYPE_I64 = 27, +# GGML_TYPE_F64 = 28, +# GGML_TYPE_IQ1_M = 29, +# GGML_TYPE_COUNT, +# }; +GGML_TYPE_F32 = 0 +GGML_TYPE_F16 = 1 +GGML_TYPE_Q4_0 = 2 +GGML_TYPE_Q4_1 = 3 +GGML_TYPE_Q5_0 = 6 +GGML_TYPE_Q5_1 = 7 +GGML_TYPE_Q8_0 = 8 +GGML_TYPE_Q8_1 = 9 +GGML_TYPE_Q2_K = 10 +GGML_TYPE_Q3_K = 11 +GGML_TYPE_Q4_K = 12 +GGML_TYPE_Q5_K = 13 +GGML_TYPE_Q6_K = 14 +GGML_TYPE_Q8_K = 15 +GGML_TYPE_IQ2_XXS = 16 +GGML_TYPE_IQ2_XS = 17 +GGML_TYPE_IQ3_XXS = 18 +GGML_TYPE_IQ1_S = 19 +GGML_TYPE_IQ4_NL = 20 +GGML_TYPE_IQ3_S = 21 +GGML_TYPE_IQ2_S = 22 +GGML_TYPE_IQ4_XS = 23 +GGML_TYPE_I8 = 24 +GGML_TYPE_I16 = 25 +GGML_TYPE_I32 = 26 +GGML_TYPE_I64 = 27 +GGML_TYPE_F64 = 28 +GGML_TYPE_IQ1_M = 29 +GGML_TYPE_COUNT = 30 # from ggml-backend.h # typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index dace8d547..c24fca652 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -175,6 +175,9 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler=chat_handler, # Speculative Decoding draft_model=draft_model, + # KV Cache Quantization + type_k=settings.type_k, + type_v=settings.type_v, # Tokenizer tokenizer=tokenizer, # Misc diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index daa913fac..9ebdd0d8c 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -159,6 +159,15 @@ class ModelSettings(BaseSettings): default=10, description="Number of tokens to predict using the draft model.", ) + # KV Cache Quantization + type_k: Optional[int] = Field( + default=None, + description="Type of the key cache quantization.", + ) + type_v: Optional[int] = Field( + default=None, + description="Type of the value cache quantization.", + ) # Misc verbose: bool = Field( default=True, description="Whether to print debug information." From a0f373e31009d678312ea8d9a3c70bac96232490 Mon Sep 17 00:00:00 2001 From: lawfordp2017 Date: Mon, 1 Apr 2024 08:21:00 -0600 Subject: [PATCH 260/624] fix: Changed local API doc references to hosted (#1317) --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 3323f3899..1e2b64018 100644 --- a/README.md +++ b/README.md @@ -321,7 +321,7 @@ For OpenAI API v1 compatibility, you use the [`create_chat_completion_openai_v1` ### JSON and JSON Schema Mode -To constrain chat responses to only valid JSON or a specific JSON Schema use the `response_format` argument in [`create_chat_completion`](http://localhost:8000/api-reference/#llama_cpp.Llama.create_chat_completion). +To constrain chat responses to only valid JSON or a specific JSON Schema use the `response_format` argument in [`create_chat_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion). #### JSON Mode @@ -529,7 +529,7 @@ llama = Llama( ### Embeddings -To generate text embeddings use [`create_embedding`](http://localhost:8000/api-reference/#llama_cpp.Llama.create_embedding). +To generate text embeddings use [`create_embedding`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_embedding). ```python import llama_cpp From 45bf5ae5829aab6d38f5920b620f653fb9261f16 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Apr 2024 10:28:22 -0400 Subject: [PATCH 261/624] chore: Bump version --- CHANGELOG.md | 11 ++++++++++- llama_cpp/__init__.py | 2 +- 2 files changed, 11 insertions(+), 2 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index a85eaa436..8786c01c8 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.58] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c +- feat: add support for KV cache quantization options by @Limour-dev in #1307 +- feat: Add logprobs support to chat completions by @windspirit95 in #1311 +- fix: set LLAMA_METAL_EMBED_LIBRARY=on on MacOS arm64 by @bretello in #1289 +- feat: Add tools/functions variables to Jinja2ChatFormatter, add function response formatting for all simple chat formats by @CISC in #1273 +- fix: Changed local API doc references to hosted by by @lawfordp2017 in #1317 + ## [0.2.57] - feat: Update llama.cpp to ggerganov/llama.cpp@ac9ee6a4ad740bc1ee484ede43e9f92b5af244c1 @@ -24,7 +33,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [0.2.55] -- feat: Update llama.cpp to ggerganov/9731134296af3a6839cd682e51d9c2109a871de5 +- feat: Update llama.cpp to ggerganov/llama.cpp@9731134296af3a6839cd682e51d9c2109a871de5 - docs: fix small typo in README: 'model know how' -> 'model knows how' by @boegel in #1244 ## [0.2.54] diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 1e802fa7e..e246595cc 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.57" \ No newline at end of file +__version__ = "0.2.58" \ No newline at end of file From 62aad610e155fdb74064316ee7e83a1251bc8fc0 Mon Sep 17 00:00:00 2001 From: Yuri Mikhailov Date: Tue, 2 Apr 2024 04:25:43 +0900 Subject: [PATCH 262/624] fix: last tokens passing to sample_repetition_penalties function (#1295) Co-authored-by: ymikhaylov Co-authored-by: Andrei --- llama_cpp/_internals.py | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 22d0bef8a..79f6543a8 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -730,12 +730,14 @@ def sample( if len(self.prev) > 0: nl_token = ctx_main.model.token_nl() nl_logit = logits_array[nl_token] - if self.params.penalty_last_n > 0: + last_tokens = self.prev[-self.params.penalty_last_n:] + last_tokens_size = min(len(last_tokens), self.params.penalty_last_n) + if last_tokens_size > 0: + last_tokens_p = (llama_cpp.llama_token * len(last_tokens))(*last_tokens) ctx_main.sample_repetition_penalties( token_data_array, - # TODO: Only create this once - (llama_cpp.llama_token * len(self.prev))(*self.prev), - self.params.penalty_last_n, + last_tokens_p, + last_tokens_size, self.params.penalty_repeat, self.params.penalty_freq, self.params.penalty_present, From e465157804b2aa734ac3b8f7957d3a92acf13fa9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 3 Apr 2024 00:55:19 -0400 Subject: [PATCH 263/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ba0c7c70a..52604860f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c +Subproject commit 52604860f93063ef98863921da697576af1c7665 From 8649d7671bd1a7c0d9cc6a5ad91c6ca286512ab3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 3 Apr 2024 15:30:31 -0400 Subject: [PATCH 264/624] fix: segfault when logits_all=False. Closes #1319 --- llama_cpp/llama.py | 18 ++++++++++-------- 1 file changed, 10 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index dcc7be758..e07d57ae4 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -535,14 +535,16 @@ def eval(self, tokens: Sequence[int]): # Save tokens self.input_ids[n_past : n_past + n_tokens] = batch # Save logits - rows = n_tokens - cols = self._n_vocab - offset = ( - 0 if self.context_params.logits_all else n_tokens - 1 - ) # NOTE: Only save the last token logits if logits_all is False - self.scores[n_past + offset : n_past + n_tokens, :].reshape(-1)[ - : - ] = self._ctx.get_logits()[offset * cols : rows * cols] + if self.context_params.logits_all: + rows = n_tokens + cols = self._n_vocab + logits = self._ctx.get_logits()[: rows * cols] + self.scores[n_past : n_past + n_tokens, :].reshape(-1)[: :] = logits + else: + rows = 1 + cols = self._n_vocab + logits = self._ctx.get_logits()[: rows * cols] + self.scores[n_past + n_tokens - 1, :].reshape(-1)[: :] = logits # Update n_tokens self.n_tokens += n_tokens From 5a930ee9a1a54d12d48cdb7ce8cc63f6bc41d1bb Mon Sep 17 00:00:00 2001 From: Andrei Date: Wed, 3 Apr 2024 15:32:13 -0400 Subject: [PATCH 265/624] feat: Binary wheels for CPU, CUDA (12.1 - 12.3), Metal (#1247) * Generate binary wheel index on release * Add total release downloads badge * Update download label * Use official cibuildwheel action * Add workflows to build CUDA and Metal wheels * Update generate index workflow * Update workflow name --- .github/workflows/build-and-release.yaml | 10 +- .github/workflows/build-wheels-cuda.yaml | 131 ++++++++++++++++++ .github/workflows/build-wheels-metal.yaml | 87 ++++++++++++ .../generate-index-from-release.yaml | 48 +++++++ README.md | 1 + scripts/release-to-pep-503.sh | 58 ++++++++ 6 files changed, 330 insertions(+), 5 deletions(-) create mode 100644 .github/workflows/build-wheels-cuda.yaml create mode 100644 .github/workflows/build-wheels-metal.yaml create mode 100644 .github/workflows/generate-index-from-release.yaml create mode 100755 scripts/release-to-pep-503.sh diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 63c81f1e3..76b5f7fcf 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -11,7 +11,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-latest, windows-latest, macOS-latest] + os: [ubuntu-20.04, windows-2019, macos-11] steps: - uses: actions/checkout@v3 @@ -23,19 +23,19 @@ jobs: with: python-version: "3.8" - - name: Install cibuildwheel - run: python -m pip install cibuildwheel==2.12.1 - - name: Install dependencies run: | python -m pip install --upgrade pip python -m pip install -e .[all] - name: Build wheels - run: python -m cibuildwheel --output-dir wheelhouse + uses: pypa/cibuildwheel@v2.16.5 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" + with: + package-dir: . + output-dir: wheelhouse - uses: actions/upload-artifact@v3 with: diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml new file mode 100644 index 000000000..a222dceb4 --- /dev/null +++ b/.github/workflows/build-wheels-cuda.yaml @@ -0,0 +1,131 @@ +name: Build Wheels (CUDA) + +on: workflow_dispatch + +permissions: + contents: write + +jobs: + define_matrix: + name: Define Build Matrix + runs-on: ubuntu-latest + outputs: + matrix: ${{ steps.set-matrix.outputs.matrix }} + defaults: + run: + shell: pwsh + + steps: + - name: Define Job Output + id: set-matrix + run: | + $matrix = @{ + 'os' = @('ubuntu-20.04', 'windows-latest') + 'pyver' = @("3.10", "3.11", "3.12") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2") + 'releasetag' = @("basic") + } + + $matrixOut = ConvertTo-Json $matrix -Compress + Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT + + build_wheels: + name: Build Wheel ${{ matrix.os }} ${{ matrix.pyver }} ${{ matrix.cuda }} ${{ matrix.releasetag == 'wheels' && 'AVX2' || matrix.releasetag }} + needs: define_matrix + runs-on: ${{ matrix.os }} + strategy: + matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }} + defaults: + run: + shell: pwsh + env: + CUDAVER: ${{ matrix.cuda }} + AVXVER: ${{ matrix.releasetag }} + + steps: + - uses: actions/checkout@v4 + with: + submodules: "recursive" + + - uses: actions/setup-python@v4 + with: + python-version: ${{ matrix.pyver }} + + - name: Setup Mamba + uses: conda-incubator/setup-miniconda@v2.2.0 + with: + activate-environment: "build" + python-version: ${{ matrix.pyver }} + miniforge-variant: Mambaforge + miniforge-version: latest + use-mamba: true + add-pip-as-python-dependency: true + auto-activate-base: false + + - name: VS Integration Cache + id: vs-integration-cache + if: runner.os == 'Windows' + uses: actions/cache@v3.3.2 + with: + path: ./MSBuildExtensions + key: cuda-${{ matrix.cuda }}-vs-integration + + - name: Get Visual Studio Integration + if: runner.os == 'Windows' && steps.vs-integration-cache.outputs.cache-hit != 'true' + run: | + if ($env:CUDAVER -eq '12.1.1') {$x = '12.1.0'} else {$x = $env:CUDAVER} + $links = (Invoke-RestMethod 'https://github.com/Jimver/cuda-toolkit/raw/dc0ca7bb29c5a92f7a963d3d5c93f8d59765136a/src/links/windows-links.ts').Trim().split().where({$_ -ne ''}) + for ($i=$q=0;$i -lt $links.count -and $q -lt 2;$i++) {if ($links[$i] -eq "'$x',") {$q++}} + Invoke-RestMethod $links[$i].Trim("'") -OutFile 'cudainstaller.zip' + & 'C:\Program Files\7-Zip\7z.exe' e cudainstaller.zip -oMSBuildExtensions -r *\MSBuildExtensions\* > $null + Remove-Item 'cudainstaller.zip' + + - name: Install Visual Studio Integration + if: runner.os == 'Windows' + run: | + $y = (gi '.\MSBuildExtensions').fullname + '\*' + (gi 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\MSBuild\Microsoft\VC\*\BuildCustomizations').fullname.foreach({cp $y $_}) + $cupath = 'CUDA_PATH_V' + $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','_') + echo "$cupath=$env:CONDA_PREFIX" >> $env:GITHUB_ENV + + - name: Install Dependencies + env: + MAMBA_DOWNLOAD_FAILFAST: "0" + MAMBA_NO_LOW_SPEED_LIMIT: "1" + run: | + $cudaVersion = $env:CUDAVER + mamba install -y 'cuda' -c nvidia/label/cuda-$cudaVersion + python -m pip install build wheel + + - name: Build Wheel + run: | + $cudaVersion = $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','') + $env:CUDA_PATH = $env:CONDA_PREFIX + $env:CUDA_HOME = $env:CONDA_PREFIX + $env:CUDA_TOOLKIT_ROOT_DIR = $env:CONDA_PREFIX + if ($IsLinux) { + $env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH + } + $env:VERBOSE = '1' + $env:CMAKE_ARGS = '-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all' + $env:CMAKE_ARGS = "-DLLAMA_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS" + if ($env:AVXVER -eq 'AVX') { + $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' + } + if ($env:AVXVER -eq 'AVX512') { + $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on' + } + if ($env:AVXVER -eq 'basic') { + $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' + } + python -m build --wheel + # write the build tag to the output + Write-Output "CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV + + - uses: softprops/action-gh-release@v1 + with: + files: dist/* + # Set tag_name to -cu + tag_name: ${{ github.ref_name }}-cu${{ env.CUDA_VERSION }} + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml new file mode 100644 index 000000000..2cca47786 --- /dev/null +++ b/.github/workflows/build-wheels-metal.yaml @@ -0,0 +1,87 @@ +name: Build Wheels (Metal) + +on: workflow_dispatch + +permissions: + contents: write + +jobs: + define_matrix: + name: Define Build Matrix + runs-on: ubuntu-latest + outputs: + matrix: ${{ steps.set-matrix.outputs.matrix }} + defaults: + run: + shell: pwsh + + steps: + - name: Define Job Output + id: set-matrix + run: | + $matrix = @{ + 'os' = @('macos-11', 'macos-12', 'macos-13') + 'pyver' = @('3.10', '3.11', '3.12') + } + + $matrixOut = ConvertTo-Json $matrix -Compress + Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT + + build_wheels: + name: ${{ matrix.os }} Python ${{ matrix.pyver }} + needs: define_matrix + runs-on: ${{ matrix.os }} + strategy: + matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }} + env: + OSVER: ${{ matrix.os }} + + steps: + - uses: actions/checkout@v4 + with: + submodules: "recursive" + + - uses: actions/setup-python@v4 + with: + python-version: ${{ matrix.pyver }} + + - name: Install Dependencies + run: | + python -m pip install build wheel cmake + + - name: Build Wheel + run: | + XCODE15PATH="/Applications/Xcode_15.0.app/Contents/Developer" + XCODE15BINPATH="${XCODE15PATH}/Toolchains/XcodeDefault.xctoolchain/usr/bin" + export CMAKE_ARGS="-DLLAMA_NATIVE=off -DLLAMA_METAL=on" + [[ "$OSVER" == "macos-13" ]] && export CC="${XCODE15BINPATH}/cc" && export CXX="${XCODE15BINPATH}/c++" && export MACOSX_DEPLOYMENT_TARGET="13.0" + [[ "$OSVER" == "macos-12" ]] && export MACOSX_DEPLOYMENT_TARGET="12.0" + [[ "$OSVER" == "macos-11" ]] && export MACOSX_DEPLOYMENT_TARGET="11.0" + + export CMAKE_OSX_ARCHITECTURES="arm64" && export ARCHFLAGS="-arch arm64" + VERBOSE=1 python -m build --wheel + + if [[ "$OSVER" == "macos-13" ]]; then + export SDKROOT="${XCODE15PATH}/Platforms/MacOSX.platform/Developer/SDKs/MacOSX14.0.sdk" + export MACOSX_DEPLOYMENT_TARGET="14.0" + VERBOSE=1 python -m build --wheel + fi + + for file in ./dist/*.whl; do cp "$file" "${file/arm64.whl/aarch64.whl}"; done + + export CMAKE_OSX_ARCHITECTURES="x86_64" && export CMAKE_ARGS="-DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_METAL=on" && export ARCHFLAGS="-arch x86_64" + VERBOSE=1 python -m build --wheel + + if [[ "$OSVER" == "macos-13" ]]; then + export SDKROOT="${XCODE15PATH}/Platforms/MacOSX.platform/Developer/SDKs/MacOSX14.0.sdk" + export MACOSX_DEPLOYMENT_TARGET="14.0" + VERBOSE=1 python -m build --wheel + fi + + - uses: softprops/action-gh-release@v1 + with: + files: dist/* + # set release name to -metal + tag_name: ${{ github.ref_name }}-metal + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml new file mode 100644 index 000000000..9042d6cfe --- /dev/null +++ b/.github/workflows/generate-index-from-release.yaml @@ -0,0 +1,48 @@ +name: Wheels Index + +on: + # Trigger on any new release + release: + types: [published] + + # Allows you to run this workflow manually from the Actions tab + workflow_dispatch: + +# Sets permissions of the GITHUB_TOKEN to allow deployment to GitHub Pages +permissions: + contents: read + pages: write + id-token: write + +# Allow only one concurrent deployment, skipping runs queued between the run in-progress and latest queued. +# However, do NOT cancel in-progress runs as we want to allow these production deployments to complete. +concurrency: + group: "pages" + cancel-in-progress: false + +jobs: + # Single deploy job since we're just deploying + deploy: + environment: + name: github-pages + url: ${{ steps.deployment.outputs.page_url }} + runs-on: ubuntu-latest + steps: + - name: Checkout + uses: actions/checkout@v4 + - name: Setup Pages + uses: actions/configure-pages@v4 + - name: Build + run: | + ./scripts/releases-to-pep-503.sh index/whl/cpu '^[v]?[0-9]+\.[0-9]+\.[0-9]+$' + ./scripts/releases-to-pep-503.sh index/whl/cu121 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu121$' + ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' + ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' + - name: Upload artifact + uses: actions/upload-pages-artifact@v3 + with: + # Upload entire repository + path: 'index' + - name: Deploy to GitHub Pages + id: deployment + uses: actions/deploy-pages@v4 diff --git a/README.md b/README.md index 1e2b64018..f251f45f9 100644 --- a/README.md +++ b/README.md @@ -6,6 +6,7 @@ [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/llama-cpp-python)](https://pypi.org/project/llama-cpp-python/) [![PyPI - License](https://img.shields.io/pypi/l/llama-cpp-python)](https://pypi.org/project/llama-cpp-python/) [![PyPI - Downloads](https://img.shields.io/pypi/dm/llama-cpp-python)](https://pypi.org/project/llama-cpp-python/) +[![Github All Releases](https://img.shields.io/github/downloads/abetlen/llama-cpp-python/total.svg?label=Github%20Downloads)]() Simple Python bindings for **@ggerganov's** [`llama.cpp`](https://github.com/ggerganov/llama.cpp) library. This package provides: diff --git a/scripts/release-to-pep-503.sh b/scripts/release-to-pep-503.sh new file mode 100755 index 000000000..00ae56721 --- /dev/null +++ b/scripts/release-to-pep-503.sh @@ -0,0 +1,58 @@ +#!/bin/bash + +# Get output directory or default to index/whl/cpu +output_dir=${1:-"index/whl/cpu"} + +# Create output directory +mkdir -p $output_dir + +# Change to output directory +pushd $output_dir + +# Create an index html file +echo "" > index.html +echo "" >> index.html +echo " " >> index.html +echo " " >> index.html +echo " llama-cpp-python" >> index.html +echo "
" >> index.html +echo " " >> index.html +echo "" >> index.html +echo "" >> index.html + +# Create llama-cpp-python directory +mkdir -p llama-cpp-python + +# Change to llama-cpp-python directory +pushd llama-cpp-python + +# Create an index html file +echo "" > index.html +echo "" >> index.html +echo " " >> index.html +echo "

Links for llama-cpp-python

" >> index.html + +# Get all releases +releases=$(curl -s https://api.github.com/repos/abetlen/llama-cpp-python/releases | jq -r .[].tag_name) + +# Get pattern from second arg or default to valid python package version pattern +pattern=${2:-"^[v]?[0-9]+\.[0-9]+\.[0-9]+$"} + +# Filter releases by pattern +releases=$(echo $releases | tr ' ' '\n' | grep -E $pattern) + +# For each release, get all assets +for release in $releases; do + assets=$(curl -s https://api.github.com/repos/abetlen/llama-cpp-python/releases/tags/$release | jq -r .assets) + echo "

$release

" >> index.html + for asset in $(echo $assets | jq -r .[].browser_download_url); do + if [[ $asset == *".whl" ]]; then + echo " $asset" >> index.html + echo "
" >> index.html + fi + done +done + +echo " " >> index.html +echo "" >> index.html +echo "" >> index.html From 5a5193636b4fe6c1b186bf3c8ee18e1b96401bed Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 3 Apr 2024 15:35:28 -0400 Subject: [PATCH 266/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 52604860f..60cdf40cc 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 52604860f93063ef98863921da697576af1c7665 +Subproject commit 60cdf40cc32f0ad4cb11e0ca8fd38f3b93d8d640 From 34081ddc5bbe2fe667ab201abcdbda64c5caa49c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 3 Apr 2024 15:38:27 -0400 Subject: [PATCH 267/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 8786c01c8..fc4e29b38 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.59] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c +- feat: Binary wheels for CPU, CUDA (12.1 - 12.3), Metal by @abetlen, @jllllll, and @oobabooga in #1247 +- fix: segfault when logits_all=False by @abetlen in 8649d7671bd1a7c0d9cc6a5ad91c6ca286512ab3 +- fix: last tokens passing to sample_repetition_penalties function by @ymikhailov in #1295 + ## [0.2.58] - feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index e246595cc..2f5219cc8 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.58" \ No newline at end of file +__version__ = "0.2.59" \ No newline at end of file From 612e78d32297002b52b264f29ae32bcd85ad54c6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 3 Apr 2024 16:15:29 -0400 Subject: [PATCH 268/624] fix(ci): use correct script name --- scripts/{release-to-pep-503.sh => releases-to-pep-503.sh} | 0 1 file changed, 0 insertions(+), 0 deletions(-) rename scripts/{release-to-pep-503.sh => releases-to-pep-503.sh} (100%) diff --git a/scripts/release-to-pep-503.sh b/scripts/releases-to-pep-503.sh similarity index 100% rename from scripts/release-to-pep-503.sh rename to scripts/releases-to-pep-503.sh From c50309e52ab90ffc71ebfc083f7397d8d79851eb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 4 Apr 2024 02:49:19 -0400 Subject: [PATCH 269/624] docs: LLAMA_CUBLAS -> LLAMA_CUDA --- README.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/README.md b/README.md index f251f45f9..5e16614f2 100644 --- a/README.md +++ b/README.md @@ -102,10 +102,10 @@ CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-
cuBLAS (CUDA) -To install with cuBLAS, set the `LLAMA_CUBLAS=on` environment variable before installing: +To install with cuBLAS, set the `LLAMA_CUDA=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python +CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python ```
@@ -569,7 +569,7 @@ python3 -m llama_cpp.server --model models/7B/llama-model.gguf Similar to Hardware Acceleration section above, you can also install with GPU (cuBLAS) support like this: ```bash -CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install 'llama-cpp-python[server]' +CMAKE_ARGS="-DLLAMA_CUDA=on" FORCE_CMAKE=1 pip install 'llama-cpp-python[server]' python3 -m llama_cpp.server --model models/7B/llama-model.gguf --n_gpu_layers 35 ``` From 1db3b58fdc7c216749401a2901fc64495032f1ff Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 4 Apr 2024 02:57:06 -0400 Subject: [PATCH 270/624] docs: Add docs explaining how to install pre-built wheels. --- README.md | 45 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+) diff --git a/README.md b/README.md index 5e16614f2..a0ef83ccf 100644 --- a/README.md +++ b/README.md @@ -44,6 +44,15 @@ This will also build `llama.cpp` from source and install it alongside this pytho If this fails, add `--verbose` to the `pip install` see the full cmake build log. +**Pre-built Wheel (New)** + +It is also possible to install a pre-built wheel with basic CPU support. + +```bash +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu +``` + ### Installation Configuration `llama.cpp` supports a number of hardware acceleration backends to speed up inference as well as backend specific options. See the [llama.cpp README](https://github.com/ggerganov/llama.cpp#build) for a full list. @@ -108,6 +117,30 @@ To install with cuBLAS, set the `LLAMA_CUDA=on` environment variable before inst CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python ``` +**Pre-built Wheel (New)** + +It is also possible to install a pre-built wheel with CUDA support. As long as your system meets some requirements: + +- CUDA Version is 12.1, 12.2 or 12.3 +- Python Version is 3.10, 3.11 or 3.12 + +```bash +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/ +``` + +Where `` is one of the following: +- `cu121`: CUDA 12.1 +- `cu122`: CUDA 12.2 +- `cu123`: CUDA 12.3 + +For example, to install the CUDA 12.1 wheel: + +```bash +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121 +``` +
@@ -119,6 +152,18 @@ To install with Metal (MPS), set the `LLAMA_METAL=on` environment variable befor CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python ``` +**Pre-built Wheel (New)** + +It is also possible to install a pre-built wheel with Metal support. As long as your system meets some requirements: + +- MacOS Version is 11.0 or later +- Python Version is 3.10, 3.11 or 3.12 + +```bash +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal +``` +
From 909ef66951fb9f3f5c724841b0c98a598a2e0802 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 4 Apr 2024 03:08:47 -0400 Subject: [PATCH 271/624] docs: Rename cuBLAS section to CUDA --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index a0ef83ccf..c4e194bfa 100644 --- a/README.md +++ b/README.md @@ -109,9 +109,9 @@ CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-
-cuBLAS (CUDA) +CUDA -To install with cuBLAS, set the `LLAMA_CUDA=on` environment variable before installing: +To install with CUDA support, set the `LLAMA_CUDA=on` environment variable before installing: ```bash CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python From 7265a5dc0eb230a646ad868bad7ed5160d815328 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 5 Apr 2024 15:14:03 +0200 Subject: [PATCH 272/624] fix(docs): incorrect tool_choice example (#1330) --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index c4e194bfa..cd41a3804 100644 --- a/README.md +++ b/README.md @@ -458,12 +458,12 @@ The high-level API supports OpenAI compatible function and tool calling. This is } } }], - tool_choice=[{ + tool_choice={ "type": "function", "function": { "name": "UserDetail" } - }] + } ) ``` From 9111b6e03aba199e00cc7f3c7b6df2a19420665c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Apr 2024 09:21:02 -0400 Subject: [PATCH 273/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 60cdf40cc..a307375c0 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 60cdf40cc32f0ad4cb11e0ca8fd38f3b93d8d640 +Subproject commit a307375c02cac45cff53cf2520330b43fecc7718 From 49bc66bfa255f0c6d421cf142ecdbfb967546a51 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Apr 2024 10:50:49 -0400 Subject: [PATCH 274/624] fix: missing logprobs in response, incorrect response type for functionary, minor type issues. Closes #1328 #1314 --- llama_cpp/llama_chat_format.py | 48 ++++++++++++++++++++-------------- 1 file changed, 29 insertions(+), 19 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 06cf9ced7..705202e8c 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -6,7 +6,7 @@ import dataclasses import random import string -from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol +from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol, cast import jinja2 @@ -338,6 +338,7 @@ def _convert_completion_to_chat_function( } ], }, + "logprobs": None, "finish_reason": "tool_calls", } ], @@ -1191,7 +1192,6 @@ def format_mistral_instruct( elif ( message["role"] == "assistant" and message["content"] is not None - and isinstance(message["content"], str) ): prompt += " [/INST]" + message["content"] + eos prompt += " [/INST]" @@ -1263,7 +1263,7 @@ def format_gemma( **kwargs: Any, ) -> ChatFormatterResponse: system_message = _get_system_message(messages) - if system_message is not None and system_message != "": + if system_message != "": logger.debug( "`role='system'` messages are not allowed on Google's Gemma models." ) @@ -1628,6 +1628,7 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): } ], }, + "logprobs": None, "finish_reason": "tool_calls", } ], @@ -1909,14 +1910,14 @@ def get_grammar(function_call): return grammar def create_completion(stop): - completion: llama_types.Completion = llama.create_completion( + completion = cast(llama_types.Completion, llama.create_completion( prompt=prompt, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p, typical_p=typical_p, - stream=stream, + stream=False, stop=stop, max_tokens=max_tokens, presence_penalty=presence_penalty, @@ -1929,7 +1930,7 @@ def create_completion(stop): model=model, logits_processor=logits_processor, grammar=grammar, - ) + )) return completion @@ -2050,7 +2051,7 @@ def create_completion(stop): assert "usage" in completion assert len(function_calls) == len(function_bodies) - tool_calls = [] + tool_calls: List[llama_types.ChatCompletionMessageToolCall] = [] for function_call, function_body in zip(function_calls, function_bodies): tool_calls.append( { @@ -2070,6 +2071,12 @@ def create_completion(stop): ) # TODO: support stream mode + function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = { + "function_call": { + "name": tool_calls[0]["function"]["name"], + "arguments": tool_calls[0]["function"]["arguments"], + } + } if len(tool_calls) == 1 else {} return llama_types.CreateChatCompletionResponse( id="chat" + completion["id"], object="chat.completion", @@ -2078,14 +2085,12 @@ def create_completion(stop): choices=[ { "index": 0, + "logprobs": None, "message": { "role": "assistant", "content": None if content == "" else content, - "function_call": { - "name": tool_calls[0]["function"]["name"], - "arguments": tool_calls[0]["function"]["arguments"], - } if len(tool_calls) > 0 else None, - "tool_calls": tool_calls if len(tool_calls) > 0 else None, + "tool_calls": tool_calls, + **function_call_dict, }, "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", } @@ -2565,8 +2570,8 @@ def chatml_function_calling( tool_name = text[len("functions.") :] tool = next((tool for tool in tools if tool["function"]["name"] == tool_name), None) if not stream: - completions = [] - completions_tool_name = [] + completions: List[llama_types.CreateCompletionResponse] = [] + completions_tool_name: List[str] = [] while tool is not None: prompt += f"functions.{tool_name}:\n" try: @@ -2603,6 +2608,7 @@ def chatml_function_calling( logits_processor=logits_processor, grammar=grammar, ) + completion_or_chunks = cast(llama_types.CreateCompletionResponse, completion_or_chunks) completions.append(completion_or_chunks) completions_tool_name.append(tool_name) prompt += completion_or_chunks["choices"][0]["text"] @@ -2631,6 +2637,7 @@ def chatml_function_calling( follow_up_gbnf_tool_grammar, verbose=llama.verbose ), ) + response = cast(llama_types.CreateCompletionResponse, response) tool_name = response["choices"][0]["text"][len("functions.") :] tool = next( @@ -2638,7 +2645,7 @@ def chatml_function_calling( ) # Merge completions - function_call = { + function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = { "function_call": { "name": tool_name, "arguments": completions[0]["choices"][0]["text"], @@ -2653,6 +2660,7 @@ def chatml_function_calling( { "finish_reason": "tool_calls", "index": 0, + "logprobs": None, "message": { "role": "assistant", "content": None, @@ -2673,20 +2681,22 @@ def chatml_function_calling( zip(completions_tool_name, completions) ) ], - **function_call + **function_call_dict }, } ], "usage": { "completion_tokens": sum( - completion["usage"]["completion_tokens"] + completion["usage"]["completion_tokens"] if "usage" in completion else 0 for completion in completions ), "prompt_tokens": sum( - completion["usage"]["prompt_tokens"] for completion in completions + completion["usage"]["prompt_tokens"] if "usage" in completion else 0 + for completion in completions ), "total_tokens": sum( - completion["usage"]["total_tokens"] for completion in completions + completion["usage"]["total_tokens"] if "usage" in completion else 0 + for completion in completions ), }, } From 1ae3abbcc3af7f4a25a3ffc40b246f18039565e8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Apr 2024 10:50:49 -0400 Subject: [PATCH 275/624] fix: missing logprobs in response, incorrect response type for functionary, minor type issues. Closes #1328 Closes #1314 --- llama_cpp/llama_chat_format.py | 48 ++++++++++++++++++++-------------- 1 file changed, 29 insertions(+), 19 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 06cf9ced7..705202e8c 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -6,7 +6,7 @@ import dataclasses import random import string -from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol +from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol, cast import jinja2 @@ -338,6 +338,7 @@ def _convert_completion_to_chat_function( } ], }, + "logprobs": None, "finish_reason": "tool_calls", } ], @@ -1191,7 +1192,6 @@ def format_mistral_instruct( elif ( message["role"] == "assistant" and message["content"] is not None - and isinstance(message["content"], str) ): prompt += " [/INST]" + message["content"] + eos prompt += " [/INST]" @@ -1263,7 +1263,7 @@ def format_gemma( **kwargs: Any, ) -> ChatFormatterResponse: system_message = _get_system_message(messages) - if system_message is not None and system_message != "": + if system_message != "": logger.debug( "`role='system'` messages are not allowed on Google's Gemma models." ) @@ -1628,6 +1628,7 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): } ], }, + "logprobs": None, "finish_reason": "tool_calls", } ], @@ -1909,14 +1910,14 @@ def get_grammar(function_call): return grammar def create_completion(stop): - completion: llama_types.Completion = llama.create_completion( + completion = cast(llama_types.Completion, llama.create_completion( prompt=prompt, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p, typical_p=typical_p, - stream=stream, + stream=False, stop=stop, max_tokens=max_tokens, presence_penalty=presence_penalty, @@ -1929,7 +1930,7 @@ def create_completion(stop): model=model, logits_processor=logits_processor, grammar=grammar, - ) + )) return completion @@ -2050,7 +2051,7 @@ def create_completion(stop): assert "usage" in completion assert len(function_calls) == len(function_bodies) - tool_calls = [] + tool_calls: List[llama_types.ChatCompletionMessageToolCall] = [] for function_call, function_body in zip(function_calls, function_bodies): tool_calls.append( { @@ -2070,6 +2071,12 @@ def create_completion(stop): ) # TODO: support stream mode + function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = { + "function_call": { + "name": tool_calls[0]["function"]["name"], + "arguments": tool_calls[0]["function"]["arguments"], + } + } if len(tool_calls) == 1 else {} return llama_types.CreateChatCompletionResponse( id="chat" + completion["id"], object="chat.completion", @@ -2078,14 +2085,12 @@ def create_completion(stop): choices=[ { "index": 0, + "logprobs": None, "message": { "role": "assistant", "content": None if content == "" else content, - "function_call": { - "name": tool_calls[0]["function"]["name"], - "arguments": tool_calls[0]["function"]["arguments"], - } if len(tool_calls) > 0 else None, - "tool_calls": tool_calls if len(tool_calls) > 0 else None, + "tool_calls": tool_calls, + **function_call_dict, }, "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", } @@ -2565,8 +2570,8 @@ def chatml_function_calling( tool_name = text[len("functions.") :] tool = next((tool for tool in tools if tool["function"]["name"] == tool_name), None) if not stream: - completions = [] - completions_tool_name = [] + completions: List[llama_types.CreateCompletionResponse] = [] + completions_tool_name: List[str] = [] while tool is not None: prompt += f"functions.{tool_name}:\n" try: @@ -2603,6 +2608,7 @@ def chatml_function_calling( logits_processor=logits_processor, grammar=grammar, ) + completion_or_chunks = cast(llama_types.CreateCompletionResponse, completion_or_chunks) completions.append(completion_or_chunks) completions_tool_name.append(tool_name) prompt += completion_or_chunks["choices"][0]["text"] @@ -2631,6 +2637,7 @@ def chatml_function_calling( follow_up_gbnf_tool_grammar, verbose=llama.verbose ), ) + response = cast(llama_types.CreateCompletionResponse, response) tool_name = response["choices"][0]["text"][len("functions.") :] tool = next( @@ -2638,7 +2645,7 @@ def chatml_function_calling( ) # Merge completions - function_call = { + function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = { "function_call": { "name": tool_name, "arguments": completions[0]["choices"][0]["text"], @@ -2653,6 +2660,7 @@ def chatml_function_calling( { "finish_reason": "tool_calls", "index": 0, + "logprobs": None, "message": { "role": "assistant", "content": None, @@ -2673,20 +2681,22 @@ def chatml_function_calling( zip(completions_tool_name, completions) ) ], - **function_call + **function_call_dict }, } ], "usage": { "completion_tokens": sum( - completion["usage"]["completion_tokens"] + completion["usage"]["completion_tokens"] if "usage" in completion else 0 for completion in completions ), "prompt_tokens": sum( - completion["usage"]["prompt_tokens"] for completion in completions + completion["usage"]["prompt_tokens"] if "usage" in completion else 0 + for completion in completions ), "total_tokens": sum( - completion["usage"]["total_tokens"] for completion in completions + completion["usage"]["total_tokens"] if "usage" in completion else 0 + for completion in completions ), }, } From f4092e6b46711c92723258fac6326074ad51daa7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Apr 2024 10:59:31 -0400 Subject: [PATCH 276/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index a307375c0..1b496a745 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit a307375c02cac45cff53cf2520330b43fecc7718 +Subproject commit 1b496a745c315022df2d919374052e6004ced8d3 From b3bfea6dbfb6ed9ce18f9a2723e0a9e4bd1da7ad Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 6 Apr 2024 01:36:53 -0400 Subject: [PATCH 277/624] fix: Always embed metal library. Closes #1332 --- CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 741514907..70f9b9993 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -18,7 +18,7 @@ if (LLAMA_BUILD) set(LLAMA_F16C "Off" CACHE BOOL "llama: enable F16C" FORCE) endif() - if (APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") + if (APPLE) set(LLAMA_METAL_EMBED_LIBRARY "On" CACHE BOOL "llama: embed metal library" FORCE) endif() From 7ca364c8bd8669add75eb4143d36651495d5c743 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 6 Apr 2024 01:37:43 -0400 Subject: [PATCH 278/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1b496a745..75cd4c772 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1b496a745c315022df2d919374052e6004ced8d3 +Subproject commit 75cd4c77292034ecec587ecb401366f57338f7c0 From 08b16afe11e7b42adec2fed0a781123383476045 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 6 Apr 2024 01:53:38 -0400 Subject: [PATCH 279/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index fc4e29b38..dcb7a816c 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.60] + +- feat: Update llama.cpp to ggerganov/llama.cpp@75cd4c77292034ecec587ecb401366f57338f7c0 +- fix: Always embed metal library by @abetlen in b3bfea6dbfb6ed9ce18f9a2723e0a9e4bd1da7ad +- fix: missing logprobs in response, incorrect response type for functionary by @abetlen in 1ae3abbcc3af7f4a25a3ffc40b246f18039565e8 +- fix(docs): incorrect tool_choice example by @CISC in #1330 + ## [0.2.59] - feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 2f5219cc8..d5db993d9 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.59" \ No newline at end of file +__version__ = "0.2.60" \ No newline at end of file From 56071c956a947355ae63fe878448fe03df3b7586 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Apr 2024 09:53:49 -0400 Subject: [PATCH 280/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 237 +++++++++++++++++++++++++++++++++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 231 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index accc02cb8..879308595 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -237,11 +237,18 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn' LLAMA_FILE_MAGIC_GGSN = 0x6767736E +#define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq' +LLAMA_FILE_MAGIC_GGSQ = 0x67677371 + # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN # define LLAMA_SESSION_VERSION 5 LLAMA_SESSION_VERSION = 5 +#define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ +LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ +#define LLAMA_STATE_SEQ_VERSION 1 +LLAMA_STATE_SEQ_VERSION = 1 # struct llama_model; llama_model_p = NewType("llama_model_p", int) @@ -1467,6 +1474,7 @@ def llama_kv_cache_clear(ctx: llama_context_p, /): # // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) +# // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails # // seq_id < 0 : match any sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) @@ -1493,6 +1501,9 @@ def llama_kv_cache_seq_rm( /, ) -> bool: """Removes all tokens that belong to the specified sequence and have positions in [p0, p1) + + Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails + seq_id < 0 : match any sequence p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" @@ -1652,7 +1663,16 @@ def llama_kv_cache_update(ctx: llama_context_p, /): # Returns the maximum size in bytes of the state (rng, logits, embedding # and kv_cache) - will often be smaller after compacting tokens -# LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); +# LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx); +@ctypes_function("llama_state_get_size", [llama_context_p_ctypes], ctypes.c_size_t) +def llama_state_get_size(ctx: llama_context_p, /) -> int: + """Returns the maximum size in bytes of the state (rng, logits, embedding + and kv_cache) - will often be smaller after compacting tokens""" + ... + + +# LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx), +# "use llama_state_get_size instead"); @ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_get_state_size(ctx: llama_context_p, /) -> int: """Returns the maximum size in bytes of the state (rng, logits, embedding @@ -1663,9 +1683,30 @@ def llama_get_state_size(ctx: llama_context_p, /) -> int: # Copies the state to the specified destination address. # Destination needs to have allocated enough memory. # Returns the number of bytes copied -# LLAMA_API size_t llama_copy_state_data( +# LLAMA_API size_t llama_state_get_data( # struct llama_context * ctx, # uint8_t * dst); +@ctypes_function( + "llama_state_get_data", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ], + ctypes.c_size_t, +) +def llama_state_get_data( + ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], / +) -> int: + """Copies the state to the specified destination address. + Destination needs to have allocated enough memory. + Returns the number of bytes copied""" + ... + + +# LLAMA_API DEPRECATED(size_t llama_copy_state_data( +# struct llama_context * ctx, +# uint8_t * dst), +# "use llama_state_get_data instead"); @ctypes_function( "llama_copy_state_data", [ @@ -1685,9 +1726,26 @@ def llama_copy_state_data( # // Set the state reading from the specified address # // Returns the number of bytes read -# LLAMA_API size_t llama_set_state_data( +# LLAMA_API size_t llama_state_set_data( # struct llama_context * ctx, # const uint8_t * src); +@ctypes_function( + "llama_state_set_data", + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], + ctypes.c_size_t, +) +def llama_state_set_data( + ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], / +) -> int: + """Set the state reading from the specified address + Returns the number of bytes read""" + ... + + +# LLAMA_API DEPRECATED(size_t llama_set_state_data( +# struct llama_context * ctx, +# const uint8_t * src), +# "use llama_state_set_data instead"); @ctypes_function( "llama_set_state_data", [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], @@ -1701,12 +1759,40 @@ def llama_set_state_data( # Save/load session file -# LLAMA_API bool llama_load_session_file( +# LLAMA_API bool llama_state_load_file( # struct llama_context * ctx, # const char * path_session, # llama_token * tokens_out, # size_t n_token_capacity, # size_t * n_token_count_out); +@ctypes_function( + "llama_state_load_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ctypes.POINTER(ctypes.c_size_t), + ], + ctypes.c_bool, +) +def llama_state_load_file( + ctx: llama_context_p, + path_session: bytes, + tokens_out: CtypesArray[llama_token], + n_token_capacity: Union[ctypes.c_size_t, int], + n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], + /, +) -> bool: ... + + +# LLAMA_API DEPRECATED(bool llama_load_session_file( +# struct llama_context * ctx, +# const char * path_session, +# llama_token * tokens_out, +# size_t n_token_capacity, +# size_t * n_token_count_out), +# "use llama_state_load_file instead"); @ctypes_function( "llama_load_session_file", [ @@ -1728,11 +1814,36 @@ def llama_load_session_file( ) -> int: ... -# LLAMA_API bool llama_save_session_file( +# LLAMA_API bool llama_state_save_file( # struct llama_context * ctx, # const char * path_session, # const llama_token * tokens, # size_t n_token_count); +@ctypes_function( + "llama_state_save_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ], + ctypes.c_bool, +) +def llama_state_save_file( + ctx: llama_context_p, + path_session: bytes, + tokens: CtypesArray[llama_token], + n_token_count: Union[ctypes.c_size_t, int], + /, +) -> bool: ... + + +# LLAMA_API DEPRECATED(bool llama_save_session_file( +# struct llama_context * ctx, +# const char * path_session, +# const llama_token * tokens, +# size_t n_token_count), +# "use llama_state_save_file instead"); @ctypes_function( "llama_save_session_file", [ @@ -1752,6 +1863,116 @@ def llama_save_session_file( ) -> int: ... +# // Get the exact size needed to copy the KV cache of a single sequence +# LLAMA_API size_t llama_state_seq_get_size( +# struct llama_context * ctx, +# llama_seq_id seq_id); +@ctypes_function( + "llama_state_seq_get_size", + [llama_context_p_ctypes, llama_seq_id], + ctypes.c_size_t, +) +def llama_state_seq_get_size(ctx: llama_context_p, seq_id: llama_seq_id, /) -> int: + """Get the exact size needed to copy the KV cache of a single sequence""" + ... + + +# // Copy the KV cache of a single sequence into the specified buffer +# LLAMA_API size_t llama_state_seq_get_data( +# struct llama_context * ctx, +# uint8_t * dst, +# llama_seq_id seq_id); +@ctypes_function( + "llama_state_seq_get_data", + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), llama_seq_id], + ctypes.c_size_t, +) +def llama_state_seq_get_data( + ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], seq_id: llama_seq_id, / +) -> int: + """Copy the KV cache of a single sequence into the specified buffer""" + ... + + +# // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence +# // Returns: +# // - Positive: Ok +# // - Zero: Failed to load +# LLAMA_API size_t llama_state_seq_set_data( +# struct llama_context * ctx, +# const uint8_t * src, +# llama_seq_id dest_seq_id); +@ctypes_function( + "llama_state_seq_set_data", + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), llama_seq_id], + ctypes.c_size_t, +) +def llama_state_seq_set_data( + ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], dest_seq_id: llama_seq_id, / +) -> int: + """Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence""" + ... + + +# LLAMA_API size_t llama_state_seq_save_file( +# struct llama_context * ctx, +# const char * filepath, +# llama_seq_id seq_id, +# const llama_token * tokens, +# size_t n_token_count); +@ctypes_function( + "llama_state_seq_save_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_seq_id, + llama_token_p, + ctypes.c_size_t, + ], + ctypes.c_size_t, +) +def llama_state_seq_save_file( + ctx: llama_context_p, + filepath: bytes, + seq_id: llama_seq_id, + tokens: CtypesArray[llama_token], + n_token_count: Union[ctypes.c_size_t, int], + /, +) -> int: + ... + + +# LLAMA_API size_t llama_state_seq_load_file( +# struct llama_context * ctx, +# const char * filepath, +# llama_seq_id dest_seq_id, +# llama_token * tokens_out, +# size_t n_token_capacity, +# size_t * n_token_count_out); +@ctypes_function( + "llama_state_seq_load_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_seq_id, + llama_token_p, + ctypes.c_size_t, + ctypes.POINTER(ctypes.c_size_t), + ], + ctypes.c_size_t, +) +def llama_state_seq_load_file( + ctx: llama_context_p, + filepath: bytes, + dest_seq_id: llama_seq_id, + tokens_out: CtypesArray[llama_token], + n_token_capacity: Union[ctypes.c_size_t, int], + n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], + /, +) -> int: + ... + + # // # // Decoding # // @@ -1930,8 +2151,9 @@ def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: ... -# // Logits for the ith token. Equivalent to: +# // Logits for the ith token. For positive indices, Equivalent to: # // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab +# // Negative indicies can be used to access logits in reverse order, -1 is the last logit. # // returns NULL for invalid ids. # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); @ctypes_function( @@ -1963,8 +2185,9 @@ def llama_get_embeddings(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float] ... -# // Get the embeddings for the ith token. Equivalent to: +# // Get the embeddings for the ith token. For positive indices, Equivalent to: # // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd +# // Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding. # // shape: [n_embd] (1-dimensional) # // returns NULL for invalid ids. # LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 75cd4c772..400d5d722 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 75cd4c77292034ecec587ecb401366f57338f7c0 +Subproject commit 400d5d722d7edf7de0cf24a18c42b183c65047d2 From 889d0e8981641b603259e9a9ee40ce3b3ee8db51 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 10 Apr 2024 02:25:58 -0400 Subject: [PATCH 281/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 31 ++++++++++++++++++++++--------- vendor/llama.cpp | 2 +- 2 files changed, 23 insertions(+), 10 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 879308595..99ae7dea8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2271,6 +2271,20 @@ def llama_token_eos(model: llama_model_p, /) -> int: ... +# LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification +@ctypes_function("llama_token_cls", [llama_model_p_ctypes], llama_token) +def llama_token_cls(model: llama_model_p, /) -> int: + """classification""" + ... + + +# LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator +@ctypes_function("llama_token_sep", [llama_model_p_ctypes], llama_token) +def llama_token_sep(model: llama_model_p, /) -> int: + """sentence separator""" + ... + + # LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line @ctypes_function("llama_token_nl", [llama_model_p_ctypes], llama_token) def llama_token_nl(model: llama_model_p, /) -> int: @@ -2326,16 +2340,16 @@ def llama_token_eot(model: llama_model_p, /) -> int: ... # /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. # /// @return Returns the number of tokens on success, no more than n_tokens_max # /// @return Returns a negative number on failure - the number of tokens that would have been returned -# /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. -# /// Does not insert a leading space. +# /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated +# /// as plaintext. Does not insert a leading space. # LLAMA_API int32_t llama_tokenize( # const struct llama_model * model, # const char * text, # int32_t text_len, # llama_token * tokens, # int32_t n_tokens_max, -# bool add_bos, -# bool special); +# bool add_special, +# bool parse_special); @ctypes_function( "llama_tokenize", [ @@ -2355,8 +2369,8 @@ def llama_tokenize( text_len: Union[ctypes.c_int, int], tokens: CtypesArray[llama_token], n_tokens_max: Union[ctypes.c_int, int], - add_bos: Union[ctypes.c_bool, bool], - special: Union[ctypes.c_bool, bool], + add_special: Union[ctypes.c_bool, bool], + parse_special: Union[ctypes.c_bool, bool], /, ) -> int: """Convert the provided text into tokens. @@ -2367,9 +2381,8 @@ def llama_tokenize( text_len: The length of the text. tokens: The tokens pointer must be large enough to hold the resulting tokens. n_max_tokens: The maximum number of tokens to return. - add_bos: Whether to add a beginning-of-sentence token. - special: Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. - Does not insert a leading space. + add_special: Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. Does not insert a leading space. + parse_special: Allow parsing special tokens. Returns: Returns the number of tokens on success, no more than n_tokens_max diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 400d5d722..ba5e134e0 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 400d5d722d7edf7de0cf24a18c42b183c65047d2 +Subproject commit ba5e134e073ec6837078c874aba44a702944a676 From 1347e1d050fc5a9a32ffe0bb3e22858da28003bd Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 10 Apr 2024 02:40:41 -0400 Subject: [PATCH 282/624] feat: Add typechecking for ctypes structure attributes --- llama_cpp/llama_cpp.py | 216 ++++++++++++++++++++++++++++++++++------- 1 file changed, 180 insertions(+), 36 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 99ae7dea8..2450d11e3 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -237,7 +237,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn' LLAMA_FILE_MAGIC_GGSN = 0x6767736E -#define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq' +# define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq' LLAMA_FILE_MAGIC_GGSQ = 0x67677371 # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN @@ -245,9 +245,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_SESSION_VERSION 5 LLAMA_SESSION_VERSION = 5 -#define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ +# define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ -#define LLAMA_STATE_SEQ_VERSION 1 +# define LLAMA_STATE_SEQ_VERSION 1 LLAMA_STATE_SEQ_VERSION = 1 # struct llama_model; @@ -431,6 +431,11 @@ class llama_token_data(ctypes.Structure): logit (float): log-odds of the token p (float): probability of the token""" + if TYPE_CHECKING: + id: llama_token + logit: float + p: float + _fields_ = [ ("id", llama_token), ("logit", ctypes.c_float), @@ -454,6 +459,11 @@ class llama_token_data_array(ctypes.Structure): size (int): size of the array sorted (bool): whether the array is sorted""" + if TYPE_CHECKING: + data: CtypesArray[llama_token_data] + size: int + sorted: bool + _fields_ = [ ("data", llama_token_data_p), ("size", ctypes.c_size_t), @@ -515,6 +525,15 @@ class llama_batch(ctypes.Structure): logits (ctypes.Array[ctypes.ctypes.c_int8]): if zero, the logits for the respective token will not be output """ + if TYPE_CHECKING: + n_tokens: int + token: CtypesArray[llama_token] + embd: CtypesArray[ctypes.c_float] + pos: CtypesArray[CtypesArray[llama_pos]] + n_seq_id: CtypesArray[ctypes.c_int] + seq_id: CtypesArray[CtypesArray[llama_seq_id]] + logits: CtypesArray[ctypes.c_int8] + _fields_ = [ ("n_tokens", ctypes.c_int32), ("token", ctypes.POINTER(llama_token)), @@ -609,6 +628,18 @@ class llama_model_params(ctypes.Structure): use_mmap (bool): use mmap if possible use_mlock (bool): force system to keep model in RAM""" + if TYPE_CHECKING: + n_gpu_layers: int + split_mode: int + main_gpu: int + tensor_split: CtypesArray[ctypes.c_float] + progress_callback: Callable[[float, ctypes.c_void_p], bool] + progress_callback_user_data: ctypes.c_void_p + kv_overrides: CtypesArray[llama_model_kv_override] + vocab_only: bool + use_mmap: bool + use_mlock: bool + _fields_ = [ ("n_gpu_layers", ctypes.c_int32), ("split_mode", ctypes.c_int), @@ -696,6 +727,34 @@ class llama_context_params(ctypes.Structure): abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback """ + if TYPE_CHECKING: + seed: int + n_ctx: int + n_batch: int + n_ubatch: int + n_seq_max: int + n_threads: int + n_threads_batch: int + rope_scaling_type: int + pooling_type: int + rope_freq_base: float + rope_freq_scale: float + yarn_ext_factor: float + yarn_attn_factor: float + yarn_beta_fast: float + yarn_beta_slow: float + yarn_orig_ctx: int + defrag_thold: float + cb_eval: Callable[[ctypes.c_void_p, bool], bool] + cb_eval_user_data: ctypes.c_void_p + type_k: int + type_v: int + logits_all: bool + embeddings: bool + offload_kqv: bool + abort_callback: Callable[[ctypes.c_void_p], bool] + abort_callback_data: ctypes.c_void_p + _fields_ = [ ("seed", ctypes.c_uint32), ("n_ctx", ctypes.c_uint32), @@ -771,6 +830,18 @@ class llama_model_quantize_params(ctypes.Structure): kv_overrides (ctypes.c_void_p): pointer to vector containing overrides """ + if TYPE_CHECKING: + nthread: int + ftype: int + output_tensor_type: int + token_embedding_type: int + allow_requantize: bool + quantize_output_tensor: bool + only_copy: bool + pure: bool + imatrix: ctypes.c_void_p + kv_overrides: ctypes.c_void_p + _fields_ = [ ("nthread", ctypes.c_int32), ("ftype", ctypes.c_int), @@ -828,6 +899,10 @@ class llama_model_quantize_params(ctypes.Structure): # uint32_t value; // Unicode code point or rule ID # } llama_grammar_element; class llama_grammar_element(ctypes.Structure): + if TYPE_CHECKING: + type: int + value: int + _fields_ = [ ("type", ctypes.c_int), ("value", ctypes.c_uint32), @@ -851,6 +926,17 @@ class llama_grammar_element(ctypes.Structure): # int32_t n_eval; # }; class llama_timings(ctypes.Structure): + if TYPE_CHECKING: + t_start_ms: float + t_end_ms: float + t_load_ms: float + t_sample_ms: float + t_p_eval_ms: float + t_eval_ms: float + n_sample: int + n_p_eval: int + n_eval: int + _fields_ = [ ("t_start_ms", ctypes.c_double), ("t_end_ms", ctypes.c_double), @@ -951,7 +1037,8 @@ def llama_backend_init(): [ctypes.c_int], None, ) -def llama_numa_init(numa: int, /): ... +def llama_numa_init(numa: int, /): + ... # // Call once at the end of the program - currently only used for MPI @@ -976,7 +1063,8 @@ def llama_backend_free(): ) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / -) -> Optional[llama_model_p]: ... +) -> Optional[llama_model_p]: + ... # LLAMA_API void llama_free_model(struct llama_model * model); @@ -985,7 +1073,8 @@ def llama_load_model_from_file( [llama_model_p_ctypes], None, ) -def llama_free_model(model: llama_model_p, /): ... +def llama_free_model(model: llama_model_p, /): + ... # LLAMA_API struct llama_context * llama_new_context_with_model( @@ -998,7 +1087,8 @@ def llama_free_model(model: llama_model_p, /): ... ) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / -) -> Optional[llama_context_p]: ... +) -> Optional[llama_context_p]: + ... # // Frees all allocated memory @@ -1019,82 +1109,98 @@ def llama_free(ctx: llama_context_p, /): [], ctypes.c_int64, ) -def llama_time_us() -> int: ... +def llama_time_us() -> int: + ... # LLAMA_API size_t llama_max_devices(void); @ctypes_function("llama_max_devices", [], ctypes.c_size_t) -def llama_max_devices() -> int: ... +def llama_max_devices() -> int: + ... # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) -def llama_supports_mmap() -> bool: ... +def llama_supports_mmap() -> bool: + ... # LLAMA_API bool llama_supports_mlock (void); @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) -def llama_supports_mlock() -> bool: ... +def llama_supports_mlock() -> bool: + ... # LLAMA_API bool llama_supports_gpu_offload(void); @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) -def llama_supports_gpu_offload() -> bool: ... +def llama_supports_gpu_offload() -> bool: + ... # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) -def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: + ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ctx(ctx: llama_context_p, /) -> int: ... +def llama_n_ctx(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_batch(ctx: llama_context_p, /) -> int: ... +def llama_n_batch(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); @ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ubatch(ctx: llama_context_p, /) -> int: ... +def llama_n_ubatch(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); @ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... +def llama_n_seq_max(ctx: llama_context_p, /) -> int: + ... # LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: ... +def llama_vocab_type(model: llama_model_p, /) -> int: + ... # LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: ... +def llama_rope_type(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: ... +def llama_n_vocab(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_ctx_train(model: llama_model_p, /) -> int: ... +def llama_n_ctx_train(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_embd(model: llama_model_p, /) -> int: ... +def llama_n_embd(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_layer (const struct llama_model * model); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_layer(model: llama_model_p, /) -> int: ... +def llama_n_layer(model: llama_model_p, /) -> int: + ... # // Get the model's RoPE frequency scaling factor @@ -1358,6 +1464,9 @@ class llama_kv_cache_view_cell(ctypes.Structure): pos (llama_pos): The position for this cell. Takes KV cache shifts into account. May be negative if the cell is not populated.""" + if TYPE_CHECKING: + pos: llama_pos + _fields_ = [("pos", llama_pos)] @@ -1394,6 +1503,16 @@ class llama_kv_cache_view_cell(ctypes.Structure): # llama_seq_id * cells_sequences; # }; class llama_kv_cache_view(ctypes.Structure): + if TYPE_CHECKING: + n_cells: int + n_max_seq: int + token_count: int + used_cells: int + max_contiguous: int + max_contiguous_idx: int + cells: CtypesArray[llama_kv_cache_view_cell] + cells_sequences: CtypesArray[llama_seq_id] + _fields_ = [ ("n_cells", ctypes.c_int32), ("n_max_seq", ctypes.c_int32), @@ -1783,7 +1902,8 @@ def llama_state_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> bool: ... +) -> bool: + ... # LLAMA_API DEPRECATED(bool llama_load_session_file( @@ -1811,7 +1931,8 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: ... +) -> int: + ... # LLAMA_API bool llama_state_save_file( @@ -1835,7 +1956,8 @@ def llama_state_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> bool: ... +) -> bool: + ... # LLAMA_API DEPRECATED(bool llama_save_session_file( @@ -1860,7 +1982,8 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: ... +) -> int: + ... # // Get the exact size needed to copy the KV cache of a single sequence @@ -2233,7 +2356,8 @@ def llama_get_embeddings_seq( ) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / -) -> bytes: ... +) -> bytes: + ... # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); @@ -2242,7 +2366,8 @@ def llama_token_get_text( ) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / -) -> float: ... +) -> float: + ... # LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); @@ -2251,7 +2376,8 @@ def llama_token_get_score( ) def llama_token_get_type( model: llama_model_p, token: Union[llama_token, int], / -) -> int: ... +) -> int: + ... # // Special tokens @@ -2318,17 +2444,20 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: ... +def llama_token_middle(model: llama_model_p, /) -> int: + ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: ... +def llama_token_suffix(model: llama_model_p, /) -> int: + ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: ... +def llama_token_eot(model: llama_model_p, /) -> int: + ... # // @@ -2459,7 +2588,8 @@ def llama_chat_apply_template( chat: CtypesArray[llama_chat_message], n_msg: int, /, -) -> int: ... +) -> int: + ... # // @@ -2989,6 +3119,12 @@ def llama_grammar_accept_token( # bool eob; // Callback should set this to true when a beam is at end-of-beam. # }; class llama_beam_view(ctypes.Structure): + if TYPE_CHECKING: + tokens: CtypesArray[llama_token] + n_tokens: int + p: float + eob: bool + _fields_ = [ ("tokens", llama_token_p), ("n_tokens", ctypes.c_size_t), @@ -3008,6 +3144,12 @@ class llama_beam_view(ctypes.Structure): # bool last_call; // True iff this is the last callback invocation. # }; class llama_beams_state(ctypes.Structure): + if TYPE_CHECKING: + beam_views: CtypesArray[llama_beam_view] + n_beams: int + common_prefix_length: int + last_call: bool + _fields_ = [ ("beam_views", ctypes.POINTER(llama_beam_view)), ("n_beams", ctypes.c_size_t), @@ -3060,7 +3202,8 @@ def llama_beam_search( n_past: Union[ctypes.c_int, int], n_predict: Union[ctypes.c_int, int], /, -): ... +): + ... # /// @details Build a split GGUF final path for this chunk. @@ -3179,4 +3322,5 @@ def llama_log_set( [ctypes.c_void_p, llama_context_p_ctypes], None, ) -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): ... +def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): + ... From 060bfa64d529ade2af9b1f4e207a3937bbc4138f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 10 Apr 2024 02:47:01 -0400 Subject: [PATCH 283/624] feat: Add support for yaml based configs --- examples/batch-processing/server.py | 30 +++++++++++++++++++++++++++++ llama_cpp/server/__main__.py | 11 ++++++++++- llama_cpp/server/app.py | 10 +++++++++- pyproject.toml | 1 + 4 files changed, 50 insertions(+), 2 deletions(-) create mode 100644 examples/batch-processing/server.py diff --git a/examples/batch-processing/server.py b/examples/batch-processing/server.py new file mode 100644 index 000000000..d3536697a --- /dev/null +++ b/examples/batch-processing/server.py @@ -0,0 +1,30 @@ +"""llama-cpp-python server from scratch in a single file. +""" + +# import llama_cpp + +# path = b"../../models/Qwen1.5-0.5B-Chat-GGUF/qwen1_5-0_5b-chat-q8_0.gguf" + +# model_params = llama_cpp.llama_model_default_params() +# model = llama_cpp.llama_load_model_from_file(path, model_params) + +# if model is None: +# raise RuntimeError(f"Failed to load model from file: {path}") + + +# ctx_params = llama_cpp.llama_context_default_params() +# ctx = llama_cpp.llama_new_context_with_model(model, ctx_params) + +# if ctx is None: +# raise RuntimeError("Failed to create context") + + +from fastapi import FastAPI + +app = FastAPI() + +import openai.types.chat as types + +@app.post("/v1/chat/completions") +def create_chat_completions(): + return {"message": "Hello World"} diff --git a/llama_cpp/server/__main__.py b/llama_cpp/server/__main__.py index fadfc5fb4..a6f1f4e9c 100644 --- a/llama_cpp/server/__main__.py +++ b/llama_cpp/server/__main__.py @@ -59,7 +59,16 @@ def main(): if not os.path.exists(config_file): raise ValueError(f"Config file {config_file} not found!") with open(config_file, "rb") as f: - config_file_settings = ConfigFileSettings.model_validate_json(f.read()) + # Check if yaml file + if config_file.endswith(".yaml") or config_file.endswith(".yml"): + import yaml + import json + + config_file_settings = ConfigFileSettings.model_validate_json( + json.dumps(yaml.safe_load(f)) + ) + else: + config_file_settings = ConfigFileSettings.model_validate_json(f.read()) server_settings = ServerSettings.model_validate(config_file_settings) model_settings = config_file_settings.models else: diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 815ed3c5e..8211323a4 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -97,7 +97,15 @@ def create_app( if not os.path.exists(config_file): raise ValueError(f"Config file {config_file} not found!") with open(config_file, "rb") as f: - config_file_settings = ConfigFileSettings.model_validate_json(f.read()) + # Check if yaml file + if config_file.endswith(".yaml") or config_file.endswith(".yml"): + import yaml + + config_file_settings = ConfigFileSettings.model_validate_json( + json.dumps(yaml.safe_load(f)) + ) + else: + config_file_settings = ConfigFileSettings.model_validate_json(f.read()) server_settings = ServerSettings.model_validate(config_file_settings) model_settings = config_file_settings.models diff --git a/pyproject.toml b/pyproject.toml index 2f3d3ced0..e2bbb4b0c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -35,6 +35,7 @@ server = [ "pydantic-settings>=2.0.1", "sse-starlette>=1.6.1", "starlette-context>=0.3.6,<0.4", + "PyYAML>=5.1", ] test = [ "pytest>=7.4.0", From bb65b4d76411112c6fb0bf759efd746f99ef3c6b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 10 Apr 2024 03:41:55 -0400 Subject: [PATCH 284/624] fix: pass correct type to chat handlers for chat completion logprobs --- llama_cpp/llama.py | 3 ++- llama_cpp/llama_chat_format.py | 24 ++++++++++++++++-------- 2 files changed, 18 insertions(+), 9 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index e07d57ae4..466dc22c5 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1664,7 +1664,8 @@ def create_chat_completion( top_k=top_k, min_p=min_p, typical_p=typical_p, - logprobs=top_logprobs if logprobs else None, + logprobs=logprobs, + top_logprobs=top_logprobs, stream=stream, stop=stop, seed=seed, diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 705202e8c..519d2f50a 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -77,6 +77,8 @@ def __call__( mirostat_eta: float = 0.1, logits_processor: Optional[llama.LogitsProcessorList] = None, grammar: Optional[llama.LlamaGrammar] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, **kwargs, # type: ignore ) -> Union[ llama_types.CreateChatCompletionResponse, @@ -338,7 +340,7 @@ def _convert_completion_to_chat_function( } ], }, - "logprobs": None, + "logprobs": completion["choices"][0]["logprobs"], "finish_reason": "tool_calls", } ], @@ -391,7 +393,7 @@ def _stream_response_to_function_stream( { "index": 0, "finish_reason": None, - "logprobs": None, + "logprobs": chunk["choices"][0]["logprobs"], "delta": { "role": None, "content": None, @@ -426,7 +428,7 @@ def _stream_response_to_function_stream( { "index": 0, "finish_reason": None, - "logprobs": None, + "logprobs": chunk["choices"][0]["logprobs"], "delta": { "role": None, "content": None, @@ -491,7 +493,6 @@ def chat_completion_handler( temperature: float = 0.2, top_p: float = 0.95, top_k: int = 40, - logprobs: int = 0, min_p: float = 0.05, typical_p: float = 1.0, stream: bool = False, @@ -512,6 +513,8 @@ def chat_completion_handler( logits_processor: Optional[llama.LogitsProcessorList] = None, grammar: Optional[llama.LlamaGrammar] = None, logit_bias: Optional[Dict[str, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, **kwargs, # type: ignore ) -> Union[ llama_types.CreateChatCompletionResponse, @@ -581,7 +584,7 @@ def chat_completion_handler( top_k=top_k, min_p=min_p, typical_p=typical_p, - logprobs=logprobs, + logprobs=top_logprobs if logprobs else None, stream=stream, stop=stop, seed=seed, @@ -1628,7 +1631,7 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): } ], }, - "logprobs": None, + "logprobs": completion["choices"][0]["logprobs"], "finish_reason": "tool_calls", } ], @@ -2085,7 +2088,7 @@ def create_completion(stop): choices=[ { "index": 0, - "logprobs": None, + "logprobs": completion["choices"][0]["logprobs"], "message": { "role": "assistant", "content": None if content == "" else content, @@ -2311,11 +2314,14 @@ def chatml_function_calling( model: Optional[str] = None, logits_processor: Optional[llama.LogitsProcessorList] = None, grammar: Optional[llama.LlamaGrammar] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, **kwargs, # type: ignore ) -> Union[ llama_types.CreateChatCompletionResponse, Iterator[llama_types.CreateChatCompletionStreamResponse], ]: + print(logprobs) function_calling_template = ( "{% for message in messages %}" "<|im_start|>{{ message.role }}\n" @@ -2437,6 +2443,7 @@ def chatml_function_calling( model=model, logits_processor=logits_processor, grammar=grammar, + logprobs=top_logprobs if logprobs else None, ), stream=stream, ) @@ -2549,6 +2556,7 @@ def chatml_function_calling( typical_p=typical_p, stream=stream, stop=["<|im_end|>"], + logprobs=top_logprobs if logprobs else None, max_tokens=None, presence_penalty=presence_penalty, frequency_penalty=frequency_penalty, @@ -2660,7 +2668,7 @@ def chatml_function_calling( { "finish_reason": "tool_calls", "index": 0, - "logprobs": None, + "logprobs": completion["choices"][0]["logprobs"], "message": { "role": "assistant", "content": None, From ef29235d453ecf552f015f6ec04270b233bb22c9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 10 Apr 2024 03:44:46 -0400 Subject: [PATCH 285/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index dcb7a816c..c67498ecd 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.61] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba5e134e073ec6837078c874aba44a702944a676 +- fix: pass correct type to chat handlers for chat completion logprobs by @abetlen in bb65b4d76411112c6fb0bf759efd746f99ef3c6b +- feat: Add support for yaml based server configs by @abetlen in 060bfa64d529ade2af9b1f4e207a3937bbc4138f +- feat: Add typechecking for ctypes structure attributes by @abetlen in 1347e1d050fc5a9a32ffe0bb3e22858da28003bd + ## [0.2.60] - feat: Update llama.cpp to ggerganov/llama.cpp@75cd4c77292034ecec587ecb401366f57338f7c0 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index d5db993d9..2382db9b2 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.60" \ No newline at end of file +__version__ = "0.2.61" \ No newline at end of file From 2e9ffd28fd99958a58fccad6db4614d8c6b555a5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 12 Apr 2024 21:09:12 -0400 Subject: [PATCH 286/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ba5e134e0..ab9a3240a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ba5e134e073ec6837078c874aba44a702944a676 +Subproject commit ab9a3240a9da941fdef5cd4a25f2b97c2f5a67aa From 90dceaba8a9938ca4d7e7d2cb0997e613bad7040 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 14 Apr 2024 11:35:57 -0400 Subject: [PATCH 287/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ab9a3240a..f184dd920 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ab9a3240a9da941fdef5cd4a25f2b97c2f5a67aa +Subproject commit f184dd920852d6d372b754f871ee06cfe6f977ad From a420f9608bbd3b76e8bfbb6cdcf4d3fa69efe5c0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 14 Apr 2024 19:14:09 -0400 Subject: [PATCH 288/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f184dd920..1958f7e06 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f184dd920852d6d372b754f871ee06cfe6f977ad +Subproject commit 1958f7e06ca2d2e3ab5698cc67513ba359144d8e From c96b2daebf7b5dbe3cfd9194b249ab579a28d633 Mon Sep 17 00:00:00 2001 From: ddh0 <40664579+ddh0@users.noreply.github.com> Date: Wed, 17 Apr 2024 09:04:33 -0500 Subject: [PATCH 289/624] feat: Use all available CPUs for batch processing (#1345) --- llama_cpp/llama.py | 4 +--- llama_cpp/server/settings.py | 2 +- 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 466dc22c5..dfac9bb8a 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -262,9 +262,7 @@ def __init__( self.n_batch = min(n_ctx, n_batch) # ??? self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1) - self.n_threads_batch = n_threads_batch or max( - multiprocessing.cpu_count() // 2, 1 - ) + self.n_threads_batch = n_threads_batch or multiprocessing.cpu_count() # Context Params self.context_params = llama_cpp.llama_context_default_params() diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 9ebdd0d8c..811c6ca66 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -70,7 +70,7 @@ class ModelSettings(BaseSettings): description="The number of threads to use.", ) n_threads_batch: int = Field( - default=max(multiprocessing.cpu_count() // 2, 1), + default=max(multiprocessing.cpu_count(), 1), ge=0, description="The number of threads to use when batch processing.", ) From 9842cbf99d048a121cff03f01340592ad4b7fc1e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Apr 2024 10:06:15 -0400 Subject: [PATCH 290/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1958f7e06..8dd1ec8b3 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1958f7e06ca2d2e3ab5698cc67513ba359144d8e +Subproject commit 8dd1ec8b3ffbfa2d26e82e672cea89f5eeb2f141 From 4924455decd79273c8c695a8ff796306ac0df30d Mon Sep 17 00:00:00 2001 From: tc-wolf <50339167+tc-wolf@users.noreply.github.com> Date: Wed, 17 Apr 2024 09:06:50 -0500 Subject: [PATCH 291/624] feat: Make saved state more compact on-disk (#1296) * State load/save changes - Only store up to `n_tokens` logits instead of full `(n_ctx, n_vocab)` sized array. - Difference between ~350MB and ~1500MB for example prompt with ~300 tokens (makes sense lol) - Auto-formatting changes * Back out formatting changes --- llama_cpp/llama.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index dfac9bb8a..5a0111bbe 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -18,6 +18,7 @@ Iterator, Deque, Callable, + Dict, ) from collections import deque from pathlib import Path @@ -1791,7 +1792,7 @@ def save_state(self) -> LlamaState: file=sys.stderr, ) return LlamaState( - scores=self.scores.copy(), + scores=self._scores.copy(), input_ids=self.input_ids.copy(), n_tokens=self.n_tokens, llama_state=bytes(llama_state_compact), @@ -1800,7 +1801,9 @@ def save_state(self) -> LlamaState: def load_state(self, state: LlamaState) -> None: assert self._ctx.ctx is not None - self.scores = state.scores.copy() + # Only filling in up to `n_tokens` and then zero-ing out the rest + self.scores[: state.n_tokens, :] = state.scores.copy() + self.scores[state.n_tokens :, :] = 0.0 self.input_ids = state.input_ids.copy() self.n_tokens = state.n_tokens state_size = state.llama_state_size @@ -1951,7 +1954,6 @@ def from_pretrained( local_dir_use_symlinks=local_dir_use_symlinks, cache_dir=cache_dir, local_files_only=True, - ) else: model_path = os.path.join(local_dir, filename) From b73c73c0c67f559fd9c7d620ad3d4e24d5c4bc4c Mon Sep 17 00:00:00 2001 From: khimaros Date: Wed, 17 Apr 2024 14:08:19 +0000 Subject: [PATCH 292/624] feat: add `disable_ping_events` flag (#1257) for backward compatibility, this is false by default it can be set to true to disable EventSource pings which are not supported by some OpenAI clients. fixes https://github.com/abetlen/llama-cpp-python/issues/1256 --- llama_cpp/server/app.py | 12 ++++++++++++ llama_cpp/server/settings.py | 4 ++++ 2 files changed, 16 insertions(+) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 8211323a4..b6ed9b1b6 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -87,6 +87,13 @@ def get_llama_proxy(): llama_outer_lock.release() +_ping_message_factory = None + +def set_ping_message_factory(factory): + global _ping_message_factory + _ping_message_factory = factory + + def create_app( settings: Settings | None = None, server_settings: ServerSettings | None = None, @@ -138,6 +145,9 @@ def create_app( assert model_settings is not None set_llama_proxy(model_settings=model_settings) + if server_settings.disable_ping_events: + set_ping_message_factory(lambda: bytes()) + return app @@ -302,6 +312,7 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: iterator=iterator(), ), sep="\n", + ping_message_factory=_ping_message_factory, ) else: return iterator_or_completion @@ -470,6 +481,7 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: iterator=iterator(), ), sep="\n", + ping_message_factory=_ping_message_factory, ) else: return iterator_or_completion diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 811c6ca66..934aecd91 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -195,6 +195,10 @@ class ServerSettings(BaseSettings): default=True, description="Whether to interrupt requests when a new request is received.", ) + disable_ping_events: bool = Field( + default=False, + description="Disable EventSource pings (may be needed for some clients).", + ) class Settings(ServerSettings, ModelSettings): From 610a592f708f2a5a8b0e1d6f0900f6337b9beb39 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Lucca=20Zen=C3=B3bio?= Date: Wed, 17 Apr 2024 11:10:21 -0300 Subject: [PATCH 293/624] feat: Update json to grammar (#1350) * feat: improve function calling * feat:grammar --- llama_cpp/llama_chat_format.py | 2 +- llama_cpp/llama_grammar.py | 612 +++++++++++++++++++++++++++------ 2 files changed, 514 insertions(+), 100 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 519d2f50a..eb98cbf39 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2709,4 +2709,4 @@ def chatml_function_calling( }, } - raise ValueError("Automatic streaming tool choice is not supported") + raise ValueError("Automatic streaming tool choice is not supported") \ No newline at end of file diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 9cc48a93b..8c0f8aa09 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -5,11 +5,12 @@ import sys from ctypes import * # type: ignore from enum import Enum -from itertools import islice +from itertools import islice, groupby from typing import ( Any, Callable, Dict, + Set, Generic, List, Optional, @@ -1391,139 +1392,552 @@ def print_grammar(file: TextIO, state: parse_state) -> None: # whitespace. Also maybe improves generation quality? SPACE_RULE = '" "?' -PRIMITIVE_RULES = { - "boolean": '("true" | "false") space', - "number": '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space', - "integer": '("-"? ([0-9] | [1-9] [0-9]*)) space', - "string": r""" "\"" ( - [^"\\] | - "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) - )* "\"" space """, - "null": '"null" space', -} INVALID_RULE_CHARS_RE = re.compile(r"[^a-zA-Z0-9-]+") GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]') GRAMMAR_LITERAL_ESCAPES = {"\r": "\\r", "\n": "\\n", '"': '\\"'} +# whitespace is constrained to a single space char to prevent model "running away" in +# whitespace. Also maybe improves generation quality? +SPACE_RULE = '" "?' + + +def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False): + if not separator_rule: + if min_items == 0 and max_items == 1: + return f'{item_rule}?' + elif min_items == 1 and max_items is None: + return f'{item_rule}+' + + result = '' + + if min_items > 0: + if item_rule_is_literal and separator_rule is None: + result = '"' + (item_rule[1:-1] * min_items) + '"' + else: + result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items) + + def opt_repetitions(up_to_n, prefix_with_sep=False): + ''' + - n=4, no sep: '(a (a (a (a)?)?)?)?' + - n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?' + - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?' + ''' + + content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule + if up_to_n == 0: + return '' + elif up_to_n == 1: + return f'({content})?' + elif separator_rule and not prefix_with_sep: + return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?' + else: + return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n) + + if min_items > 0 and max_items != min_items: + result += ' ' + + if max_items is not None: + result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0) + else: + item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})' + + if min_items == 0 and separator_rule: + result = f'({item_rule} {item_operator}*)?' + else: + result += f'{item_operator}*' + + return result + + + +class BuiltinRule: + def __init__(self, content: str, deps: list = None): + self.content = content + self.deps = deps or [] + +_up_to_15_digits = _build_repetition('[0-9]', 0, 15) + +PRIMITIVE_RULES = { + 'boolean' : BuiltinRule('("true" | "false") space', []), + 'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []), + 'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []), + 'number' : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']), + 'integer' : BuiltinRule('("-"? integral-part) space', ['integral-part']), + 'value' : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']), + 'object' : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']), + 'array' : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']), + 'uuid' : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []), + 'char' : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []), + 'string' : BuiltinRule(r'"\"" char* "\"" space', ['char']), + 'null' : BuiltinRule('"null" space', []), +} + +# TODO: support "uri", "email" string formats +STRING_FORMAT_RULES = { + 'date' : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []), + 'time' : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []), + 'date-time' : BuiltinRule('date "T" time', ['date', 'time']), + 'date-string' : BuiltinRule('"\\"" date "\\"" space', ['date']), + 'time-string' : BuiltinRule('"\\"" time "\\"" space', ['time']), + 'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']), +} + +DOTALL = '[\\U00000000-\\U0010FFFF]' +DOT = '[^\\x0A\\x0D]' + +RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()]) + + +NON_LITERAL_SET = set('|.()[]{}*+?') +ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?') + + + class SchemaConverter: def __init__(self, prop_order): self._prop_order = prop_order self._rules = {"space": SPACE_RULE} self._defs: Dict[str, Any] = {} + self._refs = {} + self._refs_being_resolved = set() - def _format_literal(self, literal: str): - escaped: str = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal) + def _format_literal(self, literal): + escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal ) return f'"{escaped}"' - def _add_rule(self, name: str, rule: str): - esc_name = INVALID_RULE_CHARS_RE.sub("-", name) + def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str: + ''' + not_literal('a') -> '[^a]' + not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?' + ''' + assert len(literal) > 0, 'Empty literal not supported' + def recurse(i: int): + c = literal[i] + if maybe_escaped_underscores and c == '_': + yield f'[^{c}\\\\]' + yield ' | ' + yield f'"\\\\"? "{c}"' + else: + yield f'[^{c}]' + if i < len(literal) - 1: + yield ' | ' + yield self._format_literal(c) + yield ' (' + yield from recurse(i + 1) + yield ')?' + + return ''.join(('(', *recurse(0), ')')) + + def _add_rule(self, name, rule): + esc_name = INVALID_RULE_CHARS_RE.sub('-', name) if esc_name not in self._rules or self._rules[esc_name] == rule: key = esc_name else: i = 0 - while f"{esc_name}{i}" in self._rules: + while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule: i += 1 - key = f"{esc_name}{i}" + key = f'{esc_name}{i}' self._rules[key] = rule return key - def visit(self, schema: Dict[str, Any], name: str) -> str: - rule_name = name or "root" - - if "$defs" in schema: - # add defs to self._defs for later inlining - for def_name, def_schema in schema["$defs"].items(): - self._defs[def_name] = def_schema + def resolve_refs(self, schema: dict, url: str): + ''' + Resolves all $ref fields in the given schema, fetching any remote schemas, + replacing $ref with absolute reference URL and populating self._refs with the + respective referenced (sub)schema dictionaries. + ''' + def visit(n: dict): + if isinstance(n, list): + return [visit(x) for x in n] + elif isinstance(n, dict): + ref = n.get('$ref') + if ref is not None and ref not in self._refs: + if ref.startswith('https://'): + assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)' + import requests + + frag_split = ref.split('#') + base_url = frag_split[0] + + target = self._refs.get(base_url) + if target is None: + target = self.resolve_refs(requests.get(ref).json(), base_url) + self._refs[base_url] = target + + if len(frag_split) == 1 or frag_split[-1] == '': + return target + elif ref.startswith('#/'): + target = schema + ref = f'{url}{ref}' + n['$ref'] = ref + else: + raise ValueError(f'Unsupported ref {ref}') + + for sel in ref.split('#')[-1].split('/')[1:]: + assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}' + target = target[sel] + + self._refs[ref] = target + else: + for v in n.values(): + visit(v) + + return n + return visit(schema) + + def _generate_union_rule(self, name, alt_schemas): + return ' | '.join(( + self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') + for i, alt_schema in enumerate(alt_schemas) + )) + + def _visit_pattern(self, pattern, name): + ''' + Transforms a regular expression pattern into a GBNF rule. + + Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions + Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md + + Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. + + Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which + we define sub-rules to keep the output lean. + ''' + + assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"' + pattern = pattern[1:-1] + sub_rule_ids = {} + + i = 0 + length = len(pattern) + + def to_rule(s: Tuple[str, bool]) -> str: + (txt, is_literal) = s + return "\"" + txt + "\"" if is_literal else txt + + def transform() -> Tuple[str, bool]: + ''' + Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. + ''' + nonlocal i + nonlocal pattern + nonlocal sub_rule_ids + + start = i + # For each component of this sequence, store its string representation and whether it's a literal. + # We only need a flat structure here to apply repetition operators to the last item, and + # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially + # (GBNF's syntax is luckily very close to regular expressions!) + seq: list[Tuple[str, bool]] = [] + + def get_dot(): + if self._dotall: + rule = DOTALL + else: + # Accept any character... except \n and \r line break chars (\x0A and \xOD) + rule = DOT + return self._add_rule(f'dot', rule) + + def join_seq(): + nonlocal seq + ret = [] + for is_literal, g in groupby(seq, lambda x: x[1]): + if is_literal: + ret.append((''.join(x[0] for x in g), True)) + else: + ret.extend(g) + if len(ret) == 1: + return ret[0] + return (' '.join(to_rule(x) for x in seq), False) + + while i < length: + c = pattern[i] + if c == '.': + seq.append((get_dot(), False)) + i += 1 + elif c == '(': + i += 1 + if i < length: + assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' + seq.append((f'({to_rule(transform())})', False)) + elif c == ')': + i += 1 + assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}' + return join_seq() + elif c == '[': + square_brackets = c + i += 1 + while i < length and pattern[i] != ']': + if pattern[i] == '\\': + square_brackets += pattern[i:i+2] + i += 2 + else: + square_brackets += pattern[i] + i += 1 + assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}' + square_brackets += ']' + i += 1 + seq.append((square_brackets, False)) + elif c == '|': + seq.append(('|', False)) + i += 1 + elif c in ('*', '+', '?'): + seq[-1] = (to_rule(seq[-1]) + c, False) + i += 1 + elif c == '{': + curly_brackets = c + i += 1 + while i < length and pattern[i] != '}': + curly_brackets += pattern[i] + i += 1 + assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}' + curly_brackets += '}' + i += 1 + nums = [s.strip() for s in curly_brackets[1:-1].split(',')] + min_times = 0 + max_times = None + try: + if len(nums) == 1: + min_times = int(nums[0]) + max_times = min_times + else: + assert len(nums) == 2 + min_times = int(nums[0]) if nums[0] else 0 + max_times = int(nums[1]) if nums[1] else None + except ValueError: + raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/') + + (sub, sub_is_literal) = seq[-1] + + if not sub_is_literal: + id = sub_rule_ids.get(sub) + if id is None: + id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub) + sub_rule_ids[sub] = id + sub = id + + seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False) + else: + literal = '' + while i < length: + if pattern[i] == '\\' and i < length - 1: + next = pattern[i + 1] + if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS: + i += 1 + literal += pattern[i] + i += 1 + else: + literal += pattern[i:i+2] + i += 2 + elif pattern[i] == '"' and not self._raw_pattern: + literal += '\\"' + i += 1 + elif pattern[i] not in NON_LITERAL_SET and \ + (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET): + literal += pattern[i] + i += 1 + else: + break + if literal: + seq.append((literal, True)) + + return join_seq() + + return self._add_rule( + name, + to_rule(transform()) if self._raw_pattern \ + else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space") + + + def _resolve_ref(self, ref): + ref_name = ref.split('/')[-1] + if ref_name not in self._rules and ref not in self._refs_being_resolved: + self._refs_being_resolved.add(ref) + resolved = self._refs[ref] + ref_name = self.visit(resolved, ref_name) + self._refs_being_resolved.remove(ref) + return ref_name + + def _generate_constant_rule(self, value): + return self._format_literal(json.dumps(value)) + + def visit(self, schema, name): + schema_type = schema.get('type') + schema_format = schema.get('format') + rule_name = name + '-' if name in RESERVED_NAMES else name or 'root' + + if (ref := schema.get('$ref')) is not None: + return self._add_rule(rule_name, self._resolve_ref(ref)) + + elif 'oneOf' in schema or 'anyOf' in schema: + return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf'])) + + elif isinstance(schema_type, list): + return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type])) + + elif 'const' in schema: + return self._add_rule(rule_name, self._generate_constant_rule(schema['const'])) + + elif 'enum' in schema: + rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) + return self._add_rule(rule_name, rule) - if "oneOf" in schema or "anyOf" in schema: - rule = " | ".join( - ( - self.visit(alt_schema, f'{name}{"-" if name else ""}{i}') - for i, alt_schema in enumerate( - schema.get("oneOf") or schema["anyOf"] - ) - ) + elif schema_type in (None, 'object') and \ + ('properties' in schema or \ + ('additionalProperties' in schema and schema['additionalProperties'] is not True)): + required = set(schema.get('required', [])) + properties = list(schema.get('properties', {}).items()) + return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties'))) + + elif schema_type in (None, 'object') and 'allOf' in schema: + required = set() + properties = [] + hybrid_name = name + def add_component(comp_schema, is_required): + if (ref := comp_schema.get('$ref')) is not None: + comp_schema = self._refs[ref] + + if 'properties' in comp_schema: + for prop_name, prop_schema in comp_schema['properties'].items(): + properties.append((prop_name, prop_schema)) + if is_required: + required.add(prop_name) + + for t in schema['allOf']: + if 'anyOf' in t: + for tt in t['anyOf']: + add_component(tt, is_required=False) + else: + add_component(t, is_required=True) + + return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[])) + + elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema): + items = schema.get('items') or schema['prefixItems'] + if isinstance(items, list): + return self._add_rule( + rule_name, + '"[" space ' + + ' "," space '.join( + self.visit(item, f'{name}{"-" if name else ""}tuple-{i}') + for i, item in enumerate(items)) + + ' "]" space') + else: + item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item') + min_items = schema.get("minItems", 0) + max_items = schema.get("maxItems") + return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space') + + elif schema_type in (None, 'string') and 'pattern' in schema: + return self._visit_pattern(schema['pattern'], rule_name) + + elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''): + return self._add_primitive( + 'root' if rule_name == 'root' else schema_format, + PRIMITIVE_RULES['uuid'] ) - return self._add_rule(rule_name, rule) - elif "const" in schema: - return self._add_rule(rule_name, self._format_literal(schema["const"])) + elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES: + prim_name = f'{schema_format}-string' + return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name])) - elif "enum" in schema: - rule = " | ".join((self._format_literal(v) for v in schema["enum"])) - return self._add_rule(rule_name, rule) + elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema): + char_rule = self._add_primitive('char', PRIMITIVE_RULES['char']) + min_len = schema.get('minLength', 0) + max_len = schema.get('maxLength') - elif "$ref" in schema: - ref = schema["$ref"] - assert ref.startswith("#/$defs/"), f"Unrecognized schema: {schema}" - # inline $defs - def_name = ref[len("#/$defs/") :] - def_schema = self._defs[def_name] - return self.visit(def_schema, f'{name}{"-" if name else ""}{def_name}') - - - schema_type: Optional[str] = schema.get("type") # type: ignore - assert isinstance(schema_type, str), f"Unrecognized schema: {schema}" - - if schema_type == "object" and "properties" in schema: - # TODO: `required` keyword - if self._prop_order: - prop_order = self._prop_order - prop_pairs = sorted( - schema["properties"].items(), - # sort by position in prop_order (if specified) then by key - key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), - ) - else: - prop_pairs = schema["properties"].items() + return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space') - rule = '"{" space' - for i, (prop_name, prop_schema) in enumerate(prop_pairs): - prop_rule_name = self.visit( - prop_schema, f'{name}{"-" if name else ""}{prop_name}' - ) - if i > 0: - rule += ' "," space' - rule += rf' {self._format_literal(prop_name)} space ":" space {prop_rule_name}' - rule += ' "}" space' - - return self._add_rule(rule_name, rule) + elif (schema_type == 'object') or (len(schema) == 0): + return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object'])) - elif schema_type == "array" and "items" in schema: - # TODO `prefixItems` keyword - item_rule_name = self.visit( - schema["items"], f'{name}{"-" if name else ""}item' + else: + assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}' + # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero + return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type]) + + def _add_primitive(self, name: str, rule: BuiltinRule): + n = self._add_rule(name, rule.content) + + for dep in rule.deps: + dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep) + assert dep_rule, f'Rule {dep} not known' + if dep not in self._rules: + self._add_primitive(dep, dep_rule) + return n + + def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]): + prop_order = self._prop_order + # sort by position in prop_order (if specified) then by original order + sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))] + + prop_kv_rule_names = {} + for prop_name, prop_schema in properties: + prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}') + prop_kv_rule_names[prop_name] = self._add_rule( + f'{name}{"-" if name else ""}{prop_name}-kv', + fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}' ) - list_item_operator = f'("," space {item_rule_name})' - successive_items = "" - min_items = schema.get("minItems", 0) - if min_items > 0: - first_item = f"({item_rule_name})" - successive_items = list_item_operator * (min_items - 1) - min_items -= 1 - else: - first_item = f"({item_rule_name})?" - max_items = schema.get("maxItems") - if max_items is not None and max_items > min_items: - successive_items += (list_item_operator + "?") * (max_items - min_items - 1) - else: - successive_items += list_item_operator + "*" - rule = f'"[" space {first_item} {successive_items} "]" space' - return self._add_rule(rule_name, rule) + required_props = [k for k in sorted_props if k in required] + optional_props = [k for k in sorted_props if k not in required] + + if additional_properties == True or isinstance(additional_properties, dict): + sub_name = f'{name}{"-" if name else ""}additional' + value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value') + prop_kv_rule_names["*"] = self._add_rule( + f'{sub_name}-kv', + self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}' + ) + optional_props.append("*") + + rule = '"{" space ' + rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props) + + if optional_props: + rule += ' (' + if required_props: + rule += ' "," space ( ' + + def get_recursive_refs(ks, first_is_optional): + [k, *rest] = ks + kv_rule_name = prop_kv_rule_names[k] + if k == '*': + res = self._add_rule( + f'{name}{"-" if name else ""}additional-kvs', + f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*' + ) + elif first_is_optional: + res = f'( "," space {kv_rule_name} )?' + else: + res = kv_rule_name + if len(rest) > 0: + res += ' ' + self._add_rule( + f'{name}{"-" if name else ""}{k}-rest', + get_recursive_refs(rest, first_is_optional=True) + ) + return res - else: - assert schema_type in PRIMITIVE_RULES, f"Unrecognized schema: {schema}" - return self._add_rule( - "root" if rule_name == "root" else schema_type, - PRIMITIVE_RULES[schema_type], + rule += ' | '.join( + get_recursive_refs(optional_props[i:], first_is_optional=False) + for i in range(len(optional_props)) ) + if required_props: + rule += ' )' + rule += ' )?' + + rule += ' "}" space' + + return rule def format_grammar(self): - return "\n".join((f"{name} ::= {rule}" for name, rule in self._rules.items())) + return '\n'.join( + f'{name} ::= {rule}' + for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0]) + ) def json_schema_to_gbnf(schema: str, prop_order: Optional[List[str]] = None): From fa4bb0cf81594d3c5bfd0cdf883fc88f820cbc54 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Apr 2024 16:18:16 -0400 Subject: [PATCH 294/624] Revert "feat: Update json to grammar (#1350)" This reverts commit 610a592f708f2a5a8b0e1d6f0900f6337b9beb39. --- llama_cpp/llama_chat_format.py | 2 +- llama_cpp/llama_grammar.py | 612 ++++++--------------------------- 2 files changed, 100 insertions(+), 514 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index eb98cbf39..519d2f50a 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2709,4 +2709,4 @@ def chatml_function_calling( }, } - raise ValueError("Automatic streaming tool choice is not supported") \ No newline at end of file + raise ValueError("Automatic streaming tool choice is not supported") diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 8c0f8aa09..9cc48a93b 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -5,12 +5,11 @@ import sys from ctypes import * # type: ignore from enum import Enum -from itertools import islice, groupby +from itertools import islice from typing import ( Any, Callable, Dict, - Set, Generic, List, Optional, @@ -1392,552 +1391,139 @@ def print_grammar(file: TextIO, state: parse_state) -> None: # whitespace. Also maybe improves generation quality? SPACE_RULE = '" "?' +PRIMITIVE_RULES = { + "boolean": '("true" | "false") space', + "number": '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space', + "integer": '("-"? ([0-9] | [1-9] [0-9]*)) space', + "string": r""" "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space """, + "null": '"null" space', +} INVALID_RULE_CHARS_RE = re.compile(r"[^a-zA-Z0-9-]+") GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]') GRAMMAR_LITERAL_ESCAPES = {"\r": "\\r", "\n": "\\n", '"': '\\"'} -# whitespace is constrained to a single space char to prevent model "running away" in -# whitespace. Also maybe improves generation quality? -SPACE_RULE = '" "?' - - -def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False): - if not separator_rule: - if min_items == 0 and max_items == 1: - return f'{item_rule}?' - elif min_items == 1 and max_items is None: - return f'{item_rule}+' - - result = '' - - if min_items > 0: - if item_rule_is_literal and separator_rule is None: - result = '"' + (item_rule[1:-1] * min_items) + '"' - else: - result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items) - - def opt_repetitions(up_to_n, prefix_with_sep=False): - ''' - - n=4, no sep: '(a (a (a (a)?)?)?)?' - - n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?' - - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?' - ''' - - content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule - if up_to_n == 0: - return '' - elif up_to_n == 1: - return f'({content})?' - elif separator_rule and not prefix_with_sep: - return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?' - else: - return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n) - - if min_items > 0 and max_items != min_items: - result += ' ' - - if max_items is not None: - result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0) - else: - item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})' - - if min_items == 0 and separator_rule: - result = f'({item_rule} {item_operator}*)?' - else: - result += f'{item_operator}*' - - return result - - - -class BuiltinRule: - def __init__(self, content: str, deps: list = None): - self.content = content - self.deps = deps or [] - -_up_to_15_digits = _build_repetition('[0-9]', 0, 15) - -PRIMITIVE_RULES = { - 'boolean' : BuiltinRule('("true" | "false") space', []), - 'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []), - 'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []), - 'number' : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']), - 'integer' : BuiltinRule('("-"? integral-part) space', ['integral-part']), - 'value' : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']), - 'object' : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']), - 'array' : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']), - 'uuid' : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []), - 'char' : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []), - 'string' : BuiltinRule(r'"\"" char* "\"" space', ['char']), - 'null' : BuiltinRule('"null" space', []), -} - -# TODO: support "uri", "email" string formats -STRING_FORMAT_RULES = { - 'date' : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []), - 'time' : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []), - 'date-time' : BuiltinRule('date "T" time', ['date', 'time']), - 'date-string' : BuiltinRule('"\\"" date "\\"" space', ['date']), - 'time-string' : BuiltinRule('"\\"" time "\\"" space', ['time']), - 'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']), -} - -DOTALL = '[\\U00000000-\\U0010FFFF]' -DOT = '[^\\x0A\\x0D]' - -RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()]) - - -NON_LITERAL_SET = set('|.()[]{}*+?') -ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?') - - - class SchemaConverter: def __init__(self, prop_order): self._prop_order = prop_order self._rules = {"space": SPACE_RULE} self._defs: Dict[str, Any] = {} - self._refs = {} - self._refs_being_resolved = set() - def _format_literal(self, literal): - escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal + def _format_literal(self, literal: str): + escaped: str = GRAMMAR_LITERAL_ESCAPE_RE.sub( + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal) ) return f'"{escaped}"' - def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str: - ''' - not_literal('a') -> '[^a]' - not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?' - ''' - assert len(literal) > 0, 'Empty literal not supported' - def recurse(i: int): - c = literal[i] - if maybe_escaped_underscores and c == '_': - yield f'[^{c}\\\\]' - yield ' | ' - yield f'"\\\\"? "{c}"' - else: - yield f'[^{c}]' - if i < len(literal) - 1: - yield ' | ' - yield self._format_literal(c) - yield ' (' - yield from recurse(i + 1) - yield ')?' - - return ''.join(('(', *recurse(0), ')')) - - def _add_rule(self, name, rule): - esc_name = INVALID_RULE_CHARS_RE.sub('-', name) + def _add_rule(self, name: str, rule: str): + esc_name = INVALID_RULE_CHARS_RE.sub("-", name) if esc_name not in self._rules or self._rules[esc_name] == rule: key = esc_name else: i = 0 - while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule: + while f"{esc_name}{i}" in self._rules: i += 1 - key = f'{esc_name}{i}' + key = f"{esc_name}{i}" self._rules[key] = rule return key - def resolve_refs(self, schema: dict, url: str): - ''' - Resolves all $ref fields in the given schema, fetching any remote schemas, - replacing $ref with absolute reference URL and populating self._refs with the - respective referenced (sub)schema dictionaries. - ''' - def visit(n: dict): - if isinstance(n, list): - return [visit(x) for x in n] - elif isinstance(n, dict): - ref = n.get('$ref') - if ref is not None and ref not in self._refs: - if ref.startswith('https://'): - assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)' - import requests - - frag_split = ref.split('#') - base_url = frag_split[0] - - target = self._refs.get(base_url) - if target is None: - target = self.resolve_refs(requests.get(ref).json(), base_url) - self._refs[base_url] = target - - if len(frag_split) == 1 or frag_split[-1] == '': - return target - elif ref.startswith('#/'): - target = schema - ref = f'{url}{ref}' - n['$ref'] = ref - else: - raise ValueError(f'Unsupported ref {ref}') - - for sel in ref.split('#')[-1].split('/')[1:]: - assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}' - target = target[sel] - - self._refs[ref] = target - else: - for v in n.values(): - visit(v) - - return n - return visit(schema) - - def _generate_union_rule(self, name, alt_schemas): - return ' | '.join(( - self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') - for i, alt_schema in enumerate(alt_schemas) - )) - - def _visit_pattern(self, pattern, name): - ''' - Transforms a regular expression pattern into a GBNF rule. - - Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions - Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md - - Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. - - Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which - we define sub-rules to keep the output lean. - ''' - - assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"' - pattern = pattern[1:-1] - sub_rule_ids = {} - - i = 0 - length = len(pattern) - - def to_rule(s: Tuple[str, bool]) -> str: - (txt, is_literal) = s - return "\"" + txt + "\"" if is_literal else txt - - def transform() -> Tuple[str, bool]: - ''' - Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. - ''' - nonlocal i - nonlocal pattern - nonlocal sub_rule_ids - - start = i - # For each component of this sequence, store its string representation and whether it's a literal. - # We only need a flat structure here to apply repetition operators to the last item, and - # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially - # (GBNF's syntax is luckily very close to regular expressions!) - seq: list[Tuple[str, bool]] = [] - - def get_dot(): - if self._dotall: - rule = DOTALL - else: - # Accept any character... except \n and \r line break chars (\x0A and \xOD) - rule = DOT - return self._add_rule(f'dot', rule) - - def join_seq(): - nonlocal seq - ret = [] - for is_literal, g in groupby(seq, lambda x: x[1]): - if is_literal: - ret.append((''.join(x[0] for x in g), True)) - else: - ret.extend(g) - if len(ret) == 1: - return ret[0] - return (' '.join(to_rule(x) for x in seq), False) - - while i < length: - c = pattern[i] - if c == '.': - seq.append((get_dot(), False)) - i += 1 - elif c == '(': - i += 1 - if i < length: - assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' - seq.append((f'({to_rule(transform())})', False)) - elif c == ')': - i += 1 - assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}' - return join_seq() - elif c == '[': - square_brackets = c - i += 1 - while i < length and pattern[i] != ']': - if pattern[i] == '\\': - square_brackets += pattern[i:i+2] - i += 2 - else: - square_brackets += pattern[i] - i += 1 - assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}' - square_brackets += ']' - i += 1 - seq.append((square_brackets, False)) - elif c == '|': - seq.append(('|', False)) - i += 1 - elif c in ('*', '+', '?'): - seq[-1] = (to_rule(seq[-1]) + c, False) - i += 1 - elif c == '{': - curly_brackets = c - i += 1 - while i < length and pattern[i] != '}': - curly_brackets += pattern[i] - i += 1 - assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}' - curly_brackets += '}' - i += 1 - nums = [s.strip() for s in curly_brackets[1:-1].split(',')] - min_times = 0 - max_times = None - try: - if len(nums) == 1: - min_times = int(nums[0]) - max_times = min_times - else: - assert len(nums) == 2 - min_times = int(nums[0]) if nums[0] else 0 - max_times = int(nums[1]) if nums[1] else None - except ValueError: - raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/') - - (sub, sub_is_literal) = seq[-1] - - if not sub_is_literal: - id = sub_rule_ids.get(sub) - if id is None: - id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub) - sub_rule_ids[sub] = id - sub = id - - seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False) - else: - literal = '' - while i < length: - if pattern[i] == '\\' and i < length - 1: - next = pattern[i + 1] - if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS: - i += 1 - literal += pattern[i] - i += 1 - else: - literal += pattern[i:i+2] - i += 2 - elif pattern[i] == '"' and not self._raw_pattern: - literal += '\\"' - i += 1 - elif pattern[i] not in NON_LITERAL_SET and \ - (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET): - literal += pattern[i] - i += 1 - else: - break - if literal: - seq.append((literal, True)) - - return join_seq() - - return self._add_rule( - name, - to_rule(transform()) if self._raw_pattern \ - else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space") - - - def _resolve_ref(self, ref): - ref_name = ref.split('/')[-1] - if ref_name not in self._rules and ref not in self._refs_being_resolved: - self._refs_being_resolved.add(ref) - resolved = self._refs[ref] - ref_name = self.visit(resolved, ref_name) - self._refs_being_resolved.remove(ref) - return ref_name - - def _generate_constant_rule(self, value): - return self._format_literal(json.dumps(value)) - - def visit(self, schema, name): - schema_type = schema.get('type') - schema_format = schema.get('format') - rule_name = name + '-' if name in RESERVED_NAMES else name or 'root' - - if (ref := schema.get('$ref')) is not None: - return self._add_rule(rule_name, self._resolve_ref(ref)) - - elif 'oneOf' in schema or 'anyOf' in schema: - return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf'])) - - elif isinstance(schema_type, list): - return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type])) - - elif 'const' in schema: - return self._add_rule(rule_name, self._generate_constant_rule(schema['const'])) - - elif 'enum' in schema: - rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) - return self._add_rule(rule_name, rule) + def visit(self, schema: Dict[str, Any], name: str) -> str: + rule_name = name or "root" - elif schema_type in (None, 'object') and \ - ('properties' in schema or \ - ('additionalProperties' in schema and schema['additionalProperties'] is not True)): - required = set(schema.get('required', [])) - properties = list(schema.get('properties', {}).items()) - return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties'))) - - elif schema_type in (None, 'object') and 'allOf' in schema: - required = set() - properties = [] - hybrid_name = name - def add_component(comp_schema, is_required): - if (ref := comp_schema.get('$ref')) is not None: - comp_schema = self._refs[ref] - - if 'properties' in comp_schema: - for prop_name, prop_schema in comp_schema['properties'].items(): - properties.append((prop_name, prop_schema)) - if is_required: - required.add(prop_name) - - for t in schema['allOf']: - if 'anyOf' in t: - for tt in t['anyOf']: - add_component(tt, is_required=False) - else: - add_component(t, is_required=True) - - return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[])) - - elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema): - items = schema.get('items') or schema['prefixItems'] - if isinstance(items, list): - return self._add_rule( - rule_name, - '"[" space ' + - ' "," space '.join( - self.visit(item, f'{name}{"-" if name else ""}tuple-{i}') - for i, item in enumerate(items)) + - ' "]" space') - else: - item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item') - min_items = schema.get("minItems", 0) - max_items = schema.get("maxItems") - return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space') - - elif schema_type in (None, 'string') and 'pattern' in schema: - return self._visit_pattern(schema['pattern'], rule_name) - - elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''): - return self._add_primitive( - 'root' if rule_name == 'root' else schema_format, - PRIMITIVE_RULES['uuid'] + if "$defs" in schema: + # add defs to self._defs for later inlining + for def_name, def_schema in schema["$defs"].items(): + self._defs[def_name] = def_schema + + if "oneOf" in schema or "anyOf" in schema: + rule = " | ".join( + ( + self.visit(alt_schema, f'{name}{"-" if name else ""}{i}') + for i, alt_schema in enumerate( + schema.get("oneOf") or schema["anyOf"] + ) + ) ) + return self._add_rule(rule_name, rule) - elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES: - prim_name = f'{schema_format}-string' - return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name])) + elif "const" in schema: + return self._add_rule(rule_name, self._format_literal(schema["const"])) - elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema): - char_rule = self._add_primitive('char', PRIMITIVE_RULES['char']) - min_len = schema.get('minLength', 0) - max_len = schema.get('maxLength') + elif "enum" in schema: + rule = " | ".join((self._format_literal(v) for v in schema["enum"])) + return self._add_rule(rule_name, rule) - return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space') + elif "$ref" in schema: + ref = schema["$ref"] + assert ref.startswith("#/$defs/"), f"Unrecognized schema: {schema}" + # inline $defs + def_name = ref[len("#/$defs/") :] + def_schema = self._defs[def_name] + return self.visit(def_schema, f'{name}{"-" if name else ""}{def_name}') + + + schema_type: Optional[str] = schema.get("type") # type: ignore + assert isinstance(schema_type, str), f"Unrecognized schema: {schema}" + + if schema_type == "object" and "properties" in schema: + # TODO: `required` keyword + if self._prop_order: + prop_order = self._prop_order + prop_pairs = sorted( + schema["properties"].items(), + # sort by position in prop_order (if specified) then by key + key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), + ) + else: + prop_pairs = schema["properties"].items() - elif (schema_type == 'object') or (len(schema) == 0): - return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object'])) + rule = '"{" space' + for i, (prop_name, prop_schema) in enumerate(prop_pairs): + prop_rule_name = self.visit( + prop_schema, f'{name}{"-" if name else ""}{prop_name}' + ) + if i > 0: + rule += ' "," space' + rule += rf' {self._format_literal(prop_name)} space ":" space {prop_rule_name}' + rule += ' "}" space' - else: - assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}' - # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero - return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type]) - - def _add_primitive(self, name: str, rule: BuiltinRule): - n = self._add_rule(name, rule.content) - - for dep in rule.deps: - dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep) - assert dep_rule, f'Rule {dep} not known' - if dep not in self._rules: - self._add_primitive(dep, dep_rule) - return n - - def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]): - prop_order = self._prop_order - # sort by position in prop_order (if specified) then by original order - sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))] - - prop_kv_rule_names = {} - for prop_name, prop_schema in properties: - prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}') - prop_kv_rule_names[prop_name] = self._add_rule( - f'{name}{"-" if name else ""}{prop_name}-kv', - fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}' - ) - required_props = [k for k in sorted_props if k in required] - optional_props = [k for k in sorted_props if k not in required] - - if additional_properties == True or isinstance(additional_properties, dict): - sub_name = f'{name}{"-" if name else ""}additional' - value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value') - prop_kv_rule_names["*"] = self._add_rule( - f'{sub_name}-kv', - self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}' - ) - optional_props.append("*") - - rule = '"{" space ' - rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props) - - if optional_props: - rule += ' (' - if required_props: - rule += ' "," space ( ' - - def get_recursive_refs(ks, first_is_optional): - [k, *rest] = ks - kv_rule_name = prop_kv_rule_names[k] - if k == '*': - res = self._add_rule( - f'{name}{"-" if name else ""}additional-kvs', - f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*' - ) - elif first_is_optional: - res = f'( "," space {kv_rule_name} )?' - else: - res = kv_rule_name - if len(rest) > 0: - res += ' ' + self._add_rule( - f'{name}{"-" if name else ""}{k}-rest', - get_recursive_refs(rest, first_is_optional=True) - ) - return res + return self._add_rule(rule_name, rule) - rule += ' | '.join( - get_recursive_refs(optional_props[i:], first_is_optional=False) - for i in range(len(optional_props)) + elif schema_type == "array" and "items" in schema: + # TODO `prefixItems` keyword + item_rule_name = self.visit( + schema["items"], f'{name}{"-" if name else ""}item' ) - if required_props: - rule += ' )' - rule += ' )?' - - rule += ' "}" space' + list_item_operator = f'("," space {item_rule_name})' + successive_items = "" + min_items = schema.get("minItems", 0) + if min_items > 0: + first_item = f"({item_rule_name})" + successive_items = list_item_operator * (min_items - 1) + min_items -= 1 + else: + first_item = f"({item_rule_name})?" + max_items = schema.get("maxItems") + if max_items is not None and max_items > min_items: + successive_items += (list_item_operator + "?") * (max_items - min_items - 1) + else: + successive_items += list_item_operator + "*" + rule = f'"[" space {first_item} {successive_items} "]" space' + return self._add_rule(rule_name, rule) - return rule + else: + assert schema_type in PRIMITIVE_RULES, f"Unrecognized schema: {schema}" + return self._add_rule( + "root" if rule_name == "root" else schema_type, + PRIMITIVE_RULES[schema_type], + ) def format_grammar(self): - return '\n'.join( - f'{name} ::= {rule}' - for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0]) - ) + return "\n".join((f"{name} ::= {rule}" for name, rule in self._rules.items())) def json_schema_to_gbnf(schema: str, prop_order: Optional[List[str]] = None): From 4f4266495549f691194abb357b792898fd12695e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Lucca=20Zen=C3=B3bio?= Date: Thu, 18 Apr 2024 02:36:25 -0300 Subject: [PATCH 295/624] feat: update grammar schema converter to match llama.cpp (#1353) * feat: improve function calling * feat:grammar * fix * fix * fix --- llama_cpp/llama_chat_format.py | 2 +- llama_cpp/llama_grammar.py | 627 +++++++++++++++++++++++++++------ 2 files changed, 523 insertions(+), 106 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 519d2f50a..eb98cbf39 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2709,4 +2709,4 @@ def chatml_function_calling( }, } - raise ValueError("Automatic streaming tool choice is not supported") + raise ValueError("Automatic streaming tool choice is not supported") \ No newline at end of file diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 9cc48a93b..6c7b57a59 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -5,11 +5,12 @@ import sys from ctypes import * # type: ignore from enum import Enum -from itertools import islice +from itertools import islice, groupby from typing import ( Any, Callable, Dict, + Set, Generic, List, Optional, @@ -1391,145 +1392,561 @@ def print_grammar(file: TextIO, state: parse_state) -> None: # whitespace. Also maybe improves generation quality? SPACE_RULE = '" "?' -PRIMITIVE_RULES = { - "boolean": '("true" | "false") space', - "number": '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space', - "integer": '("-"? ([0-9] | [1-9] [0-9]*)) space', - "string": r""" "\"" ( - [^"\\] | - "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) - )* "\"" space """, - "null": '"null" space', -} INVALID_RULE_CHARS_RE = re.compile(r"[^a-zA-Z0-9-]+") GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]') GRAMMAR_LITERAL_ESCAPES = {"\r": "\\r", "\n": "\\n", '"': '\\"'} +# whitespace is constrained to a single space char to prevent model "running away" in +# whitespace. Also maybe improves generation quality? +SPACE_RULE = '" "?' + + +def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False): + if not separator_rule: + if min_items == 0 and max_items == 1: + return f'{item_rule}?' + elif min_items == 1 and max_items is None: + return f'{item_rule}+' + + result = '' + + if min_items > 0: + if item_rule_is_literal and separator_rule is None: + result = '"' + (item_rule[1:-1] * min_items) + '"' + else: + result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items) + + def opt_repetitions(up_to_n, prefix_with_sep=False): + ''' + - n=4, no sep: '(a (a (a (a)?)?)?)?' + - n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?' + - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?' + ''' + + content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule + if up_to_n == 0: + return '' + elif up_to_n == 1: + return f'({content})?' + elif separator_rule and not prefix_with_sep: + return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?' + else: + return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n) + + if min_items > 0 and max_items != min_items: + result += ' ' + + if max_items is not None: + result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0) + else: + item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})' + + if min_items == 0 and separator_rule: + result = f'({item_rule} {item_operator}*)?' + else: + result += f'{item_operator}*' + + return result + + + +class BuiltinRule: + def __init__(self, content: str, deps: list = None): + self.content = content + self.deps = deps or [] + +_up_to_15_digits = _build_repetition('[0-9]', 0, 15) + +PRIMITIVE_RULES = { + 'boolean' : BuiltinRule('("true" | "false") space', []), + 'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []), + 'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []), + 'number' : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']), + 'integer' : BuiltinRule('("-"? integral-part) space', ['integral-part']), + 'value' : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']), + 'object' : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']), + 'array' : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']), + 'uuid' : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []), + 'char' : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []), + 'string' : BuiltinRule(r'"\"" char* "\"" space', ['char']), + 'null' : BuiltinRule('"null" space', []), +} + +# TODO: support "uri", "email" string formats +STRING_FORMAT_RULES = { + 'date' : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []), + 'time' : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []), + 'date-time' : BuiltinRule('date "T" time', ['date', 'time']), + 'date-string' : BuiltinRule('"\\"" date "\\"" space', ['date']), + 'time-string' : BuiltinRule('"\\"" time "\\"" space', ['time']), + 'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']), +} + +DOTALL = '[\\U00000000-\\U0010FFFF]' +DOT = '[^\\x0A\\x0D]' + +RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()]) + + +NON_LITERAL_SET = set('|.()[]{}*+?') +ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?') + + + class SchemaConverter: - def __init__(self, prop_order): + def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern): self._prop_order = prop_order - self._rules = {"space": SPACE_RULE} - self._defs: Dict[str, Any] = {} - - def _format_literal(self, literal: str): - escaped: str = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal) + self._allow_fetch = allow_fetch + self._dotall = dotall + self._raw_pattern = raw_pattern + self._rules = { + 'space': SPACE_RULE, + } + self._refs = {} + self._refs_being_resolved = set() + + def _format_literal(self, literal): + escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal ) return f'"{escaped}"' - def _add_rule(self, name: str, rule: str): - esc_name = INVALID_RULE_CHARS_RE.sub("-", name) + def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str: + ''' + not_literal('a') -> '[^a]' + not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?' + ''' + assert len(literal) > 0, 'Empty literal not supported' + def recurse(i: int): + c = literal[i] + if maybe_escaped_underscores and c == '_': + yield f'[^{c}\\\\]' + yield ' | ' + yield f'"\\\\"? "{c}"' + else: + yield f'[^{c}]' + if i < len(literal) - 1: + yield ' | ' + yield self._format_literal(c) + yield ' (' + yield from recurse(i + 1) + yield ')?' + + return ''.join(('(', *recurse(0), ')')) + + def _add_rule(self, name, rule): + esc_name = INVALID_RULE_CHARS_RE.sub('-', name) if esc_name not in self._rules or self._rules[esc_name] == rule: key = esc_name else: i = 0 - while f"{esc_name}{i}" in self._rules: + while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule: i += 1 - key = f"{esc_name}{i}" + key = f'{esc_name}{i}' self._rules[key] = rule return key - def visit(self, schema: Dict[str, Any], name: str) -> str: - rule_name = name or "root" - - if "$defs" in schema: - # add defs to self._defs for later inlining - for def_name, def_schema in schema["$defs"].items(): - self._defs[def_name] = def_schema + def resolve_refs(self, schema: dict, url: str): + ''' + Resolves all $ref fields in the given schema, fetching any remote schemas, + replacing $ref with absolute reference URL and populating self._refs with the + respective referenced (sub)schema dictionaries. + ''' + def visit(n: dict): + if isinstance(n, list): + return [visit(x) for x in n] + elif isinstance(n, dict): + ref = n.get('$ref') + if ref is not None and ref not in self._refs: + if ref.startswith('https://'): + assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)' + import requests + + frag_split = ref.split('#') + base_url = frag_split[0] + + target = self._refs.get(base_url) + if target is None: + target = self.resolve_refs(requests.get(ref).json(), base_url) + self._refs[base_url] = target + + if len(frag_split) == 1 or frag_split[-1] == '': + return target + elif ref.startswith('#/'): + target = schema + ref = f'{url}{ref}' + n['$ref'] = ref + else: + raise ValueError(f'Unsupported ref {ref}') + + for sel in ref.split('#')[-1].split('/')[1:]: + assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}' + target = target[sel] + + self._refs[ref] = target + else: + for v in n.values(): + visit(v) + + return n + return visit(schema) + + def _generate_union_rule(self, name, alt_schemas): + return ' | '.join(( + self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') + for i, alt_schema in enumerate(alt_schemas) + )) + + def _visit_pattern(self, pattern, name): + ''' + Transforms a regular expression pattern into a GBNF rule. + + Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions + Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md + + Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. + + Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which + we define sub-rules to keep the output lean. + ''' + + assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"' + pattern = pattern[1:-1] + sub_rule_ids = {} + + i = 0 + length = len(pattern) + + def to_rule(s: Tuple[str, bool]) -> str: + (txt, is_literal) = s + return "\"" + txt + "\"" if is_literal else txt + + def transform() -> Tuple[str, bool]: + ''' + Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. + ''' + nonlocal i + nonlocal pattern + nonlocal sub_rule_ids + + start = i + # For each component of this sequence, store its string representation and whether it's a literal. + # We only need a flat structure here to apply repetition operators to the last item, and + # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially + # (GBNF's syntax is luckily very close to regular expressions!) + seq: list[Tuple[str, bool]] = [] + + def get_dot(): + if self._dotall: + rule = DOTALL + else: + # Accept any character... except \n and \r line break chars (\x0A and \xOD) + rule = DOT + return self._add_rule(f'dot', rule) + + def join_seq(): + nonlocal seq + ret = [] + for is_literal, g in groupby(seq, lambda x: x[1]): + if is_literal: + ret.append((''.join(x[0] for x in g), True)) + else: + ret.extend(g) + if len(ret) == 1: + return ret[0] + return (' '.join(to_rule(x) for x in seq), False) + + while i < length: + c = pattern[i] + if c == '.': + seq.append((get_dot(), False)) + i += 1 + elif c == '(': + i += 1 + if i < length: + assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' + seq.append((f'({to_rule(transform())})', False)) + elif c == ')': + i += 1 + assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}' + return join_seq() + elif c == '[': + square_brackets = c + i += 1 + while i < length and pattern[i] != ']': + if pattern[i] == '\\': + square_brackets += pattern[i:i+2] + i += 2 + else: + square_brackets += pattern[i] + i += 1 + assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}' + square_brackets += ']' + i += 1 + seq.append((square_brackets, False)) + elif c == '|': + seq.append(('|', False)) + i += 1 + elif c in ('*', '+', '?'): + seq[-1] = (to_rule(seq[-1]) + c, False) + i += 1 + elif c == '{': + curly_brackets = c + i += 1 + while i < length and pattern[i] != '}': + curly_brackets += pattern[i] + i += 1 + assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}' + curly_brackets += '}' + i += 1 + nums = [s.strip() for s in curly_brackets[1:-1].split(',')] + min_times = 0 + max_times = None + try: + if len(nums) == 1: + min_times = int(nums[0]) + max_times = min_times + else: + assert len(nums) == 2 + min_times = int(nums[0]) if nums[0] else 0 + max_times = int(nums[1]) if nums[1] else None + except ValueError: + raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/') + + (sub, sub_is_literal) = seq[-1] + + if not sub_is_literal: + id = sub_rule_ids.get(sub) + if id is None: + id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub) + sub_rule_ids[sub] = id + sub = id + + seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False) + else: + literal = '' + while i < length: + if pattern[i] == '\\' and i < length - 1: + next = pattern[i + 1] + if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS: + i += 1 + literal += pattern[i] + i += 1 + else: + literal += pattern[i:i+2] + i += 2 + elif pattern[i] == '"' and not self._raw_pattern: + literal += '\\"' + i += 1 + elif pattern[i] not in NON_LITERAL_SET and \ + (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET): + literal += pattern[i] + i += 1 + else: + break + if literal: + seq.append((literal, True)) + + return join_seq() + + return self._add_rule( + name, + to_rule(transform()) if self._raw_pattern \ + else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space") + + + def _resolve_ref(self, ref): + ref_name = ref.split('/')[-1] + if ref_name not in self._rules and ref not in self._refs_being_resolved: + self._refs_being_resolved.add(ref) + resolved = self._refs[ref] + ref_name = self.visit(resolved, ref_name) + self._refs_being_resolved.remove(ref) + return ref_name + + def _generate_constant_rule(self, value): + return self._format_literal(json.dumps(value)) + + def visit(self, schema, name): + schema_type = schema.get('type') + schema_format = schema.get('format') + rule_name = name + '-' if name in RESERVED_NAMES else name or 'root' + + if (ref := schema.get('$ref')) is not None: + return self._add_rule(rule_name, self._resolve_ref(ref)) + + elif 'oneOf' in schema or 'anyOf' in schema: + return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf'])) + + elif isinstance(schema_type, list): + return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type])) + + elif 'const' in schema: + return self._add_rule(rule_name, self._generate_constant_rule(schema['const'])) + + elif 'enum' in schema: + rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) + return self._add_rule(rule_name, rule) - if "oneOf" in schema or "anyOf" in schema: - rule = " | ".join( - ( - self.visit(alt_schema, f'{name}{"-" if name else ""}{i}') - for i, alt_schema in enumerate( - schema.get("oneOf") or schema["anyOf"] - ) - ) + elif schema_type in (None, 'object') and \ + ('properties' in schema or \ + ('additionalProperties' in schema and schema['additionalProperties'] is not True)): + required = set(schema.get('required', [])) + properties = list(schema.get('properties', {}).items()) + return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties'))) + + elif schema_type in (None, 'object') and 'allOf' in schema: + required = set() + properties = [] + hybrid_name = name + def add_component(comp_schema, is_required): + if (ref := comp_schema.get('$ref')) is not None: + comp_schema = self._refs[ref] + + if 'properties' in comp_schema: + for prop_name, prop_schema in comp_schema['properties'].items(): + properties.append((prop_name, prop_schema)) + if is_required: + required.add(prop_name) + + for t in schema['allOf']: + if 'anyOf' in t: + for tt in t['anyOf']: + add_component(tt, is_required=False) + else: + add_component(t, is_required=True) + + return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[])) + + elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema): + items = schema.get('items') or schema['prefixItems'] + if isinstance(items, list): + return self._add_rule( + rule_name, + '"[" space ' + + ' "," space '.join( + self.visit(item, f'{name}{"-" if name else ""}tuple-{i}') + for i, item in enumerate(items)) + + ' "]" space') + else: + item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item') + min_items = schema.get("minItems", 0) + max_items = schema.get("maxItems") + return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space') + + elif schema_type in (None, 'string') and 'pattern' in schema: + return self._visit_pattern(schema['pattern'], rule_name) + + elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''): + return self._add_primitive( + 'root' if rule_name == 'root' else schema_format, + PRIMITIVE_RULES['uuid'] ) - return self._add_rule(rule_name, rule) - elif "const" in schema: - return self._add_rule(rule_name, self._format_literal(schema["const"])) + elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES: + prim_name = f'{schema_format}-string' + return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name])) - elif "enum" in schema: - rule = " | ".join((self._format_literal(v) for v in schema["enum"])) - return self._add_rule(rule_name, rule) + elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema): + char_rule = self._add_primitive('char', PRIMITIVE_RULES['char']) + min_len = schema.get('minLength', 0) + max_len = schema.get('maxLength') - elif "$ref" in schema: - ref = schema["$ref"] - assert ref.startswith("#/$defs/"), f"Unrecognized schema: {schema}" - # inline $defs - def_name = ref[len("#/$defs/") :] - def_schema = self._defs[def_name] - return self.visit(def_schema, f'{name}{"-" if name else ""}{def_name}') - - - schema_type: Optional[str] = schema.get("type") # type: ignore - assert isinstance(schema_type, str), f"Unrecognized schema: {schema}" - - if schema_type == "object" and "properties" in schema: - # TODO: `required` keyword - if self._prop_order: - prop_order = self._prop_order - prop_pairs = sorted( - schema["properties"].items(), - # sort by position in prop_order (if specified) then by key - key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), - ) - else: - prop_pairs = schema["properties"].items() + return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space') - rule = '"{" space' - for i, (prop_name, prop_schema) in enumerate(prop_pairs): - prop_rule_name = self.visit( - prop_schema, f'{name}{"-" if name else ""}{prop_name}' - ) - if i > 0: - rule += ' "," space' - rule += rf' {self._format_literal(prop_name)} space ":" space {prop_rule_name}' - rule += ' "}" space' - - return self._add_rule(rule_name, rule) + elif (schema_type == 'object') or (len(schema) == 0): + return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object'])) - elif schema_type == "array" and "items" in schema: - # TODO `prefixItems` keyword - item_rule_name = self.visit( - schema["items"], f'{name}{"-" if name else ""}item' + else: + assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}' + # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero + return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type]) + + def _add_primitive(self, name: str, rule: BuiltinRule): + n = self._add_rule(name, rule.content) + + for dep in rule.deps: + dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep) + assert dep_rule, f'Rule {dep} not known' + if dep not in self._rules: + self._add_primitive(dep, dep_rule) + return n + + def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]): + prop_order = self._prop_order + # sort by position in prop_order (if specified) then by original order + sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))] + + prop_kv_rule_names = {} + for prop_name, prop_schema in properties: + prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}') + prop_kv_rule_names[prop_name] = self._add_rule( + f'{name}{"-" if name else ""}{prop_name}-kv', + fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}' ) - list_item_operator = f'("," space {item_rule_name})' - successive_items = "" - min_items = schema.get("minItems", 0) - if min_items > 0: - first_item = f"({item_rule_name})" - successive_items = list_item_operator * (min_items - 1) - min_items -= 1 - else: - first_item = f"({item_rule_name})?" - max_items = schema.get("maxItems") - if max_items is not None and max_items > min_items: - successive_items += (list_item_operator + "?") * (max_items - min_items - 1) - else: - successive_items += list_item_operator + "*" - rule = f'"[" space {first_item} {successive_items} "]" space' - return self._add_rule(rule_name, rule) + required_props = [k for k in sorted_props if k in required] + optional_props = [k for k in sorted_props if k not in required] + + if additional_properties == True or isinstance(additional_properties, dict): + sub_name = f'{name}{"-" if name else ""}additional' + value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value') + prop_kv_rule_names["*"] = self._add_rule( + f'{sub_name}-kv', + self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}' + ) + optional_props.append("*") + + rule = '"{" space ' + rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props) + + if optional_props: + rule += ' (' + if required_props: + rule += ' "," space ( ' + + def get_recursive_refs(ks, first_is_optional): + [k, *rest] = ks + kv_rule_name = prop_kv_rule_names[k] + if k == '*': + res = self._add_rule( + f'{name}{"-" if name else ""}additional-kvs', + f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*' + ) + elif first_is_optional: + res = f'( "," space {kv_rule_name} )?' + else: + res = kv_rule_name + if len(rest) > 0: + res += ' ' + self._add_rule( + f'{name}{"-" if name else ""}{k}-rest', + get_recursive_refs(rest, first_is_optional=True) + ) + return res - else: - assert schema_type in PRIMITIVE_RULES, f"Unrecognized schema: {schema}" - return self._add_rule( - "root" if rule_name == "root" else schema_type, - PRIMITIVE_RULES[schema_type], + rule += ' | '.join( + get_recursive_refs(optional_props[i:], first_is_optional=False) + for i in range(len(optional_props)) ) + if required_props: + rule += ' )' + rule += ' )?' - def format_grammar(self): - return "\n".join((f"{name} ::= {rule}" for name, rule in self._rules.items())) + rule += ' "}" space' + return rule + def format_grammar(self): + return '\n'.join( + f'{name} ::= {rule}' + for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0]) + ) def json_schema_to_gbnf(schema: str, prop_order: Optional[List[str]] = None): prop_order = prop_order or [] schema = json.loads(schema) prop_order = {name: idx for idx, name in enumerate(prop_order)} - converter = SchemaConverter(prop_order) + converter = SchemaConverter(prop_order=prop_order, allow_fetch=False, dotall=False, raw_pattern=False) + schema = converter.resolve_refs(schema, "stdin") converter.visit(schema, "") return converter.format_grammar() From a128c80500046f31d6411f75438d0caf66060ace Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Apr 2024 01:39:45 -0400 Subject: [PATCH 296/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8dd1ec8b3..3b8f1ec4b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8dd1ec8b3ffbfa2d26e82e672cea89f5eeb2f141 +Subproject commit 3b8f1ec4b18770531d0b1d792f3edf08254e4f0c From 893a27a7364366cc195dcc50aadfa75d95bb9319 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 18 Apr 2024 01:43:39 -0400 Subject: [PATCH 297/624] chore: Bump version --- CHANGELOG.md | 8 ++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index c67498ecd..cf1efeec4 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.62] + +- feat: Update llama.cpp to ggerganov/llama.cpp@3b8f1ec4b18770531d0b1d792f3edf08254e4f0c +- feat: update grammar schema converter to match llama.cpp by @themrzmaster in #1353 +- feat: add disable_ping_events flag by @khimaros in #1257 +- feat: Make saved state more compact on-disk by @tc-wolf in #1296 +- feat: Use all available CPUs for batch processing by @ddh0 in #1345 + ## [0.2.61] - feat: Update llama.cpp to ggerganov/llama.cpp@ba5e134e073ec6837078c874aba44a702944a676 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 2382db9b2..828390a7c 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.61" \ No newline at end of file +__version__ = "0.2.62" \ No newline at end of file From d17c1887a364201d292e2959c45a7f144ff240b2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 19 Apr 2024 23:58:16 -0400 Subject: [PATCH 298/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 3b8f1ec4b..0e4802b2e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 3b8f1ec4b18770531d0b1d792f3edf08254e4f0c +Subproject commit 0e4802b2ecbaab04b4f829fde4a3096ca19c84b5 From cc81afebf04d26ca1ac3cf72f23f18da6ab58588 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 20 Apr 2024 00:00:53 -0400 Subject: [PATCH 299/624] feat: Add stopping_criteria to ChatFormatter, allow stopping on arbitrary token ids, fixes llama3 instruct --- llama_cpp/llama.py | 5 ++++- llama_cpp/llama_chat_format.py | 22 +++++++++++++++++++++- 2 files changed, 25 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 5a0111bbe..818be827d 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -426,7 +426,10 @@ def __init__( print(f"Using chat bos_token: {bos_token}", file=sys.stderr) self.chat_handler = llama_chat_format.Jinja2ChatFormatter( - template=template, eos_token=eos_token, bos_token=bos_token + template=template, + eos_token=eos_token, + bos_token=bos_token, + stop_token_ids=[eos_token_id], ).to_chat_handler() if self.chat_format is None and self.chat_handler is None: diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index eb98cbf39..189ccb0c6 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -10,6 +10,9 @@ import jinja2 +import numpy as np +import numpy.typing as npt + import llama_cpp.llama as llama import llama_cpp.llama_types as llama_types import llama_cpp.llama_grammar as llama_grammar @@ -150,6 +153,7 @@ class ChatFormatterResponse: prompt: str stop: Optional[Union[str, List[str]]] = None + stopping_criteria: Optional[llama.StoppingCriteriaList] = None class ChatFormatter(Protocol): @@ -173,12 +177,14 @@ def __init__( eos_token: str, bos_token: str, add_generation_prompt: bool = True, + stop_token_ids: Optional[List[int]] = None, ): """A chat formatter that uses jinja2 templates to format the prompt.""" self.template = template self.eos_token = eos_token self.bos_token = bos_token self.add_generation_prompt = add_generation_prompt + self.stop_token_ids = set(stop_token_ids) if stop_token_ids is not None else None self._environment = jinja2.Environment( loader=jinja2.BaseLoader(), @@ -211,7 +217,16 @@ def raise_exception(message: str): tool_choice=tool_choice, ) - return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token]) + stopping_criteria = None + if self.stop_token_ids is not None: + def stop_on_last_token( + tokens: npt.NDArray[np.intc], + logits: npt.NDArray[np.single] + ) -> bool: + return tokens[-1] in self.stop_token_ids + stopping_criteria = llama.StoppingCriteriaList([stop_on_last_token]) + + return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token], stopping_criteria=stopping_criteria) def to_chat_handler(self) -> LlamaChatCompletionHandler: return chat_formatter_to_chat_completion_handler(self) @@ -533,6 +548,10 @@ def chat_completion_handler( rstop = result.stop if isinstance(result.stop, list) else [result.stop] stop = stop + rstop + stopping_criteria = None + if result.stopping_criteria is not None: + stopping_criteria = result.stopping_criteria + if response_format is not None and response_format["type"] == "json_object": grammar = _grammar_for_response_format(response_format, verbose=llama.verbose) @@ -598,6 +617,7 @@ def chat_completion_handler( mirostat_eta=mirostat_eta, model=model, logits_processor=logits_processor, + stopping_criteria=stopping_criteria, grammar=grammar, logit_bias=logit_bias, ) From 02812148635bf6337ffc7d1abb34093f4065df88 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 20 Apr 2024 00:09:37 -0400 Subject: [PATCH 300/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index cf1efeec4..4c7e267fe 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.63] + +- feat: Update llama.cpp to ggerganov/llama.cpp@0e4802b2ecbaab04b4f829fde4a3096ca19c84b5 +- feat: Add stopping_criteria to ChatFormatter, allow stopping on arbitrary token ids, fixes llama3 instruct by @abetlen in cc81afebf04d26ca1ac3cf72f23f18da6ab58588 + ## [0.2.62] - feat: Update llama.cpp to ggerganov/llama.cpp@3b8f1ec4b18770531d0b1d792f3edf08254e4f0c diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 828390a7c..f2c7aba5b 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.62" \ No newline at end of file +__version__ = "0.2.63" \ No newline at end of file From 159cc4e5d924804c9776355011d5378dc8d5d9f4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 21 Apr 2024 20:46:40 -0400 Subject: [PATCH 301/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 14 +++++++------- llama_cpp/llama_cpp.py | 28 +++++++++++++++++++++++++--- vendor/llama.cpp | 2 +- 3 files changed, 33 insertions(+), 11 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 79f6543a8..ff2d657e4 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -181,20 +181,20 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): ) return list(tokens[:n_tokens]) - def token_to_piece(self, token: int) -> bytes: + def token_to_piece(self, token: int, special: bool = False) -> bytes: assert self.model is not None buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32) + llama_cpp.llama_token_to_piece(self.model, token, buf, 32, special) return bytes(buf) - def detokenize(self, tokens: List[int]) -> bytes: + def detokenize(self, tokens: List[int], special: bool = False) -> bytes: assert self.model is not None output = b"" size = 32 buffer = (ctypes.c_char * size)() for token in tokens: n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token), buffer, size + self.model, llama_cpp.llama_token(token), buffer, size, special ) assert n <= size output += bytes(buffer[:n]) @@ -597,13 +597,13 @@ def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> li return list(result) -def _token_to_piece(model: _LlamaModel, token: int) -> str: +def _token_to_piece(model: _LlamaModel, token: int, special: bool = False) -> str: assert model.model is not None result = (ctypes.c_char * 8)(0) - n_tokens = llama_cpp.llama_token_to_piece(model.model, token, result, len(result)) + n_tokens = llama_cpp.llama_token_to_piece(model.model, token, result, len(result), special) if n_tokens < 0: result = (ctypes.c_char * -n_tokens)(0) - check = llama_cpp.llama_token_to_piece(model.model, token, result, len(result)) + check = llama_cpp.llama_token_to_piece(model.model, token, result, len(result), special) if check != -n_tokens: raise RuntimeError(f"Failed to get piece: token={token}") else: diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 2450d11e3..c2b909ec4 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2380,6 +2380,18 @@ def llama_token_get_type( ... +# // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) +# LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token); +@ctypes_function( + "llama_token_is_eog", [llama_model_p_ctypes, llama_token], ctypes.c_bool +) +def llama_token_is_eog( + model: llama_model_p, token: Union[llama_token, int], / +) -> bool: + """Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)""" + ... + + # // Special tokens @@ -2434,7 +2446,7 @@ def llama_add_eos_token(model: llama_model_p, /) -> int: ... -# // codellama infill tokens +# // Codellama infill tokens # LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix @ctypes_function("llama_token_prefix", [llama_model_p_ctypes], llama_token) def llama_token_prefix(model: llama_model_p) -> int: @@ -2524,11 +2536,13 @@ def llama_tokenize( # // Uses the vocabulary in the provided context. # // Does not write null terminator to the buffer. # // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. +# // @param special If true, special tokens are rendered in the output. # LLAMA_API int32_t llama_token_to_piece( # const struct llama_model * model, # llama_token token, # char * buf, -# int32_t length); +# int32_t length, +# bool special); @ctypes_function( "llama_token_to_piece", [ @@ -2536,6 +2550,7 @@ def llama_tokenize( llama_token, ctypes.c_char_p, ctypes.c_int32, + ctypes.c_bool, ], ctypes.c_int32, ) @@ -2544,13 +2559,20 @@ def llama_token_to_piece( token: Union[llama_token, int], buf: Union[ctypes.c_char_p, bytes, CtypesArray[ctypes.c_char]], length: Union[ctypes.c_int, int], + special: Union[ctypes.c_bool, bool], /, ) -> int: """Token Id -> Piece. Uses the vocabulary in the provided context. Does not write null terminator to the buffer. User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. - """ + + Args: + model: The model to use for tokenization. + token: The token to convert. + buf: The buffer to write the token to. + length: The length of the buffer. + special: If true, special tokens are rendered in the output.""" ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 3b8f1ec4b..5cf5e7d49 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 3b8f1ec4b18770531d0b1d792f3edf08254e4f0c +Subproject commit 5cf5e7d490dfdd2e70bface2d35dfd14aa44b4fb From d40a250ef3cfaa8224d12c83776a2f1de96ae3d1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 22 Apr 2024 00:35:47 -0400 Subject: [PATCH 302/624] feat: Use new llama_token_is_eog in create_completions --- llama_cpp/llama.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 818be827d..0a576d4cf 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1034,7 +1034,8 @@ def logit_bias_processor( logits_processor=logits_processor, grammar=grammar, ): - if token == self._token_eos: + assert self._model.model is not None + if llama_cpp.llama_token_is_eog(self._model.model, token): text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" break From 617d536e1c4f9e84e551828897f5e2b8004539cb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Apr 2024 02:31:40 -0400 Subject: [PATCH 303/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5cf5e7d49..4e96a812b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5cf5e7d490dfdd2e70bface2d35dfd14aa44b4fb +Subproject commit 4e96a812b3ce7322a29a3008db2ed73d9087b176 From 8559e8ce88b7c7343004eeccb7333b806034b01c Mon Sep 17 00:00:00 2001 From: abk16 Date: Tue, 23 Apr 2024 06:33:29 +0000 Subject: [PATCH 304/624] feat: Add Llama-3 chat format (#1371) * feat: Add Llama-3 chat format * feat: Auto-detect Llama-3 chat format from gguf template * feat: Update llama.cpp to b2715 Includes proper Llama-3 <|eot_id|> token handling. --------- Co-authored-by: Andrei Betlen --- llama_cpp/llama_chat_format.py | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 189ccb0c6..17b570a0f 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -35,6 +35,9 @@ # Source: https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1/blob/main/tokenizer_config.json MIXTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" +# Source: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/blob/main/tokenizer_config.json +LLAMA3_INSTRUCT_CHAT_TEMPLATE = "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}" + ### Chat Completion Handler ### @@ -729,6 +732,9 @@ def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[s metadata["tokenizer.chat_template"] == MIXTRAL_INSTRUCT_CHAT_TEMPLATE): return "mistral-instruct" + if metadata["tokenizer.chat_template"] == LLAMA3_INSTRUCT_CHAT_TEMPLATE: + return "llama-3" + return None @@ -920,6 +926,26 @@ def format_llama2( return ChatFormatterResponse(prompt=_prompt) +# Chat format for Llama-3 models, see more details at: +# https://github.com/meta-llama/llama3/blob/main/llama/tokenizer.py#L202-L229 +@register_chat_format("llama-3") +def format_llama3( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + _roles = dict( + system="<|start_header_id|>system<|end_header_id|>\n\n", + user="<|start_header_id|>user<|end_header_id|>\n\n", + assistant="<|start_header_id|>assistant<|end_header_id|>\n\n", + ) + _begin_token = "<|begin_of_text|>" + _sep = "<|eot_id|>" + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_no_colon_single(_begin_token, _messages, _sep) + return ChatFormatterResponse(prompt=_prompt, stop=_sep) + + @register_chat_format("alpaca") def format_alpaca( messages: List[llama_types.ChatCompletionRequestMessage], From 507c1da0663e80d01a5a608355f523c0a791971d Mon Sep 17 00:00:00 2001 From: Geza Velkey Date: Tue, 23 Apr 2024 08:34:15 +0200 Subject: [PATCH 305/624] fix: Update scikit-build-core build dependency avoid bug in 0.9.1 (#1370) cmake [options] cmake [options] cmake [options] -S -B Specify a source directory to (re-)generate a build system for it in the current working directory. Specify an existing build directory to re-generate its build system. Run 'cmake --help' for more information. issue --- pyproject.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/pyproject.toml b/pyproject.toml index e2bbb4b0c..8345cb1f0 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,5 +1,5 @@ [build-system] -requires = ["scikit-build-core[pyproject]>=0.5.1"] +requires = ["scikit-build-core[pyproject]>=0.9.2"] build-backend = "scikit_build_core.build" [project] From 53ebcc8bb5beb912bda4baa44ae27a1e3a7eabd0 Mon Sep 17 00:00:00 2001 From: Sean Bailey <34511443+sean-bailey@users.noreply.github.com> Date: Tue, 23 Apr 2024 02:35:38 -0400 Subject: [PATCH 306/624] feat(server): Provide ability to dynamically allocate all threads if desired using `-1` (#1364) --- llama_cpp/server/settings.py | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 934aecd91..eab5a8aac 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -3,7 +3,7 @@ import multiprocessing from typing import Optional, List, Literal, Union -from pydantic import Field +from pydantic import Field, root_validator from pydantic_settings import BaseSettings import llama_cpp @@ -67,12 +67,12 @@ class ModelSettings(BaseSettings): n_threads: int = Field( default=max(multiprocessing.cpu_count() // 2, 1), ge=1, - description="The number of threads to use.", + description="The number of threads to use. Use -1 for max cpu threads", ) n_threads_batch: int = Field( default=max(multiprocessing.cpu_count(), 1), ge=0, - description="The number of threads to use when batch processing.", + description="The number of threads to use when batch processing. Use -1 for max cpu threads", ) rope_scaling_type: int = Field( default=llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED @@ -173,6 +173,16 @@ class ModelSettings(BaseSettings): default=True, description="Whether to print debug information." ) + @root_validator(pre=True) # pre=True to ensure this runs before any other validation + def set_dynamic_defaults(cls, values): + # If n_threads or n_threads_batch is -1, set it to multiprocessing.cpu_count() + cpu_count = multiprocessing.cpu_count() + if values.get('n_threads', 0) == -1: + values['n_threads'] = cpu_count + if values.get('n_threads_batch', 0) == -1: + values['n_threads_batch'] = cpu_count + return values + class ServerSettings(BaseSettings): """Server settings used to configure the FastAPI and Uvicorn server.""" From 611781f5319719a3d05fefccbbf0cc321742a026 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Apr 2024 02:48:09 -0400 Subject: [PATCH 307/624] ci: Build arm64 wheels. Closes #1342 --- .github/workflows/build-and-release.yaml | 31 +++++++++++++++++++++++- 1 file changed, 30 insertions(+), 1 deletion(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 76b5f7fcf..07742f1dc 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -41,6 +41,35 @@ jobs: with: path: ./wheelhouse/*.whl + build_arm64_wheels: + name: Build arm64 wheels + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v4 + with: + submodules: "recursive" + + - name: Set up QEMU + uses: docker/setup-qemu-action@v3 + with: + platforms: linux/arm64 + + - name: Build wheels + uses: pypa/cibuildwheel@v2.16.5 + env: + CIBW_SKIP: "*musllinux* pp*" + CIBW_REPAIR_WHEEL_COMMAND: "" + CIBW_ARCHS: "aarch64" + CIBW_BUILD: "cp38-* cp39-* cp310-* cp311-* cp312-*" + with: + output-dir: wheelhouse/ + + - name: Upload wheels as artifacts + uses: actions/upload-artifact@v4 + with: + name: wheels-${{ matrix.version }} + path: wheelhouse/*.whl + build_sdist: name: Build source distribution runs-on: ubuntu-latest @@ -65,7 +94,7 @@ jobs: release: name: Release - needs: [build_wheels, build_sdist] + needs: [build_wheels, build_arm64_wheels, build_sdist] runs-on: ubuntu-latest steps: From c50d3300d2a09c98765be7f2c05b7e4fd0b4232e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Apr 2024 02:53:20 -0400 Subject: [PATCH 308/624] chore: Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 4c7e267fe..25b883552 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.64] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4e96a812b3ce7322a29a3008db2ed73d9087b176 +- feat: Add `llama-3` chat format by @andreabak in #1371 +- feat: Use new llama_token_is_eog in create_completions by @abetlen in d40a250ef3cfaa8224d12c83776a2f1de96ae3d1 +- feat(server): Provide ability to dynamically allocate all threads if desired using -1 by @sean-bailey in #1364 +- ci: Build arm64 wheels by @gaby in 611781f5319719a3d05fefccbbf0cc321742a026 +- fix: Update scikit-build-core build dependency avoid bug in 0.9.1 by @evelkey in #1370 + ## [0.2.63] - feat: Update llama.cpp to ggerganov/llama.cpp@0e4802b2ecbaab04b4f829fde4a3096ca19c84b5 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index f2c7aba5b..f73645874 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.63" \ No newline at end of file +__version__ = "0.2.64" \ No newline at end of file From 2a9979fce16ea647918473b85b6f03effb7e1e4c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Apr 2024 02:48:26 -0400 Subject: [PATCH 309/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 12 +++++++++--- vendor/llama.cpp | 2 +- 2 files changed, 10 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index c2b909ec4..7f5d2658d 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1167,13 +1167,19 @@ def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... -# LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +@ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) +def llama_pooling_type(ctx: llama_context_p, /) -> int: + ... + + +# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) def llama_vocab_type(model: llama_model_p, /) -> int: ... -# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); +# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) def llama_rope_type(model: llama_model_p, /) -> int: ... @@ -3091,7 +3097,7 @@ def llama_sample_token_greedy( ... -# /// @details Randomly selects a token from the candidates based on their probabilities. +# /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx. # LLAMA_API llama_token llama_sample_token( # struct llama_context * ctx, # llama_token_data_array * candidates); diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4e96a812b..784e11dea 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4e96a812b3ce7322a29a3008db2ed73d9087b176 +Subproject commit 784e11dea1f5ce9638851b2b0dddb107e2a609c8 From de37420fcf52f759768cec51ddbd5cc768cd9ef3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Apr 2024 03:08:32 -0400 Subject: [PATCH 310/624] fix(ci): Fix python macos test runners issue --- .github/workflows/test.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 77df54697..06af61f00 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -57,7 +57,7 @@ jobs: build-macos: - runs-on: macos-latest + runs-on: macos-13 strategy: matrix: python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] From 266abfc1a36a4b33c0da3f3689676827df3e6a88 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Apr 2024 03:09:46 -0400 Subject: [PATCH 311/624] fix(ci): Fix metal tests as well --- .github/workflows/test.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 06af61f00..6b0e7e8c6 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -107,7 +107,7 @@ jobs: build-macos-metal: - runs-on: macos-latest + runs-on: macos-13 steps: - uses: actions/checkout@v3 From 7f52335c50be425d7dce6302ae38ecff87b0ee74 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Apr 2024 21:21:29 -0400 Subject: [PATCH 312/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++++ vendor/llama.cpp | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 7f5d2658d..3b96adc4c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -811,6 +811,7 @@ class llama_context_params(ctypes.Structure): # bool quantize_output_tensor; // quantize output.weight # bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored # bool pure; // quantize all tensors to the default type +# bool keep_split; // quantize to the same number of shards # void * imatrix; // pointer to importance matrix data # void * kv_overrides; // pointer to vector containing overrides # } llama_model_quantize_params; @@ -826,6 +827,7 @@ class llama_model_quantize_params(ctypes.Structure): quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored pure (bool): quantize all tensors to the default type + keep_split (bool): quantize to the same number of shards imatrix (ctypes.c_void_p): pointer to importance matrix data kv_overrides (ctypes.c_void_p): pointer to vector containing overrides """ @@ -839,6 +841,7 @@ class llama_model_quantize_params(ctypes.Structure): quantize_output_tensor: bool only_copy: bool pure: bool + keep_split: bool imatrix: ctypes.c_void_p kv_overrides: ctypes.c_void_p @@ -851,6 +854,7 @@ class llama_model_quantize_params(ctypes.Structure): ("quantize_output_tensor", ctypes.c_bool), ("only_copy", ctypes.c_bool), ("pure", ctypes.c_bool), + ("keep_split", ctypes.c_bool), ("imatrix", ctypes.c_void_p), ("kv_overrides", ctypes.c_void_p), ] diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 784e11dea..46e12c469 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 784e11dea1f5ce9638851b2b0dddb107e2a609c8 +Subproject commit 46e12c4692a37bdd31a0432fc5153d7d22bc7f72 From fcfea66857c73c9b95326835debb80dbc0b76f17 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Apr 2024 21:21:48 -0400 Subject: [PATCH 313/624] fix: pydantic deprecation warning --- llama_cpp/server/settings.py | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index eab5a8aac..0c858f9fe 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -2,8 +2,10 @@ import multiprocessing -from typing import Optional, List, Literal, Union -from pydantic import Field, root_validator +from typing import Optional, List, Literal, Union, Dict, cast +from typing_extensions import Self + +from pydantic import Field, model_validator from pydantic_settings import BaseSettings import llama_cpp @@ -173,15 +175,16 @@ class ModelSettings(BaseSettings): default=True, description="Whether to print debug information." ) - @root_validator(pre=True) # pre=True to ensure this runs before any other validation - def set_dynamic_defaults(cls, values): + @model_validator(mode="before") # pre=True to ensure this runs before any other validation + def set_dynamic_defaults(self) -> Self: # If n_threads or n_threads_batch is -1, set it to multiprocessing.cpu_count() cpu_count = multiprocessing.cpu_count() + values = cast(Dict[str, int], self) if values.get('n_threads', 0) == -1: values['n_threads'] = cpu_count if values.get('n_threads_batch', 0) == -1: values['n_threads_batch'] = cpu_count - return values + return self class ServerSettings(BaseSettings): From f6ed21f9a26874625b020996fed1f107c1990f50 Mon Sep 17 00:00:00 2001 From: Douglas Hanley Date: Thu, 25 Apr 2024 20:32:44 -0500 Subject: [PATCH 314/624] feat: Allow for possibly non-pooled embeddings (#1380) * allow for possibly non-pooled embeddings * add more to embeddings section in README.md --------- Co-authored-by: Andrei --- README.md | 8 +++++- llama_cpp/_internals.py | 14 ++++++++++ llama_cpp/llama.py | 57 +++++++++++++++++++++++++++------------- llama_cpp/llama_cpp.py | 6 +++++ llama_cpp/llama_types.py | 2 +- 5 files changed, 67 insertions(+), 20 deletions(-) diff --git a/README.md b/README.md index cd41a3804..b5e7d2035 100644 --- a/README.md +++ b/README.md @@ -575,7 +575,7 @@ llama = Llama( ### Embeddings -To generate text embeddings use [`create_embedding`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_embedding). +To generate text embeddings use [`create_embedding`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_embedding) or [`embed`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.embed). Note that you must pass `embedding=True` to the constructor upon model creation for these to work properly. ```python import llama_cpp @@ -589,6 +589,12 @@ embeddings = llm.create_embedding("Hello, world!") embeddings = llm.create_embedding(["Hello, world!", "Goodbye, world!"]) ``` +There are two primary notions of embeddings in a Transformer-style model: *token level* and *sequence level*. Sequence level embeddings are produced by "pooling" token level embeddings together, usually by averaging them or using the first token. + +Models that are explicitly geared towards embeddings will usually return sequence level embeddings by default, one for each input string. Non-embedding models such as those designed for text generation will typically return only token level embeddings, one for each token in each sequence. Thus the dimensionality of the return type will be one higher for token level embeddings. + +It is possible to control pooling behavior in some cases using the `pooling_type` flag on model creation. You can ensure token level embeddings from any model using `LLAMA_POOLING_TYPE_NONE`. The reverse, getting a generation oriented model to yield sequence level embeddings is currently not possible, but you can always do the pooling manually. + ### Adjusting the Context Window The context window of the Llama models determines the maximum number of tokens that can be processed at once. By default, this is set to 512 tokens, but can be adjusted based on your requirements. diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index ff2d657e4..cc3d98928 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -273,6 +273,10 @@ def n_ctx(self) -> int: assert self.ctx is not None return llama_cpp.llama_n_ctx(self.ctx) + def pooling_type(self) -> int: + assert self.ctx is not None + return llama_cpp.llama_pooling_type(self.ctx) + def kv_cache_clear(self): assert self.ctx is not None llama_cpp.llama_kv_cache_clear(self.ctx) @@ -641,6 +645,16 @@ def _should_add_bos(model: _LlamaModel) -> bool: return llama_cpp.llama_vocab_type(model.model) == llama_cpp.LLAMA_VOCAB_TYPE_SPM +# Embedding functions + + +def _normalize_embedding(embedding): + norm = float(np.linalg.norm(embedding)) + if norm == 0.0: + return embedding + return [v / norm for v in embedding] + + # Python wrappers over common/sampling structs diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 0a576d4cf..481842b65 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -50,6 +50,7 @@ _LlamaTokenDataArray, # type: ignore _LlamaSamplingParams, # type: ignore _LlamaSamplingContext, # type: ignore + _normalize_embedding, # type: ignore ) from ._logger import set_verbose from ._utils import suppress_stdout_stderr @@ -760,7 +761,7 @@ def create_embedding( input = input if isinstance(input, list) else [input] # get numeric embeddings - embeds: List[List[float]] + embeds: Union[List[List[float]], List[List[List[float]]]] total_tokens: int embeds, total_tokens = self.embed(input, return_count=True) # type: ignore @@ -787,7 +788,7 @@ def create_embedding( def embed( self, input: Union[str, List[str]], - normalize: bool = True, + normalize: bool = False, truncate: bool = True, return_count: bool = False, ): @@ -803,6 +804,10 @@ def embed( n_embd = self.n_embd() n_batch = self.n_batch + # get pooling information + pooling_type = self.pooling_type() + logits_all = pooling_type == llama_cpp.LLAMA_POOLING_TYPE_NONE + if self.context_params.embeddings == False: raise RuntimeError( "Llama model must be created with embedding=True to call this method" @@ -820,29 +825,37 @@ def embed( self._batch.reset() # decode and fetch embeddings - data: List[List[float]] = [] + data: Union[List[List[float]], List[List[List[float]]]] = [] - def decode_batch(n_seq: int): + def decode_batch(seq_sizes: List[int]): assert self._ctx.ctx is not None llama_cpp.llama_kv_cache_clear(self._ctx.ctx) self._ctx.decode(self._batch) self._batch.reset() # store embeddings - for i in range(n_seq): - ptr = llama_cpp.llama_get_embeddings_seq( - self._ctx.ctx, i - ) - if not ptr: - raise RuntimeError("Failed to get embeddings from sequence pooling type is not set") - embedding: List[float] = ptr[:n_embd] - if normalize: - norm = float(np.linalg.norm(embedding)) - embedding = [v / norm for v in embedding] - data.append(embedding) + if pooling_type == llama_cpp.LLAMA_POOLING_TYPE_NONE: + pos: int = 0 + for i, size in enumerate(seq_sizes): + ptr = llama_cpp.llama_get_embeddings(self._ctx.ctx) + embedding: List[List[float]] = [ + ptr[pos + j * n_embd : pos + (j + 1) * n_embd] for j in range(size) + ] + if normalize: + embedding = [_normalize_embedding(e) for e in embedding] + data.append(embedding) + pos += size + else: + for i in range(len(seq_sizes)): + ptr = llama_cpp.llama_get_embeddings_seq(self._ctx.ctx, i) + embedding: List[float] = ptr[:n_embd] + if normalize: + embedding = _normalize_embedding(embedding) + data.append(embedding) # init state total_tokens = 0 + s_batch = [] t_batch = 0 p_batch = 0 @@ -863,17 +876,21 @@ def decode_batch(n_seq: int): # time to eval batch if t_batch + n_tokens > n_batch: - decode_batch(p_batch) + decode_batch(s_batch) + s_batch = [] t_batch = 0 p_batch = 0 # add to batch - self._batch.add_sequence(tokens, p_batch, False) + self._batch.add_sequence(tokens, p_batch, logits_all) + + # update batch stats + s_batch.append(n_tokens) t_batch += n_tokens p_batch += 1 # hanlde last batch - decode_batch(p_batch) + decode_batch(s_batch) if self.verbose: llama_cpp.llama_print_timings(self._ctx.ctx) @@ -1845,6 +1862,10 @@ def token_nl(self) -> int: """Return the newline token.""" return self._model.token_nl() + def pooling_type(self) -> str: + """Return the pooling type.""" + return self._ctx.pooling_type() + @staticmethod def logits_to_logprobs( logits: Union[npt.NDArray[np.single], List], axis: int = -1 diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 3b96adc4c..609dd369f 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1189,6 +1189,12 @@ def llama_rope_type(model: llama_model_p, /) -> int: ... +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_model * model); +@ctypes_function("llama_pooling_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_pooling_type(model: llama_model_p, /) -> int: + ... + + # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_vocab(model: llama_model_p, /) -> int: diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 87e000f14..4677785ae 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -24,7 +24,7 @@ class EmbeddingUsage(TypedDict): class Embedding(TypedDict): index: int object: str - embedding: List[float] + embedding: Union[List[float], List[List[float]]] class CreateEmbeddingResponse(TypedDict): From 173ebc78782b2eaed40b3b29989502cb6ecdea44 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 25 Apr 2024 21:36:09 -0400 Subject: [PATCH 315/624] fix: Remove duplicate pooling_type definition and add misisng n_vocab definition in bindings --- llama_cpp/llama_cpp.py | 6 ------ 1 file changed, 6 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 609dd369f..3b96adc4c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1189,12 +1189,6 @@ def llama_rope_type(model: llama_model_p, /) -> int: ... -# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_model * model); -@ctypes_function("llama_pooling_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_pooling_type(model: llama_model_p, /) -> int: - ... - - # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_vocab(model: llama_model_p, /) -> int: From 65edc906712eb205afaedc7d69f27e33be162228 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 26 Apr 2024 10:11:31 -0400 Subject: [PATCH 316/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 25b883552..806876ce2 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.65] + +- feat: Update llama.cpp to ggerganov/llama.cpp@46e12c4692a37bdd31a0432fc5153d7d22bc7f72 +- feat: Allow for possibly non-pooled embeddings by @iamlemec in #1380 + ## [0.2.64] - feat: Update llama.cpp to ggerganov/llama.cpp@4e96a812b3ce7322a29a3008db2ed73d9087b176 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index f73645874..6db6333ba 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.64" \ No newline at end of file +__version__ = "0.2.65" \ No newline at end of file From 9e7f738220a28de2c0b3fdac8ceceb5a2c4d02ec Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Sun, 28 Apr 2024 02:47:07 +0200 Subject: [PATCH 317/624] ci: Update dependabot.yml (#1391) Add github-actions update --- .github/dependabot.yml | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/.github/dependabot.yml b/.github/dependabot.yml index 91abb11fd..c58c9ae57 100644 --- a/.github/dependabot.yml +++ b/.github/dependabot.yml @@ -9,3 +9,7 @@ updates: directory: "/" # Location of package manifests schedule: interval: "weekly" + - package-ecosystem: "github-actions" + directory: "/" + schedule: + interval: "weekly" From c58b56123d8659195609098e14eacd13b7374742 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Sun, 28 Apr 2024 02:47:49 +0200 Subject: [PATCH 318/624] ci: Update action versions in build-wheels-metal.yaml (#1390) * Bump actions/setup-python@v4 to v5 * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml --- .github/workflows/build-wheels-metal.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index 2cca47786..fc798c84e 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -41,7 +41,7 @@ jobs: with: submodules: "recursive" - - uses: actions/setup-python@v4 + - uses: actions/setup-python@v5 with: python-version: ${{ matrix.pyver }} @@ -78,7 +78,7 @@ jobs: VERBOSE=1 python -m build --wheel fi - - uses: softprops/action-gh-release@v1 + - uses: softprops/action-gh-release@v2 with: files: dist/* # set release name to -metal From e6bbfb863c38e7575e4fe87a823ac6ce2e15c27c Mon Sep 17 00:00:00 2001 From: iyubondyrev <76585902+iyubondyrev@users.noreply.github.com> Date: Sun, 28 Apr 2024 02:48:47 +0200 Subject: [PATCH 319/624] examples: fix quantize example (#1387) @iyubondyrev thank you! --- examples/low_level_api/quantize.py | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/examples/low_level_api/quantize.py b/examples/low_level_api/quantize.py index 8bd03f88a..cdb038a71 100644 --- a/examples/low_level_api/quantize.py +++ b/examples/low_level_api/quantize.py @@ -4,14 +4,16 @@ def main(args): + fname_inp = args.fname_inp.encode("utf-8") + fname_out = args.fname_out.encode("utf-8") if not os.path.exists(fname_inp): raise RuntimeError(f"Input file does not exist ({fname_inp})") if os.path.exists(fname_out): raise RuntimeError(f"Output file already exists ({fname_out})") - fname_inp = args.fname_inp.encode("utf-8") - fname_out = args.fname_out.encode("utf-8") - itype = args.itype - return_code = llama_cpp.llama_model_quantize(fname_inp, fname_out, itype) + ftype = args.type + args = llama_cpp.llama_model_quantize_default_params() + args.ftype = ftype + return_code = llama_cpp.llama_model_quantize(fname_inp, fname_out, args) if return_code != 0: raise RuntimeError("Failed to quantize model") @@ -20,6 +22,7 @@ def main(args): parser = argparse.ArgumentParser() parser.add_argument("fname_inp", type=str, help="Path to input model") parser.add_argument("fname_out", type=str, help="Path to output model") - parser.add_argument("type", type=int, help="Type of quantization (2: q4_0, 3: q4_1)") + parser.add_argument("type", type=int, help="Type of quantization (2: q4_0, 3: q4_1), see llama_cpp.py for enum") args = parser.parse_args() main(args) + From f178636e1b2a30b12498aa656011779796b9ba11 Mon Sep 17 00:00:00 2001 From: Jeffrey Fong Date: Sun, 28 Apr 2024 08:49:52 +0800 Subject: [PATCH 320/624] fix: Functionary bug fixes (#1385) * fix completion tokens tracking, prompt forming * fix 'function_call' and 'tool_calls' depending on 'functions' and 'tools', incompatibility with python 3.8 * Updated README * fix for openai server compatibility --------- Co-authored-by: Andrei --- README.md | 2 + llama_cpp/llama_chat_format.py | 96 +++++++++++++++++++--------------- 2 files changed, 57 insertions(+), 41 deletions(-) diff --git a/README.md b/README.md index b5e7d2035..a33524cab 100644 --- a/README.md +++ b/README.md @@ -484,6 +484,8 @@ Due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is requi tokenizer=LlamaHFTokenizer.from_pretrained("meetkai/functionary-small-v2.2-GGUF") ) ``` + +**NOTE**: There is no need to provide the default system messages used in Functionary as they are added automatically in the Functionary chat handler. Thus, the messages should contain just the chat messages and/or system messages that provide additional context for the model (e.g.: datetime, etc.).
### Multi-modal Models diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 17b570a0f..71aac8061 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1828,27 +1828,35 @@ def prepare_messages_for_inference( version: Literal["v1", "v2"], functions: Optional[List[llama_types.ChatCompletionFunctions]] = None, tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Union[Dict, str] = "auto", ): all_messages: List[llama_types.ChatCompletionRequestMessage] = [] - if functions is not None: + if tool_choice == "none": all_messages.append( llama_types.ChatCompletionRequestSystemMessage( - role="system", content=generate_schema_from_functions(functions) + role="system", content=generate_schema_from_functions([]) ) ) - elif tools is not None: - all_messages.append( - llama_types.ChatCompletionRequestSystemMessage( - role="system", - content=generate_schema_from_functions( - [ - tool["function"] - for tool in tools - if tool["type"] == "function" - ] - ), + else: + if functions is not None: + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=generate_schema_from_functions(functions) + ) + ) + elif tools is not None and tool_choice != "none": + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", + content=generate_schema_from_functions( + [ + tool["function"] + for tool in tools + if tool["type"] == "function" + ] + ), + ) ) - ) all_messages.append( llama_types.ChatCompletionRequestSystemMessage( @@ -1888,7 +1896,7 @@ def prepare_messages_for_inference( function_call = "auto" prompt = prepare_messages_for_inference( - messages, tokenizer, version, functions, tools + messages, tokenizer, version, functions, tools, function_call ) # If no tools/functions are provided @@ -1985,17 +1993,12 @@ def create_completion(stop): content = "" function_calls, function_bodies = [], [] + completion_tokens = 0 if version == "v1": # If no or "auto" tool_choice/function_call if isinstance(function_call, str) and function_call == "auto": stops = ["\n", END_ASSISTANT_TOKEN] - # If tool_choice/function_call is "none" - elif isinstance(function_call, str) and function_call == "none": - prompt = prepare_messages_for_inference( - messages, tokenizer, version, [], [] - ) - stops = END_ASSISTANT_TOKEN # If tool_choice/function_call is provided elif isinstance(function_call, dict): prompt += f"{START_FUNCTION_CALL_TOKEN}{function_call['name']}:\n" @@ -2009,12 +2012,15 @@ def create_completion(stop): completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + # If the generation does not involve a function call if ( START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN not in completion_text ): + completion["usage"]["completion_tokens"] = completion_tokens return _convert_completion_to_chat(completion, stream=stream) # type: ignore # If the generation involves a function call in completion, generate the parameters elif ( @@ -2032,23 +2038,14 @@ def create_completion(stop): ) grammar = get_grammar(function_calls[-1]) completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) + completion_tokens += completion["usage"]["completion_tokens"] function_bodies.append(completion["choices"][0]["text"].strip()) # If the prompt involves a function call, just append generated parameters to function_bodies else: function_bodies.append(completion_text.strip()) else: - # If tool_choice/function_call is "none" - if isinstance(function_call, str) and function_call == "none": - prompt = ( - prepare_messages_for_inference(messages, tokenizer, version, [], []) - + "all\n<|content|>" - ) - stops = [STOP_TOKEN, FROM_TOKEN] - completion = create_completion(stop=stops) - completion["choices"][0]["text"] = completion["choices"][0]["text"].strip() - return _convert_completion_to_chat(completion, stream=stream) # type: ignore # If tool_choice/function_call is provided - elif isinstance(function_call, dict): + if isinstance(function_call, dict): prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" function_call = function_call["name"] function_calls.append(function_call) @@ -2056,6 +2053,7 @@ def create_completion(stop): stops = [STOP_TOKEN, FROM_TOKEN] completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] function_bodies.append(completion_text.strip()) # If "auto" or no tool_choice/function_call elif isinstance(function_call, str) and function_call == "auto": @@ -2065,6 +2063,7 @@ def create_completion(stop): stops = CONTENT_TOKEN completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] function_name = completion_text.strip() if function_name == "all": prompt += "all\n<|content|>" @@ -2077,12 +2076,23 @@ def create_completion(stop): stops = [RECIPIENT_TOKEN, STOP_TOKEN] completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] if function_name == "all": - content += completion_text.removesuffix("\n<|from|>assistant\n").removesuffix("\n<|from|> assistant\n") + if completion_text.endswith("\n<|from|>assistant\n"): + content += completion_text[:-len("\n<|from|>assistant\n")] + if completion_text.endswith("\n<|from|> assistant\n"): + content += completion_text[-len("\n<|from|> assistant\n")] + else: + content += completion_text content = content.lstrip() # Check whether the model wants to generate another turn if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text: - cleaned_completion_text = completion_text.removesuffix("\n<|from|>assistant\n").removesuffix("\n<|from|> assistant\n").strip() + if completion_text.endswith("\n<|from|>assistant\n"): + cleaned_completion_text = completion_text[:-len("\n<|from|>assistant\n")].strip() + elif completion_text.endswith("\n<|from|> assistant\n"): + cleaned_completion_text = completion_text[-len("\n<|from|> assistant\n")].strip() + else: + cleaned_completion_text = completion_text.strip() prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" else: break @@ -2092,6 +2102,7 @@ def create_completion(stop): prompt += completion_text.strip() grammar = None completion = create_completion(stop=stops) + completion_tokens += completion["usage"]["completion_tokens"] if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]: prompt += "\n<|from|>assistant\n<|recipient|>" else: @@ -2120,12 +2131,16 @@ def create_completion(stop): ) # TODO: support stream mode - function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = { - "function_call": { - "name": tool_calls[0]["function"]["name"], - "arguments": tool_calls[0]["function"]["arguments"], - } - } if len(tool_calls) == 1 else {} + function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = {} + if len(tool_calls) > 0: + if tools is not None: + function_call_dict["tool_calls"] = tool_calls + else: + function_call_dict["function_call"] = { + "name": tool_calls[0]["function"]["name"], + "arguments": tool_calls[0]["function"]["arguments"], + } + completion["usage"]["completion_tokens"] = completion_tokens return llama_types.CreateChatCompletionResponse( id="chat" + completion["id"], object="chat.completion", @@ -2138,7 +2153,6 @@ def create_completion(stop): "message": { "role": "assistant", "content": None if content == "" else content, - "tool_calls": tool_calls, **function_call_dict, }, "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", From 17bdfc818f50988ddb6dfc5a30cbec68571e4c56 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat, 27 Apr 2024 20:50:28 -0400 Subject: [PATCH 321/624] chore(deps): bump conda-incubator/setup-miniconda from 2.2.0 to 3.0.4 (#1397) Bumps [conda-incubator/setup-miniconda](https://github.com/conda-incubator/setup-miniconda) from 2.2.0 to 3.0.4. - [Release notes](https://github.com/conda-incubator/setup-miniconda/releases) - [Changelog](https://github.com/conda-incubator/setup-miniconda/blob/main/CHANGELOG.md) - [Commits](https://github.com/conda-incubator/setup-miniconda/compare/v2.2.0...v3.0.4) --- updated-dependencies: - dependency-name: conda-incubator/setup-miniconda dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index a222dceb4..9647036b8 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -52,7 +52,7 @@ jobs: python-version: ${{ matrix.pyver }} - name: Setup Mamba - uses: conda-incubator/setup-miniconda@v2.2.0 + uses: conda-incubator/setup-miniconda@v3.0.4 with: activate-environment: "build" python-version: ${{ matrix.pyver }} From 27038db3d6cbab301dc96e21d7f24fc53c5ccbf5 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat, 27 Apr 2024 20:50:39 -0400 Subject: [PATCH 322/624] chore(deps): bump actions/cache from 3.3.2 to 4.0.2 (#1398) Bumps [actions/cache](https://github.com/actions/cache) from 3.3.2 to 4.0.2. - [Release notes](https://github.com/actions/cache/releases) - [Changelog](https://github.com/actions/cache/blob/main/RELEASES.md) - [Commits](https://github.com/actions/cache/compare/v3.3.2...v4.0.2) --- updated-dependencies: - dependency-name: actions/cache dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 9647036b8..274b94669 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -65,7 +65,7 @@ jobs: - name: VS Integration Cache id: vs-integration-cache if: runner.os == 'Windows' - uses: actions/cache@v3.3.2 + uses: actions/cache@v4.0.2 with: path: ./MSBuildExtensions key: cuda-${{ matrix.cuda }}-vs-integration From 79318ba1d181f5765d3abfab00d27ec5c42a3e8d Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat, 27 Apr 2024 20:50:50 -0400 Subject: [PATCH 323/624] chore(deps): bump docker/login-action from 2 to 3 (#1399) Bumps [docker/login-action](https://github.com/docker/login-action) from 2 to 3. - [Release notes](https://github.com/docker/login-action/releases) - [Commits](https://github.com/docker/login-action/compare/v2...v3) --- updated-dependencies: - dependency-name: docker/login-action dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-docker.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-docker.yaml b/.github/workflows/build-docker.yaml index 750b91e1f..726f87233 100644 --- a/.github/workflows/build-docker.yaml +++ b/.github/workflows/build-docker.yaml @@ -23,7 +23,7 @@ jobs: uses: docker/setup-buildx-action@v2 - name: Login to GitHub Container Registry - uses: docker/login-action@v2 + uses: docker/login-action@v3 with: registry: ghcr.io username: ${{ github.repository_owner }} From 7074c4d256e06c31a6386bb099bad1e891284334 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat, 27 Apr 2024 20:51:02 -0400 Subject: [PATCH 324/624] chore(deps): bump docker/build-push-action from 4 to 5 (#1400) Bumps [docker/build-push-action](https://github.com/docker/build-push-action) from 4 to 5. - [Release notes](https://github.com/docker/build-push-action/releases) - [Commits](https://github.com/docker/build-push-action/compare/v4...v5) --- updated-dependencies: - dependency-name: docker/build-push-action dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-docker.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-docker.yaml b/.github/workflows/build-docker.yaml index 726f87233..68cc06f15 100644 --- a/.github/workflows/build-docker.yaml +++ b/.github/workflows/build-docker.yaml @@ -31,7 +31,7 @@ jobs: - name: Build and push id: docker_build - uses: docker/build-push-action@v4 + uses: docker/build-push-action@v5 with: context: . file: "docker/simple/Dockerfile" From c07db99e5b34b9d791ac9d14cdde46928f4694ae Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat, 27 Apr 2024 20:51:13 -0400 Subject: [PATCH 325/624] chore(deps): bump pypa/cibuildwheel from 2.16.5 to 2.17.0 (#1401) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.16.5 to 2.17.0. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.16.5...v2.17.0) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 07742f1dc..2137da350 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.16.5 + uses: pypa/cibuildwheel@v2.17.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -55,7 +55,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.16.5 + uses: pypa/cibuildwheel@v2.17.0 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" From c9b85bf09859161d3ab9de49ed448297488e0c89 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 27 Apr 2024 23:41:54 -0400 Subject: [PATCH 326/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 152 ++++++++++++++++++----------------------- vendor/llama.cpp | 2 +- 2 files changed, 69 insertions(+), 85 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 3b96adc4c..d00dfcb95 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -552,19 +552,25 @@ class llama_batch(ctypes.Structure): # LLAMA_KV_OVERRIDE_TYPE_INT, # LLAMA_KV_OVERRIDE_TYPE_FLOAT, # LLAMA_KV_OVERRIDE_TYPE_BOOL, +# LLAMA_KV_OVERRIDE_TYPE_STR, # }; LLAMA_KV_OVERRIDE_TYPE_INT = 0 LLAMA_KV_OVERRIDE_TYPE_FLOAT = 1 LLAMA_KV_OVERRIDE_TYPE_BOOL = 2 +LLAMA_KV_OVERRIDE_TYPE_STR = 3 # struct llama_model_kv_override { -# char key[128]; # enum llama_model_kv_override_type tag; + +# char key[128]; + + # union { -# int64_t int_value; -# double float_value; -# bool bool_value; +# int64_t val_i64; +# double val_f64; +# bool val_bool; +# char val_str[128]; # }; # }; class llama_model_kv_override_value(ctypes.Union): @@ -572,16 +578,28 @@ class llama_model_kv_override_value(ctypes.Union): ("int_value", ctypes.c_int64), ("float_value", ctypes.c_double), ("bool_value", ctypes.c_bool), + ("str_value", ctypes.c_char * 128), ] + if TYPE_CHECKING: + int_value: int + float_value: float + bool_value: bool + str_value: bytes + class llama_model_kv_override(ctypes.Structure): _fields_ = [ - ("key", ctypes.c_char * 128), ("tag", ctypes.c_int), + ("key", ctypes.c_char * 128), ("value", llama_model_kv_override_value), ] + if TYPE_CHECKING: + tag: int + key: bytes + value: Union[int, float, bool, bytes] + # struct llama_model_params { # int32_t n_gpu_layers; // number of layers to store in VRAM @@ -609,9 +627,10 @@ class llama_model_kv_override(ctypes.Structure): # // Keep the booleans together to avoid misalignment during copy-by-value. -# bool vocab_only; // only load the vocabulary, no weights -# bool use_mmap; // use mmap if possible -# bool use_mlock; // force system to keep model in RAM +# bool vocab_only; // only load the vocabulary, no weights +# bool use_mmap; // use mmap if possible +# bool use_mlock; // force system to keep model in RAM +# bool check_tensors; // validate model tensor data # }; class llama_model_params(ctypes.Structure): """Parameters for llama_model @@ -626,7 +645,8 @@ class llama_model_params(ctypes.Structure): kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data vocab_only (bool): only load the vocabulary, no weights use_mmap (bool): use mmap if possible - use_mlock (bool): force system to keep model in RAM""" + use_mlock (bool): force system to keep model in RAM + check_tensors (bool): validate model tensor data""" if TYPE_CHECKING: n_gpu_layers: int @@ -639,6 +659,7 @@ class llama_model_params(ctypes.Structure): vocab_only: bool use_mmap: bool use_mlock: bool + check_tensors: bool _fields_ = [ ("n_gpu_layers", ctypes.c_int32), @@ -651,6 +672,7 @@ class llama_model_params(ctypes.Structure): ("vocab_only", ctypes.c_bool), ("use_mmap", ctypes.c_bool), ("use_mlock", ctypes.c_bool), + ("check_tensors", ctypes.c_bool), ] @@ -1041,8 +1063,7 @@ def llama_backend_init(): [ctypes.c_int], None, ) -def llama_numa_init(numa: int, /): - ... +def llama_numa_init(numa: int, /): ... # // Call once at the end of the program - currently only used for MPI @@ -1067,8 +1088,7 @@ def llama_backend_free(): ) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / -) -> Optional[llama_model_p]: - ... +) -> Optional[llama_model_p]: ... # LLAMA_API void llama_free_model(struct llama_model * model); @@ -1077,8 +1097,7 @@ def llama_load_model_from_file( [llama_model_p_ctypes], None, ) -def llama_free_model(model: llama_model_p, /): - ... +def llama_free_model(model: llama_model_p, /): ... # LLAMA_API struct llama_context * llama_new_context_with_model( @@ -1091,8 +1110,7 @@ def llama_free_model(model: llama_model_p, /): ) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / -) -> Optional[llama_context_p]: - ... +) -> Optional[llama_context_p]: ... # // Frees all allocated memory @@ -1113,104 +1131,87 @@ def llama_free(ctx: llama_context_p, /): [], ctypes.c_int64, ) -def llama_time_us() -> int: - ... +def llama_time_us() -> int: ... # LLAMA_API size_t llama_max_devices(void); @ctypes_function("llama_max_devices", [], ctypes.c_size_t) -def llama_max_devices() -> int: - ... +def llama_max_devices() -> int: ... # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) -def llama_supports_mmap() -> bool: - ... +def llama_supports_mmap() -> bool: ... # LLAMA_API bool llama_supports_mlock (void); @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) -def llama_supports_mlock() -> bool: - ... +def llama_supports_mlock() -> bool: ... # LLAMA_API bool llama_supports_gpu_offload(void); @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) -def llama_supports_gpu_offload() -> bool: - ... +def llama_supports_gpu_offload() -> bool: ... # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) -def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: - ... +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ctx(ctx: llama_context_p, /) -> int: - ... +def llama_n_ctx(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_batch(ctx: llama_context_p, /) -> int: - ... +def llama_n_batch(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); @ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ubatch(ctx: llama_context_p, /) -> int: - ... +def llama_n_ubatch(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); @ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_seq_max(ctx: llama_context_p, /) -> int: - ... +def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... # LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) -def llama_pooling_type(ctx: llama_context_p, /) -> int: - ... +def llama_pooling_type(ctx: llama_context_p, /) -> int: ... # LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: - ... +def llama_vocab_type(model: llama_model_p, /) -> int: ... # LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: - ... +def llama_rope_type(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: - ... +def llama_n_vocab(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_ctx_train(model: llama_model_p, /) -> int: - ... +def llama_n_ctx_train(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_embd(model: llama_model_p, /) -> int: - ... +def llama_n_embd(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_layer (const struct llama_model * model); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_layer(model: llama_model_p, /) -> int: - ... +def llama_n_layer(model: llama_model_p, /) -> int: ... # // Get the model's RoPE frequency scaling factor @@ -1912,8 +1913,7 @@ def llama_state_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> bool: - ... +) -> bool: ... # LLAMA_API DEPRECATED(bool llama_load_session_file( @@ -1941,8 +1941,7 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: - ... +) -> int: ... # LLAMA_API bool llama_state_save_file( @@ -1966,8 +1965,7 @@ def llama_state_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> bool: - ... +) -> bool: ... # LLAMA_API DEPRECATED(bool llama_save_session_file( @@ -1992,8 +1990,7 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: - ... +) -> int: ... # // Get the exact size needed to copy the KV cache of a single sequence @@ -2071,8 +2068,7 @@ def llama_state_seq_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: - ... +) -> int: ... # LLAMA_API size_t llama_state_seq_load_file( @@ -2102,8 +2098,7 @@ def llama_state_seq_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: - ... +) -> int: ... # // @@ -2366,8 +2361,7 @@ def llama_get_embeddings_seq( ) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / -) -> bytes: - ... +) -> bytes: ... # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); @@ -2376,8 +2370,7 @@ def llama_token_get_text( ) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / -) -> float: - ... +) -> float: ... # LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); @@ -2386,8 +2379,7 @@ def llama_token_get_score( ) def llama_token_get_type( model: llama_model_p, token: Union[llama_token, int], / -) -> int: - ... +) -> int: ... # // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) @@ -2395,9 +2387,7 @@ def llama_token_get_type( @ctypes_function( "llama_token_is_eog", [llama_model_p_ctypes, llama_token], ctypes.c_bool ) -def llama_token_is_eog( - model: llama_model_p, token: Union[llama_token, int], / -) -> bool: +def llama_token_is_eog(model: llama_model_p, token: Union[llama_token, int], /) -> bool: """Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)""" ... @@ -2466,20 +2456,17 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: - ... +def llama_token_middle(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: - ... +def llama_token_suffix(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: - ... +def llama_token_eot(model: llama_model_p, /) -> int: ... # // @@ -2620,8 +2607,7 @@ def llama_chat_apply_template( chat: CtypesArray[llama_chat_message], n_msg: int, /, -) -> int: - ... +) -> int: ... # // @@ -3234,8 +3220,7 @@ def llama_beam_search( n_past: Union[ctypes.c_int, int], n_predict: Union[ctypes.c_int, int], /, -): - ... +): ... # /// @details Build a split GGUF final path for this chunk. @@ -3354,5 +3339,4 @@ def llama_log_set( [ctypes.c_void_p, llama_context_p_ctypes], None, ) -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): - ... +def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 46e12c469..4dba7e811 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 46e12c4692a37bdd31a0432fc5153d7d22bc7f72 +Subproject commit 4dba7e8114d84241c842b986e008af8b88d1a019 From a411612b385cef100d76145da1fbd02a7b7cc894 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 27 Apr 2024 23:42:19 -0400 Subject: [PATCH 327/624] feat: Add support for str type kv_overrides --- llama_cpp/llama.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 481842b65..96aac6668 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -73,7 +73,7 @@ def __init__( vocab_only: bool = False, use_mmap: bool = True, use_mlock: bool = False, - kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None, + kv_overrides: Optional[Dict[str, Union[bool, int, float, str]]] = None, # Context Params seed: int = llama_cpp.LLAMA_DEFAULT_SEED, n_ctx: int = 512, @@ -254,6 +254,13 @@ def __init__( elif isinstance(v, float): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT self._kv_overrides_array[i].value.float_value = v + elif isinstance(v, str): # type: ignore + v_bytes = v.encode("utf-8") + if len(v_bytes) > 128: # TODO: Make this a constant + raise ValueError(f"Value for {k} is too long: {v}") + v_bytes = v_bytes.ljust(128, b"\0") + self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR + self._kv_overrides_array[i].value.str_value[:128] = v_bytes else: raise ValueError(f"Unknown value type for {k}: {v}") From 2355ce2227476ea69f1d6142dedd2f4b7f27a86b Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Sun, 28 Apr 2024 05:44:47 +0200 Subject: [PATCH 328/624] ci: Add support for pre-built cuda 12.4.1 wheels (#1388) * Add support for cuda 12.4.1 * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml Revert --- .github/workflows/build-wheels-cuda.yaml | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 274b94669..ae9e8632c 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -22,7 +22,7 @@ jobs: $matrix = @{ 'os' = @('ubuntu-20.04', 'windows-latest') 'pyver' = @("3.10", "3.11", "3.12") - 'cuda' = @("12.1.1", "12.2.2", "12.3.2") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") 'releasetag' = @("basic") } @@ -47,7 +47,7 @@ jobs: with: submodules: "recursive" - - uses: actions/setup-python@v4 + - uses: actions/setup-python@v5 with: python-version: ${{ matrix.pyver }} @@ -74,7 +74,7 @@ jobs: if: runner.os == 'Windows' && steps.vs-integration-cache.outputs.cache-hit != 'true' run: | if ($env:CUDAVER -eq '12.1.1') {$x = '12.1.0'} else {$x = $env:CUDAVER} - $links = (Invoke-RestMethod 'https://github.com/Jimver/cuda-toolkit/raw/dc0ca7bb29c5a92f7a963d3d5c93f8d59765136a/src/links/windows-links.ts').Trim().split().where({$_ -ne ''}) + $links = (Invoke-RestMethod 'https://raw.githubusercontent.com/Jimver/cuda-toolkit/master/src/links/windows-links.ts').Trim().split().where({$_ -ne ''}) for ($i=$q=0;$i -lt $links.count -and $q -lt 2;$i++) {if ($links[$i] -eq "'$x',") {$q++}} Invoke-RestMethod $links[$i].Trim("'") -OutFile 'cudainstaller.zip' & 'C:\Program Files\7-Zip\7z.exe' e cudainstaller.zip -oMSBuildExtensions -r *\MSBuildExtensions\* > $null @@ -122,7 +122,7 @@ jobs: # write the build tag to the output Write-Output "CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV - - uses: softprops/action-gh-release@v1 + - uses: softprops/action-gh-release@v2 with: files: dist/* # Set tag_name to -cu From 0c3bc4b92813381f3ffddcc9b579e221ee48667b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Apr 2024 12:37:22 -0400 Subject: [PATCH 329/624] fix(ci): Update generate wheel index script to include cu12.3 and cu12.4 Closes #1406 --- .github/workflows/generate-index-from-release.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 9042d6cfe..858e50711 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -37,6 +37,8 @@ jobs: ./scripts/releases-to-pep-503.sh index/whl/cpu '^[v]?[0-9]+\.[0-9]+\.[0-9]+$' ./scripts/releases-to-pep-503.sh index/whl/cu121 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu121$' ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' + ./scripts/releases-to-pep-503.sh index/whl/cu123 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu123$' + ./scripts/releases-to-pep-503.sh index/whl/cu124 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' - name: Upload artifact uses: actions/upload-pages-artifact@v3 From 03c654a3d9889bcf8a0402c3b7478ad85fbaf111 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Tue, 30 Apr 2024 04:52:23 +0200 Subject: [PATCH 330/624] ci(fix): Workflow actions updates and fix arm64 wheels not included in release (#1392) * Update test.yaml Bump actions/checkout@v3 to v4 Bump action/setup-python@v4 to v5 * Update test-pypi.yaml Bum actions/setup-python@v4 to v5 * Update build-and-release.yaml Bump softprops/action-gh-release@v1 to v2 Bump actions/checkout@v3 to v4 Bump actions/setup-python@v3 to v5 * Update publish.yaml Bump actions/checkout@v3 to v4 Bump actions/sertup-python@v4 to v5 * Update publish-to-test.yaml Bump actions/checkout@v3 to v4 Bump actions/setup-python @v4 to v5 * Update test-pypi.yaml Add Python 3.12 * Update build-and-release.yaml * Update build-docker.yaml Bump docker/setup-qemu-action@v2 to v3 Bump docker/setup-buildx-action@v2 to v3 * Update build-and-release.yaml * Update build-and-release.yaml --- .github/workflows/build-and-release.yaml | 19 +++++++++---------- .github/workflows/build-docker.yaml | 6 +++--- .github/workflows/publish-to-test.yaml | 6 +++--- .github/workflows/publish.yaml | 4 ++-- .github/workflows/test-pypi.yaml | 14 +++++++------- .github/workflows/test.yaml | 18 +++++++++--------- 6 files changed, 33 insertions(+), 34 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 2137da350..76bf708fe 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -14,12 +14,12 @@ jobs: os: [ubuntu-20.04, windows-2019, macos-11] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" # Used to host cibuildwheel - - uses: actions/setup-python@v3 + - uses: actions/setup-python@v5 with: python-version: "3.8" @@ -62,23 +62,22 @@ jobs: CIBW_ARCHS: "aarch64" CIBW_BUILD: "cp38-* cp39-* cp310-* cp311-* cp312-*" with: - output-dir: wheelhouse/ + output-dir: wheelhouse - name: Upload wheels as artifacts - uses: actions/upload-artifact@v4 + uses: actions/upload-artifact@v3 with: - name: wheels-${{ matrix.version }} - path: wheelhouse/*.whl + path: ./wheelhouse/*.whl build_sdist: name: Build source distribution runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" - - uses: actions/setup-python@v3 + - uses: actions/setup-python@v5 with: python-version: "3.8" - name: Install dependencies @@ -102,8 +101,8 @@ jobs: with: name: artifact path: dist - - uses: softprops/action-gh-release@v1 + - uses: softprops/action-gh-release@v2 with: files: dist/* env: - GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} \ No newline at end of file + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/build-docker.yaml b/.github/workflows/build-docker.yaml index 68cc06f15..4ebe3bb6d 100644 --- a/.github/workflows/build-docker.yaml +++ b/.github/workflows/build-docker.yaml @@ -12,15 +12,15 @@ jobs: runs-on: ubuntu-latest steps: - name: Checkout - uses: actions/checkout@v3 + uses: actions/checkout@v4 with: submodules: "recursive" - name: Set up QEMU - uses: docker/setup-qemu-action@v2 + uses: docker/setup-qemu-action@v3 - name: Set up Docker Buildx - uses: docker/setup-buildx-action@v2 + uses: docker/setup-buildx-action@v3 - name: Login to GitHub Container Registry uses: docker/login-action@v3 diff --git a/.github/workflows/publish-to-test.yaml b/.github/workflows/publish-to-test.yaml index 47e7c40b1..2bf0ea9ba 100644 --- a/.github/workflows/publish-to-test.yaml +++ b/.github/workflows/publish-to-test.yaml @@ -16,11 +16,11 @@ jobs: runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" - name: Set up Python - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: "3.8" - name: Append Dev Version to __version__ @@ -40,4 +40,4 @@ jobs: uses: pypa/gh-action-pypi-publish@release/v1 with: password: ${{ secrets.TEST_PYPI_API_TOKEN }} - repository-url: https://test.pypi.org/legacy/ \ No newline at end of file + repository-url: https://test.pypi.org/legacy/ diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index 1afdd667d..bc4ec9063 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -10,11 +10,11 @@ jobs: runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" - name: Set up Python - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: "3.8" - name: Install dependencies diff --git a/.github/workflows/test-pypi.yaml b/.github/workflows/test-pypi.yaml index cc6a3a725..aa8e8fa0b 100644 --- a/.github/workflows/test-pypi.yaml +++ b/.github/workflows/test-pypi.yaml @@ -8,11 +8,11 @@ jobs: runs-on: ubuntu-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"] + python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - name: Install dependencies @@ -28,11 +28,11 @@ jobs: runs-on: windows-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"] + python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - name: Install dependencies @@ -48,11 +48,11 @@ jobs: runs-on: macos-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"] + python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - name: Install dependencies @@ -61,4 +61,4 @@ jobs: python3 -m pip install --verbose llama-cpp-python[all] - name: Test with pytest run: | - python3 -c "import llama_cpp" \ No newline at end of file + python3 -c "import llama_cpp" diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 6b0e7e8c6..292343c2b 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -21,7 +21,7 @@ jobs: with: submodules: "recursive" - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - name: Install dependencies @@ -40,11 +40,11 @@ jobs: python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - name: Install dependencies @@ -63,11 +63,11 @@ jobs: python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - name: Install dependencies @@ -83,11 +83,11 @@ jobs: # runs-on: ubuntu-latest # steps: - # - uses: actions/checkout@v3 + # - uses: actions/checkout@v4 # with: # submodules: "recursive" # - name: Set up Python 3.8 - # uses: actions/setup-python@v4 + # uses: actions/setup-python@v5 # with: # python-version: "3.8" # - name: Set up OpenCL & CLBlast @@ -110,11 +110,11 @@ jobs: runs-on: macos-13 steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: submodules: "recursive" - name: Set up Python 3.8 - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: "3.8" - name: Install dependencies From 32c000f3ec5d0417612256a92297362c4c21c4ce Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon, 29 Apr 2024 22:52:58 -0400 Subject: [PATCH 331/624] chore(deps): bump softprops/action-gh-release from 1 to 2 (#1408) Bumps [softprops/action-gh-release](https://github.com/softprops/action-gh-release) from 1 to 2. - [Release notes](https://github.com/softprops/action-gh-release/releases) - [Changelog](https://github.com/softprops/action-gh-release/blob/master/CHANGELOG.md) - [Commits](https://github.com/softprops/action-gh-release/compare/v1...v2) --- updated-dependencies: - dependency-name: softprops/action-gh-release dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> From be43018e09ef795fc8b214a4d13fd702dc670a74 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon, 29 Apr 2024 22:53:21 -0400 Subject: [PATCH 332/624] chore(deps): bump actions/configure-pages from 4 to 5 (#1411) Bumps [actions/configure-pages](https://github.com/actions/configure-pages) from 4 to 5. - [Release notes](https://github.com/actions/configure-pages/releases) - [Commits](https://github.com/actions/configure-pages/compare/v4...v5) --- updated-dependencies: - dependency-name: actions/configure-pages dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/generate-index-from-release.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 858e50711..500c4613c 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -31,7 +31,7 @@ jobs: - name: Checkout uses: actions/checkout@v4 - name: Setup Pages - uses: actions/configure-pages@v4 + uses: actions/configure-pages@v5 - name: Build run: | ./scripts/releases-to-pep-503.sh index/whl/cpu '^[v]?[0-9]+\.[0-9]+\.[0-9]+$' From df2b5b5d4485c485b6043d193f10092452d760d6 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon, 29 Apr 2024 22:53:42 -0400 Subject: [PATCH 333/624] chore(deps): bump actions/upload-artifact from 3 to 4 (#1412) Bumps [actions/upload-artifact](https://github.com/actions/upload-artifact) from 3 to 4. - [Release notes](https://github.com/actions/upload-artifact/releases) - [Commits](https://github.com/actions/upload-artifact/compare/v3...v4) --- updated-dependencies: - dependency-name: actions/upload-artifact dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 76bf708fe..78bcb87d4 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -37,7 +37,7 @@ jobs: package-dir: . output-dir: wheelhouse - - uses: actions/upload-artifact@v3 + - uses: actions/upload-artifact@v4 with: path: ./wheelhouse/*.whl @@ -65,7 +65,7 @@ jobs: output-dir: wheelhouse - name: Upload wheels as artifacts - uses: actions/upload-artifact@v3 + uses: actions/upload-artifact@v4 with: path: ./wheelhouse/*.whl @@ -87,7 +87,7 @@ jobs: - name: Build source distribution run: | python -m build --sdist - - uses: actions/upload-artifact@v3 + - uses: actions/upload-artifact@v4 with: path: ./dist/*.tar.gz From 97fb860eba42e1018abc9b603f8e4602e84dd153 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 29 Apr 2024 23:34:55 -0400 Subject: [PATCH 334/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 21 +++++++++++++++++++++ tests/test_llama.py | 2 +- vendor/llama.cpp | 2 +- 3 files changed, 23 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d00dfcb95..9c8f77889 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -284,6 +284,27 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa """BERT tokenizer based on WordPiece""" +# // pre-tokenization types +# enum llama_vocab_pre_type { +# LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0, +# LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3, +# LLAMA_VOCAB_PRE_TYPE_FALCON = 4, +# LLAMA_VOCAB_PRE_TYPE_MPT = 5, +# LLAMA_VOCAB_PRE_TYPE_STARCODER = 6, +# LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, +# }; +LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 +LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3 +LLAMA_VOCAB_PRE_TYPE_FALCON = 4 +LLAMA_VOCAB_PRE_TYPE_MPT = 5 +LLAMA_VOCAB_PRE_TYPE_STARCODER = 6 +LLAMA_VOCAB_PRE_TYPE_GPT2 = 7 + + # // note: these values should be synchronized with ggml_rope # // TODO: maybe move this enum to ggml.h (ggml_rope_type) # enum llama_rope_type { diff --git a/tests/test_llama.py b/tests/test_llama.py index fa2f6dfc1..469ef91ca 100644 --- a/tests/test_llama.py +++ b/tests/test_llama.py @@ -6,7 +6,7 @@ import llama_cpp -MODEL = "./vendor/llama.cpp/models/ggml-vocab-llama.gguf" +MODEL = "./vendor/llama.cpp/models/ggml-vocab-llama-spm.gguf" def test_llama_cpp_tokenization(): diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4dba7e811..8843a98c2 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4dba7e8114d84241c842b986e008af8b88d1a019 +Subproject commit 8843a98c2ba97a25e93319a104f9ddfaf83ce4c4 From fe2da0953895000e20655d2c991308f702508494 Mon Sep 17 00:00:00 2001 From: Andrei Date: Tue, 30 Apr 2024 01:35:38 -0400 Subject: [PATCH 335/624] feat: Generic Chat Formats, Tool Calling, and Huggingface Pull Support for Multimodal Models (Obsidian, LLaVA1.6, Moondream) (#1147) * Test dummy image tags in chat templates * Format and improve types for llava_cpp.py * Add from_pretrained support to llava chat format. * Refactor llava chat format to use a jinja2 * Revert chat format test * Add moondream support (wip) * Update moondream chat format * Update moondream chat format * Update moondream prompt * Add function calling support * Cache last image embed * Add Llava1.6 support * Add nanollava support * Add obisidian support * Remove unnecessary import * Re-order multimodal chat formats * Logits all no longer required for multi-modal models * Update README.md * Update docs * Update README * Fix typo * Update README * Fix typo --- README.md | 41 ++- docs/server.md | 2 + llama_cpp/llama_chat_format.py | 631 +++++++++++++++++++++++++++------ llama_cpp/llava_cpp.py | 111 ++++-- llama_cpp/server/model.py | 71 +++- 5 files changed, 711 insertions(+), 145 deletions(-) diff --git a/README.md b/README.md index a33524cab..d34caf409 100644 --- a/README.md +++ b/README.md @@ -490,14 +490,15 @@ Due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is requi ### Multi-modal Models -`llama-cpp-python` supports the llava1.5 family of multi-modal models which allow the language model to -read information from both text and images. +`llama-cpp-python` supports such as llava1.5 which allow the language model to read information from both text and images. You'll first need to download one of the available multi-modal models in GGUF format: - [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) - [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) - [bakllava-1-7b](https://huggingface.co/mys/ggml_bakllava-1) +- [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) +- [moondream2](https://huggingface.co/vikhyatk/moondream2) Then you'll need to use a custom chat handler to load the clip model and process the chat messages and images. @@ -509,7 +510,6 @@ Then you'll need to use a custom chat handler to load the clip model and process model_path="./path/to/llava/llama-model.gguf", chat_handler=chat_handler, n_ctx=2048, # n_ctx should be increased to accomodate the image embedding - logits_all=True,# needed to make llava work ) >>> llm.create_chat_completion( messages = [ @@ -517,14 +517,45 @@ Then you'll need to use a custom chat handler to load the clip model and process { "role": "user", "content": [ - {"type": "image_url", "image_url": {"url": "https://.../image.png"}}, - {"type" : "text", "text": "Describe this image in detail please."} + {"type" : "text", "text": "What's in this image?"}, + {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" } } ] } ] ) ``` +You can also pull the model from the Hugging Face Hub using the `from_pretrained` method. + +```python +>>> from llama_cpp import Llama +>>> from llama_cpp.llama_chat_format import MoondreamChatHandler +>>> chat_handler = MoondreamChatHandler.from_pretrained( + repo_id="vikhyatk/moondream2", + filename="*mmproj*", +) +>>> llm = Llama.from_pretrained( + repo_id="vikhyatk/moondream2" + filename="*text-model*", + chat_handler=chat_handler, + n_ctx=2048, # n_ctx should be increased to accomodate the image embedding +) +>>> llm.create_chat_completion( + messages = [ + { + "role": "user", + "content": [ + {"type" : "text", "text": "What's in this image?"}, + {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" } } + + ] + } + ] +) +``` + +**Note**: Multi-modal models also support tool calling and JSON mode. +
Loading a Local Image diff --git a/docs/server.md b/docs/server.md index c66c9cce0..cd6f86c51 100644 --- a/docs/server.md +++ b/docs/server.md @@ -98,6 +98,8 @@ You'll first need to download one of the available multi-modal models in GGUF fo - [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) - [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) - [bakllava-1-7b](https://huggingface.co/mys/ggml_bakllava-1) +- [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) +- [moondream2](https://huggingface.co/vikhyatk/moondream2) Then when you run the server you'll need to also specify the path to the clip model used for image embedding and the `llava-1-5` chat_format diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 71aac8061..cbd8d203c 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -6,6 +6,8 @@ import dataclasses import random import string + +from contextlib import ExitStack from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol, cast import jinja2 @@ -2163,42 +2165,80 @@ def create_completion(stop): class Llava15ChatHandler: - _clip_free = None + DEFAULT_SYSTEM_MESSAGE = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." + + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.role == 'user' %}" + "{% if message.content is string %}" + "\nUSER: {{ message.content }}" + "{% elif message.content is iterable %}" + "\nUSER: " + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "{% endif %}" + "{% if message.role == 'assistant' and message.content is not none %}" + "\nASSISTANT: {{ message.content }}" + "{% endif %}" + "{% endfor %}" + "{% if add_generation_prompt %}" + "\nASSISTANT: " + "{% endif %}" + ) def __init__(self, clip_model_path: str, verbose: bool = False): import llama_cpp.llava_cpp as llava_cpp - self._llava_cpp = llava_cpp self.clip_model_path = clip_model_path self.verbose = verbose - self._clip_free = self._llava_cpp._libllava.clip_free # type: ignore + + self._llava_cpp = llava_cpp # TODO: Fix + self._exit_stack = ExitStack() + self._last_image_embed: Optional[llava_cpp.CtypesPointer[llava_cpp.llava_image_embed]] = None + self._last_image_hash: Optional[int] = None if not os.path.exists(clip_model_path): raise ValueError(f"Clip model path does not exist: {clip_model_path}") with suppress_stdout_stderr(disable=self.verbose): - self.clip_ctx = self._llava_cpp.clip_model_load( + clip_ctx = self._llava_cpp.clip_model_load( self.clip_model_path.encode(), 0 ) - def __del__(self): - with suppress_stdout_stderr(disable=self.verbose): - if self.clip_ctx is not None and self._clip_free is not None: - self._clip_free(self.clip_ctx) - self.clip_ctx = None + if clip_ctx is None: + raise ValueError(f"Failed to load clip model: {clip_model_path}") + + self.clip_ctx = clip_ctx - def load_image(self, image_url: str) -> bytes: - if image_url.startswith("data:"): - import base64 + def clip_free(): + with suppress_stdout_stderr(disable=self.verbose): + self._llava_cpp.clip_free(self.clip_ctx) + + self._exit_stack.callback(clip_free) + + def last_image_embed_free(): + with suppress_stdout_stderr(disable=self.verbose): + if self._last_image_embed is not None: + self._llava_cpp.llava_image_embed_free(self._last_image_embed) + self._last_image_embed = None - image_bytes = base64.b64decode(image_url.split(",")[1]) - return image_bytes - else: - import urllib.request + self._exit_stack.callback(last_image_embed_free) - with urllib.request.urlopen(image_url) as f: - image_bytes = f.read() - return image_bytes + def load_image(self, image_url: str) -> bytes: + return self._load_image(image_url) def __call__( self, @@ -2216,6 +2256,7 @@ def __call__( typical_p: float = 1.0, stream: bool = False, stop: Optional[Union[str, List[str]]] = [], + seed: Optional[int] = None, response_format: Optional[ llama_types.ChatCompletionRequestResponseFormat ] = None, @@ -2230,121 +2271,477 @@ def __call__( model: Optional[str] = None, logits_processor: Optional[llama.LogitsProcessorList] = None, grammar: Optional[llama.LlamaGrammar] = None, + logit_bias: Optional[Dict[str, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, **kwargs, # type: ignore ) -> Union[ llama_types.CreateChatCompletionResponse, Iterator[llama_types.CreateChatCompletionStreamResponse], ]: - assert ( - llama.context_params.logits_all is True - ) # BUG: logits_all=True is required for llava assert self.clip_ctx is not None + system_prompt = _get_system_message(messages) - system_prompt = ( - system_prompt - if system_prompt != "" - else "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." - ) - user_role = "\nUSER:" - assistant_role = "\nASSISTANT:" - llama.reset() - llama.eval(llama.tokenize(system_prompt.encode("utf8"), add_bos=True)) - for message in messages: - if message["role"] == "user" and message["content"] is not None: - if isinstance(message["content"], str): - llama.eval( - llama.tokenize( - f"{user_role} {message['content']}".encode("utf8"), - add_bos=False, - ) - ) - else: - assert isinstance(message["content"], list) - llama.eval( - llama.tokenize(f"{user_role} ".encode("utf8"), add_bos=False) - ) - for content in message["content"]: - if content["type"] == "text": - llama.eval( - llama.tokenize( - f"{content['text']}".encode("utf8"), add_bos=False - ) - ) - if content["type"] == "image_url": - image_bytes = ( - self.load_image(content["image_url"]["url"]) - if isinstance(content["image_url"], dict) - else self.load_image(content["image_url"]) - ) - import array - - data_array = array.array("B", image_bytes) - c_ubyte_ptr = ( - ctypes.c_ubyte * len(data_array) - ).from_buffer(data_array) - with suppress_stdout_stderr(disable=self.verbose): - embed = ( - self._llava_cpp.llava_image_embed_make_with_bytes( - self.clip_ctx, - llama.context_params.n_threads, - c_ubyte_ptr, - len(image_bytes), - ) - ) - try: - n_past = ctypes.c_int(llama.n_tokens) - n_past_p = ctypes.pointer(n_past) - with suppress_stdout_stderr(disable=self.verbose): - self._llava_cpp.llava_eval_image_embed( - llama.ctx, - embed, - llama.n_batch, - n_past_p, - ) - assert llama.n_ctx() >= n_past.value - llama.n_tokens = n_past.value - finally: - with suppress_stdout_stderr(disable=self.verbose): - self._llava_cpp.llava_image_embed_free(embed) - if message["role"] == "assistant" and message["content"] is not None: - llama.eval( - llama.tokenize( - f"ASSISTANT: {message['content']}".encode("utf8"), add_bos=False + if system_prompt == "": + messages = [llama_types.ChatCompletionRequestSystemMessage(role="system", content=self.DEFAULT_SYSTEM_MESSAGE)] + messages + + image_urls = self.get_image_urls(messages) + template = jinja2.Template(self.CHAT_FORMAT) + text = template.render(messages=messages, add_generation_prompt=True) + split_text = self.split_text_on_image_urls(text, image_urls) + + def embed_image_bytes(image_bytes: bytes): + if self._last_image_embed is not None and self._last_image_hash is not None and hash(image_bytes) == self._last_image_hash: + return self._last_image_embed + with suppress_stdout_stderr(disable=self.verbose): + embed = ( + self._llava_cpp.llava_image_embed_make_with_bytes( + self.clip_ctx, + llama.context_params.n_threads_batch, + (ctypes.c_uint8 * len(image_bytes)).from_buffer(bytearray(image_bytes)), + len(image_bytes), ) ) - assert llama.n_ctx() >= llama.n_tokens - llama.eval(llama.tokenize(f"{assistant_role}".encode("utf8"), add_bos=False)) - assert llama.n_ctx() >= llama.n_tokens + self._last_image_embed = embed + self._last_image_hash = hash(image_bytes) + return embed + + # Evaluate prompt + llama.reset() + for i, (type_, value) in enumerate(split_text): + if type_ == "text": + tokens = llama.tokenize(value.encode("utf8"), add_bos=i == 0) + if llama.n_tokens + len(tokens) > llama.n_ctx(): + raise ValueError("Prompt exceeds n_ctx") # TODO: Fix + llama.eval(tokens) + else: + image_bytes = self.load_image(value) + embed = embed_image_bytes(image_bytes) + if llama.n_tokens + embed.contents.n_image_pos > llama.n_ctx(): + raise ValueError("Prompt exceeds n_ctx") # TODO: Fix + n_past = ctypes.c_int(llama.n_tokens) + n_past_p = ctypes.pointer(n_past) + with suppress_stdout_stderr(disable=self.verbose): + self._llava_cpp.llava_eval_image_embed( + llama.ctx, + embed, + llama.n_batch, + n_past_p, + ) + llama.n_tokens = n_past.value + # Get prompt tokens to avoid a cache miss prompt = llama.input_ids[: llama.n_tokens].tolist() if response_format is not None and response_format["type"] == "json_object": grammar = _grammar_for_response_format(response_format) - return _convert_completion_to_chat( - llama.create_completion( - prompt=prompt, - temperature=temperature, - top_p=top_p, - top_k=top_k, - min_p=min_p, - typical_p=typical_p, - stream=stream, - stop=stop, - max_tokens=max_tokens, - presence_penalty=presence_penalty, - frequency_penalty=frequency_penalty, - repeat_penalty=repeat_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - model=model, - logits_processor=logits_processor, - grammar=grammar, - ), + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + tool = None + if tool_choice is not None and isinstance(tool_choice, dict) and tools is not None: + name = tool_choice["function"]["name"] + tool = next((t for t in tools if t["function"]["name"] == name), None) + if tool is None: + raise ValueError(f"Tool choice '{name}' not found in tools.") + schema = tool["function"]["parameters"] + try: + # create grammar from json schema + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(schema), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + logprobs=top_logprobs if logprobs else None, stream=stream, + stop=stop, + seed=seed, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + logit_bias=logit_bias, ) + if tool is not None: + tool_name = tool["function"]["name"] + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) + return _convert_completion_to_chat(completion_or_chunks, stream=stream) + + @staticmethod + def _load_image(image_url: str) -> bytes: + # TODO: Add Pillow support for other image formats beyond (jpg, png) + if image_url.startswith("data:"): + import base64 + + image_bytes = base64.b64decode(image_url.split(",")[1]) + return image_bytes + else: + import urllib.request + + with urllib.request.urlopen(image_url) as f: + image_bytes = f.read() + return image_bytes + + @staticmethod + def get_image_urls(messages: List[llama_types.ChatCompletionRequestMessage]): + image_urls: List[str] = [] + for message in messages: + if message["role"] == "user": + if message["content"] is None: + continue + for content in message["content"]: + if isinstance(content, dict) and "type" in content: + if content["type"] == "image_url": + if ( + isinstance(content["image_url"], dict) + and "url" in content["image_url"] + ): + image_urls.append(content["image_url"]["url"]) + else: + image_urls.append(content["image_url"]) + return image_urls + + @staticmethod + def split_text_on_image_urls(text: str, image_urls: List[str]): + def find_first(s: str, substrs: List[str]): + for i, substr in enumerate(substrs): + pos = s.find(substr) + if pos != -1: + return pos, i + return None, None + + split_text: List[Tuple[Literal["text", "image_url"], str]] = [] + remaining = text + while remaining: + # Find first image_url + pos, i = find_first(remaining, image_urls) + if pos is not None and i is not None: + if pos > 0: + split_text.append(("text", remaining[:pos])) + split_text.append(("image_url", image_urls[i])) + remaining = remaining[pos + len(image_urls[i]) :] + else: + split_text.append(("text", remaining)) + remaining = "" + return split_text + + @classmethod + def from_pretrained( + cls, + repo_id: str, + filename: Optional[str], + local_dir: Optional[Union[str, os.PathLike[str]]] = None, + local_dir_use_symlinks: Union[bool, Literal["auto"]] = "auto", + cache_dir: Optional[Union[str, os.PathLike[str]]] = None, + **kwargs: Any, + ) -> "Llava15ChatHandler": + import fnmatch + from pathlib import Path + try: + from huggingface_hub import hf_hub_download, HfFileSystem # type: ignore + from huggingface_hub.utils import validate_repo_id # type: ignore + except ImportError: + raise ImportError( + "Llama.from_pretrained requires the huggingface-hub package. " + "You can install it with `pip install huggingface-hub`." + ) + + validate_repo_id(repo_id) + + hffs = HfFileSystem() + + files = [ + file["name"] if isinstance(file, dict) else file + for file in hffs.ls(repo_id) # type: ignore + ] + + # split each file into repo_id, subfolder, filename + file_list: List[str] = [] + for file in files: + rel_path = Path(file).relative_to(repo_id) + file_list.append(str(rel_path)) + + matching_files = [file for file in file_list if fnmatch.fnmatch(file, filename)] # type: ignore + + if len(matching_files) == 0: + raise ValueError( + f"No file found in {repo_id} that match {filename}\n\n" + f"Available Files:\n{json.dumps(file_list)}" + ) + + if len(matching_files) > 1: + raise ValueError( + f"Multiple files found in {repo_id} matching {filename}\n\n" + f"Available Files:\n{json.dumps(files)}" + ) + + (matching_file,) = matching_files + + subfolder = str(Path(matching_file).parent) + filename = Path(matching_file).name + + # download the file + hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=cast(Union[str, Path, None], local_dir), + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cast(Union[str, Path, None], cache_dir), + ) + + if local_dir is None: + model_path = hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cast(Union[str, Path, None], cache_dir), + local_files_only=True, + ) + else: + model_path = os.path.join(local_dir, filename) + + return cls( + clip_model_path=model_path, + **kwargs, + ) + +class ObsidianChatHandler(Llava15ChatHandler): + # Prompt Format + # The model followed ChatML format. However, with ### as the seperator + + # <|im_start|>user + # What is this sign about?\n + # ### + # <|im_start|>assistant + # The sign is about bullying, and it is placed on a black background with a red background. + # ### + + CHAT_FORMAT = ( + "{% for message in messages %}" + # System message + "{% if message.role == 'system' %}" + "<|im_start|>system\n" + "{{ message.content }}\n" + "###\n" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "<|im_start|>user\n" + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.content is iterable %}" + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% if content.type == 'image_url' %}" + "{{ content.image_url }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "###\n" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + "<|im_start|>assistant\n" + "{{ message.content }}" + "###\n" + "{% endif %}" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "<|im_start|>assistant\n" + "{% endif %}" + ) + +class MoondreamChatHandler(Llava15ChatHandler): + # Chat Format: + # f"\n\n{chat_history}Question: {question}\n\nAnswer:" + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if message.role == 'user' %}" + "{% if message.content is iterable %}" + "{% for content in message.content %}" + + # + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}\n\n" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}\n\n" + "{% endif %}" + "{% endif %}" + + # Question: + "{% if content.type == 'text' %}" + "Question: {{ content.text }}\n\n" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + + # Question: + "{% if message.content is string %}" + "Question: {{ message.content }}\n\n" + "{% endif %}" + + "{% endif %}" + + # Answer: + "{% if message.role == 'assistant' %}" + "Answer:{{ message.content }}\n\n" + "{% endif %}" + "{% endfor %}" + + # Generation prompt + "{% if add_generation_prompt %}" + "Answer:" + "{% endif %}" + ) + +class Llava16ChatHandler(Llava15ChatHandler): + DEFAULT_SYSTEM_MESSAGE = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. " + + # Example prompt + # "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: \nWhat is shown in this image? ASSISTANT:" + + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.role == 'user' %}" + "{% if message.content is iterable %}" + "{% for content in message.content %}" + + # + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}\n" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}\n" + "{% endif %}" + "{% endif %}" + + # Question: + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + + # Question: + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + + "{% endif %}" + + # Answer: + "{% if message.role == 'assistant' %}" + "{{ message.content }}" + "{% endif %}" + "{% endfor %}" + + # Generation prompt + "{% if add_generation_prompt %}" + "Answer:" + "{% endif %}" + ) + +class NanoLlavaChatHandler(Llava15ChatHandler): + # Prompt Format + # The model follow the ChatML standard, however, without \n at the end of <|im_end|>: + + # <|im_start|>system + # Answer the question<|im_end|><|im_start|>user + # + # What is the picture about?<|im_end|><|im_start|>assistant + + CHAT_FORMAT = ( + "{% for message in messages %}" + # System message + "{% if message.role == 'system' %}" + "<|im_start|>system\n" + "{{ message.content }}" + "<|im_end|>" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "<|im_start|>user\n" + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.content is iterable %}" + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% if content.type == 'image_url' %}" + "{{ content.image_url }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "<|im_end|>" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + "<|im_start|>assistant\n" + "{{ message.content }}" + "<|im_end|>" + "{% endif %}" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "<|im_start|>assistant\n" + "{% endif %}" + ) @register_chat_completion_handler("chatml-function-calling") diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index 543c87d0b..3ded1f25d 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import sys import os import ctypes @@ -14,10 +16,22 @@ Structure, ) import pathlib -from typing import List, Union, NewType, Optional, TypeVar, Callable, Any +from typing import ( + List, + Union, + NewType, + Optional, + TypeVar, + Callable, + Any, + TYPE_CHECKING, + Generic, +) +from typing_extensions import TypeAlias import llama_cpp.llama_cpp as llama_cpp + # Load the library def _load_shared_library(lib_base_name: str): # Construct the paths to the possible shared library names @@ -62,7 +76,7 @@ def _load_shared_library(lib_base_name: str): for _lib_path in _lib_paths: if _lib_path.exists(): try: - return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore + return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore except Exception as e: raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") @@ -79,8 +93,27 @@ def _load_shared_library(lib_base_name: str): # ctypes helper +if TYPE_CHECKING: + CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore + + CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore + + CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore + + CtypesVoidPointer: TypeAlias = ctypes.c_void_p + + class CtypesRef(Generic[CtypesCData]): + pass + + CtypesPointerOrRef: TypeAlias = Union[ + CtypesPointer[CtypesCData], CtypesRef[CtypesCData] + ] + + CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore + F = TypeVar("F", bound=Callable[..., Any]) + def ctypes_function_for_shared_library(lib: ctypes.CDLL): def ctypes_function( name: str, argtypes: List[Any], restype: Any, enabled: bool = True @@ -111,6 +144,7 @@ def decorator(f: F) -> F: clip_ctx_p = NewType("clip_ctx_p", int) clip_ctx_p_ctypes = c_void_p + # struct llava_image_embed { # float * embed; # int n_image_pos; @@ -121,36 +155,72 @@ class llava_image_embed(Structure): ("n_image_pos", c_int), ] + # /** sanity check for clip <-> llava embed size match */ # LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip); -@ctypes_function("llava_validate_embed_size", [llama_cpp.llama_context_p_ctypes, clip_ctx_p_ctypes], c_bool) -def llava_validate_embed_size(ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p, /) -> bool: - ... +@ctypes_function( + "llava_validate_embed_size", + [llama_cpp.llama_context_p_ctypes, clip_ctx_p_ctypes], + c_bool, +) +def llava_validate_embed_size( + ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p, / +) -> bool: ... # /** build an image embed from image file bytes */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); -@ctypes_function("llava_image_embed_make_with_bytes", [clip_ctx_p_ctypes, c_int, POINTER(c_uint8), c_int], POINTER(llava_image_embed)) -def llava_image_embed_make_with_bytes(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_bytes: bytes, image_bytes_length: Union[c_int, int], /) -> "_Pointer[llava_image_embed]": - ... +@ctypes_function( + "llava_image_embed_make_with_bytes", + [clip_ctx_p_ctypes, c_int, POINTER(c_uint8), c_int], + POINTER(llava_image_embed), +) +def llava_image_embed_make_with_bytes( + ctx_clip: clip_ctx_p, + n_threads: Union[c_int, int], + image_bytes: CtypesArray[c_uint8], + image_bytes_length: Union[c_int, int], + /, +) -> "_Pointer[llava_image_embed]": ... + # /** build an image embed from a path to an image filename */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); -@ctypes_function("llava_image_embed_make_with_filename", [clip_ctx_p_ctypes, c_int, c_char_p], POINTER(llava_image_embed)) -def llava_image_embed_make_with_filename(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes, /) -> "_Pointer[llava_image_embed]": - ... +@ctypes_function( + "llava_image_embed_make_with_filename", + [clip_ctx_p_ctypes, c_int, c_char_p], + POINTER(llava_image_embed), +) +def llava_image_embed_make_with_filename( + ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes, / +) -> "_Pointer[llava_image_embed]": ... + # LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); # /** free an embedding made with llava_image_embed_make_* */ @ctypes_function("llava_image_embed_free", [POINTER(llava_image_embed)], None) -def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): - ... +def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): ... + # /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ # LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); -@ctypes_function("llava_eval_image_embed", [llama_cpp.llama_context_p_ctypes, POINTER(llava_image_embed), c_int, POINTER(c_int)], c_bool) -def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointer[llava_image_embed]", n_batch: Union[c_int, int], n_past: "_Pointer[c_int]", /) -> bool: - ... +@ctypes_function( + "llava_eval_image_embed", + [ + llama_cpp.llama_context_p_ctypes, + POINTER(llava_image_embed), + c_int, + POINTER(c_int), + ], + c_bool, +) +def llava_eval_image_embed( + ctx_llama: llama_cpp.llama_context_p, + embed: "_Pointer[llava_image_embed]", + n_batch: Union[c_int, int], + n_past: "_Pointer[c_int]", + /, +) -> bool: ... ################################################ @@ -161,11 +231,12 @@ def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointe # /** load mmproj model */ # CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); @ctypes_function("clip_model_load", [c_char_p, c_int], clip_ctx_p_ctypes) -def clip_model_load(fname: bytes, verbosity: Union[c_int, int], /) -> Optional[clip_ctx_p]: - ... +def clip_model_load( + fname: bytes, verbosity: Union[c_int, int], / +) -> Optional[clip_ctx_p]: ... + # /** free mmproj model */ # CLIP_API void clip_free(struct clip_ctx * ctx); @ctypes_function("clip_free", [clip_ctx_p_ctypes], None) -def clip_free(ctx: clip_ctx_p, /): - ... +def clip_free(ctx: clip_ctx_p, /): ... diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index c24fca652..1ad592d21 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -72,9 +72,74 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler = None if settings.chat_format == "llava-1-5": assert settings.clip_model_path is not None, "clip model not found" - chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler( - clip_model_path=settings.clip_model_path, verbose=settings.verbose - ) + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.Llava15ChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "obsidian": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.ObsidianChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.ObsidianChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "llava-1-6": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.Llava16ChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.Llava16ChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "moondream": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.MoondreamChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.MoondreamChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "nanollava": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.NanoLlavaChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.NanoLlavaChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) elif settings.chat_format == "hf-autotokenizer": assert ( settings.hf_pretrained_model_name_or_path is not None From 26c7876ba0b31bd0e9f7bfc8f341b8b5b3d81fa7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 01:48:40 -0400 Subject: [PATCH 336/624] chore: Bump version --- CHANGELOG.md | 11 +++++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 12 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 806876ce2..26df77fb3 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.66] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8843a98c2ba97a25e93319a104f9ddfaf83ce4c4 +- feat: Generic Chat Formats, Tool Calling, and Huggingface Pull Support for Multimodal Models (Obsidian, LLaVA1.6, Moondream) by @abetlen in #1147 +- ci(fix): Workflow actions updates and fix arm64 wheels not included in release by @Smartappli in #1392 +- ci: Add support for pre-built cuda 12.4.1 wheels by @Smartappli in #1388 +- feat: Add support for str type kv_overrides by @abetlen in a411612b385cef100d76145da1fbd02a7b7cc894 +- fix: Functionary bug fixes by @jeffrey-fong in #1385 +- examples: fix quantize example by @iyubondyrev in #1387 +- ci: Update dependabot.yml by @Smartappli in #1391 + ## [0.2.65] - feat: Update llama.cpp to ggerganov/llama.cpp@46e12c4692a37bdd31a0432fc5153d7d22bc7f72 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 6db6333ba..49657c973 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.65" \ No newline at end of file +__version__ = "0.2.66" \ No newline at end of file From d03f15bb73a1d520970357b702a9e7d4cc2a7a62 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 02:58:55 -0400 Subject: [PATCH 337/624] fix(ci): Fix bug in use of upload-artifact failing to merge multiple artifacts into a single release. --- .github/workflows/build-and-release.yaml | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 78bcb87d4..bb3c9f8d9 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -39,9 +39,10 @@ jobs: - uses: actions/upload-artifact@v4 with: + name: wheels path: ./wheelhouse/*.whl - build_arm64_wheels: + build_wheels_arm64: name: Build arm64 wheels runs-on: ubuntu-latest steps: @@ -67,6 +68,7 @@ jobs: - name: Upload wheels as artifacts uses: actions/upload-artifact@v4 with: + name: wheels_arm64 path: ./wheelhouse/*.whl build_sdist: @@ -89,18 +91,25 @@ jobs: python -m build --sdist - uses: actions/upload-artifact@v4 with: + name: sdist path: ./dist/*.tar.gz release: name: Release - needs: [build_wheels, build_arm64_wheels, build_sdist] + needs: [build_wheels, build_wheels_arm64, build_sdist] runs-on: ubuntu-latest steps: + - name: Merge Artifacts + uses: actions/upload-artifact/merge@v4 + with: + name: release + - uses: actions/download-artifact@v3 with: - name: artifact + name: release path: dist + - uses: softprops/action-gh-release@v2 with: files: dist/* From 3489ef09d3775f4a87fb7114f619e8ba9cb6b656 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 03:08:46 -0400 Subject: [PATCH 338/624] fix: Ensure image renders before text in chat formats regardless of message content order. --- llama_cpp/llama_chat_format.py | 52 ++++++++++++++++++++++++++-------- 1 file changed, 40 insertions(+), 12 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index cbd8d203c..63eaf8a03 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2175,12 +2175,11 @@ class Llava15ChatHandler: "{% if message.role == 'user' %}" "{% if message.content is string %}" "\nUSER: {{ message.content }}" - "{% elif message.content is iterable %}" + "{% endif %}" + "{% if message.content is iterable %}" "\nUSER: " + "{% for content in message.content %}" - "{% if content.type == 'text' %}" - "{{ content.text }}" - "{% endif %}" "{% if content.type == 'image_url' and content.image_url is string %}" "{{ content.image_url }}" "{% endif %}" @@ -2188,6 +2187,13 @@ class Llava15ChatHandler: "{{ content.image_url.url }}" "{% endif %}" "{% endfor %}" + + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" "{% endif %}" "{% if message.role == 'assistant' and message.content is not none %}" @@ -2575,14 +2581,22 @@ class ObsidianChatHandler(Llava15ChatHandler): "{{ message.content }}" "{% endif %}" "{% if message.content is iterable %}" + + "{% for content in message.content %}" + "{% if content.type == 'image_url' and content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endfor %}" + "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" - "{% if content.type == 'image_url' %}" - "{{ content.image_url }}" - "{% endif %}" "{% endfor %}" + "{% endif %}" "###\n" "{% endif %}" @@ -2606,9 +2620,9 @@ class MoondreamChatHandler(Llava15ChatHandler): "{% for message in messages %}" "{% if message.role == 'user' %}" "{% if message.content is iterable %}" - "{% for content in message.content %}" # + "{% for content in message.content %}" "{% if content.type == 'image_url' %}" "{% if content.image_url is string %}" "{{ content.image_url }}\n\n" @@ -2617,12 +2631,15 @@ class MoondreamChatHandler(Llava15ChatHandler): "{{ content.image_url.url }}\n\n" "{% endif %}" "{% endif %}" + "{% endfor %}" # Question: + "{% for content in message.content %}" "{% if content.type == 'text' %}" "Question: {{ content.text }}\n\n" "{% endif %}" "{% endfor %}" + "{% endif %}" # Question: @@ -2657,9 +2674,9 @@ class Llava16ChatHandler(Llava15ChatHandler): "{% endif %}" "{% if message.role == 'user' %}" "{% if message.content is iterable %}" - "{% for content in message.content %}" # + "{% for content in message.content %}" "{% if content.type == 'image_url' %}" "{% if content.image_url is string %}" "{{ content.image_url }}\n" @@ -2668,12 +2685,15 @@ class Llava16ChatHandler(Llava15ChatHandler): "{{ content.image_url.url }}\n" "{% endif %}" "{% endif %}" + "{% endfor %}" # Question: + "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" "{% endfor %}" + "{% endif %}" # Question: @@ -2719,14 +2739,22 @@ class NanoLlavaChatHandler(Llava15ChatHandler): "{{ message.content }}" "{% endif %}" "{% if message.content is iterable %}" + + "{% for content in message.content %}" + "{% if content.type == 'image_url' and content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endfor %}" + "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" - "{% if content.type == 'image_url' %}" - "{{ content.image_url }}" - "{% endif %}" "{% endfor %}" + "{% endif %}" "<|im_end|>" "{% endif %}" From f417cce28a3f3386206dabe98e87c038d05b228d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 03:11:02 -0400 Subject: [PATCH 339/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 26df77fb3..4b2c34b22 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.67] + +- fix: Ensure image renders before text in chat formats regardless of message content order by @abetlen in 3489ef09d3775f4a87fb7114f619e8ba9cb6b656 +- fix(ci): Fix bug in use of upload-artifact failing to merge multiple artifacts into a single release by @abetlen in d03f15bb73a1d520970357b702a9e7d4cc2a7a62 + ## [0.2.66] - feat: Update llama.cpp to ggerganov/llama.cpp@8843a98c2ba97a25e93319a104f9ddfaf83ce4c4 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 49657c973..286baf445 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.66" \ No newline at end of file +__version__ = "0.2.67" \ No newline at end of file From c8cd8c17c6f178285579ce1f06347edb7324088c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 03:12:46 -0400 Subject: [PATCH 340/624] docs: Update README to include CUDA 12.4 wheels --- README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index d34caf409..310873f54 100644 --- a/README.md +++ b/README.md @@ -121,7 +121,7 @@ CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python It is also possible to install a pre-built wheel with CUDA support. As long as your system meets some requirements: -- CUDA Version is 12.1, 12.2 or 12.3 +- CUDA Version is 12.1, 12.2, 12.3, or 12.4 - Python Version is 3.10, 3.11 or 3.12 ```bash @@ -133,6 +133,7 @@ Where `` is one of the following: - `cu121`: CUDA 12.1 - `cu122`: CUDA 12.2 - `cu123`: CUDA 12.3 +- `cu124`: CUDA 12.4 For example, to install the CUDA 12.1 wheel: From 6332527a698f5aa2f524773f42a0510ba5f650a0 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Tue, 30 Apr 2024 15:16:14 +0200 Subject: [PATCH 341/624] fix(ci): Fix build-and-release.yaml (#1413) * Update build-and-release.yaml * Update build-and-release.yaml --- .github/workflows/build-and-release.yaml | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index bb3c9f8d9..8fbd68f4a 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -39,7 +39,7 @@ jobs: - uses: actions/upload-artifact@v4 with: - name: wheels + name: wheels-${{ matrix.os }} path: ./wheelhouse/*.whl build_wheels_arm64: @@ -100,14 +100,9 @@ jobs: runs-on: ubuntu-latest steps: - - name: Merge Artifacts - uses: actions/upload-artifact/merge@v4 + - uses: actions/download-artifact@v4 with: - name: release - - - uses: actions/download-artifact@v3 - with: - name: release + merge-multiple: true path: dist - uses: softprops/action-gh-release@v2 From 8c2b24d5aafffffadf37c3067ca12a45b59d95c4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 09:27:55 -0400 Subject: [PATCH 342/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 10 +++++++--- vendor/llama.cpp | 2 +- 2 files changed, 8 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 9c8f77889..46aa51662 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -242,8 +242,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 5 -LLAMA_SESSION_VERSION = 5 +# define LLAMA_SESSION_VERSION 6 +LLAMA_SESSION_VERSION = 6 # define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ @@ -730,6 +730,7 @@ class llama_model_params(ctypes.Structure): # bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) # bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU +# bool flash_attn; // whether to use flash attention # // Abort callback @@ -766,6 +767,7 @@ class llama_context_params(ctypes.Structure): logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) embeddings (bool): if true, extract embeddings (together with logits) offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU + flash_attn (bool): whether to use flash attention abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback """ @@ -795,6 +797,7 @@ class llama_context_params(ctypes.Structure): logits_all: bool embeddings: bool offload_kqv: bool + flash_attn: bool abort_callback: Callable[[ctypes.c_void_p], bool] abort_callback_data: ctypes.c_void_p @@ -823,6 +826,7 @@ class llama_context_params(ctypes.Structure): ("logits_all", ctypes.c_bool), ("embeddings", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), + ("flash_attn", ctypes.c_bool), ("abort_callback", ggml_abort_callback), ("abort_callback_data", ctypes.c_void_p), ] @@ -1615,7 +1619,7 @@ def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: ... -# // Clear the KV cache +# // Clear the KV cache - both cell info is erased and KV data is zeroed # LLAMA_API void llama_kv_cache_clear( # struct llama_context * ctx); @ctypes_function("llama_kv_cache_clear", [llama_context_p_ctypes], None) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8843a98c2..77e15bec6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8843a98c2ba97a25e93319a104f9ddfaf83ce4c4 +Subproject commit 77e15bec6217a39be59b9cc83d6b9afb6b0d8167 From 22d77eefd2edaf0148f53374d0cac74d0e25d06e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 09:29:16 -0400 Subject: [PATCH 343/624] feat: Add option to enable `flash_attn` to Lllama params and ModelSettings --- llama_cpp/llama.py | 4 ++++ llama_cpp/server/settings.py | 3 +++ 2 files changed, 7 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 96aac6668..172f4c646 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -92,6 +92,7 @@ def __init__( logits_all: bool = False, embedding: bool = False, offload_kqv: bool = True, + flash_attn: bool = False, # Sampling Params last_n_tokens_size: int = 64, # LoRA Params @@ -168,6 +169,7 @@ def __init__( logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs. embedding: Embedding mode only. offload_kqv: Offload K, Q, V to GPU. + flash_attn: Use flash attention. last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque. lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model. lora_path: Path to a LoRA file to apply to the model. @@ -310,6 +312,7 @@ def __init__( ) # Must be set to True for speculative decoding self.context_params.embeddings = embedding # TODO: Rename to embeddings self.context_params.offload_kqv = offload_kqv + self.context_params.flash_attn = flash_attn # KV cache quantization if type_k is not None: self.context_params.type_k = type_k @@ -1774,6 +1777,7 @@ def __getstate__(self): logits_all=self.context_params.logits_all, embedding=self.context_params.embeddings, offload_kqv=self.context_params.offload_kqv, + flash_offload=self.context_params.flash_offload, # Sampling Params last_n_tokens_size=self.last_n_tokens_size, # LoRA Params diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 0c858f9fe..ed05a889f 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -96,6 +96,9 @@ class ModelSettings(BaseSettings): offload_kqv: bool = Field( default=True, description="Whether to offload kqv to the GPU." ) + flash_attn: bool = Field( + default=False, description="Whether to use flash attention." + ) # Sampling Params last_n_tokens_size: int = Field( default=64, From 29b6e9a5c832b7d148044b45fb3cd6f3f923d96b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 09:32:47 -0400 Subject: [PATCH 344/624] fix: wrong parameter for flash attention in pickle __getstate__ --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 172f4c646..f927f0ca2 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1777,7 +1777,7 @@ def __getstate__(self): logits_all=self.context_params.logits_all, embedding=self.context_params.embeddings, offload_kqv=self.context_params.offload_kqv, - flash_offload=self.context_params.flash_offload, + flash_attn=self.context_params.flash_attn, # Sampling Params last_n_tokens_size=self.last_n_tokens_size, # LoRA Params From b14dd98922c7f18468ae202eadbaf58fe17f5320 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 09:39:56 -0400 Subject: [PATCH 345/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 4b2c34b22..c9681f355 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.68] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ +- feat: Add option to enable flash_attn to Lllama params and ModelSettings by @abetlen in 22d77eefd2edaf0148f53374d0cac74d0e25d06e +- fix(ci): Fix build-and-release.yaml by @Smartappli in #1413 + ## [0.2.67] - fix: Ensure image renders before text in chat formats regardless of message content order by @abetlen in 3489ef09d3775f4a87fb7114f619e8ba9cb6b656 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 286baf445..63c622551 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.67" \ No newline at end of file +__version__ = "0.2.68" \ No newline at end of file From 26478ab293bde6374e2c350fed7b07c3027939b6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 10:11:38 -0400 Subject: [PATCH 346/624] docs: Update README.md --- README.md | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/README.md b/README.md index 310873f54..fce8eda57 100644 --- a/README.md +++ b/README.md @@ -529,19 +529,22 @@ Then you'll need to use a custom chat handler to load the clip model and process You can also pull the model from the Hugging Face Hub using the `from_pretrained` method. ```python ->>> from llama_cpp import Llama ->>> from llama_cpp.llama_chat_format import MoondreamChatHandler ->>> chat_handler = MoondreamChatHandler.from_pretrained( +from llama_cpp import Llama +from llama_cpp.llama_chat_format import MoondreamChatHandler + +chat_handler = MoondreamChatHandler.from_pretrained( repo_id="vikhyatk/moondream2", filename="*mmproj*", ) ->>> llm = Llama.from_pretrained( - repo_id="vikhyatk/moondream2" + +llm = Llama.from_pretrained( + repo_id="vikhyatk/moondream2", filename="*text-model*", chat_handler=chat_handler, n_ctx=2048, # n_ctx should be increased to accomodate the image embedding ) ->>> llm.create_chat_completion( + +respoonse = llm.create_chat_completion( messages = [ { "role": "user", @@ -553,6 +556,7 @@ You can also pull the model from the Hugging Face Hub using the `from_pretrained } ] ) +print(response["choices"][0]["text"]) ``` **Note**: Multi-modal models also support tool calling and JSON mode. From 945c62c567a862e1c631a86bf88b86ebacefd40d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 10:15:04 -0400 Subject: [PATCH 347/624] docs: Change all examples from interpreter style to script style. --- README.md | 74 ++++++++++++++++++++++++++++++------------------------- 1 file changed, 40 insertions(+), 34 deletions(-) diff --git a/README.md b/README.md index fce8eda57..640671460 100644 --- a/README.md +++ b/README.md @@ -277,20 +277,26 @@ The high-level API provides a simple managed interface through the [`Llama`](htt Below is a short example demonstrating how to use the high-level API to for basic text completion: ```python ->>> from llama_cpp import Llama ->>> llm = Llama( +from llama_cpp import Llama + +llm = Llama( model_path="./models/7B/llama-model.gguf", # n_gpu_layers=-1, # Uncomment to use GPU acceleration # seed=1337, # Uncomment to set a specific seed # n_ctx=2048, # Uncomment to increase the context window ) ->>> output = llm( +output = llm( "Q: Name the planets in the solar system? A: ", # Prompt max_tokens=32, # Generate up to 32 tokens, set to None to generate up to the end of the context window stop=["Q:", "\n"], # Stop generating just before the model would generate a new question echo=True # Echo the prompt back in the output ) # Generate a completion, can also call create_completion ->>> print(output) +print(output) +``` + +By default `llama-cpp-python` generates completions in an OpenAI compatible format: + +```python { "id": "cmpl-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx", "object": "text_completion", @@ -345,12 +351,12 @@ The model will will format the messages into a single prompt using the following Set `verbose=True` to see the selected chat format. ```python ->>> from llama_cpp import Llama ->>> llm = Llama( +from llama_cpp import Llama +llm = Llama( model_path="path/to/llama-2/llama-model.gguf", chat_format="llama-2" ) ->>> llm.create_chat_completion( +llm.create_chat_completion( messages = [ {"role": "system", "content": "You are an assistant who perfectly describes images."}, { @@ -375,9 +381,9 @@ To constrain chat responses to only valid JSON or a specific JSON Schema use the The following example will constrain the response to valid JSON strings only. ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") ->>> llm.create_chat_completion( +from llama_cpp import Llama +llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") +llm.create_chat_completion( messages=[ { "role": "system", @@ -397,9 +403,9 @@ The following example will constrain the response to valid JSON strings only. To constrain the response further to a specific JSON Schema add the schema to the `schema` property of the `response_format` argument. ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") ->>> llm.create_chat_completion( +from llama_cpp import Llama +llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") +llm.create_chat_completion( messages=[ { "role": "system", @@ -424,9 +430,9 @@ To constrain the response further to a specific JSON Schema add the schema to th The high-level API supports OpenAI compatible function and tool calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat format. ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/chatml/llama-model.gguf", chat_format="chatml-function-calling") ->>> llm.create_chat_completion( +from llama_cpp import Llama +llm = Llama(model_path="path/to/chatml/llama-model.gguf", chat_format="chatml-function-calling") +llm.create_chat_completion( messages = [ { "role": "system", @@ -476,9 +482,9 @@ The various gguf-converted files for this set of models can be found [here](http Due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is required to provide HF Tokenizer for functionary. The `LlamaHFTokenizer` class can be initialized and passed into the Llama class. This will override the default llama.cpp tokenizer used in Llama class. The tokenizer files are already included in the respective HF repositories hosting the gguf files. ```python ->>> from llama_cpp import Llama ->>> from llama_cpp.llama_tokenizer import LlamaHFTokenizer ->>> llm = Llama.from_pretrained( +from llama_cpp import Llama +from llama_cpp.llama_tokenizer import LlamaHFTokenizer +llm = Llama.from_pretrained( repo_id="meetkai/functionary-small-v2.2-GGUF", filename="functionary-small-v2.2.q4_0.gguf", chat_format="functionary-v2", @@ -504,15 +510,15 @@ You'll first need to download one of the available multi-modal models in GGUF fo Then you'll need to use a custom chat handler to load the clip model and process the chat messages and images. ```python ->>> from llama_cpp import Llama ->>> from llama_cpp.llama_chat_format import Llava15ChatHandler ->>> chat_handler = Llava15ChatHandler(clip_model_path="path/to/llava/mmproj.bin") ->>> llm = Llama( +from llama_cpp import Llama +from llama_cpp.llama_chat_format import Llava15ChatHandler +chat_handler = Llava15ChatHandler(clip_model_path="path/to/llava/mmproj.bin") +llm = Llama( model_path="./path/to/llava/llama-model.gguf", chat_handler=chat_handler, n_ctx=2048, # n_ctx should be increased to accomodate the image embedding ) ->>> llm.create_chat_completion( +llm.create_chat_completion( messages = [ {"role": "system", "content": "You are an assistant who perfectly describes images."}, { @@ -709,18 +715,18 @@ The entire low-level API can be found in [llama_cpp/llama_cpp.py](https://github Below is a short example demonstrating how to use the low-level API to tokenize a prompt: ```python ->>> import llama_cpp ->>> import ctypes ->>> llama_cpp.llama_backend_init(False) # Must be called once at the start of each program ->>> params = llama_cpp.llama_context_default_params() +import llama_cpp +import ctypes +llama_cpp.llama_backend_init(False) # Must be called once at the start of each program +params = llama_cpp.llama_context_default_params() # use bytes for char * params ->>> model = llama_cpp.llama_load_model_from_file(b"./models/7b/llama-model.gguf", params) ->>> ctx = llama_cpp.llama_new_context_with_model(model, params) ->>> max_tokens = params.n_ctx +model = llama_cpp.llama_load_model_from_file(b"./models/7b/llama-model.gguf", params) +ctx = llama_cpp.llama_new_context_with_model(model, params) +max_tokens = params.n_ctx # use ctypes arrays for array params ->>> tokens = (llama_cpp.llama_token * int(max_tokens))() ->>> n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, llama_cpp.c_bool(True)) ->>> llama_cpp.llama_free(ctx) +tokens = (llama_cpp.llama_token * int(max_tokens))() +n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, llama_cpp.c_bool(True)) +llama_cpp.llama_free(ctx) ``` Check out the [examples folder](examples/low_level_api) for more examples of using the low-level API. From 3226b3c5ef93998f5c412fb926ff136b50816bd7 Mon Sep 17 00:00:00 2001 From: Jonathan Soma Date: Tue, 30 Apr 2024 14:33:23 -0400 Subject: [PATCH 348/624] fix: UTF-8 handling with grammars (#1415) Use Python's built-in UTF-8 handling to get code points --- llama_cpp/llama_grammar.py | 16 +++++----------- 1 file changed, 5 insertions(+), 11 deletions(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 6c7b57a59..d9a3823b4 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -556,17 +556,11 @@ def add_rule( # } def decode_utf8(src: const_char_p) -> Tuple[int, const_char_p]: """Decodes a UTF-8 character from the source string.""" - lookup = (1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4) - first_byte = ord(src[0]) # type: int - highbits = first_byte >> 4 # type: int - len = lookup[highbits] # type: int - mask = (1 << (8 - len)) - 1 # type: int - value = first_byte & mask # type: int - end = src + len # type: const_char_p # may overrun! - pos = src + 1 # type: const_char_p - while pos < end and pos[0]: - value = (value << 6) + (ord(pos[0]) & 0x3F) - pos += 1 + # Get the codepoint of the first character + value = ord(src[0]) + # Move the pointer ahead one character + pos = src + 1 + return value, pos From f116175a5a7c84569c88cad231855c1e6e59ff6e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 15:45:34 -0400 Subject: [PATCH 349/624] fix: Suppress all logs when verbose=False, use hardcoded fileno's to work in colab notebooks. Closes #796 Closes #729 --- llama_cpp/_internals.py | 8 +++++--- llama_cpp/_utils.py | 25 +++++++++++-------------- 2 files changed, 16 insertions(+), 17 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index cc3d98928..d7409f63a 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -15,6 +15,7 @@ from .llama_types import * from .llama_grammar import LlamaGrammar +from ._utils import suppress_stdout_stderr import llama_cpp.llama_cpp as llama_cpp @@ -47,9 +48,10 @@ def __init__( if not os.path.exists(path_model): raise ValueError(f"Model path does not exist: {path_model}") - self.model = llama_cpp.llama_load_model_from_file( - self.path_model.encode("utf-8"), self.params - ) + with suppress_stdout_stderr(disable=verbose): + self.model = llama_cpp.llama_load_model_from_file( + self.path_model.encode("utf-8"), self.params + ) if self.model is None: raise ValueError(f"Failed to load model from file: {path_model}") diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index 4a106470b..781b26501 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -1,13 +1,15 @@ import os import sys -import sys from typing import Any, Dict # Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor outnull_file = open(os.devnull, "w") errnull_file = open(os.devnull, "w") +STDOUT_FILENO = 1 +STDERR_FILENO = 2 + class suppress_stdout_stderr(object): # NOTE: these must be "saved" here to avoid exceptions when using # this context manager inside of a __del__ method @@ -22,12 +24,8 @@ def __enter__(self): if self.disable: return self - # Check if sys.stdout and sys.stderr have fileno method - if not hasattr(self.sys.stdout, 'fileno') or not hasattr(self.sys.stderr, 'fileno'): - return self # Return the instance without making changes - - self.old_stdout_fileno_undup = self.sys.stdout.fileno() - self.old_stderr_fileno_undup = self.sys.stderr.fileno() + self.old_stdout_fileno_undup = STDOUT_FILENO + self.old_stderr_fileno_undup = STDERR_FILENO self.old_stdout_fileno = self.os.dup(self.old_stdout_fileno_undup) self.old_stderr_fileno = self.os.dup(self.old_stderr_fileno_undup) @@ -47,15 +45,14 @@ def __exit__(self, *_): return # Check if sys.stdout and sys.stderr have fileno method - if hasattr(self.sys.stdout, 'fileno') and hasattr(self.sys.stderr, 'fileno'): - self.sys.stdout = self.old_stdout - self.sys.stderr = self.old_stderr + self.sys.stdout = self.old_stdout + self.sys.stderr = self.old_stderr - self.os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) - self.os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) + self.os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) + self.os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) - self.os.close(self.old_stdout_fileno) - self.os.close(self.old_stderr_fileno) + self.os.close(self.old_stdout_fileno) + self.os.close(self.old_stderr_fileno) class MetaSingleton(type): From 946156fb6c09389cc184cb5209f96aac54c08b85 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 15:46:45 -0400 Subject: [PATCH 350/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 77e15bec6..f364eb6fb 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 77e15bec6217a39be59b9cc83d6b9afb6b0d8167 +Subproject commit f364eb6fb5d46118a76fa045f487318de4c24961 From 4f01c452b6c738dc56eacac3758119b12c57ea94 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Apr 2024 15:50:30 -0400 Subject: [PATCH 351/624] fix: Change default verbose value of verbose in image chat format handlers to True to match Llama --- llama_cpp/llama_chat_format.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 63eaf8a03..0af410a97 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2205,7 +2205,7 @@ class Llava15ChatHandler: "{% endif %}" ) - def __init__(self, clip_model_path: str, verbose: bool = False): + def __init__(self, clip_model_path: str, verbose: bool = True): import llama_cpp.llava_cpp as llava_cpp self.clip_model_path = clip_model_path From 31b1d95a6c19f5b615a3286069f181a415f872e8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 2 May 2024 11:32:18 -0400 Subject: [PATCH 352/624] feat: Add llama-3-vision-alpha chat format --- llama_cpp/llama_chat_format.py | 64 ++++++++++++++++++++++++++++++++-- llama_cpp/server/model.py | 14 ++++++++ 2 files changed, 76 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 0af410a97..a17b86b30 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2165,7 +2165,7 @@ def create_completion(stop): class Llava15ChatHandler: - DEFAULT_SYSTEM_MESSAGE = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." + DEFAULT_SYSTEM_MESSAGE: Optional[str] = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." CHAT_FORMAT = ( "{% for message in messages %}" @@ -2288,7 +2288,7 @@ def __call__( assert self.clip_ctx is not None system_prompt = _get_system_message(messages) - if system_prompt == "": + if system_prompt == "" and self.DEFAULT_SYSTEM_MESSAGE is not None: messages = [llama_types.ChatCompletionRequestSystemMessage(role="system", content=self.DEFAULT_SYSTEM_MESSAGE)] + messages image_urls = self.get_image_urls(messages) @@ -2771,6 +2771,66 @@ class NanoLlavaChatHandler(Llava15ChatHandler): "{% endif %}" ) +class Llama3VisionAlpha(Llava15ChatHandler): + # question = "" + q + + # prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{question}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" + DEFAULT_SYSTEM_MESSAGE = None + + CHAT_FORMAT = ( + "{% for message in messages %}" + + "<|start_header_id|>" + + "{% if message.role == 'user' %}" + + "user<|end_header_id|>\n\n" + + "{% if message.content is iterable %}" + + # + "{% for content in message.content %}" + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + + # Question: + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + + "{% endif %}" + + # Question: + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + + "{% endif %}" + + # Answer: + "{% if message.role == 'assistant' %}" + "assistant<|end_header_id|>\n\n" + "{{ message.content }}" + "{% endif %}" + + "<|eot_id|>" + + "{% endfor %}" + + # Generation prompt + "{% if add_generation_prompt %}" + "<|start_header_id|>assistant<|end_header_id|>\n\n" + "{% endif %}" + ) @register_chat_completion_handler("chatml-function-calling") def chatml_function_calling( diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 1ad592d21..e102fadbd 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -140,6 +140,20 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler = llama_cpp.llama_chat_format.NanoLlavaChatHandler( clip_model_path=settings.clip_model_path, verbose=settings.verbose ) + elif settings.chat_format == "llama-3-vision-alpha": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.Llama3VisionAlpha.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.Llama3VisionAlpha( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) elif settings.chat_format == "hf-autotokenizer": assert ( settings.hf_pretrained_model_name_or_path is not None From d75dea18db61becae72784b9d750897b7cdcda26 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 2 May 2024 12:00:44 -0400 Subject: [PATCH 353/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f364eb6fb..6ecf3189e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f364eb6fb5d46118a76fa045f487318de4c24961 +Subproject commit 6ecf3189e00a1e8e737a78b6d10e1d7006e050a2 From 21171223967896b9e941007841ebd1513b3cc4b9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 2 May 2024 12:07:09 -0400 Subject: [PATCH 354/624] chore: Bump version --- CHANGELOG.md | 10 +++++++++- llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 2 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index c9681f355..9b995509e 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,9 +7,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.69] + +- feat: Update llama.cpp to ggerganov/llama.cpp@6ecf3189e00a1e8e737a78b6d10e1d7006e050a2 +- feat: Add llama-3-vision-alpha chat format by @abetlen in 31b1d95a6c19f5b615a3286069f181a415f872e8 +- fix: Change default verbose value of verbose in image chat format handlers to True to match Llama by @abetlen in 4f01c452b6c738dc56eacac3758119b12c57ea94 +- fix: Suppress all logs when verbose=False, use hardcoded fileno's to work in colab notebooks by @abetlen in f116175a5a7c84569c88cad231855c1e6e59ff6e +- fix: UTF-8 handling with grammars by @jsoma in #1415 + ## [0.2.68] -- feat: Update llama.cpp to ggerganov/llama.cpp@ +- feat: Update llama.cpp to ggerganov/llama.cpp@77e15bec6217a39be59b9cc83d6b9afb6b0d8167 - feat: Add option to enable flash_attn to Lllama params and ModelSettings by @abetlen in 22d77eefd2edaf0148f53374d0cac74d0e25d06e - fix(ci): Fix build-and-release.yaml by @Smartappli in #1413 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 63c622551..95c881966 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.68" \ No newline at end of file +__version__ = "0.2.69" \ No newline at end of file From 2138561fab5e60672c63b6c446b62a8bc26e17c4 Mon Sep 17 00:00:00 2001 From: Daniel Thuerck Date: Fri, 3 May 2024 18:17:07 +0200 Subject: [PATCH 355/624] fix(server): Propagate `flash_attn` to model load. (#1424) --- llama_cpp/server/model.py | 1 + 1 file changed, 1 insertion(+) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index e102fadbd..f00292410 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -242,6 +242,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: logits_all=settings.logits_all, embedding=settings.embedding, offload_kqv=settings.offload_kqv, + flash_attn=settings.flash_attn, # Sampling Params last_n_tokens_size=settings.last_n_tokens_size, # LoRA Params From 0a454bebe67d12a446981eb16028c168ca5faa81 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 3 May 2024 15:23:06 -0400 Subject: [PATCH 356/624] feat(server): Remove temperature bounds checks for server. Closes #1384 --- llama_cpp/server/types.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index ce9c87a69..a20b3940f 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -18,8 +18,6 @@ temperature_field = Field( default=0.8, - ge=0.0, - le=2.0, description="Adjust the randomness of the generated text.\n\n" + "Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.", ) From 9f7a85571ae80d3b6ddbd3e1bae407b9f1e3448a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 3 May 2024 19:07:50 -0400 Subject: [PATCH 357/624] fix: Use memmove to copy str_value kv_override. Closes #1417 --- llama_cpp/llama.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index f927f0ca2..17576a69b 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -262,7 +262,12 @@ def __init__( raise ValueError(f"Value for {k} is too long: {v}") v_bytes = v_bytes.ljust(128, b"\0") self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR - self._kv_overrides_array[i].value.str_value[:128] = v_bytes + # copy min(v_bytes, 128) to str_value + ctypes.memmove( + self._kv_overrides_array[i].value.str_value, + v_bytes, + min(len(v_bytes), 128), + ) else: raise ValueError(f"Unknown value type for {k}: {v}") From 1f56c648c3a4aed2b4fe3edb4dc745fae3e7d8ae Mon Sep 17 00:00:00 2001 From: Jeffrey Fong Date: Sat, 4 May 2024 22:11:20 +0800 Subject: [PATCH 358/624] feat: Implement streaming for Functionary v2 + Bug fixes (#1419) * set up streaming for v2 * assert v2 streaming, fix tool_call vs function_call * fix streaming with tool_choice/function_call * make functions return 1 function call only when 'auto' * fix --------- Co-authored-by: Andrei --- llama_cpp/llama_chat_format.py | 576 +++++++++++++++++++++++++-------- 1 file changed, 443 insertions(+), 133 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index a17b86b30..3ab94e0d3 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1894,6 +1894,8 @@ def prepare_messages_for_inference( function_call = ( tool_choice if isinstance(tool_choice, str) else tool_choice["function"] ) + elif function_call is not None: + pass else: function_call = "auto" @@ -1930,11 +1932,10 @@ def prepare_messages_for_inference( logits_processor=logits_processor, grammar=grammar, ) - completion_or_completion_chunks["choices"][0]["text"] = completion_or_completion_chunks["choices"][0]["text"].lstrip() + if stream is False: + completion_or_completion_chunks["choices"][0]["text"] = completion_or_completion_chunks["choices"][0]["text"].lstrip() return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore - assert stream is False # TODO: support stream mode - def get_grammar(function_call): function_body = None for function in functions or []: @@ -1968,7 +1969,7 @@ def get_grammar(function_call): return grammar - def create_completion(stop): + def create_completion(prompt, stop, grammar): completion = cast(llama_types.Completion, llama.create_completion( prompt=prompt, temperature=temperature, @@ -1976,7 +1977,7 @@ def create_completion(stop): top_k=top_k, min_p=min_p, typical_p=typical_p, - stream=False, + stream=stream, stop=stop, max_tokens=max_tokens, presence_penalty=presence_penalty, @@ -1996,172 +1997,481 @@ def create_completion(stop): content = "" function_calls, function_bodies = [], [] completion_tokens = 0 - - if version == "v1": - # If no or "auto" tool_choice/function_call - if isinstance(function_call, str) and function_call == "auto": - stops = ["\n", END_ASSISTANT_TOKEN] - # If tool_choice/function_call is provided - elif isinstance(function_call, dict): - prompt += f"{START_FUNCTION_CALL_TOKEN}{function_call['name']}:\n" - stops = END_FUNCTION_CALL_TOKEN - function_call = function_call["name"] - function_calls.append(function_call) - grammar = get_grammar(function_call) - else: - prompt = prompt - stops = ["\n", END_ASSISTANT_TOKEN] - - completion = create_completion(stop=stops) - completion_text = completion["choices"][0]["text"] - completion_tokens += completion["usage"]["completion_tokens"] + + def generate_streaming(tools, functions, function_call, prompt): + assert version == "v2", "Streaming for v1 is not supported" + + chunk_id, chunk_created = None, None - - # If the generation does not involve a function call - if ( - START_FUNCTION_CALL_TOKEN not in prompt - and START_FUNCTION_CALL_TOKEN not in completion_text - ): - completion["usage"]["completion_tokens"] = completion_tokens - return _convert_completion_to_chat(completion, stream=stream) # type: ignore - # If the generation involves a function call in completion, generate the parameters - elif ( - START_FUNCTION_CALL_TOKEN not in prompt - and START_FUNCTION_CALL_TOKEN in completion_text - ): - prompt += ( - completion_text.replace( - f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN - ) - + "\n" - ) - function_calls.append( - completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() - ) - grammar = get_grammar(function_calls[-1]) - completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) - completion_tokens += completion["usage"]["completion_tokens"] - function_bodies.append(completion["choices"][0]["text"].strip()) - # If the prompt involves a function call, just append generated parameters to function_bodies - else: - function_bodies.append(completion_text.strip()) - else: # If tool_choice/function_call is provided if isinstance(function_call, dict): prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" - function_call = function_call["name"] - function_calls.append(function_call) - grammar = get_grammar(function_call) + grammar = get_grammar(function_call["name"]) stops = [STOP_TOKEN, FROM_TOKEN] - completion = create_completion(stop=stops) - completion_text = completion["choices"][0]["text"] - completion_tokens += completion["usage"]["completion_tokens"] - function_bodies.append(completion_text.strip()) + tool_id = "".join([random.choice(string.ascii_letters + string.digits) for _ in range(24)]) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion_text = "" + first = True + for chunk in completion: + # Yield the tool/function name first + if first: + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": 0, + "id": "call_" + tool_id, + "type": "function", + "function": {"name": function_call["name"], "arguments": ""}, + } + ] + } + else: + func_call_dict = {"function_call": {"name": function_call["name"], "arguments": ""}} + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk["id"], + object="chat.completion.chunk", + created=chunk["created"], + model=chunk["model"], + choices=[ + {"index": 0, "logprobs": None, "delta": {"role": None, "content": None, **func_call_dict}} + ], + ) + first = False + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": 0, + "id": "call_" + tool_id, + "type": "function", + "function": { + "name": None, + "arguments": chunk["choices"][0]["text"].rstrip(), + }, + } + ] + } + else: + func_call_dict = {"function_call": {"name": None, "arguments": chunk["choices"][0]["text"].rstrip()}} + if len(chunk["choices"][0]["text"].rstrip()) > 0: + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk["id"], + object="chat.completion.chunk", + created=chunk["created"], + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": chunk["choices"][0]["logprobs"], + "delta": { + "role": None, + "content": None, + **func_call_dict, + }, + } + ], + ) + # Yield tool_call/function_call stop message + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk["id"], + object="chat.completion.chunk", + created=chunk["created"], + model=chunk["model"], + choices=[ + { + "index": 0, + "finish_reason": "tool_calls" if tools is not None else "function_call", + "logprobs": None, + "delta": { + "role": None, "content": None, "function_call": None, "tool_calls": None + }, + } + ], + ) # If "auto" or no tool_choice/function_call elif isinstance(function_call, str) and function_call == "auto": + tool_index = 0 while True: # Generate function name first grammar = None stops = CONTENT_TOKEN - completion = create_completion(stop=stops) - completion_text = completion["choices"][0]["text"] - completion_tokens += completion["usage"]["completion_tokens"] + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion_text = "" + for chunk in completion: + completion_text += chunk["choices"][0]["text"] + if chunk_id is None: + chunk_id = chunk["id"] + if chunk_created is None: + chunk_created = chunk["created"] function_name = completion_text.strip() if function_name == "all": prompt += "all\n<|content|>" + # Yield the first empty message for content + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + model=chunk["model"], + created=chunk_created, + object="chat.completion.chunk", + choices=[ + { + "index": 0, + "delta": {"role": "assistant", "content": ""}, + "logprobs": None, + "finish_reason": None, + } + ], + ) else: - function_call = completion_text.strip() - prompt += f"{function_call}\n<|content|>" - function_calls.append(function_call) - grammar = get_grammar(function_call) + prompt += f"{function_name}\n<|content|>" + grammar = get_grammar(function_name) + tool_id = "".join([random.choice(string.ascii_letters + string.digits) for _ in range(24)]) + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": tool_index, + "id": "call_" + tool_id, + "type": "function", + "function": {"name": function_name, "arguments": ""}, + } + ] + } + else: + func_call_dict = {"function_call": {"name": function_name, "arguments": ""}} + # Stream function name + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": chunk["choices"][0]["logprobs"], + "delta": { + "role": "assistant", + "content": None, + **func_call_dict, + }, + } + ], + ) # Generate content stops = [RECIPIENT_TOKEN, STOP_TOKEN] - completion = create_completion(stop=stops) - completion_text = completion["choices"][0]["text"] - completion_tokens += completion["usage"]["completion_tokens"] + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) if function_name == "all": - if completion_text.endswith("\n<|from|>assistant\n"): - content += completion_text[:-len("\n<|from|>assistant\n")] - if completion_text.endswith("\n<|from|> assistant\n"): - content += completion_text[-len("\n<|from|> assistant\n")] - else: - content += completion_text - content = content.lstrip() + completion_text = "" + stop_sequence, buffer, is_end = "\n<|from|>assistant\n<|recipient|>", [], False + for i, chunk in enumerate(completion): + completion_text += chunk["choices"][0]["text"] + if is_end: + buffer.append(chunk["choices"][0]["text"].strip(" ")) + if stop_sequence.startswith("".join(buffer)): + continue + else: + buffer.pop() + while len(buffer) > 0: + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": chunk["choices"][0]["logprobs"], + "delta": { + "role": "assistant", "content": buffer.pop(0) + }, + } + ], + ) + is_end = False + elif chunk["choices"][0]["text"] == "\n": + is_end = True + buffer.append(chunk["choices"][0]["text"].strip(" ")) + continue + + if len(buffer) == 0 and len(chunk["choices"][0]["text"]) > 0: + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": chunk["choices"][0]["logprobs"], + "delta": { + "role": "assistant", + "content": chunk["choices"][0]["text"] if i > 0 else chunk["choices"][0]["text"].lstrip() + }, + } + ], + ) # Check whether the model wants to generate another turn if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text: if completion_text.endswith("\n<|from|>assistant\n"): cleaned_completion_text = completion_text[:-len("\n<|from|>assistant\n")].strip() elif completion_text.endswith("\n<|from|> assistant\n"): - cleaned_completion_text = completion_text[-len("\n<|from|> assistant\n")].strip() + cleaned_completion_text = completion_text[:-len("\n<|from|> assistant\n")].strip() else: cleaned_completion_text = completion_text.strip() prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" else: + # Yield stop message + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + model=chunk["model"], + created=chunk_created, + object="chat.completion.chunk", + choices=[ + { + "index": 0, + "delta": {}, + "logprobs": None, + "finish_reason": "stop", + } + ], + ) break else: - function_bodies.append(completion_text.strip()) # Check whether the model wants to generate another turn + completion_text = "" + for chunk in completion: + completion_text += chunk["choices"][0]["text"] + if len(chunk["choices"][0]["text"].rstrip()) > 0: + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": tool_index, + "id": "call_" + tool_id, + "type": "function", + "function": { + "name": None, + "arguments": chunk["choices"][0]["text"].rstrip(), + }, + } + ] + } + else: + func_call_dict = {"function_call": {"name": None, "arguments": chunk["choices"][0]["text"].rstrip()}} + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": chunk["choices"][0]["logprobs"], + "delta": { + "role": None, + "content": None, + **func_call_dict, + }, + } + ], + ) prompt += completion_text.strip() grammar = None - completion = create_completion(stop=stops) - completion_tokens += completion["usage"]["completion_tokens"] - if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]: + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion_text += "".join([chunk["choices"][0]["text"] for chunk in completion]) + if ("<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text) and tools is not None: prompt += "\n<|from|>assistant\n<|recipient|>" + tool_index += 1 else: + # Yield tool_call/function_call stop message + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "finish_reason": "tool_calls" if tools is not None else "function_call", + "logprobs": None, + "delta": { + "role": None, "content": None, "function_call": None, "tool_calls": None + }, + } + ], + ) break - - assert "usage" in completion - assert len(function_calls) == len(function_bodies) - - tool_calls: List[llama_types.ChatCompletionMessageToolCall] = [] - for function_call, function_body in zip(function_calls, function_bodies): - tool_calls.append( - { - "id": "call_" - + "".join( - [ - random.choice(string.ascii_letters + string.digits) - for _ in range(24) - ] - ), - "type": "function", - "function": { - "name": function_call, - "arguments": function_body, - }, - } + + if stream is not False: + return generate_streaming( + tools=tools, functions=functions, function_call=function_call, prompt=prompt ) + else: + if version == "v1": + # If no or "auto" tool_choice/function_call + if isinstance(function_call, str) and function_call == "auto": + stops = ["\n", END_ASSISTANT_TOKEN] + # If tool_choice/function_call is provided + elif isinstance(function_call, dict): + prompt += f"{START_FUNCTION_CALL_TOKEN}{function_call['name']}:\n" + stops = END_FUNCTION_CALL_TOKEN + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + else: + prompt = prompt + stops = ["\n", END_ASSISTANT_TOKEN] - # TODO: support stream mode - function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = {} - if len(tool_calls) > 0: - if tools is not None: - function_call_dict["tool_calls"] = tool_calls + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + + + # If the generation does not involve a function call + if ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN not in completion_text + ): + completion["usage"]["completion_tokens"] = completion_tokens + return _convert_completion_to_chat(completion, stream=stream) # type: ignore + # If the generation involves a function call in completion, generate the parameters + elif ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN in completion_text + ): + prompt += ( + completion_text.replace( + f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN + ) + + "\n" + ) + function_calls.append( + completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() + ) + grammar = get_grammar(function_calls[-1]) + completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) + completion_tokens += completion["usage"]["completion_tokens"] + function_bodies.append(completion["choices"][0]["text"].strip()) + # If the prompt involves a function call, just append generated parameters to function_bodies + else: + function_bodies.append(completion_text.strip()) else: - function_call_dict["function_call"] = { - "name": tool_calls[0]["function"]["name"], - "arguments": tool_calls[0]["function"]["arguments"], - } - completion["usage"]["completion_tokens"] = completion_tokens - return llama_types.CreateChatCompletionResponse( - id="chat" + completion["id"], - object="chat.completion", - created=completion["created"], - model=completion["model"], - choices=[ - { - "index": 0, - "logprobs": completion["choices"][0]["logprobs"], - "message": { - "role": "assistant", - "content": None if content == "" else content, - **function_call_dict, - }, - "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", - } - ], - usage=completion["usage"], - ) + # If tool_choice/function_call is provided + if isinstance(function_call, dict): + prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + stops = [STOP_TOKEN, FROM_TOKEN] + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + function_bodies.append(completion_text.strip()) + # If "auto" or no tool_choice/function_call + elif isinstance(function_call, str) and function_call == "auto": + while True: + # Generate function name first + grammar = None + stops = CONTENT_TOKEN + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + function_name = completion_text.strip() + if function_name == "all": + prompt += "all\n<|content|>" + else: + function_call = completion_text.strip() + prompt += f"{function_call}\n<|content|>" + function_calls.append(function_call) + grammar = get_grammar(function_call) + # Generate content + stops = [RECIPIENT_TOKEN, STOP_TOKEN] + completion = create_completion(stop=stops) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + if function_name == "all": + if completion_text.endswith("\n<|from|>assistant\n"): + content += completion_text[:-len("\n<|from|>assistant\n")] + if completion_text.endswith("\n<|from|> assistant\n"): + content += completion_text[-len("\n<|from|> assistant\n")] + else: + content += completion_text + content = content.lstrip() + # Check whether the model wants to generate another turn + if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text: + if completion_text.endswith("\n<|from|>assistant\n"): + cleaned_completion_text = completion_text[:-len("\n<|from|>assistant\n")].strip() + elif completion_text.endswith("\n<|from|> assistant\n"): + cleaned_completion_text = completion_text[-len("\n<|from|> assistant\n")].strip() + else: + cleaned_completion_text = completion_text.strip() + prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" + else: + break + else: + function_bodies.append(completion_text.strip()) + # Check whether the model wants to generate another turn + prompt += completion_text.strip() + grammar = None + completion = create_completion(stop=stops) + completion_tokens += completion["usage"]["completion_tokens"] + if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]: + prompt += "\n<|from|>assistant\n<|recipient|>" + else: + break + + assert "usage" in completion + assert len(function_calls) == len(function_bodies) + + tool_calls: List[llama_types.ChatCompletionMessageToolCall] = [] + for function_call, function_body in zip(function_calls, function_bodies): + tool_calls.append( + { + "id": "call_" + + "".join( + [ + random.choice(string.ascii_letters + string.digits) + for _ in range(24) + ] + ), + "type": "function", + "function": { + "name": function_call, + "arguments": function_body, + }, + } + ) + + # TODO: support stream mode + function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = {} + if len(tool_calls) > 0: + if tools is not None: + function_call_dict["tool_calls"] = tool_calls + else: + function_call_dict["function_call"] = { + "name": tool_calls[0]["function"]["name"], + "arguments": tool_calls[0]["function"]["arguments"], + } + completion["usage"]["completion_tokens"] = completion_tokens + return llama_types.CreateChatCompletionResponse( + id="chat" + completion["id"], + object="chat.completion", + created=completion["created"], + model=completion["model"], + choices=[ + { + "index": 0, + "logprobs": completion["choices"][0]["logprobs"], + "message": { + "role": "assistant", + "content": None if content == "" else content, + **function_call_dict, + }, + "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", + } + ], + usage=completion["usage"], + ) class Llava15ChatHandler: From e0d7674e62bdc5b906d2461238993ea3a022f61f Mon Sep 17 00:00:00 2001 From: Noam Gat Date: Sat, 4 May 2024 17:14:59 +0300 Subject: [PATCH 359/624] fix: detokenization case where first token does not start with a leading space (#1375) * Fix tokenization edge case where llama output does not start with a space See this notebook: https://colab.research.google.com/drive/1Ooz11nFPk19zyJdMDx42CeesU8aWZMdI#scrollTo=oKpHw5PZ30uC * Update _internals.py Fixing to compare to b' ' instead of (str)' ' --------- Co-authored-by: Andrei --- llama_cpp/_internals.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index d7409f63a..b404601d3 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -203,7 +203,7 @@ def detokenize(self, tokens: List[int], special: bool = False) -> bytes: # NOTE: Llama1 models automatically added a space at the start of the prompt # this line removes a leading space if the first token is a beginning of sentence token return ( - output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output + output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() and output[0:1] == b' ' else output ) # Extra @@ -812,4 +812,4 @@ def sample( def accept(self, ctx_main: _LlamaContext, id: int, apply_grammar: bool): if apply_grammar and self.grammar is not None: ctx_main.grammar_accept_token(self.grammar, id) - self.prev.append(id) \ No newline at end of file + self.prev.append(id) From 3e2597eac888e6e12c7bc7021016ca5104db83ba Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 5 May 2024 12:12:27 -0400 Subject: [PATCH 360/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 6 +++++- vendor/llama.cpp | 2 +- 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 46aa51662..9e934e052 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -294,6 +294,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_MPT = 5, # LLAMA_VOCAB_PRE_TYPE_STARCODER = 6, # LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, +# LLAMA_VOCAB_PRE_TYPE_REFACT = 8, +# LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -303,6 +305,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_MPT = 5 LLAMA_VOCAB_PRE_TYPE_STARCODER = 6 LLAMA_VOCAB_PRE_TYPE_GPT2 = 7 +LLAMA_VOCAB_PRE_TYPE_REFACT = 8 +LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9 # // note: these values should be synchronized with ggml_rope @@ -494,7 +498,7 @@ class llama_token_data_array(ctypes.Structure): llama_token_data_array_p = ctypes.POINTER(llama_token_data_array) -# typedef bool (*llama_progress_callback)(float progress, void *ctx); +# typedef bool (*llama_progress_callback)(float progress, void * user_data); llama_progress_callback = ctypes.CFUNCTYPE( ctypes.c_bool, ctypes.c_float, ctypes.c_void_p ) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6ecf3189e..628b29910 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6ecf3189e00a1e8e737a78b6d10e1d7006e050a2 +Subproject commit 628b299106d1e9476fdecb3cbe546bf5c60f1b89 From 36668331074c79f40494b51e4d541cd03601b9f2 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Sun, 5 May 2024 18:42:28 +0200 Subject: [PATCH 361/624] feat(ci): Add docker checks and check deps more frequently (#1426) * Update dependabot.yml Add github-actions update * Update dependabot.yml * Update dependabot.yml --- .github/dependabot.yml | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/.github/dependabot.yml b/.github/dependabot.yml index c58c9ae57..6bf90273a 100644 --- a/.github/dependabot.yml +++ b/.github/dependabot.yml @@ -8,8 +8,12 @@ updates: - package-ecosystem: "pip" # See documentation for possible values directory: "/" # Location of package manifests schedule: - interval: "weekly" + interval: "daily" - package-ecosystem: "github-actions" directory: "/" schedule: - interval: "weekly" + interval: "daily" + - package-ecosystem: "docker" + directory: "/" + schedule: + interval: "daily" From 0318702cdc860999ee70f277425edbbfe0e60419 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 5 May 2024 12:49:31 -0400 Subject: [PATCH 362/624] feat(server): Add support for setting root_path. Closes #1420 --- llama_cpp/server/app.py | 1 + llama_cpp/server/settings.py | 4 ++++ 2 files changed, 5 insertions(+) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index b6ed9b1b6..4cf10d1f6 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -132,6 +132,7 @@ def create_app( middleware=middleware, title="🦙 llama.cpp Python API", version=llama_cpp.__version__, + root_path=server_settings.root_path, ) app.add_middleware( CORSMiddleware, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index ed05a889f..a3e185007 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -215,6 +215,10 @@ class ServerSettings(BaseSettings): default=False, description="Disable EventSource pings (may be needed for some clients).", ) + root_path: str = Field( + default="", + description="The root path for the server. Useful when running behind a reverse proxy.", + ) class Settings(ServerSettings, ModelSettings): From a50d24e3a7ce92d5a5256738afd3a4450d1d85e7 Mon Sep 17 00:00:00 2001 From: Bruno Alvisio Date: Tue, 7 May 2024 23:19:35 -0700 Subject: [PATCH 363/624] fix: chat_format log where auto-detected format prints `None` (#1434) --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 17576a69b..a8c3f9a98 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -451,7 +451,7 @@ def __init__( if self.chat_format is None and self.chat_handler is None: self.chat_format = "llama-2" if self.verbose: - print(f"Using fallback chat format: {chat_format}", file=sys.stderr) + print(f"Using fallback chat format: {self.chat_format}", file=sys.stderr) @property def ctx(self) -> llama_cpp.llama_context_p: From 07966b9ba7afa8a7c0a04a54fc07b8984b131385 Mon Sep 17 00:00:00 2001 From: Ikko Eltociear Ashimine Date: Wed, 8 May 2024 15:20:20 +0900 Subject: [PATCH 364/624] docs: update README.md (#1432) accomodate -> accommodate --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 640671460..5f3ff6824 100644 --- a/README.md +++ b/README.md @@ -516,7 +516,7 @@ chat_handler = Llava15ChatHandler(clip_model_path="path/to/llava/mmproj.bin") llm = Llama( model_path="./path/to/llava/llama-model.gguf", chat_handler=chat_handler, - n_ctx=2048, # n_ctx should be increased to accomodate the image embedding + n_ctx=2048, # n_ctx should be increased to accommodate the image embedding ) llm.create_chat_completion( messages = [ @@ -547,7 +547,7 @@ llm = Llama.from_pretrained( repo_id="vikhyatk/moondream2", filename="*text-model*", chat_handler=chat_handler, - n_ctx=2048, # n_ctx should be increased to accomodate the image embedding + n_ctx=2048, # n_ctx should be increased to accommodate the image embedding ) respoonse = llm.create_chat_completion( From 903b28adf5846b7a76999993c218adef4946c0fe Mon Sep 17 00:00:00 2001 From: Sarunas Kalade Date: Wed, 8 May 2024 07:21:27 +0100 Subject: [PATCH 365/624] fix: adding missing args in create_completion for functionary chat handler (#1430) --- llama_cpp/llama_chat_format.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 3ab94e0d3..9da6b9800 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2322,7 +2322,7 @@ def generate_streaming(tools, functions, function_call, prompt): prompt = prompt stops = ["\n", END_ASSISTANT_TOKEN] - completion = create_completion(stop=stops) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] @@ -2349,7 +2349,7 @@ def generate_streaming(tools, functions, function_call, prompt): completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() ) grammar = get_grammar(function_calls[-1]) - completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) + completion = create_completion(prompt=prompt, stop=END_FUNCTION_CALL_TOKEN, grammar=grammar) completion_tokens += completion["usage"]["completion_tokens"] function_bodies.append(completion["choices"][0]["text"].strip()) # If the prompt involves a function call, just append generated parameters to function_bodies @@ -2363,7 +2363,7 @@ def generate_streaming(tools, functions, function_call, prompt): function_calls.append(function_call) grammar = get_grammar(function_call) stops = [STOP_TOKEN, FROM_TOKEN] - completion = create_completion(stop=stops) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] function_bodies.append(completion_text.strip()) @@ -2373,7 +2373,7 @@ def generate_streaming(tools, functions, function_call, prompt): # Generate function name first grammar = None stops = CONTENT_TOKEN - completion = create_completion(stop=stops) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] function_name = completion_text.strip() @@ -2386,7 +2386,7 @@ def generate_streaming(tools, functions, function_call, prompt): grammar = get_grammar(function_call) # Generate content stops = [RECIPIENT_TOKEN, STOP_TOKEN] - completion = create_completion(stop=stops) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] if function_name == "all": @@ -2413,7 +2413,7 @@ def generate_streaming(tools, functions, function_call, prompt): # Check whether the model wants to generate another turn prompt += completion_text.strip() grammar = None - completion = create_completion(stop=stops) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_tokens += completion["usage"]["completion_tokens"] if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]: prompt += "\n<|from|>assistant\n<|recipient|>" @@ -3564,4 +3564,4 @@ def chatml_function_calling( }, } - raise ValueError("Automatic streaming tool choice is not supported") \ No newline at end of file + raise ValueError("Automatic streaming tool choice is not supported") From 228949c1f7e1f88c2e24393f45fa4156df7dc0cc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 May 2024 02:22:15 -0400 Subject: [PATCH 366/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 9e934e052..d8075e457 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -296,6 +296,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, # LLAMA_VOCAB_PRE_TYPE_REFACT = 8, # LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, +# LLAMA_VOCAB_PRE_TYPE_OLMO = 10, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -307,6 +308,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_GPT2 = 7 LLAMA_VOCAB_PRE_TYPE_REFACT = 8 LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9 +LLAMA_VOCAB_PRE_TYPE_OLMO = 10 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 628b29910..c0e6fbf8c 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 628b299106d1e9476fdecb3cbe546bf5c60f1b89 +Subproject commit c0e6fbf8c380718102bd25fcb8d2e55f8f9480d1 From 4a7122d22f7a49bc0d055c1dbb892d489812ba10 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Wed, 8 May 2024 08:26:22 +0200 Subject: [PATCH 367/624] feat: fill-in-middle support (#1386) * Proper fill-in-middle support Use prefix/middle/suffix tokens when metadata is present in GGUF, like f.ex. in [this](https://huggingface.co/CISCai/CodeQwen1.5-7B-Chat-SOTA-GGUF) one. * fall back to internal prefix/middle/suffix id In some cases llama.cpp will make a guess at fim tokens, use them if there's no metadata. * typo-- * don't insert special tokens that are not there in suffix Note: add_bos is misnamed, it's actually add_special and can cause several special tokens to be added to the token list (the special parameter is actually parse_special). * don't add/parse any special tokens when using fim I've left original behavior when no fim tokens are found, but this should perhaps be re-evaluated. * don't append suffix to prompt_tokens unless fim tokens are detected * make sure we only do this for fim --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 47 ++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 41 insertions(+), 6 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index a8c3f9a98..5acc112d1 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -955,18 +955,53 @@ def _create_completion( completion_id: str = f"cmpl-{str(uuid.uuid4())}" created: int = int(time.time()) + prefix_token_id: int = int(self.metadata.get("tokenizer.ggml.prefix_token_id", self._model.token_prefix())) + middle_token_id: int = int(self.metadata.get("tokenizer.ggml.middle_token_id", self._model.token_middle())) + suffix_token_id: int = int(self.metadata.get("tokenizer.ggml.suffix_token_id", self._model.token_suffix())) # If prompt is empty, initialize completion with BOS token to avoid # detokenization including a space at the beginning of the completion completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()] # Add blank space to start of prompt to match OG llama tokenizer prompt_tokens: List[int] = ( ( - self.tokenize(prompt.encode("utf-8"), special=True) - if prompt != "" - else [self.token_bos()] + [prefix_token_id] + if prefix_token_id >= 0 and suffix is not None + else [] + ) + + + ( + ( + self.tokenize(prompt.encode("utf-8"), add_bos=(prefix_token_id < 0 or suffix is None), special=(prefix_token_id < 0 or suffix is None)) + if prompt != "" + else ( + [] + if prefix_token_id >= 0 and suffix is not None + else [self.token_bos()] + ) + ) + if isinstance(prompt, str) + else prompt + ) + + + ( + ( + [suffix_token_id] + + + ( + self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False) + if suffix + else [] + ) + ) + if suffix_token_id >= 0 and suffix is not None + else [] + ) + + + ( + [middle_token_id] + if middle_token_id >= 0 and suffix is not None + else [] ) - if isinstance(prompt, str) - else prompt ) text: bytes = b"" returned_tokens: int = 0 @@ -1346,7 +1381,7 @@ def logit_bias_processor( if echo: text_str = prompt + text_str - if suffix is not None: + if suffix_token_id < 0 and suffix is not None: text_str = text_str + suffix logprobs_or_none: Optional[CompletionLogprobs] = None From 9ce5cb376a12a56028aec1fd3b0edc55949b996f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 May 2024 02:36:42 -0400 Subject: [PATCH 368/624] chore: Bump version --- CHANGELOG.md | 15 +++++++++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 16 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 9b995509e..911d7fbea 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.70] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ +- feat: fill-in-middle support by @CISC in #1386 +- fix: adding missing args in create_completion for functionary chat handler by @skalade in #1430 +- docs: update README.md @eltociear in #1432 +- fix: chat_format log where auto-detected format prints None by @balvisio in #1434 +- feat(server): Add support for setting root_path by @abetlen in 0318702cdc860999ee70f277425edbbfe0e60419 +- feat(ci): Add docker checks and check deps more frequently by @Smartappli in #1426 +- fix: detokenization case where first token does not start with a leading space by @noamgat in #1375 +- feat: Implement streaming for Functionary v2 + Bug fixes by @jeffrey-fong in #1419 +- fix: Use memmove to copy str_value kv_override by @abetlen in 9f7a85571ae80d3b6ddbd3e1bae407b9f1e3448a +- feat(server): Remove temperature bounds checks for server by @abetlen in 0a454bebe67d12a446981eb16028c168ca5faa81 +- fix(server): Propagate flash_attn to model load by @dthuerck in #1424 + ## [0.2.69] - feat: Update llama.cpp to ggerganov/llama.cpp@6ecf3189e00a1e8e737a78b6d10e1d7006e050a2 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 95c881966..6a36717fd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.69" \ No newline at end of file +__version__ = "0.2.70" \ No newline at end of file From 2a39b9957508c46b49caf83eba3925964545fc2f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 May 2024 08:42:22 -0400 Subject: [PATCH 369/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 11 +++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d8075e457..461122a44 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -296,7 +296,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, # LLAMA_VOCAB_PRE_TYPE_REFACT = 8, # LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, -# LLAMA_VOCAB_PRE_TYPE_OLMO = 10, +# LLAMA_VOCAB_PRE_TYPE_QWEN2 = 10, +# LLAMA_VOCAB_PRE_TYPE_OLMO = 11, +# LLAMA_VOCAB_PRE_TYPE_DBRX = 12, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -308,7 +310,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_GPT2 = 7 LLAMA_VOCAB_PRE_TYPE_REFACT = 8 LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9 -LLAMA_VOCAB_PRE_TYPE_OLMO = 10 +LLAMA_VOCAB_PRE_TYPE_QWEN2 = 10 +LLAMA_VOCAB_PRE_TYPE_OLMO = 11 +LLAMA_VOCAB_PRE_TYPE_DBRX = 12 # // note: these values should be synchronized with ggml_rope @@ -377,6 +381,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -409,6 +414,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_MOSTLY_IQ2_S = 28 LLAMA_FTYPE_MOSTLY_IQ2_M = 29 LLAMA_FTYPE_MOSTLY_IQ4_XS = 30 +LLAMA_FTYPE_MOSTLY_IQ1_M = 31 +LLAMA_FTYPE_MOSTLY_BF16 = 32 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c0e6fbf8c..911b3900d 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c0e6fbf8c380718102bd25fcb8d2e55f8f9480d1 +Subproject commit 911b3900dded9a1cfe0f0e41b82c7a29baf3a217 From 77122638b4153e31d9f277b3d905c2900b536632 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 May 2024 13:12:31 -0400 Subject: [PATCH 370/624] fix: Make leading bos_token optional for image chat formats, fix nanollava system message --- llama_cpp/llama_chat_format.py | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 9da6b9800..beff6ca9c 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2603,7 +2603,12 @@ def __call__( image_urls = self.get_image_urls(messages) template = jinja2.Template(self.CHAT_FORMAT) - text = template.render(messages=messages, add_generation_prompt=True) + text = template.render( + messages=messages, + add_generation_prompt=True, + eos_token=llama.detokenize([llama.token_eos()]), + bos_token=llama.detokenize([llama.token_bos()]), + ) split_text = self.split_text_on_image_urls(text, image_urls) def embed_image_bytes(image_bytes: bytes): @@ -2624,9 +2629,9 @@ def embed_image_bytes(image_bytes: bytes): # Evaluate prompt llama.reset() - for i, (type_, value) in enumerate(split_text): + for type_, value in split_text: if type_ == "text": - tokens = llama.tokenize(value.encode("utf8"), add_bos=i == 0) + tokens = llama.tokenize(value.encode("utf8"), add_bos=False, special=True) if llama.n_tokens + len(tokens) > llama.n_ctx(): raise ValueError("Prompt exceeds n_ctx") # TODO: Fix llama.eval(tokens) @@ -2644,6 +2649,8 @@ def embed_image_bytes(image_bytes: bytes): llama.n_batch, n_past_p, ) + # Required to avoid issues with hf tokenizer + llama.input_ids[llama.n_tokens : n_past.value] = -1 llama.n_tokens = n_past.value # Get prompt tokens to avoid a cache miss @@ -3033,6 +3040,7 @@ class NanoLlavaChatHandler(Llava15ChatHandler): # Answer the question<|im_end|><|im_start|>user # # What is the picture about?<|im_end|><|im_start|>assistant + DEFAULT_SYSTEM_MESSAGE = "Answer the question" CHAT_FORMAT = ( "{% for message in messages %}" From 3757328b703b2cd32dcbd5853271e3a8c8599fe7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 May 2024 22:16:18 -0400 Subject: [PATCH 371/624] fix: free last image embed in llava chat handler --- llama_cpp/llama_chat_format.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index beff6ca9c..84de989a5 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2615,6 +2615,11 @@ def embed_image_bytes(image_bytes: bytes): if self._last_image_embed is not None and self._last_image_hash is not None and hash(image_bytes) == self._last_image_hash: return self._last_image_embed with suppress_stdout_stderr(disable=self.verbose): + # Free the previous image embed + if self._last_image_embed is not None: + self._llava_cpp.llava_image_embed_free(self._last_image_embed) + self._last_image_embed = None + self._last_image_hash = None embed = ( self._llava_cpp.llava_image_embed_make_with_bytes( self.clip_ctx, From bf66a283e899e367c82bdf13b89179fb14ed176e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 9 May 2024 03:02:52 -0400 Subject: [PATCH 372/624] chore: Bump version --- CHANGELOG.md | 8 +++++++- llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 2 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 911d7fbea..ec70b6d6d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,9 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.71] + +- feat: Update llama.cpp to ggerganov/llama.cpp@911b3900dded9a1cfe0f0e41b82c7a29baf3a217 +- fix: Make leading bos_token optional for image chat formats, fix nanollava system message by @abetlen in 77122638b4153e31d9f277b3d905c2900b536632 +- fix: free last image embed in llava chat handler by @abetlen in 3757328b703b2cd32dcbd5853271e3a8c8599fe7 + ## [0.2.70] -- feat: Update llama.cpp to ggerganov/llama.cpp@ +- feat: Update llama.cpp to ggerganov/llama.cpp@c0e6fbf8c380718102bd25fcb8d2e55f8f9480d1 - feat: fill-in-middle support by @CISC in #1386 - fix: adding missing args in create_completion for functionary chat handler by @skalade in #1430 - docs: update README.md @eltociear in #1432 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 6a36717fd..c77d2a95a 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.70" \ No newline at end of file +__version__ = "0.2.71" \ No newline at end of file From 5ab40e6167e64ecb06c2a4d8ae798391bf9a5893 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Thu, 9 May 2024 15:49:09 +0200 Subject: [PATCH 373/624] feat: Support multiple chat templates - step 1 (#1396) * Support multiple chat templates - step 1 As a first step, allow user to to select template from metadata with chat_format parameter in the form of `chat_template.name`. * register chat templates to self.chat_formats instead of globally * Don't expose internal chat handlers yet --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 50 ++++++++++++++++++++++++++-------------------- 1 file changed, 28 insertions(+), 22 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 5acc112d1..4212669eb 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -378,6 +378,7 @@ def __init__( self.chat_format = chat_format self.chat_handler = chat_handler + self._chat_handlers: Dict[str, llama_chat_format.LlamaChatCompletionHandler] = {} self.draft_model = draft_model @@ -409,10 +410,33 @@ def __init__( if self.verbose: print(f"Model metadata: {self.metadata}", file=sys.stderr) + eos_token_id = int(self.metadata.get("tokenizer.ggml.eos_token_id", self.token_eos())) + bos_token_id = int(self.metadata.get("tokenizer.ggml.bos_token_id", self.token_bos())) + + eos_token = self._model.token_get_text(eos_token_id) + bos_token = self._model.token_get_text(bos_token_id) + + # Unfortunately the llama.cpp API does not return metadata arrays, so we can't get template names from tokenizer.chat_templates + template_choices = dict((name[10:], template) for name, template in self.metadata.items() if name.startswith("tokenizer.chat_template.")) + + if "tokenizer.chat_template" in self.metadata: + template_choices["chat_template.default"] = self.metadata["tokenizer.chat_template"] + + if self.verbose and template_choices: + print(f"Available chat formats from metadata: {', '.join(template_choices.keys())}", file=sys.stderr) + + for name, template in template_choices.items(): + self._chat_handlers[name] = llama_chat_format.Jinja2ChatFormatter( + template=template, + eos_token=eos_token, + bos_token=bos_token, + stop_token_ids=[eos_token_id], + ).to_chat_handler() + if ( self.chat_format is None and self.chat_handler is None - and "tokenizer.chat_template" in self.metadata + and "chat_template.default" in template_choices ): chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata( self.metadata @@ -423,30 +447,12 @@ def __init__( if self.verbose: print(f"Guessed chat format: {chat_format}", file=sys.stderr) else: - template = self.metadata["tokenizer.chat_template"] - try: - eos_token_id = int(self.metadata["tokenizer.ggml.eos_token_id"]) - except: - eos_token_id = self.token_eos() - try: - bos_token_id = int(self.metadata["tokenizer.ggml.bos_token_id"]) - except: - bos_token_id = self.token_bos() - - eos_token = self._model.token_get_text(eos_token_id) - bos_token = self._model.token_get_text(bos_token_id) - if self.verbose: - print(f"Using gguf chat template: {template}", file=sys.stderr) + print(f"Using gguf chat template: {template_choices['chat_template.default']}", file=sys.stderr) print(f"Using chat eos_token: {eos_token}", file=sys.stderr) print(f"Using chat bos_token: {bos_token}", file=sys.stderr) - self.chat_handler = llama_chat_format.Jinja2ChatFormatter( - template=template, - eos_token=eos_token, - bos_token=bos_token, - stop_token_ids=[eos_token_id], - ).to_chat_handler() + self.chat_format = "chat_template.default" if self.chat_format is None and self.chat_handler is None: self.chat_format = "llama-2" @@ -1719,7 +1725,7 @@ def create_chat_completion( Returns: Generated chat completion or a stream of chat completion chunks. """ - handler = self.chat_handler or llama_chat_format.get_chat_completion_handler( + handler = self.chat_handler or self._chat_handlers.get(self.chat_format) or llama_chat_format.get_chat_completion_handler( self.chat_format ) return handler( From b454f40a9a1787b2b5659cd2cb00819d983185df Mon Sep 17 00:00:00 2001 From: Patrick Peng Date: Fri, 10 May 2024 12:47:56 +0800 Subject: [PATCH 374/624] Merge pull request from GHSA-56xg-wfcc-g829 Co-authored-by: Andrei --- llama_cpp/llama_chat_format.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 84de989a5..e0c59c9f1 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -11,6 +11,7 @@ from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol, cast import jinja2 +from jinja2.sandbox import ImmutableSandboxedEnvironment import numpy as np import numpy.typing as npt @@ -191,7 +192,7 @@ def __init__( self.add_generation_prompt = add_generation_prompt self.stop_token_ids = set(stop_token_ids) if stop_token_ids is not None else None - self._environment = jinja2.Environment( + self._environment = ImmutableSandboxedEnvironment( loader=jinja2.BaseLoader(), trim_blocks=True, lstrip_blocks=True, From 561e880654aa4e1311b22577e9149962be0d8ba0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 10 May 2024 06:49:40 +0200 Subject: [PATCH 375/624] fix(security): Render all jinja templates in immutable sandbox (#1441) Chat templates are rendered with ImmutableSandboxedEnvironment in transformers so no need to do otherwise here. Co-authored-by: Andrei --- llama_cpp/llama_chat_format.py | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index e0c59c9f1..d795b3049 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -685,8 +685,7 @@ def hf_tokenizer_config_to_chat_formatter( assert isinstance(tokenizer_config["eos_token"], str) eos_token = tokenizer_config["eos_token"] - env = jinja2.Environment( - loader=jinja2.BaseLoader(), + env = ImmutableSandboxedEnvironment( trim_blocks=True, lstrip_blocks=True, ).from_string(chat_template) @@ -2603,7 +2602,10 @@ def __call__( messages = [llama_types.ChatCompletionRequestSystemMessage(role="system", content=self.DEFAULT_SYSTEM_MESSAGE)] + messages image_urls = self.get_image_urls(messages) - template = jinja2.Template(self.CHAT_FORMAT) + template = ImmutableSandboxedEnvironment( + trim_blocks=True, + lstrip_blocks=True, + ).from_string(self.CHAT_FORMAT) text = template.render( messages=messages, add_generation_prompt=True, @@ -3242,8 +3244,7 @@ def chatml_function_calling( "{% endfor %}" "{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}" ) - template_renderer = jinja2.Environment( - loader=jinja2.BaseLoader(), + template_renderer = ImmutableSandboxedEnvironment( autoescape=jinja2.select_autoescape(["html", "xml"]), undefined=jinja2.StrictUndefined, ).from_string(function_calling_template) From 4badac3a60c6cef16d1249cae6b0c2af94cb9415 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 10 May 2024 00:56:19 -0400 Subject: [PATCH 376/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index ec70b6d6d..578d08d69 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.72] + +- fix(security): Remote Code Execution by Server-Side Template Injection in Model Metadata by @retr0reg in b454f40a9a1787b2b5659cd2cb00819d983185df +- fix(security): Update remaining jinja chat templates to use immutable sandbox by @CISC in #1441 + ## [0.2.71] - feat: Update llama.cpp to ggerganov/llama.cpp@911b3900dded9a1cfe0f0e41b82c7a29baf3a217 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index c77d2a95a..0165379ae 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.71" \ No newline at end of file +__version__ = "0.2.72" \ No newline at end of file From ac55d0a175115d1e719672ce1cb1bec776c738b1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 10 May 2024 02:38:10 -0400 Subject: [PATCH 377/624] fix: Clear kv cache to avoid kv bug when image is evaluated first --- llama_cpp/llama_chat_format.py | 1 + 1 file changed, 1 insertion(+) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index d795b3049..e2d7e2785 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2637,6 +2637,7 @@ def embed_image_bytes(image_bytes: bytes): # Evaluate prompt llama.reset() + llama._ctx.kv_cache_clear() for type_, value in split_text: if type_ == "text": tokens = llama.tokenize(value.encode("utf8"), add_bos=False, special=True) From eafb6ec5e8b4540e0b58e44b4deaaf592eb5014a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 10 May 2024 08:39:55 -0400 Subject: [PATCH 378/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 911b3900d..25c6e82e7 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 911b3900dded9a1cfe0f0e41b82c7a29baf3a217 +Subproject commit 25c6e82e7a1ad25a42b0894e87d9b5c557409516 From 73165021bbe888d58ce1adc66a2bcd6a0ecf10dd Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 10 May 2024 09:44:18 -0400 Subject: [PATCH 379/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 578d08d69..9f36b44f1 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.73] + +- feat: Update llama.cpp to ggerganov/llama.cpp@25c6e82e7a1ad25a42b0894e87d9b5c557409516 +- fix: Clear kv cache at beginning of image chat formats to avoid bug when image is evaluated first by @abetlen in ac55d0a175115d1e719672ce1cb1bec776c738b1 + ## [0.2.72] - fix(security): Remote Code Execution by Server-Side Template Injection in Model Metadata by @retr0reg in b454f40a9a1787b2b5659cd2cb00819d983185df diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 0165379ae..d6e8d0bc9 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.72" \ No newline at end of file +__version__ = "0.2.73" \ No newline at end of file From 7f59856fa6f3e23f07e12fc15aeb9359dc6c3bb4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 10 May 2024 10:18:47 -0400 Subject: [PATCH 380/624] fix: Enable CUDA backend for llava. Closes #1324 --- CMakeLists.txt | 3 ++- Makefile | 2 +- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 70f9b9993..dcaddb678 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -51,8 +51,9 @@ if (LLAMA_BUILD) ) if (LLAVA_BUILD) - if (LLAMA_CUBLAS) + if (LLAMA_CUBLAS OR LLAMA_CUDA) add_compile_definitions(GGML_USE_CUBLAS) + add_compile_definitions(GGML_USE_CUDA) endif() if (LLAMA_METAL) diff --git a/Makefile b/Makefile index 4ae011074..83085d270 100644 --- a/Makefile +++ b/Makefile @@ -16,7 +16,7 @@ build.debug: CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Debug" python3 -m pip install --verbose --config-settings=cmake.verbose=true --config-settings=logging.level=INFO --config-settings=install.strip=false --editable . build.cuda: - CMAKE_ARGS="-DLLAMA_CUBLAS=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DLLAMA_CUDA=on" python3 -m pip install --verbose -e . build.opencl: CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e . From 1547202b77c0932e36b6b3b8319c54d62ac2d127 Mon Sep 17 00:00:00 2001 From: Peng Yu Date: Fri, 10 May 2024 10:35:51 -0400 Subject: [PATCH 381/624] docs: Fix typo in README.md (#1444) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 5f3ff6824..d5ced8606 100644 --- a/README.md +++ b/README.md @@ -550,7 +550,7 @@ llm = Llama.from_pretrained( n_ctx=2048, # n_ctx should be increased to accommodate the image embedding ) -respoonse = llm.create_chat_completion( +response = llm.create_chat_completion( messages = [ { "role": "user", From 9dc5e20fb6b0992b4eab6608b36548346ac3eb25 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 12 May 2024 10:30:23 -0400 Subject: [PATCH 382/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 911b3900d..b228aba91 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 911b3900dded9a1cfe0f0e41b82c7a29baf3a217 +Subproject commit b228aba91ac2cd9eb90e9d423ba1d0d20e0117e2 From 3c19faa0d4dc249e5dfdd9e3b79fe7d911ea0bfc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 12 May 2024 10:32:52 -0400 Subject: [PATCH 383/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 9f36b44f1..2782be023 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.74] + +- feat: Update llama.cpp to ggerganov/llama.cpp@b228aba91ac2cd9eb90e9d423ba1d0d20e0117e2 +- fix: Enable CUDA backend for llava by @abetlen in 7f59856fa6f3e23f07e12fc15aeb9359dc6c3bb4 +- docs: Fix typo in README.md by @yupbank in #1444 + ## [0.2.73] - feat: Update llama.cpp to ggerganov/llama.cpp@25c6e82e7a1ad25a42b0894e87d9b5c557409516 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index d6e8d0bc9..b7b8e47c6 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.73" \ No newline at end of file +__version__ = "0.2.74" \ No newline at end of file From 3f8e17af63db75116bb5a93b2d12cf3407e8561e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 12 May 2024 11:45:55 -0400 Subject: [PATCH 384/624] fix(ci): Use version without extra platform tag in pep503 index --- scripts/releases-to-pep-503.sh | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/scripts/releases-to-pep-503.sh b/scripts/releases-to-pep-503.sh index 00ae56721..44fbbf3cf 100755 --- a/scripts/releases-to-pep-503.sh +++ b/scripts/releases-to-pep-503.sh @@ -44,7 +44,9 @@ releases=$(echo $releases | tr ' ' '\n' | grep -E $pattern) # For each release, get all assets for release in $releases; do assets=$(curl -s https://api.github.com/repos/abetlen/llama-cpp-python/releases/tags/$release | jq -r .assets) - echo "

$release

" >> index.html + # Get release version from release ie v0.1.0-cu121 -> v0.1.0 + release_version=$(echo $release | grep -oE "^[v]?[0-9]+\.[0-9]+\.[0-9]+") + echo "

$release_version

" >> index.html for asset in $(echo $assets | jq -r .[].browser_download_url); do if [[ $asset == *".whl" ]]; then echo " $asset" >> index.html From 43ba1526c8e4a903ea060c00c4dcc1ad12227a91 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 13 May 2024 09:39:08 -0400 Subject: [PATCH 385/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index b228aba91..1c570d8be 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit b228aba91ac2cd9eb90e9d423ba1d0d20e0117e2 +Subproject commit 1c570d8beeebad95872dc738ea542a4a0022f78a From 50f5c74ecf0fe53e13187c2572487a7e7ca344fd Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 14 May 2024 09:30:04 -0400 Subject: [PATCH 386/624] Update llama.cpp --- llama_cpp/llama_cpp.py | 6 ++++++ vendor/llama.cpp | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 461122a44..8da3f140c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -648,6 +648,9 @@ class llama_model_kv_override(ctypes.Structure): # // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() # const float * tensor_split; +# // comma separated list of RPC servers to use for offloading +# const char * rpc_servers; + # // Called with a progress value between 0.0 and 1.0. Pass NULL to disable. # // If the provided progress_callback returns true, model loading continues. # // If it returns false, model loading is immediately aborted. @@ -674,6 +677,7 @@ class llama_model_params(ctypes.Structure): split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored tensor_split (ctypes.Array[ctypes.ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() + rpc_servers (ctypes.c_char_p): comma separated list of RPC servers to use for offloading progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. progress_callback_user_data (ctypes.ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data @@ -687,6 +691,7 @@ class llama_model_params(ctypes.Structure): split_mode: int main_gpu: int tensor_split: CtypesArray[ctypes.c_float] + rpc_servers: ctypes.c_char_p progress_callback: Callable[[float, ctypes.c_void_p], bool] progress_callback_user_data: ctypes.c_void_p kv_overrides: CtypesArray[llama_model_kv_override] @@ -700,6 +705,7 @@ class llama_model_params(ctypes.Structure): ("split_mode", ctypes.c_int), ("main_gpu", ctypes.c_int32), ("tensor_split", ctypes.POINTER(ctypes.c_float)), + ("rpc_servers", ctypes.c_char_p), ("progress_callback", llama_progress_callback), ("progress_callback_user_data", ctypes.c_void_p), ("kv_overrides", ctypes.POINTER(llama_model_kv_override)), diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1c570d8be..1265c670f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1c570d8beeebad95872dc738ea542a4a0022f78a +Subproject commit 1265c670fd8e41e1947352c96c5179adda97fb2c From 4b54f7933050cbe9107fd949008185f92cf7de8f Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue, 14 May 2024 09:35:52 -0400 Subject: [PATCH 387/624] chore(deps): bump pypa/cibuildwheel from 2.17.0 to 2.18.0 (#1453) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.17.0 to 2.18.0. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.17.0...v2.18.0) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 8fbd68f4a..b50da4845 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.17.0 + uses: pypa/cibuildwheel@v2.18.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -56,7 +56,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.17.0 + uses: pypa/cibuildwheel@v2.18.0 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" From 389e09c2f55a9ac659fd64fada56cf750a7a378b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Tue, 14 May 2024 15:44:09 +0200 Subject: [PATCH 388/624] misc: Remove unnecessary metadata lookups (#1448) Special tokens are already mapped from metadata by llama.cpp --- llama_cpp/llama.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 4212669eb..281ee7918 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -410,8 +410,8 @@ def __init__( if self.verbose: print(f"Model metadata: {self.metadata}", file=sys.stderr) - eos_token_id = int(self.metadata.get("tokenizer.ggml.eos_token_id", self.token_eos())) - bos_token_id = int(self.metadata.get("tokenizer.ggml.bos_token_id", self.token_bos())) + eos_token_id = self.token_eos() + bos_token_id = self.token_bos() eos_token = self._model.token_get_text(eos_token_id) bos_token = self._model.token_get_text(bos_token_id) @@ -961,9 +961,9 @@ def _create_completion( completion_id: str = f"cmpl-{str(uuid.uuid4())}" created: int = int(time.time()) - prefix_token_id: int = int(self.metadata.get("tokenizer.ggml.prefix_token_id", self._model.token_prefix())) - middle_token_id: int = int(self.metadata.get("tokenizer.ggml.middle_token_id", self._model.token_middle())) - suffix_token_id: int = int(self.metadata.get("tokenizer.ggml.suffix_token_id", self._model.token_suffix())) + prefix_token_id: int = self._model.token_prefix() + middle_token_id: int = self._model.token_middle() + suffix_token_id: int = self._model.token_suffix() # If prompt is empty, initialize completion with BOS token to avoid # detokenization including a space at the beginning of the completion completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()] From 5212fb08ae69a721b2ced4e5e8b96ce642219e16 Mon Sep 17 00:00:00 2001 From: twaka Date: Tue, 14 May 2024 22:50:53 +0900 Subject: [PATCH 389/624] feat: add MinTokensLogitProcessor and min_tokens argument to server (#1333) * implement min_tokens * set default to 0 * pass min_tokens * fix * remove copy * implement MinTokensLogitsProcessor * format * fix condition --- llama_cpp/llama.py | 16 ++++++++++++++++ llama_cpp/server/app.py | 20 ++++++++++++++++++++ llama_cpp/server/types.py | 8 ++++++++ 3 files changed, 44 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 281ee7918..714532945 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2084,3 +2084,19 @@ def __call__( self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single] ) -> bool: return any([stopping_criteria(input_ids, logits) for stopping_criteria in self]) + + +class MinTokensLogitsProcessor(LogitsProcessor): + def __init__(self, min_tokens: int, token_eos: int): + self.min_tokens = min_tokens + self.token_eos = token_eos + self.prompt_tokens = None + + def __call__( + self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single] + ) -> npt.NDArray[np.single]: + if self.prompt_tokens is None: + self.prompt_tokens = len(input_ids) + if len(input_ids) - self.prompt_tokens < self.min_tokens: + scores[self.token_eos] = -np.inf + return scores diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 4cf10d1f6..4cda4af7a 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -275,6 +275,7 @@ async def create_completion( "best_of", "logit_bias_type", "user", + "min_tokens", } kwargs = body.model_dump(exclude=exclude) @@ -288,6 +289,15 @@ async def create_completion( if body.grammar is not None: kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) + if body.min_tokens > 0: + _min_tokens_logits_processor = llama_cpp.LogitsProcessorList( + [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())] + ) + if "logits_processor" not in kwargs: + kwargs["logits_processor"] = _min_tokens_logits_processor + else: + kwargs["logits_processor"].extend(_min_tokens_logits_processor) + iterator_or_completion: Union[ llama_cpp.CreateCompletionResponse, Iterator[llama_cpp.CreateCompletionStreamResponse], @@ -445,6 +455,7 @@ async def create_chat_completion( "n", "logit_bias_type", "user", + "min_tokens", } kwargs = body.model_dump(exclude=exclude) llama = llama_proxy(body.model) @@ -458,6 +469,15 @@ async def create_chat_completion( if body.grammar is not None: kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) + if body.min_tokens > 0: + _min_tokens_logits_processor = llama_cpp.LogitsProcessorList( + [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())] + ) + if "logits_processor" not in kwargs: + kwargs["logits_processor"] = _min_tokens_logits_processor + else: + kwargs["logits_processor"].extend(_min_tokens_logits_processor) + iterator_or_completion: Union[ llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index a20b3940f..a75f9e55b 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -16,6 +16,12 @@ default=16, ge=1, description="The maximum number of tokens to generate." ) +min_tokens_field = Field( + default=0, + ge=0, + description="The minimum number of tokens to generate. It may return fewer tokens if another condition is met (e.g. max_tokens, stop).", +) + temperature_field = Field( default=0.8, description="Adjust the randomness of the generated text.\n\n" @@ -111,6 +117,7 @@ class CreateCompletionRequest(BaseModel): max_tokens: Optional[int] = Field( default=16, ge=0, description="The maximum number of tokens to generate." ) + min_tokens: int = min_tokens_field temperature: float = temperature_field top_p: float = top_p_field min_p: float = min_p_field @@ -206,6 +213,7 @@ class CreateChatCompletionRequest(BaseModel): default=None, description="The maximum number of tokens to generate. Defaults to inf", ) + min_tokens: int = min_tokens_field logprobs: Optional[bool] = Field( default=False, description="Whether to output the logprobs or not. Default is True" From ca8e3c967d4fc3b1b7763f7134e78eccad549919 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 15 May 2024 23:59:17 -0400 Subject: [PATCH 390/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1265c670f..13ad16af1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1265c670fd8e41e1947352c96c5179adda97fb2c +Subproject commit 13ad16af1231ab2d245d35df3295bcfa23de1305 From d99a6ba607a4885fb00e63e967964aa41bdbbbcb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 16 May 2024 00:37:27 -0400 Subject: [PATCH 391/624] fix: segfault for models without eos / bos tokens. Closes #1463 --- llama_cpp/llama.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 714532945..043fb2a6e 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -413,8 +413,8 @@ def __init__( eos_token_id = self.token_eos() bos_token_id = self.token_bos() - eos_token = self._model.token_get_text(eos_token_id) - bos_token = self._model.token_get_text(bos_token_id) + eos_token = self._model.token_get_text(eos_token_id) if eos_token_id != -1 else "" + bos_token = self._model.token_get_text(bos_token_id) if bos_token_id != -1 else "" # Unfortunately the llama.cpp API does not return metadata arrays, so we can't get template names from tokenizer.chat_templates template_choices = dict((name[10:], template) for name, template in self.metadata.items() if name.startswith("tokenizer.chat_template.")) From b564d05806381fd56f2a6d5488ac24be931e0e5b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 16 May 2024 00:41:21 -0400 Subject: [PATCH 392/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 2782be023..55db68e43 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.75] + +- feat: Update llama.cpp to ggerganov/llama.cpp@13ad16af1231ab2d245d35df3295bcfa23de1305 +- fix: segfault for models without eos / bos tokens by @abetlen in d99a6ba607a4885fb00e63e967964aa41bdbbbcb +- feat: add MinTokensLogitProcessor and min_tokens argument to server by @twaka in #1333 +- misc: Remove unnecessary metadata lookups by @CISC in #1448 + ## [0.2.74] - feat: Update llama.cpp to ggerganov/llama.cpp@b228aba91ac2cd9eb90e9d423ba1d0d20e0117e2 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index b7b8e47c6..8f9305534 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.74" \ No newline at end of file +__version__ = "0.2.75" \ No newline at end of file From 03f171e81069932fda296c85ebafc6fdf11025d0 Mon Sep 17 00:00:00 2001 From: Radoslav Gerganov Date: Fri, 17 May 2024 20:27:26 +0300 Subject: [PATCH 393/624] example: LLM inference with Ray Serve (#1465) --- examples/ray/README.md | 19 +++++++++++++++++++ examples/ray/llm.py | 20 ++++++++++++++++++++ examples/ray/requirements.txt | 3 +++ 3 files changed, 42 insertions(+) create mode 100644 examples/ray/README.md create mode 100755 examples/ray/llm.py create mode 100644 examples/ray/requirements.txt diff --git a/examples/ray/README.md b/examples/ray/README.md new file mode 100644 index 000000000..6e338ba17 --- /dev/null +++ b/examples/ray/README.md @@ -0,0 +1,19 @@ +This is an example of doing LLM inference with [Ray](https://docs.ray.io/en/latest/index.html) and [Ray Serve](https://docs.ray.io/en/latest/serve/index.html). + +First, install the requirements: + +```bash +$ pip install -r requirements.txt +``` + +Deploy a GGUF model to Ray Serve with the following command: + +```bash +$ serve run llm:llm_builder model_path='../models/mistral-7b-instruct-v0.2.Q4_K_M.gguf' +``` + +This will start an API endpoint at `http://localhost:8000/`. You can query the model like this: + +```bash +$ curl -k -d '{"prompt": "tell me a joke", "max_tokens": 128}' -X POST http://localhost:8000 +``` diff --git a/examples/ray/llm.py b/examples/ray/llm.py new file mode 100755 index 000000000..27b15d21d --- /dev/null +++ b/examples/ray/llm.py @@ -0,0 +1,20 @@ +from starlette.requests import Request +from typing import Dict +from ray import serve +from ray.serve import Application +from llama_cpp import Llama + +@serve.deployment +class LlamaDeployment: + def __init__(self, model_path: str): + self._llm = Llama(model_path=model_path) + + async def __call__(self, http_request: Request) -> Dict: + input_json = await http_request.json() + prompt = input_json["prompt"] + max_tokens = input_json.get("max_tokens", 64) + return self._llm(prompt, max_tokens=max_tokens) + + +def llm_builder(args: Dict[str, str]) -> Application: + return LlamaDeployment.bind(args["model_path"]) diff --git a/examples/ray/requirements.txt b/examples/ray/requirements.txt new file mode 100644 index 000000000..a409fb882 --- /dev/null +++ b/examples/ray/requirements.txt @@ -0,0 +1,3 @@ +ray[serve] +--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu +llama-cpp-python From d8a3b013c3b5193656b527750557d17e61c09057 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 18 May 2024 01:19:19 -0400 Subject: [PATCH 394/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 13ad16af1..05834841d 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 13ad16af1231ab2d245d35df3295bcfa23de1305 +Subproject commit 05834841dcb4f922983ea976539c70472272df9a From 5a595f035a094574a9bcf153e6696369f63fb585 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 22 May 2024 02:40:31 -0400 Subject: [PATCH 395/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 14 ++++++++------ vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 7 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 8da3f140c..88fa07186 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -296,9 +296,10 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, # LLAMA_VOCAB_PRE_TYPE_REFACT = 8, # LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, -# LLAMA_VOCAB_PRE_TYPE_QWEN2 = 10, -# LLAMA_VOCAB_PRE_TYPE_OLMO = 11, -# LLAMA_VOCAB_PRE_TYPE_DBRX = 12, +# LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10, +# LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11, +# LLAMA_VOCAB_PRE_TYPE_OLMO = 12, +# LLAMA_VOCAB_PRE_TYPE_DBRX = 13, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -310,9 +311,10 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_GPT2 = 7 LLAMA_VOCAB_PRE_TYPE_REFACT = 8 LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9 -LLAMA_VOCAB_PRE_TYPE_QWEN2 = 10 -LLAMA_VOCAB_PRE_TYPE_OLMO = 11 -LLAMA_VOCAB_PRE_TYPE_DBRX = 12 +LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10 +LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11 +LLAMA_VOCAB_PRE_TYPE_OLMO = 12 +LLAMA_VOCAB_PRE_TYPE_DBRX = 13 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 05834841d..201cc11af 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 05834841dcb4f922983ea976539c70472272df9a +Subproject commit 201cc11afa0a1950e1f632390b2ac6c937a0d8f0 From 087cc0b036e6a52fda0df6fb4575e908fa0e32fd Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 24 May 2024 01:43:36 -0400 Subject: [PATCH 396/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 16 ++++++++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 17 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 88fa07186..4284019cb 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -2265,6 +2265,22 @@ def llama_set_n_threads( ... +# // Get the number of threads used for generation of a single token. +# LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx); +@ctypes_function("llama_n_threads", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_threads(ctx: llama_context_p, /) -> int: + """Get the number of threads used for generation of a single token""" + ... + + +# // Get the number of threads used for prompt and batch processing (multiple token). +# LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx); +@ctypes_function("llama_n_threads_batch", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_threads_batch(ctx: llama_context_p, /) -> int: + """Get the number of threads used for prompt and batch processing (multiple token)""" + ... + + # // Set whether to use causal attention or not # // If set to true, the model will only attend to the past tokens # LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn); diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 201cc11af..0df0aa8e4 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 201cc11afa0a1950e1f632390b2ac6c937a0d8f0 +Subproject commit 0df0aa8e43c3378975269a51f9b876c8692e70da From 5cae1040e35326dc514552bb9423c3718a37bf44 Mon Sep 17 00:00:00 2001 From: Linghan Zhong Date: Fri, 24 May 2024 00:49:44 -0500 Subject: [PATCH 397/624] feat: Improve Llama.eval performance by avoiding list conversion (#1476) Co-authored-by: Andrei --- llama_cpp/llama.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 043fb2a6e..6dad650c6 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -562,12 +562,12 @@ def eval(self, tokens: Sequence[int]): if self.context_params.logits_all: rows = n_tokens cols = self._n_vocab - logits = self._ctx.get_logits()[: rows * cols] + logits = np.ctypeslib.as_array(self._ctx.get_logits(), shape=(rows * cols, )) self.scores[n_past : n_past + n_tokens, :].reshape(-1)[: :] = logits else: rows = 1 cols = self._n_vocab - logits = self._ctx.get_logits()[: rows * cols] + logits = np.ctypeslib.as_array(self._ctx.get_logits(), shape=(rows * cols, )) self.scores[n_past + n_tokens - 1, :].reshape(-1)[: :] = logits # Update n_tokens self.n_tokens += n_tokens From a4c9ab885d3c5b95ca881e72666bef876df97844 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 24 May 2024 01:59:25 -0400 Subject: [PATCH 398/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 55db68e43..3e11407c6 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.76] + +- feat: Update llama.cpp to ggerganov/llama.cpp@0df0aa8e43c3378975269a51f9b876c8692e70da +- feat: Improve Llama.eval performance by avoiding list conversion by @thoughtp0lice in #1476 +- example: LLM inference with Ray Serve by @rgerganov in #1465 + ## [0.2.75] - feat: Update llama.cpp to ggerganov/llama.cpp@13ad16af1231ab2d245d35df3295bcfa23de1305 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 8f9305534..fcc50d75b 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.75" \ No newline at end of file +__version__ = "0.2.76" \ No newline at end of file From ec43e8920f2231639bea33d3db6651841bcec45c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 24 May 2024 11:54:15 -0400 Subject: [PATCH 399/624] docs: Update multi-modal model section --- README.md | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index d5ced8606..3255e6956 100644 --- a/README.md +++ b/README.md @@ -499,13 +499,16 @@ llm = Llama.from_pretrained( `llama-cpp-python` supports such as llava1.5 which allow the language model to read information from both text and images. -You'll first need to download one of the available multi-modal models in GGUF format: - -- [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) -- [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) -- [bakllava-1-7b](https://huggingface.co/mys/ggml_bakllava-1) -- [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) -- [moondream2](https://huggingface.co/vikhyatk/moondream2) +Below are the supported multi-modal models and their respective chat handlers (Python API) and chat formats (Server API). + +| Model | `LlamaChatHandler` | `chat_format` | +| --- | --- | --- | +| [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) | `Llava15ChatHandler` | `llava-1-5` | +| [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) | `Llava15ChatHandler` | `llava-1-5` | +| [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) | `Llava16ChatHandler` | `llava-1-6` | +| [moondream2](https://huggingface.co/vikhyatk/moondream2) | `MoondreamChatHandler` | `moondream2` | +| [nanollava](https://huggingface.co/abetlen/nanollava) | `NanollavaChatHandler` | `nanollava` | +| [llama-3-vision-alpha](https://huggingface.co/abetlen/llama-3-vision-alpha) | `Llama3VisionAlphaChatHandler` | `llama-3-vision-alpha` | Then you'll need to use a custom chat handler to load the clip model and process the chat messages and images. From 9e8d7d55bd9873d10c23a78a954a3063a5e995c0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 24 May 2024 11:55:01 -0400 Subject: [PATCH 400/624] fix(docs): Fix link typo --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 3255e6956..07445eb7f 100644 --- a/README.md +++ b/README.md @@ -507,8 +507,8 @@ Below are the supported multi-modal models and their respective chat handlers (P | [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) | `Llava15ChatHandler` | `llava-1-5` | | [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) | `Llava16ChatHandler` | `llava-1-6` | | [moondream2](https://huggingface.co/vikhyatk/moondream2) | `MoondreamChatHandler` | `moondream2` | -| [nanollava](https://huggingface.co/abetlen/nanollava) | `NanollavaChatHandler` | `nanollava` | -| [llama-3-vision-alpha](https://huggingface.co/abetlen/llama-3-vision-alpha) | `Llama3VisionAlphaChatHandler` | `llama-3-vision-alpha` | +| [nanollava](https://huggingface.co/abetlen/nanollava-gguf) | `NanollavaChatHandler` | `nanollava` | +| [llama-3-vision-alpha](https://huggingface.co/abetlen/llama-3-vision-alpha-gguf) | `Llama3VisionAlphaChatHandler` | `llama-3-vision-alpha` | Then you'll need to use a custom chat handler to load the clip model and process the chat messages and images. From 2d8996414706c0bb69fae06ddf7f092f27790fe7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 24 May 2024 11:55:41 -0400 Subject: [PATCH 401/624] docs: Fix table formatting --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 07445eb7f..342e924a0 100644 --- a/README.md +++ b/README.md @@ -502,7 +502,7 @@ llm = Llama.from_pretrained( Below are the supported multi-modal models and their respective chat handlers (Python API) and chat formats (Server API). | Model | `LlamaChatHandler` | `chat_format` | -| --- | --- | --- | +|:--- |:--- |:--- | | [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) | `Llava15ChatHandler` | `llava-1-5` | | [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) | `Llava15ChatHandler` | `llava-1-5` | | [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) | `Llava16ChatHandler` | `llava-1-6` | From 454c9bb1cb3a58ecc37b58abeff6f245e3b8f316 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 27 May 2024 10:51:57 -0400 Subject: [PATCH 402/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 21 +++++++++++++++++---- vendor/llama.cpp | 2 +- 2 files changed, 18 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 4284019cb..d9b5087da 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -300,6 +300,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11, # LLAMA_VOCAB_PRE_TYPE_OLMO = 12, # LLAMA_VOCAB_PRE_TYPE_DBRX = 13, +# LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -315,6 +316,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11 LLAMA_VOCAB_PRE_TYPE_OLMO = 12 LLAMA_VOCAB_PRE_TYPE_DBRX = 13 +LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 # // note: these values should be synchronized with ggml_rope @@ -718,6 +720,8 @@ class llama_model_params(ctypes.Structure): ] +# // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations +# // https://github.com/ggerganov/llama.cpp/pull/7544 # struct llama_context_params { # uint32_t seed; // RNG seed, -1 for random # uint32_t n_ctx; // text context, 0 = from model @@ -744,15 +748,14 @@ class llama_model_params(ctypes.Structure): # ggml_backend_sched_eval_callback cb_eval; # void * cb_eval_user_data; -# enum ggml_type type_k; // data type for K cache -# enum ggml_type type_v; // data type for V cache +# enum ggml_type type_k; // data type for K cache [EXPERIMENTAL] +# enum ggml_type type_v; // data type for V cache [EXPERIMENTAL] # // Keep the booleans together to avoid misalignment during copy-by-value. # bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) # bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU -# bool flash_attn; // whether to use flash attention - +# bool flash_attn; // whether to use flash attention [EXPERIMENTAL] # // Abort callback # // if it returns true, execution of llama_decode() will be aborted @@ -2454,6 +2457,16 @@ def llama_token_is_eog(model: llama_model_p, token: Union[llama_token, int], /) ... +# // Identify if Token Id is a control token or a render-able token +# LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token); +@ctypes_function( + "llama_token_is_control", [llama_model_p_ctypes, llama_token], ctypes.c_bool +) +def llama_token_is_control(model: llama_model_p, token: Union[llama_token, int], /) -> bool: + """Identify if Token Id is a control token or a render-able token""" + ... + + # // Special tokens diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 0df0aa8e4..5487593bc 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 0df0aa8e43c3378975269a51f9b876c8692e70da +Subproject commit 5487593bc7ee0b65b9d2e2985b4b61dc77043101 From c564007ff6c93d3408031db9562de13d50112707 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon, 27 May 2024 10:57:17 -0400 Subject: [PATCH 403/624] chore(deps): bump pypa/cibuildwheel from 2.18.0 to 2.18.1 (#1472) updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Andrei --- .github/workflows/build-and-release.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index b50da4845..957912d09 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.18.0 + uses: pypa/cibuildwheel@v2.18.1 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -56,7 +56,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.18.0 + uses: pypa/cibuildwheel@v2.18.1 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" From c26004b1be20c485ea547a87b3871551f6a8774c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 28 May 2024 22:52:03 -0400 Subject: [PATCH 404/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5487593bc..504f0c340 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5487593bc7ee0b65b9d2e2985b4b61dc77043101 +Subproject commit 504f0c340f6b5e04de682f6ddefdd3b81208df5d From 2907c26906272a333ab913054f7ae7cf2bbd94ed Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 28 May 2024 22:52:28 -0400 Subject: [PATCH 405/624] misc: Update debug build to keep all debug symbols for easier gdb debugging --- Makefile | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/Makefile b/Makefile index 83085d270..3796d17e3 100644 --- a/Makefile +++ b/Makefile @@ -13,7 +13,13 @@ build: python3 -m pip install --verbose -e . build.debug: - CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Debug" python3 -m pip install --verbose --config-settings=cmake.verbose=true --config-settings=logging.level=INFO --config-settings=install.strip=false --editable . + python3 -m pip install \ + --verbose \ + --config-settings=cmake.verbose=true \ + --config-settings=logging.level=INFO \ + --config-settings=install.strip=false \ + --config-settings=cmake.args="-DCMAKE_BUILD_TYPE=Debug;-DCMAKE_C_FLAGS='-ggdb -O0';-DCMAKE_CXX_FLAGS='-ggdb -O0'" \ + --editable . build.cuda: CMAKE_ARGS="-DLLAMA_CUDA=on" python3 -m pip install --verbose -e . From df45a4b3fe46e72664bda87301b318210c6d4782 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 29 May 2024 02:02:22 -0400 Subject: [PATCH 406/624] fix: fix string value kv_overrides. Closes #1487 --- llama_cpp/llama.py | 13 ++++++++----- llama_cpp/server/model.py | 4 +++- 2 files changed, 11 insertions(+), 6 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 6dad650c6..6d872e3e6 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -6,6 +6,7 @@ import time import json import ctypes +import typing import fnmatch import multiprocessing @@ -249,13 +250,13 @@ def __init__( self._kv_overrides_array[i].key = k.encode("utf-8") if isinstance(v, bool): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL - self._kv_overrides_array[i].value.bool_value = v + self._kv_overrides_array[i].value.val_bool = v elif isinstance(v, int): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT - self._kv_overrides_array[i].value.int_value = v + self._kv_overrides_array[i].value.val_i64 = v elif isinstance(v, float): self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT - self._kv_overrides_array[i].value.float_value = v + self._kv_overrides_array[i].value.val_f64 = v elif isinstance(v, str): # type: ignore v_bytes = v.encode("utf-8") if len(v_bytes) > 128: # TODO: Make this a constant @@ -263,10 +264,12 @@ def __init__( v_bytes = v_bytes.ljust(128, b"\0") self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR # copy min(v_bytes, 128) to str_value + address = typing.cast(int, ctypes.addressof(self._kv_overrides_array[i].value) + llama_cpp.llama_model_kv_override_value.val_str.offset) + buffer_start = ctypes.cast(address, ctypes.POINTER(ctypes.c_char)) ctypes.memmove( - self._kv_overrides_array[i].value.str_value, + buffer_start, v_bytes, - min(len(v_bytes), 128), + 128, ) else: raise ValueError(f"Unknown value type for {k}: {v}") diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index f00292410..4f8371668 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -183,7 +183,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: num_pred_tokens=settings.draft_model_num_pred_tokens ) - kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None + kv_overrides: Optional[Dict[str, Union[bool, int, float, str]]] = None if settings.kv_overrides is not None: assert isinstance(settings.kv_overrides, list) kv_overrides = {} @@ -197,6 +197,8 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: kv_overrides[key] = int(value) elif value_type == "float": kv_overrides[key] = float(value) + elif value_type == "str": + kv_overrides[key] = value else: raise ValueError(f"Unknown value type {value_type}") From 91d05aba469e9ce4632995b8f41e7c3b3c3518a5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 29 May 2024 02:28:58 -0400 Subject: [PATCH 407/624] fix: adjust kv_override member names to match llama.cpp --- llama_cpp/llama_cpp.py | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d9b5087da..1668717a1 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -613,17 +613,17 @@ class llama_batch(ctypes.Structure): # }; class llama_model_kv_override_value(ctypes.Union): _fields_ = [ - ("int_value", ctypes.c_int64), - ("float_value", ctypes.c_double), - ("bool_value", ctypes.c_bool), - ("str_value", ctypes.c_char * 128), + ("val_i64", ctypes.c_int64), + ("val_f64", ctypes.c_double), + ("val_bool", ctypes.c_bool), + ("val_str", ctypes.c_char * 128), ] if TYPE_CHECKING: - int_value: int - float_value: float - bool_value: bool - str_value: bytes + val_i64: int + val_f64: float + val_bool: bool + val_str: bytes class llama_model_kv_override(ctypes.Structure): From 165b4dc6c188f8fda2fc616154e111f710484eba Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 29 May 2024 02:29:44 -0400 Subject: [PATCH 408/624] fix: Fix typo in Llama3VisionAlphaChatHandler. Closes #1488 --- llama_cpp/llama_chat_format.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index e2d7e2785..8f3b1de11 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -3098,7 +3098,7 @@ class NanoLlavaChatHandler(Llava15ChatHandler): "{% endif %}" ) -class Llama3VisionAlpha(Llava15ChatHandler): +class Llama3VisionAlphaChatHandler(Llava15ChatHandler): # question = "" + q # prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{question}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" @@ -3159,6 +3159,10 @@ class Llama3VisionAlpha(Llava15ChatHandler): "{% endif %}" ) +# alias +Llama3VisionAlpha = Llama3VisionAlphaChatHandler + + @register_chat_completion_handler("chatml-function-calling") def chatml_function_calling( llama: llama.Llama, @@ -3193,7 +3197,6 @@ def chatml_function_calling( llama_types.CreateChatCompletionResponse, Iterator[llama_types.CreateChatCompletionStreamResponse], ]: - print(logprobs) function_calling_template = ( "{% for message in messages %}" "<|im_start|>{{ message.role }}\n" From af3ed503e9ce60fe6b5365031abad4176a3536b3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 1 Jun 2024 18:09:24 -0400 Subject: [PATCH 409/624] fix: Use numpy recarray for candidates data, fixes bug with temp < 0 --- llama_cpp/_internals.py | 22 +++++++++------------- 1 file changed, 9 insertions(+), 13 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index b404601d3..e3f539b11 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -545,13 +545,12 @@ def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool): class _LlamaTokenDataArray: def __init__(self, *, n_vocab: int): self.n_vocab = n_vocab - self.candidates_data = np.array( - [], + self.candidates_data = np.recarray( + (self.n_vocab,), dtype=np.dtype( [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True ), ) - self.candidates_data.resize(3, self.n_vocab, refcheck=False) self.candidates = llama_cpp.llama_token_data_array( data=self.candidates_data.ctypes.data_as(llama_cpp.llama_token_data_p), size=self.n_vocab, @@ -561,14 +560,11 @@ def __init__(self, *, n_vocab: int): self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) def copy_logits(self, logits: npt.NDArray[np.single]): - self.candidates_data["id"][:] = self.default_candidates_data_id - self.candidates_data["logit"][:] = logits - self.candidates_data["p"][:] = self.default_candidates_data_p - self.candidates.data = self.candidates_data.ctypes.data_as( - llama_cpp.llama_token_data_p - ) - self.candidates.sorted = ctypes.c_bool(False) - self.candidates.size = ctypes.c_size_t(self.n_vocab) + self.candidates_data.id[:] = self.default_candidates_data_id + self.candidates_data.logit[:] = logits + self.candidates_data.p[:] = self.default_candidates_data_p + self.candidates.sorted = False + self.candidates.size = self.n_vocab # Python wrappers over common/common @@ -759,14 +755,14 @@ def sample( self.params.penalty_present, ) if not self.params.penalize_nl: - token_data_array.candidates_data["logit"][nl_token] = nl_logit + token_data_array.candidates_data.logit[nl_token] = nl_logit if self.grammar is not None: ctx_main.sample_grammar(token_data_array, self.grammar) if self.params.temp < 0: ctx_main.sample_softmax(token_data_array) - id = token_data_array.candidates_data["id"][0] + id = token_data_array.candidates_data.id[0] elif self.params.temp == 0: id = ctx_main.sample_token_greedy(token_data_array) else: From 6b018e00b15c18f0c72d85bdc64ee92f69b036ed Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 3 Jun 2024 11:19:10 -0400 Subject: [PATCH 410/624] misc: Improve llava error messages --- llama_cpp/llama_chat_format.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 8f3b1de11..b384749e0 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2642,13 +2642,13 @@ def embed_image_bytes(image_bytes: bytes): if type_ == "text": tokens = llama.tokenize(value.encode("utf8"), add_bos=False, special=True) if llama.n_tokens + len(tokens) > llama.n_ctx(): - raise ValueError("Prompt exceeds n_ctx") # TODO: Fix + raise ValueError(f"Prompt exceeds n_ctx: {llama.n_tokens + len(tokens)} > {llama.n_ctx()}") llama.eval(tokens) else: image_bytes = self.load_image(value) embed = embed_image_bytes(image_bytes) if llama.n_tokens + embed.contents.n_image_pos > llama.n_ctx(): - raise ValueError("Prompt exceeds n_ctx") # TODO: Fix + raise ValueError(f"Prompt exceeds n_ctx: {llama.n_tokens + embed.contents.n_image_pos} > {llama.n_ctx()}") n_past = ctypes.c_int(llama.n_tokens) n_past_p = ctypes.pointer(n_past) with suppress_stdout_stderr(disable=self.verbose): From cd3f1bb387fb351fab710d4eb10b13fd87e4aa36 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 4 Jun 2024 00:35:47 -0400 Subject: [PATCH 411/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++-- vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1668717a1..706f50e1a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1233,12 +1233,12 @@ def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... def llama_pooling_type(ctx: llama_context_p, /) -> int: ... -# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); +# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) def llama_vocab_type(model: llama_model_p, /) -> int: ... -# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); +# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) def llama_rope_type(model: llama_model_p, /) -> int: ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 504f0c340..bde7cd3cd 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 504f0c340f6b5e04de682f6ddefdd3b81208df5d +Subproject commit bde7cd3cd949c1a85d3a199498ac98e78039d46f From ae5682f500623383a3676fec17aec19330ead67e Mon Sep 17 00:00:00 2001 From: Engininja2 <139037756+Engininja2@users.noreply.github.com> Date: Mon, 3 Jun 2024 22:42:34 -0600 Subject: [PATCH 412/624] fix: Disable Windows+CUDA workaround when compiling for HIPBLAS (#1493) * Disable Windows+CUDA workaround when compiling for HIPBLAS * fix spacing * change condition to check for Windows & CUDA Co-authored-by: Andrei --------- Co-authored-by: Andrei --- CMakeLists.txt | 18 ++++++++++-------- 1 file changed, 10 insertions(+), 8 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index dcaddb678..4b9508d40 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -41,14 +41,16 @@ if (LLAMA_BUILD) RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp ) # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 - install( - FILES $ - DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ) - install( - FILES $ - DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - ) + if (WIN32 AND (LLAMA_CUDA OR LLAMA_CUBLAS)) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp + ) + install( + FILES $ + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp + ) + endif() if (LLAVA_BUILD) if (LLAMA_CUBLAS OR LLAMA_CUDA) From c3ef41ba06c6fab90592040e582985c13d185faa Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 4 Jun 2024 00:49:24 -0400 Subject: [PATCH 413/624] chore: Bump version --- CHANGELOG.md | 8 ++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 3e11407c6..1b9865da0 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.77] + +- feat: Update llama.cpp to ggerganov/llama.cpp@bde7cd3cd949c1a85d3a199498ac98e78039d46f +- fix: string value kv_overrides by @abetlen in df45a4b3fe46e72664bda87301b318210c6d4782 +- fix: Fix typo in Llama3VisionAlphaChatHandler by @abetlen in 165b4dc6c188f8fda2fc616154e111f710484eba +- fix: Use numpy recarray for candidates data, fixes bug with temp < 0 by @abetlen in af3ed503e9ce60fe6b5365031abad4176a3536b3 +fix: Disable Windows+CUDA workaround when compiling for HIPBLAS by Engininja2 in #1493 + ## [0.2.76] - feat: Update llama.cpp to ggerganov/llama.cpp@0df0aa8e43c3378975269a51f9b876c8692e70da diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index fcc50d75b..3df249f5e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.76" \ No newline at end of file +__version__ = "0.2.77" \ No newline at end of file From 027f7bc67890f1de801407fbbb608c182e2ad286 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Tue, 4 Jun 2024 16:15:41 +0200 Subject: [PATCH 414/624] fix: Avoid duplicate special tokens in chat formats (#1439) * Templates sometimes have BOS in them, remove duplicate * tokenize chat format prompts before completion This is to ensure that we don't duplicate any special tokens. Hopefully I amended the existing formats correctly? * updated comment * corrected a few * add some missing internals * proper bos/eos detection * just let tokenizer do the job * typo-- * align test with new response * changed to a warning * move to another PR * Use python warnings module --------- Co-authored-by: Andrei Betlen --- llama_cpp/_internals.py | 8 ++++++++ llama_cpp/llama.py | 7 +++++++ llama_cpp/llama_chat_format.py | 17 ++++++++--------- tests/test_llama_chat_format.py | 3 ++- 4 files changed, 25 insertions(+), 10 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index e3f539b11..fb437c3f1 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -142,6 +142,14 @@ def token_eos(self) -> int: assert self.model is not None return llama_cpp.llama_token_eos(self.model) + def token_cls(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_cls(self.model) + + def token_sep(self) -> int: + assert self.model is not None + return llama_cpp.llama_token_sep(self.model) + def token_nl(self) -> int: assert self.model is not None return llama_cpp.llama_token_nl(self.model) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 6d872e3e6..45e1526ef 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -8,6 +8,7 @@ import ctypes import typing import fnmatch +import warnings import multiprocessing from typing import ( @@ -1019,6 +1020,12 @@ def _create_completion( ) model_name: str = model if model is not None else self.model_path + if prompt_tokens[:2] == [self.token_bos()] * 2: + warnings.warn( + f'Detected duplicate leading "{self._model.token_get_text(self.token_bos())}" in prompt, this will likely reduce response quality, consider removing it...', + RuntimeWarning, + ) + # NOTE: This likely doesn't work correctly for the first token in the prompt # because of the extra space added to the start of the prompt_tokens if logit_bias is not None: diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index b384749e0..975f74762 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -160,6 +160,7 @@ class ChatFormatterResponse: prompt: str stop: Optional[Union[str, List[str]]] = None stopping_criteria: Optional[llama.StoppingCriteriaList] = None + added_special: bool = False class ChatFormatter(Protocol): @@ -232,7 +233,7 @@ def stop_on_last_token( return tokens[-1] in self.stop_token_ids stopping_criteria = llama.StoppingCriteriaList([stop_on_last_token]) - return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token], stopping_criteria=stopping_criteria) + return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token], stopping_criteria=stopping_criteria, added_special=True) def to_chat_handler(self) -> LlamaChatCompletionHandler: return chat_formatter_to_chat_completion_handler(self) @@ -548,7 +549,7 @@ def chat_completion_handler( tools=tools, tool_choice=tool_choice, ) - prompt = result.prompt + prompt = llama.tokenize(result.prompt.encode("utf-8"), add_bos=not result.added_special, special=True) if result.stop is not None: stop = [] if stop is None else [stop] if isinstance(stop, str) else stop rstop = result.stop if isinstance(result.stop, list) else [result.stop] @@ -655,7 +656,7 @@ def format_autotokenizer( prompt: str = tokenizer.apply_chat_template(messages, tokenize=False) # type: ignore assert isinstance(prompt, str) # Return formatted prompt and eos token by default - return ChatFormatterResponse(prompt=prompt, stop=tokenizer.eos_token) + return ChatFormatterResponse(prompt=prompt, stop=tokenizer.eos_token, added_special=True) return format_autotokenizer @@ -708,7 +709,7 @@ def format_tokenizer_config( bos_token=bos_token, eos_token=eos_token, ) - return ChatFormatterResponse(prompt=prompt, stop=[eos_token, bos_token]) + return ChatFormatterResponse(prompt=prompt, stop=[eos_token, bos_token], added_special=True) return format_tokenizer_config @@ -918,7 +919,7 @@ def format_llama2( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - _system_template = "[INST] <>\n{system_message}\n<>" + _system_template = "[INST] <>\n{system_message}\n<>" _roles = dict(user="[INST]", assistant="[/INST]") _messages = _map_roles(messages, _roles) system_message = _get_system_message(messages) @@ -940,11 +941,10 @@ def format_llama3( user="<|start_header_id|>user<|end_header_id|>\n\n", assistant="<|start_header_id|>assistant<|end_header_id|>\n\n", ) - _begin_token = "<|begin_of_text|>" _sep = "<|eot_id|>" _messages = _map_roles(messages, _roles) _messages.append((_roles["assistant"], None)) - _prompt = _format_no_colon_single(_begin_token, _messages, _sep) + _prompt = _format_no_colon_single("", _messages, _sep) return ChatFormatterResponse(prompt=_prompt, stop=_sep) @@ -1229,10 +1229,9 @@ def format_mistral_instruct( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - bos = "" eos = "" stop = eos - prompt = bos + prompt = "" for message in messages: if ( message["role"] == "user" diff --git a/tests/test_llama_chat_format.py b/tests/test_llama_chat_format.py index c10aee42e..f031bf72b 100644 --- a/tests/test_llama_chat_format.py +++ b/tests/test_llama_chat_format.py @@ -21,12 +21,13 @@ def test_mistral_instruct(): response = llama_chat_format.format_mistral_instruct( messages=messages, ) + prompt = ("" if response.added_special else "") + response.prompt reference = chat_formatter.render( messages=messages, bos_token="", eos_token="", ) - assert response.prompt == reference + assert prompt == reference mistral_7b_tokenizer_config = """{ From 6e0642ca195e2716e6e1bfa05d2b76f1e2dae0a3 Mon Sep 17 00:00:00 2001 From: Asghar Ghorbani Date: Tue, 4 Jun 2024 16:18:38 +0200 Subject: [PATCH 415/624] fix: fix logprobs when BOS is not present (#1471) * Fix lobprobs when BOS is not present * Fix logprobs when bos is not available --- llama_cpp/llama.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 45e1526ef..fb26950b6 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1410,8 +1410,8 @@ def logit_bias_processor( top_logprobs: List[Optional[Dict[str, float]]] = [] if echo: - # Remove leading BOS token - all_tokens = prompt_tokens[1:] + completion_tokens + # Remove leading BOS token if exists + all_tokens = prompt_tokens[1 if prompt_tokens[0] == self.token_bos() else 0:] + completion_tokens else: all_tokens = completion_tokens From d634efcdd91a0f79ae0fac13cfdf46eb47d0602d Mon Sep 17 00:00:00 2001 From: nullname Date: Tue, 4 Jun 2024 22:38:21 +0800 Subject: [PATCH 416/624] feat: adding `rpc_servers` parameter to `Llama` class (#1477) * passthru rpc_servers params wip * enable llama rpc by default * convert string to byte * add rpc package * Revert "enable llama rpc by default" This reverts commit 832c6dd56c979514cec5df224bf2d2014dccd790. * update readme * Only set rpc_servers when provided * Add rpc servers to server options --------- Co-authored-by: Andrei Betlen --- Makefile | 3 +++ README.md | 11 +++++++++++ llama_cpp/llama.py | 7 +++++++ llama_cpp/server/model.py | 1 + llama_cpp/server/settings.py | 4 ++++ 5 files changed, 26 insertions(+) diff --git a/Makefile b/Makefile index 3796d17e3..d8fb0cc27 100644 --- a/Makefile +++ b/Makefile @@ -45,6 +45,9 @@ build.kompute: build.sycl: CMAKE_ARGS="-DLLAMA_SYCL=on" python3 -m pip install --verbose -e . +build.rpc: + CMAKE_ARGS="-DLLAMA_RPC=on" python3 -m pip install --verbose -e . + build.sdist: python3 -m build --sdist diff --git a/README.md b/README.md index 342e924a0..0f7abfb28 100644 --- a/README.md +++ b/README.md @@ -221,6 +221,17 @@ CMAKE_ARGS="-DLLAMA_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pi ```
+
+RPC + +To install with RPC support, set the `LLAMA_RPC=on` environment variable before installing: + +```bash +source /opt/intel/oneapi/setvars.sh +CMAKE_ARGS="-DLLAMA_RPC=on" pip install llama-cpp-python +``` +
+ ### Windows Notes diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index fb26950b6..bf3bd656d 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -72,6 +72,7 @@ def __init__( split_mode: int = llama_cpp.LLAMA_SPLIT_MODE_LAYER, main_gpu: int = 0, tensor_split: Optional[List[float]] = None, + rpc_servers: Optional[str] = None, vocab_only: bool = False, use_mmap: bool = True, use_mlock: bool = False, @@ -150,6 +151,7 @@ def __init__( split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options. main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_LAYER: ignored tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split. + rpc_servers: Comma separated list of RPC servers to use for offloading vocab_only: Only load the vocabulary no weights. use_mmap: Use mmap if possible. use_mlock: Force the system to keep the model in RAM. @@ -221,6 +223,11 @@ def __init__( ) # 0x7FFFFFFF is INT32 max, will be auto set to all layers self.model_params.split_mode = split_mode self.model_params.main_gpu = main_gpu + if rpc_servers is not None: + self.model_params.rpc_servers = rpc_servers.encode('utf-8') + self._rpc_servers = rpc_servers + else: + self._rpc_servers = None self.tensor_split = tensor_split self._c_tensor_split = None if self.tensor_split is not None: diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 4f8371668..d4d4acbe3 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -226,6 +226,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: use_mmap=settings.use_mmap, use_mlock=settings.use_mlock, kv_overrides=kv_overrides, + rpc_servers=settings.rpc_servers, # Context Params seed=settings.seed, n_ctx=settings.n_ctx, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index a3e185007..4d924f337 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -58,6 +58,10 @@ class ModelSettings(BaseSettings): default=None, description="List of model kv overrides in the format key=type:value where type is one of (bool, int, float). Valid true values are (true, TRUE, 1), otherwise false.", ) + rpc_servers: Optional[str] = Field( + default=None, + description="comma seperated list of rpc servers for offloading", + ) # Context Params seed: int = Field( default=llama_cpp.LLAMA_DEFAULT_SEED, description="Random seed. -1 for random." From 255e1b44955f1c81c6dccce312e162dbc5168def Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 7 Jun 2024 02:02:12 -0400 Subject: [PATCH 417/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 4 +- llama_cpp/llama_cpp.py | 135 +++++++++++----------------------------- vendor/llama.cpp | 2 +- 3 files changed, 38 insertions(+), 103 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index fb437c3f1..ba0429139 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -128,9 +128,9 @@ def token_get_score(self, token: int) -> float: assert self.model is not None return llama_cpp.llama_token_get_score(self.model, token) - def token_get_type(self, token: int) -> int: + def token_get_attr(self, token: int) -> int: assert self.model is not None - return llama_cpp.llama_token_get_type(self.model, token) + return llama_cpp.llama_token_get_attr(self.model, token) # Special tokens diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 706f50e1a..6f37fcb69 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -333,7 +333,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_ROPE_TYPE_GLM = 4 -# enum llama_token_type { +# enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file # LLAMA_TOKEN_TYPE_UNDEFINED = 0, # LLAMA_TOKEN_TYPE_NORMAL = 1, # LLAMA_TOKEN_TYPE_UNKNOWN = 2, @@ -351,6 +351,32 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_TOKEN_TYPE_BYTE = 6 +# enum llama_token_attr { +# LLAMA_TOKEN_ATTR_UNDEFINED = 0, +# LLAMA_TOKEN_ATTR_UNKNOWN = 1 << 0, +# LLAMA_TOKEN_ATTR_UNUSED = 1 << 1, +# LLAMA_TOKEN_ATTR_NORMAL = 1 << 2, +# LLAMA_TOKEN_ATTR_CONTROL = 1 << 3, // SPECIAL? +# LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4, +# LLAMA_TOKEN_ATTR_BYTE = 1 << 5, +# LLAMA_TOKEN_ATTR_NORMALIZED = 1 << 6, +# LLAMA_TOKEN_ATTR_LSTRIP = 1 << 7, +# LLAMA_TOKEN_ATTR_RSTRIP = 1 << 8, +# LLAMA_TOKEN_ATTR_SINGLE_WORD = 1 << 9, +# }; +LLAMA_TOKEN_ATTR_UNDEFINED = 0 +LLAMA_TOKEN_ATTR_UNKNOWN = 1 << 0 +LLAMA_TOKEN_ATTR_UNUSED = 1 << 1 +LLAMA_TOKEN_ATTR_NORMAL = 1 << 2 +LLAMA_TOKEN_ATTR_CONTROL = 1 << 3 +LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4 +LLAMA_TOKEN_ATTR_BYTE = 1 << 5 +LLAMA_TOKEN_ATTR_NORMALIZED = 1 << 6 +LLAMA_TOKEN_ATTR_LSTRIP = 1 << 7 +LLAMA_TOKEN_ATTR_RSTRIP = 1 << 8 +LLAMA_TOKEN_ATTR_SINGLE_WORD = 1 << 9 + + # // model file types # enum llama_ftype { # LLAMA_FTYPE_ALL_F32 = 0, @@ -959,6 +985,9 @@ class llama_model_quantize_params(ctypes.Structure): # // modifies a preceding LLAMA_GRETYPE_CHAR or # // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) # LLAMA_GRETYPE_CHAR_ALT = 6, + +# // any character (.) +# LLAMA_GRETYPE_CHAR_ANY = 7, # }; LLAMA_GRETYPE_END = 0 LLAMA_GRETYPE_ALT = 1 @@ -967,6 +996,7 @@ class llama_model_quantize_params(ctypes.Structure): LLAMA_GRETYPE_CHAR_NOT = 4 LLAMA_GRETYPE_CHAR_RNG_UPPER = 5 LLAMA_GRETYPE_CHAR_ALT = 6 +LLAMA_GRETYPE_CHAR_ANY = 7 # typedef struct llama_grammar_element { @@ -2438,11 +2468,11 @@ def llama_token_get_score( ) -> float: ... -# LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); +# LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token); @ctypes_function( - "llama_token_get_type", [llama_model_p_ctypes, llama_token], ctypes.c_int + "llama_token_get_attr", [llama_model_p_ctypes, llama_token], ctypes.c_int ) -def llama_token_get_type( +def llama_token_get_attr( model: llama_model_p, token: Union[llama_token, int], / ) -> int: ... @@ -3200,104 +3230,9 @@ def llama_grammar_accept_token( # // -# // Beam search +# // Model split # // -# struct llama_beam_view { -# const llama_token * tokens; - - -# size_t n_tokens; -# float p; // Cumulative beam probability (renormalized relative to all beams) -# bool eob; // Callback should set this to true when a beam is at end-of-beam. -# }; -class llama_beam_view(ctypes.Structure): - if TYPE_CHECKING: - tokens: CtypesArray[llama_token] - n_tokens: int - p: float - eob: bool - - _fields_ = [ - ("tokens", llama_token_p), - ("n_tokens", ctypes.c_size_t), - ("p", ctypes.c_float), - ("eob", ctypes.c_bool), - ] - - -# // Passed to beam_search_callback function. -# // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams -# // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks. -# // These pointers are valid only during the synchronous callback, so should not be saved. -# struct llama_beams_state { -# struct llama_beam_view * beam_views; -# size_t n_beams; // Number of elements in beam_views[]. -# size_t common_prefix_length; // Current max length of prefix tokens shared by all beams. -# bool last_call; // True iff this is the last callback invocation. -# }; -class llama_beams_state(ctypes.Structure): - if TYPE_CHECKING: - beam_views: CtypesArray[llama_beam_view] - n_beams: int - common_prefix_length: int - last_call: bool - - _fields_ = [ - ("beam_views", ctypes.POINTER(llama_beam_view)), - ("n_beams", ctypes.c_size_t), - ("common_prefix_length", ctypes.c_size_t), - ("last_call", ctypes.c_bool), - ] - - -# // Type of pointer to the beam_search_callback function. -# // void* callback_data is any custom data passed to llama_beam_search, that is subsequently -# // passed back to beam_search_callback. This avoids having to use global variables in the callback. -# typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); -llama_beam_search_callback_fn_t = ctypes.CFUNCTYPE( - None, ctypes.c_void_p, llama_beams_state -) - - -# /// @details Deterministically returns entire sentence constructed by a beam search. -# /// @param ctx Pointer to the llama_context. -# /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state. -# /// @param callback_data A pointer that is simply passed back to callback. -# /// @param n_beams Number of beams to use. -# /// @param n_past Number of tokens already evaluated. -# /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier. -# /// @param n_threads Number of threads as passed to llama_eval(). -# LLAMA_API void llama_beam_search( -# struct llama_context * ctx, -# llama_beam_search_callback_fn_t callback, -# void * callback_data, -# size_t n_beams, -# int32_t n_past, -# int32_t n_predict); -@ctypes_function( - "llama_beam_search", - [ - llama_context_p_ctypes, - llama_beam_search_callback_fn_t, - ctypes.c_void_p, - ctypes.c_size_t, - ctypes.c_int32, - ctypes.c_int32, - ], - None, -) -def llama_beam_search( - ctx: llama_context_p, - callback: CtypesFuncPointer, - callback_data: ctypes.c_void_p, - n_beams: Union[ctypes.c_size_t, int], - n_past: Union[ctypes.c_int, int], - n_predict: Union[ctypes.c_int, int], - /, -): ... - - # /// @details Build a split GGUF final path for this chunk. # /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" # // Returns the split_path length. diff --git a/vendor/llama.cpp b/vendor/llama.cpp index bde7cd3cd..ee459f40f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit bde7cd3cd949c1a85d3a199498ac98e78039d46f +Subproject commit ee459f40f65810a810151b24eba5b8bd174ceffe From 83d6b26e6f45c31b5b55c4b5e5570397540d1729 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 8 Jun 2024 23:14:22 -0400 Subject: [PATCH 418/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ee459f40f..5795b9418 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ee459f40f65810a810151b24eba5b8bd174ceffe +Subproject commit 5795b941827fdec6c1662986de962badff456718 From 1615eb9e5b74fe00eaa065520796b1af18c94557 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 10 Jun 2024 11:05:45 -0400 Subject: [PATCH 419/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5795b9418..fd5ea0f89 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5795b941827fdec6c1662986de962badff456718 +Subproject commit fd5ea0f897ecb3659d6c269ef6f3d833e865ead7 From 86a38ad4a02933fc7a1d86dac865b790dd7da1e6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 10 Jun 2024 11:14:33 -0400 Subject: [PATCH 420/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 1b9865da0..7b1998502 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.78] + +- feat: Update llama.cpp to ggerganov/llama.cpp@fd5ea0f897ecb3659d6c269ef6f3d833e865ead7 +- fix: Avoid duplicate special tokens in chat formats by @CISC in #1439 +- fix: fix logprobs when BOS is not present by @ghorbani in #1471 +- feat: adding rpc_servers parameter to Llama class by @chraac in #1477 + ## [0.2.77] - feat: Update llama.cpp to ggerganov/llama.cpp@bde7cd3cd949c1a85d3a199498ac98e78039d46f diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 3df249f5e..5dd2b43cd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.77" \ No newline at end of file +__version__ = "0.2.78" \ No newline at end of file From e342161371c85fa1bd68dca2727cd633480a6524 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 13 Jun 2024 03:38:11 -0400 Subject: [PATCH 421/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index fd5ea0f89..f578b86b2 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit fd5ea0f897ecb3659d6c269ef6f3d833e865ead7 +Subproject commit f578b86b2123d0f92afbaa98a031df4d4464e582 From dbcf64cf0796c27a241991ef6fd0e52283eeb462 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Thu, 13 Jun 2024 09:45:24 +0200 Subject: [PATCH 422/624] feat: Support SPM infill (#1492) * Support SPM infill * typo-- * one less layer of parenthesis necessary * new required internals * manually add bos/eos if model requires it * add bos even when unknown This is identical behaviour to llama.cpp I guess any model that doesn't use BOS is recent enough to have the add_bos_token metadata. * don't add bos/eos on non-infill pre-tokenized prompt * add tokenizer hack to remove leading space in suffix * I keep forgetting metadata are strings * check if bos exists * add example * add cls/sep instead of bos/eos for WPM vocab * simplify * color-code filtered suffix --------- Co-authored-by: Andrei Betlen --- .../high_level_api/high_level_api_infill.py | 33 +++++++++ llama_cpp/_internals.py | 8 ++ llama_cpp/llama.py | 73 ++++++++++++------- 3 files changed, 87 insertions(+), 27 deletions(-) create mode 100644 examples/high_level_api/high_level_api_infill.py diff --git a/examples/high_level_api/high_level_api_infill.py b/examples/high_level_api/high_level_api_infill.py new file mode 100644 index 000000000..27af6367e --- /dev/null +++ b/examples/high_level_api/high_level_api_infill.py @@ -0,0 +1,33 @@ +import argparse + +from llama_cpp import Llama + +parser = argparse.ArgumentParser() +parser.add_argument("-m", "--model", type=str, default="../models/7B/ggml-models.bin") +parser.add_argument("-p", "--prompt", type=str, default="def add(") +parser.add_argument("-s", "--suffix", type=str, default="\n return sum\n\n") +parser.add_argument("-i", "--spm-infill", action='store_true') +args = parser.parse_args() + +llm = Llama(model_path=args.model, n_gpu_layers=-1, spm_infill=args.spm_infill) + +output = llm.create_completion( + temperature = 0.0, + repeat_penalty = 1.0, + prompt = args.prompt, + suffix = args.suffix, +) + +# Models sometimes repeat suffix in response, attempt to filter that +response = output["choices"][0]["text"] +response_stripped = response.rstrip() +unwanted_response_suffix = args.suffix.rstrip() +unwanted_response_length = len(unwanted_response_suffix) + +filtered = False +if unwanted_response_suffix and response_stripped[-unwanted_response_length:] == unwanted_response_suffix: + response = response_stripped[:-unwanted_response_length] + filtered = True + +print(f"Fill-in-Middle completion{' (filtered)' if filtered else ''}:\n\n{args.prompt}\033[32m{response}\033[{'33' if filtered else '0'}m{args.suffix}\033[0m") + diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index ba0429139..27452c324 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -170,6 +170,14 @@ def token_eot(self) -> int: assert self.model is not None return llama_cpp.llama_token_eot(self.model) + def add_bos_token(self) -> int: + assert self.model is not None + return llama_cpp.llama_add_bos_token(self.model) + + def add_eos_token(self) -> int: + assert self.model is not None + return llama_cpp.llama_add_eos_token(self.model) + # Tokenization def tokenize(self, text: bytes, add_bos: bool, special: bool): diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index bf3bd656d..f15e0ef57 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -115,6 +115,7 @@ def __init__( type_k: Optional[int] = None, type_v: Optional[int] = None, # Misc + spm_infill: bool = False, verbose: bool = True, # Extra Params **kwargs, # type: ignore @@ -185,6 +186,7 @@ def __init__( verbose: Print verbose output to stderr. type_k: KV cache data type for K (default: f16) type_v: KV cache data type for V (default: f16) + spm_infill: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. Raises: ValueError: If the model path does not exist. @@ -343,6 +345,8 @@ def __init__( self.lora_scale = lora_scale self.lora_path = lora_path + self.spm_infill = spm_infill + if not os.path.exists(model_path): raise ValueError(f"Model path does not exist: {model_path}") @@ -972,14 +976,33 @@ def _create_completion( completion_id: str = f"cmpl-{str(uuid.uuid4())}" created: int = int(time.time()) + bos_token_id: int = self.token_bos() + cls_token_id: int = self._model.token_cls() + sep_token_id: int = self._model.token_sep() prefix_token_id: int = self._model.token_prefix() middle_token_id: int = self._model.token_middle() suffix_token_id: int = self._model.token_suffix() + add_space_prefix: bool = self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true" + bos_tokens: List[int] = [cls_token_id if cls_token_id != -1 else bos_token_id] + eos_tokens: List[int] = [sep_token_id if sep_token_id != -1 else self.token_eos()] + + if (isinstance(prompt, list) and suffix is None) or self._model.add_bos_token() == 0 or bos_tokens[:1] == [-1]: + bos_tokens = [] + + if (isinstance(prompt, list) and suffix is None) or (self._model.add_eos_token() != 1 and sep_token_id == -1): + eos_tokens = [] + + suffix_space_prefix: int = 0 + # Tokenizer hack to remove leading space + if add_space_prefix and suffix_token_id >= 0 and suffix: + suffix = "☺" + suffix + suffix_space_prefix = 2 + # If prompt is empty, initialize completion with BOS token to avoid # detokenization including a space at the beginning of the completion - completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()] + completion_tokens: List[int] = [] if len(prompt) > 0 else [bos_token_id] # Add blank space to start of prompt to match OG llama tokenizer - prompt_tokens: List[int] = ( + prefix_tokens: List[int] = ( ( [prefix_token_id] if prefix_token_id >= 0 and suffix is not None @@ -988,38 +1011,33 @@ def _create_completion( + ( ( - self.tokenize(prompt.encode("utf-8"), add_bos=(prefix_token_id < 0 or suffix is None), special=(prefix_token_id < 0 or suffix is None)) + self.tokenize(prompt.encode("utf-8"), add_bos=False, special=(prefix_token_id < 0 or suffix is None)) if prompt != "" - else ( - [] - if prefix_token_id >= 0 and suffix is not None - else [self.token_bos()] - ) + else [] ) if isinstance(prompt, str) else prompt ) - + + ) + suffix_tokens: List[int] = ( ( + [suffix_token_id] + + ( - [suffix_token_id] - + - ( - self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False) - if suffix - else [] - ) + self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)[suffix_space_prefix:] + if suffix + else [] ) - if suffix_token_id >= 0 and suffix is not None - else [] - ) - + - ( - [middle_token_id] - if middle_token_id >= 0 and suffix is not None - else [] ) + if suffix_token_id >= 0 and suffix is not None + else [] + ) + middle_tokens: List[int] = ( + [middle_token_id] + if middle_token_id >= 0 and suffix is not None + else [] ) + prompt_tokens: List[int] = bos_tokens + ((suffix_tokens + prefix_tokens + middle_tokens) if self.spm_infill else (prefix_tokens + suffix_tokens + middle_tokens)) + eos_tokens text: bytes = b"" returned_tokens: int = 0 stop = ( @@ -1176,7 +1194,7 @@ def logit_bias_processor( # not sure how to handle this branch when dealing # with CJK output, so keep it unchanged for token in remaining_tokens: - if token == self.token_bos(): + if token == bos_token_id: continue token_end_position += len(self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens])) # Check if stop sequence is in the token @@ -1303,7 +1321,7 @@ def logit_bias_processor( logprobs_or_none: Optional[CompletionLogprobs] = None if logprobs is not None: - if token == self.token_bos(): + if token == bos_token_id: continue token_str = self.detokenize([token]).decode( "utf-8", errors="ignore" @@ -1431,7 +1449,7 @@ def logit_bias_processor( for idx, (token, token_str, logprobs_token) in enumerate( zip(all_tokens, all_token_strs, all_logprobs) ): - if token == self.token_bos(): + if token == bos_token_id: continue text_offsets.append( text_offset @@ -1858,6 +1876,7 @@ def __getstate__(self): type_k=self.context_params.type_k, type_v=self.context_params.type_v, # Misc + spm_infill=self.spm_infill, verbose=self.verbose, ) From 320a5d7ea5ebcc5a8dc114d9bdf7c584537b04b0 Mon Sep 17 00:00:00 2001 From: Junpei Kawamoto Date: Thu, 13 Jun 2024 02:16:14 -0600 Subject: [PATCH 423/624] feat: Add `.close()` method to `Llama` class to explicitly free model from memory (#1513) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * feat: add explicit methods to free model This commit introduces a `close` method to both `Llama` and `_LlamaModel`, allowing users to explicitly free the model from RAM/VRAM. The previous implementation relied on the destructor of `_LlamaModel` to free the model. However, in Python, the timing of destructor calls is unclear—for instance, the `del` statement does not guarantee immediate invocation of the destructor. This commit provides an explicit method to release the model, which works immediately and allows the user to load another model without memory issues. Additionally, this commit implements a context manager in the `Llama` class, enabling the automatic closure of the `Llama` object when used with the `with` statement. * feat: Implement ContextManager in _LlamaModel, _LlamaContext, and _LlamaBatch This commit enables automatic resource management by implementing the `ContextManager` protocol in `_LlamaModel`, `_LlamaContext`, and `_LlamaBatch`. This ensures that resources are properly managed and released within a `with` statement, enhancing robustness and safety in resource handling. * feat: add ExitStack for Llama's internal class closure This update implements ExitStack to manage and close internal classes in Llama, enhancing efficient and safe resource management. * Use contextlib ExitStack and closing * Explicitly free model when closing resources on server --------- Co-authored-by: Andrei Betlen --- llama_cpp/_internals.py | 56 ++++++++++++++++++++++----------------- llama_cpp/llama.py | 21 ++++++++++----- llama_cpp/server/model.py | 3 +++ 3 files changed, 50 insertions(+), 30 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 27452c324..ee990d474 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -9,6 +9,7 @@ Sequence, ) from dataclasses import dataclass, field +from contextlib import ExitStack import numpy as np import numpy.typing as npt @@ -27,9 +28,6 @@ class _LlamaModel: """Intermediate Python wrapper for a llama.cpp llama_model. NOTE: For stability it's recommended you use the Llama class instead.""" - _llama_free_model = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - def __init__( self, *, @@ -40,8 +38,7 @@ def __init__( self.path_model = path_model self.params = params self.verbose = verbose - - self._llama_free_model = llama_cpp._lib.llama_free_model # type: ignore + self._exit_stack = ExitStack() self.model = None @@ -56,11 +53,17 @@ def __init__( if self.model is None: raise ValueError(f"Failed to load model from file: {path_model}") - def __del__(self): - if self.model is not None and self._llama_free_model is not None: - self._llama_free_model(self.model) + def free_model(): + if self.model is None: + return + llama_cpp.llama_free_model(self.model) self.model = None + self._exit_stack.callback(free_model) + + def close(self): + self._exit_stack.close() + def vocab_type(self) -> int: assert self.model is not None return llama_cpp.llama_vocab_type(self.model) @@ -257,8 +260,6 @@ class _LlamaContext: """Intermediate Python wrapper for a llama.cpp llama_context. NOTE: For stability it's recommended you use the Llama class instead.""" - _llama_free = None - def __init__( self, *, @@ -269,24 +270,28 @@ def __init__( self.model = model self.params = params self.verbose = verbose + self._exit_stack = ExitStack() - self._llama_free = llama_cpp._lib.llama_free # type: ignore self.ctx = None assert self.model.model is not None - self.ctx = llama_cpp.llama_new_context_with_model( - self.model.model, self.params - ) + self.ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params) if self.ctx is None: raise ValueError("Failed to create llama_context") - def __del__(self): - if self.ctx is not None and self._llama_free is not None: - self._llama_free(self.ctx) + def free_ctx(): + if self.ctx is None: + return + llama_cpp.llama_free(self.ctx) self.ctx = None + self._exit_stack.callback(free_ctx) + + def close(self): + self._exit_stack.close() + def n_ctx(self) -> int: assert self.ctx is not None return llama_cpp.llama_n_ctx(self.ctx) @@ -501,8 +506,6 @@ def default_params(): class _LlamaBatch: - _llama_batch_free = None - def __init__( self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True ): @@ -510,19 +513,24 @@ def __init__( self.embd = embd self.n_seq_max = n_seq_max self.verbose = verbose - - self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore + self._exit_stack = ExitStack() self.batch = None self.batch = llama_cpp.llama_batch_init( self._n_tokens, self.embd, self.n_seq_max ) - def __del__(self): - if self.batch is not None and self._llama_batch_free is not None: - self._llama_batch_free(self.batch) + def free_batch(): + if self.batch is None: + return + llama_cpp.llama_batch_free(self.batch) self.batch = None + self._exit_stack.callback(free_batch) + + def close(self): + self._exit_stack.close() + def n_tokens(self) -> int: assert self.batch is not None return self.batch.n_tokens diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index f15e0ef57..459b29f92 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -9,7 +9,9 @@ import typing import fnmatch import warnings +import contextlib import multiprocessing +from types import TracebackType from typing import ( List, @@ -21,6 +23,7 @@ Deque, Callable, Dict, + Type, ) from collections import deque from pathlib import Path @@ -350,9 +353,11 @@ def __init__( if not os.path.exists(model_path): raise ValueError(f"Model path does not exist: {model_path}") - self._model = _LlamaModel( + self._stack = contextlib.ExitStack() + + self._model = self._stack.enter_context(contextlib.closing(_LlamaModel( path_model=self.model_path, params=self.model_params, verbose=self.verbose - ) + ))) # Override tokenizer self.tokenizer_ = tokenizer or LlamaTokenizer(self) @@ -364,18 +369,18 @@ def __init__( self.context_params.n_ctx = self._model.n_ctx_train() self.context_params.n_batch = self.n_batch - self._ctx = _LlamaContext( + self._ctx = self._stack.enter_context(contextlib.closing(_LlamaContext( model=self._model, params=self.context_params, verbose=self.verbose, - ) + ))) - self._batch = _LlamaBatch( + self._batch = self._stack.enter_context(contextlib.closing(_LlamaBatch( n_tokens=self.n_batch, embd=0, n_seq_max=self.context_params.n_ctx, verbose=self.verbose, - ) + ))) if self.lora_path: if self._model.apply_lora_from_file( @@ -1959,6 +1964,10 @@ def pooling_type(self) -> str: """Return the pooling type.""" return self._ctx.pooling_type() + def close(self) -> None: + """Explicitly free the model from memory.""" + self._stack.close() + @staticmethod def logits_to_logprobs( logits: Union[npt.NDArray[np.single], List], axis: int = -1 diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index d4d4acbe3..ad39c1004 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -44,6 +44,8 @@ def __call__(self, model: Optional[str] = None) -> llama_cpp.Llama: if self._current_model is not None: return self._current_model + if self._current_model: + self._current_model.close() self._current_model = None settings = self._model_settings_dict[model] @@ -65,6 +67,7 @@ def __iter__(self): def free(self): if self._current_model: + self._current_model.close() del self._current_model @staticmethod From 5af81634cb0c47943f5653275b04f7747b89792d Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu, 13 Jun 2024 10:12:02 -0400 Subject: [PATCH 424/624] chore(deps): bump pypa/cibuildwheel from 2.18.1 to 2.19.0 (#1522) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.18.1 to 2.19.0. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.18.1...v2.19.0) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 957912d09..ebf52b337 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.18.1 + uses: pypa/cibuildwheel@v2.19.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -56,7 +56,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.18.1 + uses: pypa/cibuildwheel@v2.19.0 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" From 9e396b3ebd997e3155b5bb8ab3ab162dc26da5c5 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Thu, 13 Jun 2024 16:19:57 +0200 Subject: [PATCH 425/624] feat: Update workflows and pre-built wheels (#1416) * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * revert * Bump pyhton from 3.8 to 3.9 * Remove python 3.8 * Remove Python 3.7 and 3.8 deprecated * Bump python from 3.8 to 3.9 * Add python 3.9 * Add python 3.9, remove macos-11 deprecated, add macos-14 * Bump python 3.8 to 3.9 * Add python 3.13 * Add python 3.13 * python 3.13 remove * remove python 3.13 * remove python 3.8 * Bump macos-13 to macos-14 * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update generate-index-from-release.yaml Add avx, avx2 and avx512 * Update test.yaml * Update test-pypi.yaml * Update publish.yaml * Update publish-to-test.yaml * Update build-wheels-cuda.yaml Cuda with AVX2 by default * Update build-wheels-cuda.yaml * remove DEPRECATED 32 bits * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml Upgrade matrix os to latest version * Update build-wheels-metal.yaml * Update build-wheels-cuda.yaml * Update test.yaml * Update test-pypi.yaml * Update test.yaml Add cache: 'pip' * Update publish-to-test.yaml * Update build-wheels-metal.yaml Add cache: 'pip' * Update build-wheels-cuda.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml remove x86_64 * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * revert * Remove cpu variants --------- Co-authored-by: Andrei Betlen --- .github/workflows/build-wheels-cuda.yaml | 23 +++--- .github/workflows/build-wheels-metal.yaml | 91 +++++++++-------------- .github/workflows/publish-to-test.yaml | 9 ++- .github/workflows/publish.yaml | 8 +- .github/workflows/test-pypi.yaml | 27 ++++--- .github/workflows/test.yaml | 47 ++++++------ 6 files changed, 96 insertions(+), 109 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index ae9e8632c..b8496d850 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -20,8 +20,8 @@ jobs: id: set-matrix run: | $matrix = @{ - 'os' = @('ubuntu-20.04', 'windows-latest') - 'pyver' = @("3.10", "3.11", "3.12") + 'os' = @('ubuntu-latest', 'windows-latest') + 'pyver' = @("3.9", "3.10", "3.11", "3.12") 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") 'releasetag' = @("basic") } @@ -50,6 +50,7 @@ jobs: - uses: actions/setup-python@v5 with: python-version: ${{ matrix.pyver }} + cache: 'pip' - name: Setup Mamba uses: conda-incubator/setup-miniconda@v3.0.4 @@ -109,15 +110,15 @@ jobs: $env:VERBOSE = '1' $env:CMAKE_ARGS = '-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all' $env:CMAKE_ARGS = "-DLLAMA_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS" - if ($env:AVXVER -eq 'AVX') { - $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' - } - if ($env:AVXVER -eq 'AVX512') { - $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on' - } - if ($env:AVXVER -eq 'basic') { - $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' - } + # if ($env:AVXVER -eq 'AVX') { + $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' + # } + # if ($env:AVXVER -eq 'AVX512') { + # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on' + # } + # if ($env:AVXVER -eq 'basic') { + # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' + # } python -m build --wheel # write the build tag to the output Write-Output "CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index fc798c84e..f007eb398 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -6,81 +6,60 @@ permissions: contents: write jobs: - define_matrix: - name: Define Build Matrix - runs-on: ubuntu-latest - outputs: - matrix: ${{ steps.set-matrix.outputs.matrix }} - defaults: - run: - shell: pwsh - - steps: - - name: Define Job Output - id: set-matrix - run: | - $matrix = @{ - 'os' = @('macos-11', 'macos-12', 'macos-13') - 'pyver' = @('3.10', '3.11', '3.12') - } - - $matrixOut = ConvertTo-Json $matrix -Compress - Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT - build_wheels: - name: ${{ matrix.os }} Python ${{ matrix.pyver }} - needs: define_matrix + name: Build wheels on ${{ matrix.os }} runs-on: ${{ matrix.os }} strategy: - matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }} - env: - OSVER: ${{ matrix.os }} + matrix: + os: [macos-12, macos-13, macos-14] steps: - uses: actions/checkout@v4 with: submodules: "recursive" + # Used to host cibuildwheel - uses: actions/setup-python@v5 with: - python-version: ${{ matrix.pyver }} - - - name: Install Dependencies - run: | - python -m pip install build wheel cmake + python-version: "3.12" + cache: 'pip' - - name: Build Wheel + - name: Install dependencies run: | - XCODE15PATH="/Applications/Xcode_15.0.app/Contents/Developer" - XCODE15BINPATH="${XCODE15PATH}/Toolchains/XcodeDefault.xctoolchain/usr/bin" - export CMAKE_ARGS="-DLLAMA_NATIVE=off -DLLAMA_METAL=on" - [[ "$OSVER" == "macos-13" ]] && export CC="${XCODE15BINPATH}/cc" && export CXX="${XCODE15BINPATH}/c++" && export MACOSX_DEPLOYMENT_TARGET="13.0" - [[ "$OSVER" == "macos-12" ]] && export MACOSX_DEPLOYMENT_TARGET="12.0" - [[ "$OSVER" == "macos-11" ]] && export MACOSX_DEPLOYMENT_TARGET="11.0" + python -m pip install --upgrade pip + python -m pip install -e .[all] - export CMAKE_OSX_ARCHITECTURES="arm64" && export ARCHFLAGS="-arch arm64" - VERBOSE=1 python -m build --wheel - - if [[ "$OSVER" == "macos-13" ]]; then - export SDKROOT="${XCODE15PATH}/Platforms/MacOSX.platform/Developer/SDKs/MacOSX14.0.sdk" - export MACOSX_DEPLOYMENT_TARGET="14.0" - VERBOSE=1 python -m build --wheel - fi - - for file in ./dist/*.whl; do cp "$file" "${file/arm64.whl/aarch64.whl}"; done + - name: Build wheels + uses: pypa/cibuildwheel@v2.18.1 + env: + # disable repair + CIBW_REPAIR_WHEEL_COMMAND: "" + CIBW_ARCHS: "arm64" + CIBW_ENVIRONMENT: CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on" + CIBW_BUILD: "cp39-* cp310-* cp311-* cp312-*" + with: + package-dir: . + output-dir: wheelhouse2 - export CMAKE_OSX_ARCHITECTURES="x86_64" && export CMAKE_ARGS="-DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_METAL=on" && export ARCHFLAGS="-arch x86_64" - VERBOSE=1 python -m build --wheel + - uses: actions/upload-artifact@v4 + with: + name: wheels-mac_${{ matrix.os }} + path: ./wheelhouse2/*.whl - if [[ "$OSVER" == "macos-13" ]]; then - export SDKROOT="${XCODE15PATH}/Platforms/MacOSX.platform/Developer/SDKs/MacOSX14.0.sdk" - export MACOSX_DEPLOYMENT_TARGET="14.0" - VERBOSE=1 python -m build --wheel - fi + release: + name: Release + needs: [build_wheels] + runs-on: ubuntu-latest + steps: + - uses: actions/download-artifact@v4 + with: + merge-multiple: true + path: dist2 + - uses: softprops/action-gh-release@v2 with: - files: dist/* + files: dist2/* # set release name to -metal tag_name: ${{ github.ref_name }}-metal env: diff --git a/.github/workflows/publish-to-test.yaml b/.github/workflows/publish-to-test.yaml index 2bf0ea9ba..19613233b 100644 --- a/.github/workflows/publish-to-test.yaml +++ b/.github/workflows/publish-to-test.yaml @@ -22,7 +22,8 @@ jobs: - name: Set up Python uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.11" + cache: 'pip' - name: Append Dev Version to __version__ run: | DEV_VERSION=${{ github.event.inputs.dev_version }} @@ -31,11 +32,11 @@ jobs: sed -i 's/__version__ = \".*\"/__version__ = \"'"${NEW_VERSION}"'\"/' llama_cpp/__init__.py - name: Install dependencies run: | - python3 -m pip install --upgrade pip build - python3 -m pip install -e .[all] + python -m pip install --upgrade pip build + python -m pip install -e .[all] - name: Build source distribution run: | - python3 -m build --sdist + python -m build --sdist - name: Publish to Test PyPI uses: pypa/gh-action-pypi-publish@release/v1 with: diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index bc4ec9063..c6abb43b3 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -16,14 +16,14 @@ jobs: - name: Set up Python uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.9" - name: Install dependencies run: | - python3 -m pip install --upgrade pip build - python3 -m pip install -e .[all] + python -m pip install --upgrade pip build + python -m pip install -e .[all] - name: Build source distribution run: | - python3 -m build --sdist + python -m build --sdist - name: Publish distribution to PyPI # TODO: move to tag based releases # if: startsWith(github.ref, 'refs/tags') diff --git a/.github/workflows/test-pypi.yaml b/.github/workflows/test-pypi.yaml index aa8e8fa0b..d7131956d 100644 --- a/.github/workflows/test-pypi.yaml +++ b/.github/workflows/test-pypi.yaml @@ -8,57 +8,60 @@ jobs: runs-on: ubuntu-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + cache: 'pip' - name: Install dependencies run: | - python3 -m pip install --upgrade pip - python3 -m pip install --verbose llama-cpp-python[all] + python -m pip install --upgrade pip + python -m pip install --verbose llama-cpp-python[all] - name: Test with pytest run: | - python3 -c "import llama_cpp" + python -c "import llama_cpp" build-windows: runs-on: windows-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + cache: 'pip' - name: Install dependencies run: | - python3 -m pip install --upgrade pip - python3 -m pip install --verbose llama-cpp-python[all] + python -m pip install --upgrade pip + python -m pip install --verbose llama-cpp-python[all] - name: Test with pytest run: | - python3 -c "import llama_cpp" + python -c "import llama_cpp" build-macos: runs-on: macos-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + cache: 'pip' - name: Install dependencies run: | - python3 -m pip install --upgrade pip - python3 -m pip install --verbose llama-cpp-python[all] + python -m pip install --upgrade pip + python -m pip install --verbose llama-cpp-python[all] - name: Test with pytest run: | - python3 -c "import llama_cpp" + python -c "import llama_cpp" diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 292343c2b..6d8231d18 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -14,7 +14,7 @@ jobs: runs-on: ubuntu-latest strategy: matrix: - python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - uses: actions/checkout@v4 @@ -24,20 +24,21 @@ jobs: uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + cache: 'pip' - name: Install dependencies run: | - python3 -m pip install --upgrade pip - python3 -m pip install .[all] -v + python -m pip install --upgrade pip + python -m pip install .[all] -v - name: Test with pytest run: | - python3 -m pytest + python -m pytest build-windows: runs-on: windows-latest strategy: matrix: - python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - uses: actions/checkout@v4 @@ -47,20 +48,21 @@ jobs: uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + cache: 'pip' - name: Install dependencies run: | - python3 -m pip install --upgrade pip - python3 -m pip install .[all] -v + python -m pip install --upgrade pip + python -m pip install .[all] -v - name: Test with pytest run: | - python3 -m pytest + python -m pytest build-macos: - runs-on: macos-13 + runs-on: macos-latest strategy: matrix: - python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - uses: actions/checkout@v4 @@ -70,13 +72,14 @@ jobs: uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + cache: 'pip' - name: Install dependencies run: | - python3 -m pip install --upgrade pip - python3 -m pip install .[all] --verbose + python -m pip install --upgrade pip + python -m pip install .[all] --verbose - name: Test with pytest run: | - python3 -m pytest + python -m pytest # build-linux-opencl: @@ -98,29 +101,29 @@ jobs: # sudo apt-get install -y --no-install-recommends llvm intel-oneapi-runtime-opencl intel-oneapi-runtime-compilers libclblast-dev # - name: Install dependencies # run: | - # python3 -m pip install --upgrade pip - # CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install .[all] --verbose + # python -m pip install --upgrade pip + # CMAKE_ARGS="-DLLAMA_CLBLAST=on" python -m pip install .[all] --verbose # - name: Test with pytest # run: | - # python3 -m pytest + # python -m pytest build-macos-metal: - runs-on: macos-13 + runs-on: macos-latest steps: - uses: actions/checkout@v4 with: submodules: "recursive" - - name: Set up Python 3.8 + - name: Set up Python 3.9 uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.9" - name: Install dependencies run: | - python3 -m pip install --upgrade pip - CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install .[all] --verbose + python -m pip install --upgrade pip + CMAKE_ARGS="-DLLAMA_METAL=on" python -m pip install .[all] --verbose - name: Test with pytest run: | - python3 -m pytest + python -m pytest From 8401c6f2d133543350e14de6076fe7ec9efe6653 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 13 Jun 2024 11:31:31 -0400 Subject: [PATCH 426/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f578b86b2..172c82568 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f578b86b2123d0f92afbaa98a031df4d4464e582 +Subproject commit 172c8256840ffd882ab9992ecedbb587d9b21f15 From f4491c4903f791b74de9e98e9dc03b87fcc9dfbb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 17 Jun 2024 11:56:03 -0400 Subject: [PATCH 427/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 6f37fcb69..6d24324f0 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -301,6 +301,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_OLMO = 12, # LLAMA_VOCAB_PRE_TYPE_DBRX = 13, # LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, +# LLAMA_VOCAB_PRE_TYPE_PORO = 15, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -317,6 +318,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_OLMO = 12 LLAMA_VOCAB_PRE_TYPE_DBRX = 13 LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 +LLAMA_VOCAB_PRE_TYPE_PORO = 15 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 172c82568..b473e9508 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 172c8256840ffd882ab9992ecedbb587d9b21f15 +Subproject commit b473e95084c286780165568cf0f385f21141d68d From 4c1d74c0aecb5fe263674cc761573654b16d4c20 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 19 Jun 2024 10:07:20 -0400 Subject: [PATCH 428/624] fix: Make destructor to automatically call .close() method on Llama class. --- llama_cpp/_internals.py | 9 +++++++++ llama_cpp/llama.py | 3 +++ 2 files changed, 12 insertions(+) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index ee990d474..423c1b5bc 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -64,6 +64,9 @@ def free_model(): def close(self): self._exit_stack.close() + def __del__(self): + self.close() + def vocab_type(self) -> int: assert self.model is not None return llama_cpp.llama_vocab_type(self.model) @@ -292,6 +295,9 @@ def free_ctx(): def close(self): self._exit_stack.close() + def __del__(self): + self.close() + def n_ctx(self) -> int: assert self.ctx is not None return llama_cpp.llama_n_ctx(self.ctx) @@ -531,6 +537,9 @@ def free_batch(): def close(self): self._exit_stack.close() + def __del__(self): + self.close() + def n_tokens(self) -> int: assert self.batch is not None return self.batch.n_tokens diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 459b29f92..5f5396683 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1968,6 +1968,9 @@ def close(self) -> None: """Explicitly free the model from memory.""" self._stack.close() + def __del__(self) -> None: + self.close() + @staticmethod def logits_to_logprobs( logits: Union[npt.NDArray[np.single], List], axis: int = -1 From 554fd08e7d90f1c49cc513febdc79cc723ceb512 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 19 Jun 2024 10:07:28 -0400 Subject: [PATCH 429/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index b473e9508..9c77ec1d7 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit b473e95084c286780165568cf0f385f21141d68d +Subproject commit 9c77ec1d74874ee22bdef8f110e8e8d41389abf2 From 6c331909ca365cb5aa490288a02853f1bb1d6979 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 19 Jun 2024 10:10:01 -0400 Subject: [PATCH 430/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 7b1998502..42b14c862 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.79] + +- feat: Update llama.cpp to ggerganov/llama.cpp@9c77ec1d74874ee22bdef8f110e8e8d41389abf2 +- feat(ci): Update workflows and pre-built wheels by @Smartappli in #1416 +- feat: Add .close() method to Llama class to explicitly free model from memory by @jkawamoto in #1513 +- feat: Support SPM infill by @CISC in #1492 + ## [0.2.78] - feat: Update llama.cpp to ggerganov/llama.cpp@fd5ea0f897ecb3659d6c269ef6f3d833e865ead7 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 5dd2b43cd..71371b3b0 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.78" \ No newline at end of file +__version__ = "0.2.79" \ No newline at end of file From d98a24a25b49b02007071dc3e4bc2cc9a0212935 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 20 Jun 2024 10:50:40 -0400 Subject: [PATCH 431/624] docs: Remove references to deprecated opencl backend. Closes #1512 --- Makefile | 3 --- README.md | 11 ----------- 2 files changed, 14 deletions(-) diff --git a/Makefile b/Makefile index d8fb0cc27..90f562fa1 100644 --- a/Makefile +++ b/Makefile @@ -24,9 +24,6 @@ build.debug: build.cuda: CMAKE_ARGS="-DLLAMA_CUDA=on" python3 -m pip install --verbose -e . -build.opencl: - CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e . - build.openblas: CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" python3 -m pip install --verbose -e . diff --git a/README.md b/README.md index 0f7abfb28..4c37ba3e4 100644 --- a/README.md +++ b/README.md @@ -165,17 +165,6 @@ pip install llama-cpp-python \ --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal ``` - -
- -CLBlast (OpenCL) - -To install with CLBlast, set the `LLAMA_CLBLAST=on` environment variable before installing: - -```bash -CMAKE_ARGS="-DLLAMA_CLBLAST=on" pip install llama-cpp-python -``` -
From 5beec1a1fd28acf6e32d45a6b990ef6389d288ed Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 21 Jun 2024 12:09:14 -0400 Subject: [PATCH 432/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 9c77ec1d7..557b653dc 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 9c77ec1d74874ee22bdef8f110e8e8d41389abf2 +Subproject commit 557b653dc9ed91e8c313e87500e0050c775f81b6 From 27d53589ff884559919b4bd2c40690223feaaffa Mon Sep 17 00:00:00 2001 From: Jon Craton Date: Fri, 21 Jun 2024 12:10:15 -0400 Subject: [PATCH 433/624] docs: Update readme examples to use newer Qwen2 model (#1544) --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 4c37ba3e4..4a710d4e5 100644 --- a/README.md +++ b/README.md @@ -327,7 +327,7 @@ You'll need to install the `huggingface-hub` package to use this feature (`pip i ```python llm = Llama.from_pretrained( - repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", + repo_id="Qwen/Qwen2-0.5B-Instruct-GGUF", filename="*q8_0.gguf", verbose=False ) @@ -688,7 +688,7 @@ For possible options, see [llama_cpp/llama_chat_format.py](llama_cpp/llama_chat_ If you have `huggingface-hub` installed, you can also use the `--hf_model_repo_id` flag to load a model from the Hugging Face Hub. ```bash -python3 -m llama_cpp.server --hf_model_repo_id Qwen/Qwen1.5-0.5B-Chat-GGUF --model '*q8_0.gguf' +python3 -m llama_cpp.server --hf_model_repo_id Qwen/Qwen2-0.5B-Instruct-GGUF --model '*q8_0.gguf' ``` ### Web Server Features From 398fe81547d2b36a57730d19e66fb69184054958 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri, 21 Jun 2024 12:10:34 -0400 Subject: [PATCH 434/624] chore(deps): bump docker/build-push-action from 5 to 6 (#1539) Bumps [docker/build-push-action](https://github.com/docker/build-push-action) from 5 to 6. - [Release notes](https://github.com/docker/build-push-action/releases) - [Commits](https://github.com/docker/build-push-action/compare/v5...v6) --- updated-dependencies: - dependency-name: docker/build-push-action dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-docker.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-docker.yaml b/.github/workflows/build-docker.yaml index 4ebe3bb6d..b5c7346db 100644 --- a/.github/workflows/build-docker.yaml +++ b/.github/workflows/build-docker.yaml @@ -31,7 +31,7 @@ jobs: - name: Build and push id: docker_build - uses: docker/build-push-action@v5 + uses: docker/build-push-action@v6 with: context: . file: "docker/simple/Dockerfile" From 35c980eb2ece665a8d3b636968358bd82f3f27bd Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri, 21 Jun 2024 12:10:43 -0400 Subject: [PATCH 435/624] chore(deps): bump pypa/cibuildwheel from 2.18.1 to 2.19.1 (#1527) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.18.1 to 2.19.1. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.18.1...v2.19.1) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 4 ++-- .github/workflows/build-wheels-metal.yaml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index ebf52b337..34e9dc139 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.19.0 + uses: pypa/cibuildwheel@v2.19.1 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -56,7 +56,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.19.0 + uses: pypa/cibuildwheel@v2.19.1 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index f007eb398..11ab795c8 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -30,7 +30,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.18.1 + uses: pypa/cibuildwheel@v2.19.1 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" From 04959f1884c8ef93bd5a4aa40ff0accb8438c0c1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 21 Jun 2024 16:56:15 -0400 Subject: [PATCH 436/624] feat: Update llama_cpp.py bindings --- llama_cpp/llama_cpp.py | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 6d24324f0..1116b1f51 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -468,11 +468,13 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_POOLING_TYPE_NONE = 0, # LLAMA_POOLING_TYPE_MEAN = 1, # LLAMA_POOLING_TYPE_CLS = 2, +# LLAMA_POOLING_TYPE_LAST = 3, # }; LLAMA_POOLING_TYPE_UNSPECIFIED = -1 LLAMA_POOLING_TYPE_NONE = 0 LLAMA_POOLING_TYPE_MEAN = 1 LLAMA_POOLING_TYPE_CLS = 2 +LLAMA_POOLING_TYPE_LAST = 3 # enum llama_split_mode { # LLAMA_SPLIT_MODE_NONE = 0, // single GPU @@ -761,7 +763,6 @@ class llama_model_params(ctypes.Structure): # enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` # enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id -# // (ignored if no pooling layer) # // ref: https://github.com/ggerganov/llama.cpp/pull/2054 # float rope_freq_base; // RoPE base frequency, 0 = from model @@ -2316,6 +2317,16 @@ def llama_n_threads_batch(ctx: llama_context_p, /) -> int: ... +# // Set whether the model is in embeddings model or not +# // If true, embeddings will be returned but logits will not +# LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings); +@ctypes_function("llama_set_embeddings", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_embeddings(ctx: llama_context_p, embeddings: bool, /): + """Set whether the model is in embeddings model or not + If true, embeddings will be returned but logits will not""" + ... + + # // Set whether to use causal attention or not # // If set to true, the model will only attend to the past tokens # LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn); From 117cbb2f531eee1db284440097866be7d7b153c4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 21:28:11 -0400 Subject: [PATCH 437/624] feat: Update llama.cpp --- CMakeLists.txt | 69 ++++++++++++++++++++++-------------------- llama_cpp/llama_cpp.py | 7 ++++- vendor/llama.cpp | 2 +- 3 files changed, 44 insertions(+), 34 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 4b9508d40..353a8d996 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -5,9 +5,28 @@ project(llama_cpp) option(LLAMA_BUILD "Build llama.cpp shared library and install alongside python package" ON) option(LLAVA_BUILD "Build llava shared library and install alongside python package" ON) +if(SKBUILD_STATE STREQUAL "editable") + # Install into the source directory + # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 + set(LLAMA_CPP_PYTHON_INSTALL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp) +else() + set(LLAMA_CPP_PYTHON_INSTALL_DIR ${SKBUILD_PLATLIB_DIR}/llama_cpp) +endif() + if (LLAMA_BUILD) set(BUILD_SHARED_LIBS "On") + set(CMAKE_SKIP_BUILD_RPATH FALSE) + + # When building, don't use the install RPATH already + # (but later on when installing) + set(CMAKE_BUILD_WITH_INSTALL_RPATH FALSE) + + # Add the automatically determined parts of the RPATH + # which point to directories outside the build tree to the install RPATH + set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) + set(CMAKE_INSTALL_RPATH "${LLAMA_CPP_PYTHON_INSTALL_DIR}") + # Building llama if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") # Need to disable these llama.cpp flags on Apple x86_64, @@ -24,31 +43,26 @@ if (LLAMA_BUILD) add_subdirectory(vendor/llama.cpp) install( - TARGETS llama - LIBRARY DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RUNTIME DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ARCHIVE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - FRAMEWORK DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RESOURCE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp + TARGETS llama + LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + RUNTIME DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + ARCHIVE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) - # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 install( - TARGETS llama - LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RUNTIME DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - ARCHIVE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp + TARGETS ggml + LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + RUNTIME DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + ARCHIVE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 if (WIN32 AND (LLAMA_CUDA OR LLAMA_CUBLAS)) install( FILES $ - DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ) - install( - FILES $ - DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp + DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) endif() @@ -71,20 +85,11 @@ if (LLAMA_BUILD) endif() install( TARGETS llava_shared - LIBRARY DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RUNTIME DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ARCHIVE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - FRAMEWORK DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RESOURCE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ) - # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 - install( - TARGETS llava_shared - LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RUNTIME DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - ARCHIVE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp + LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + RUNTIME DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + ARCHIVE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) endif() endif() diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1116b1f51..9014e0268 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -273,6 +273,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_TYPE_SPM = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback # LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE # LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece +# LLAMA_VOCAB_TYPE_UGM = 4, // T5 tokenizer based on Unigram # }; LLAMA_VOCAB_TYPE_NONE = 0 """For models without vocab""" @@ -282,6 +283,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa """GPT-2 tokenizer based on byte-level BPE""" LLAMA_VOCAB_TYPE_WPM = 3 """BERT tokenizer based on WordPiece""" +LLAMA_VOCAB_TYPE_UGM = 4 +"""T5 tokenizer based on Unigram""" # // pre-tokenization types @@ -302,6 +305,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_DBRX = 13, # LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, # LLAMA_VOCAB_PRE_TYPE_PORO = 15, +# LLAMA_VOCAB_PRE_TYPE_VIKING = 16, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -319,6 +323,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_DBRX = 13 LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 LLAMA_VOCAB_PRE_TYPE_PORO = 15 +LLAMA_VOCAB_PRE_TYPE_VIKING = 16 # // note: these values should be synchronized with ggml_rope @@ -2317,7 +2322,7 @@ def llama_n_threads_batch(ctx: llama_context_p, /) -> int: ... -# // Set whether the model is in embeddings model or not +# // Set whether the model is in embeddings mode or not # // If true, embeddings will be returned but logits will not # LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings); @ctypes_function("llama_set_embeddings", [llama_context_p_ctypes, ctypes.c_bool], None) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 557b653dc..5fac350b9 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 557b653dc9ed91e8c313e87500e0050c775f81b6 +Subproject commit 5fac350b9cc49d0446fc291b9c4ad53666c77591 From bf5e0bb4b151f4ca2f5a21af68eb832a96a79d75 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 21:29:13 -0400 Subject: [PATCH 438/624] fix(server): Update `embeddings=False` by default. Embeddings should be enabled by default for embedding models. --- llama_cpp/server/settings.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 4d924f337..9d9b533d5 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -96,7 +96,7 @@ class ModelSettings(BaseSettings): default=True, description="if true, use experimental mul_mat_q kernels" ) logits_all: bool = Field(default=True, description="Whether to return logits.") - embedding: bool = Field(default=True, description="Whether to use embeddings.") + embedding: bool = Field(default=False, description="Whether to use embeddings.") offload_kqv: bool = Field( default=True, description="Whether to offload kqv to the GPU." ) From 73ddf297bea8769ae0d745bb8b88e1b059edb92e Mon Sep 17 00:00:00 2001 From: oobabooga <112222186+oobabooga@users.noreply.github.com> Date: Mon, 1 Jul 2024 22:31:25 -0300 Subject: [PATCH 439/624] fix(ci): Fix the CUDA workflow (#1551) --- .github/workflows/build-wheels-cuda.yaml | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index b8496d850..31c861905 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -20,7 +20,7 @@ jobs: id: set-matrix run: | $matrix = @{ - 'os' = @('ubuntu-latest', 'windows-latest') + 'os' = @('ubuntu-latest', 'windows-2019') 'pyver' = @("3.9", "3.10", "3.11", "3.12") 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") 'releasetag' = @("basic") @@ -43,6 +43,12 @@ jobs: AVXVER: ${{ matrix.releasetag }} steps: + - name: Add MSBuild to PATH + if: runner.os == 'Windows' + uses: microsoft/setup-msbuild@v1.1 + with: + vs-version: '[16.11,16.12)' + - uses: actions/checkout@v4 with: submodules: "recursive" @@ -85,7 +91,7 @@ jobs: if: runner.os == 'Windows' run: | $y = (gi '.\MSBuildExtensions').fullname + '\*' - (gi 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\MSBuild\Microsoft\VC\*\BuildCustomizations').fullname.foreach({cp $y $_}) + (gi 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\MSBuild\Microsoft\VC\*\BuildCustomizations').fullname.foreach({cp $y $_}) $cupath = 'CUDA_PATH_V' + $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','_') echo "$cupath=$env:CONDA_PREFIX" >> $env:GITHUB_ENV From c546c94b48061dd22ecab8fe5309ded750ae9fb9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 22:03:19 -0400 Subject: [PATCH 440/624] misc: Install shared libraries to lib subdirectory --- CMakeLists.txt | 6 ++++-- llama_cpp/llama_cpp.py | 2 +- llama_cpp/llava_cpp.py | 2 +- 3 files changed, 6 insertions(+), 4 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 353a8d996..02adc445f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -8,11 +8,13 @@ option(LLAVA_BUILD "Build llava shared library and install alongside python pack if(SKBUILD_STATE STREQUAL "editable") # Install into the source directory # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 - set(LLAMA_CPP_PYTHON_INSTALL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp) + set(LLAMA_CPP_PYTHON_INSTALL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib) else() - set(LLAMA_CPP_PYTHON_INSTALL_DIR ${SKBUILD_PLATLIB_DIR}/llama_cpp) + set(LLAMA_CPP_PYTHON_INSTALL_DIR ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib) endif() +message(STATUS "LLAMA_CPP_PYTHON_INSTALL_DIR: ${LLAMA_CPP_PYTHON_INSTALL_DIR}") + if (LLAMA_BUILD) set(BUILD_SHARED_LIBS "On") diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 9014e0268..1999b9a40 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -23,7 +23,7 @@ # Load the library def _load_shared_library(lib_base_name: str): # Construct the paths to the possible shared library names - _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) + _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" # Searching for the library in the current directory under the name "libllama" (default name # for llamacpp) and "llama" (default name for this repo) _lib_paths: List[pathlib.Path] = [] diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index 3ded1f25d..b80d85913 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -35,7 +35,7 @@ # Load the library def _load_shared_library(lib_base_name: str): # Construct the paths to the possible shared library names - _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) + _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" # Searching for the library in the current directory under the name "libllama" (default name # for llamacpp) and "llama" (default name for this repo) _lib_paths: List[pathlib.Path] = [] From 139774b8b0a0d15101c60fbe5dab5fa3e4428b95 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 22:21:34 -0400 Subject: [PATCH 441/624] fix: Update shared library rpath --- CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 02adc445f..6ad49f3d2 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -27,7 +27,7 @@ if (LLAMA_BUILD) # Add the automatically determined parts of the RPATH # which point to directories outside the build tree to the install RPATH set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) - set(CMAKE_INSTALL_RPATH "${LLAMA_CPP_PYTHON_INSTALL_DIR}") + set(CMAKE_INSTALL_RPATH "$ORIGIN") # Building llama if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") From d5f6a15a9b390807b7d7e023b1d7363294dc02f3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 23:03:26 -0400 Subject: [PATCH 442/624] fix: force $ORIGIN rpath for shared library files --- CMakeLists.txt | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 6ad49f3d2..b0553cd24 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -28,6 +28,7 @@ if (LLAMA_BUILD) # which point to directories outside the build tree to the install RPATH set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) set(CMAKE_INSTALL_RPATH "$ORIGIN") + set(CMAKE_SKIP_RPATH FALSE) # Building llama if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") @@ -52,6 +53,10 @@ if (LLAMA_BUILD) FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) + set_target_properties(llama PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE + ) install( TARGETS ggml LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} @@ -60,6 +65,10 @@ if (LLAMA_BUILD) FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) + set_target_properties(ggml PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE + ) # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 if (WIN32 AND (LLAMA_CUDA OR LLAMA_CUBLAS)) install( @@ -93,5 +102,9 @@ if (LLAMA_BUILD) FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} ) + set_target_properties(llava PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE + ) endif() endif() From e51f200f2c2faea42286828a90ad4d838830e7a2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 23:11:49 -0400 Subject: [PATCH 443/624] fix: Fix installation location for shared libraries --- CMakeLists.txt | 70 ++++++++++++++++++-------------------------------- 1 file changed, 25 insertions(+), 45 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index b0553cd24..fc58cb884 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -5,15 +5,28 @@ project(llama_cpp) option(LLAMA_BUILD "Build llama.cpp shared library and install alongside python package" ON) option(LLAVA_BUILD "Build llava shared library and install alongside python package" ON) -if(SKBUILD_STATE STREQUAL "editable") - # Install into the source directory - # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 - set(LLAMA_CPP_PYTHON_INSTALL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib) -else() - set(LLAMA_CPP_PYTHON_INSTALL_DIR ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib) -endif() - -message(STATUS "LLAMA_CPP_PYTHON_INSTALL_DIR: ${LLAMA_CPP_PYTHON_INSTALL_DIR}") +function(llama_cpp_python_install_target target) + install( + TARGETS ${target} + LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + RUNTIME DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ARCHIVE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + TARGETS ${target} + LIBRARY DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + RUNTIME DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ARCHIVE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + FRAMEWORK DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + RESOURCE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ) + set_target_properties(${target} PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE + ) +endfunction() if (LLAMA_BUILD) set(BUILD_SHARED_LIBS "On") @@ -45,30 +58,8 @@ if (LLAMA_BUILD) endif() add_subdirectory(vendor/llama.cpp) - install( - TARGETS llama - LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - RUNTIME DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - ARCHIVE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - ) - set_target_properties(llama PROPERTIES - INSTALL_RPATH "$ORIGIN" - BUILD_WITH_INSTALL_RPATH TRUE - ) - install( - TARGETS ggml - LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - RUNTIME DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - ARCHIVE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - ) - set_target_properties(ggml PROPERTIES - INSTALL_RPATH "$ORIGIN" - BUILD_WITH_INSTALL_RPATH TRUE - ) + llama_cpp_python_install_target(llama) + llama_cpp_python_install_target(ggml) # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 if (WIN32 AND (LLAMA_CUDA OR LLAMA_CUBLAS)) install( @@ -94,17 +85,6 @@ if (LLAMA_BUILD) if (WIN32) set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) endif() - install( - TARGETS llava_shared - LIBRARY DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - RUNTIME DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - ARCHIVE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - FRAMEWORK DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - RESOURCE DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} - ) - set_target_properties(llava PROPERTIES - INSTALL_RPATH "$ORIGIN" - BUILD_WITH_INSTALL_RPATH TRUE - ) + llama_cpp_python_install_target(llava_shared) endif() endif() From 73fe013a4857962c60c6cb16c330cece639fb8ac Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 23:17:02 -0400 Subject: [PATCH 444/624] fix: Fix RPATH so it works on macos --- CMakeLists.txt | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index fc58cb884..e0281b059 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -26,6 +26,19 @@ function(llama_cpp_python_install_target target) INSTALL_RPATH "$ORIGIN" BUILD_WITH_INSTALL_RPATH TRUE ) + if(UNIX) + if(APPLE) + set_target_properties(${target} PROPERTIES + INSTALL_RPATH "@loader_path" + BUILD_WITH_INSTALL_RPATH TRUE + ) + else() + set_target_properties(${target} PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE + ) + endif() + endif() endfunction() if (LLAMA_BUILD) @@ -40,7 +53,6 @@ if (LLAMA_BUILD) # Add the automatically determined parts of the RPATH # which point to directories outside the build tree to the install RPATH set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) - set(CMAKE_INSTALL_RPATH "$ORIGIN") set(CMAKE_SKIP_RPATH FALSE) # Building llama From dc20e8c342264b950c0cb3f8d7ba816fba15cfc8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 1 Jul 2024 23:28:19 -0400 Subject: [PATCH 445/624] fix: Copy dependencies for windows fix: :( fix fix fix: Copy runtime dlls on windows fix: Add explicit copy command for windows dlls fix fix fix: >:( fix fix fix fix Update path on windows check dll dependancies fix: Update PATH on win32 ci: Update test.yaml --- .github/workflows/test.yaml | 26 -------------------------- CMakeLists.txt | 29 ++++++++++++++++++++++++++--- llama_cpp/llama_cpp.py | 5 +++++ 3 files changed, 31 insertions(+), 29 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 6d8231d18..9a938ae11 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -81,32 +81,6 @@ jobs: run: | python -m pytest - # build-linux-opencl: - - # runs-on: ubuntu-latest - - # steps: - # - uses: actions/checkout@v4 - # with: - # submodules: "recursive" - # - name: Set up Python 3.8 - # uses: actions/setup-python@v5 - # with: - # python-version: "3.8" - # - name: Set up OpenCL & CLBlast - # run: | - # wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null - # echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list - # sudo apt-get update - # sudo apt-get install -y --no-install-recommends llvm intel-oneapi-runtime-opencl intel-oneapi-runtime-compilers libclblast-dev - # - name: Install dependencies - # run: | - # python -m pip install --upgrade pip - # CMAKE_ARGS="-DLLAMA_CLBLAST=on" python -m pip install .[all] --verbose - # - name: Test with pytest - # run: | - # python -m pytest - build-macos-metal: diff --git a/CMakeLists.txt b/CMakeLists.txt index e0281b059..0c4fa026f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -72,11 +72,24 @@ if (LLAMA_BUILD) add_subdirectory(vendor/llama.cpp) llama_cpp_python_install_target(llama) llama_cpp_python_install_target(ggml) + # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 - if (WIN32 AND (LLAMA_CUDA OR LLAMA_CUBLAS)) + if (WIN32) install( FILES $ - DESTINATION ${LLAMA_CPP_PYTHON_INSTALL_DIR} + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib ) endif() @@ -98,5 +111,15 @@ if (LLAMA_BUILD) set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) endif() llama_cpp_python_install_target(llava_shared) + if (WIN32) + install( + FILES $ + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ) + endif() endif() -endif() +endif() \ No newline at end of file diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1999b9a40..8da50af92 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -52,7 +52,12 @@ def _load_shared_library(lib_base_name: str): _lib_paths = [_lib.resolve()] cdll_args = dict() # type: ignore + # Add the library directory to the DLL search path on Windows (if needed) + if sys.platform == "win32": + os.add_dll_directory(str(_base_path)) + os.environ['PATH'] = str(_base_path) + os.pathsep + os.environ['PATH'] + if sys.platform == "win32" and sys.version_info >= (3, 8): os.add_dll_directory(str(_base_path)) if "CUDA_PATH" in os.environ: From 296304b60bb83689659883c9cc24f4c074dd88ff Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 02:49:20 -0400 Subject: [PATCH 446/624] fix(server): Fix bug in FastAPI streaming response where dependency was released before request completes causing SEGFAULT --- llama_cpp/server/app.py | 24 ++++++++++++++++++++++-- 1 file changed, 22 insertions(+), 2 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 4cda4af7a..f4d356b23 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -2,6 +2,7 @@ import os import json +import contextlib from threading import Lock from functools import partial @@ -156,6 +157,7 @@ async def get_event_publisher( request: Request, inner_send_chan: MemoryObjectSendStream, iterator: Iterator, + on_complete=None, ): async with inner_send_chan: try: @@ -175,6 +177,9 @@ async def get_event_publisher( with anyio.move_on_after(1, shield=True): print(f"Disconnected from client (via refresh/close) {request.client}") raise e + finally: + if on_complete: + on_complete() def _logit_bias_tokens_to_input_ids( @@ -258,8 +263,11 @@ async def authenticate( async def create_completion( request: Request, body: CreateCompletionRequest, - llama_proxy: LlamaProxy = Depends(get_llama_proxy), ) -> llama_cpp.Completion: + exit_stack = contextlib.ExitStack() + llama_proxy = await run_in_threadpool( + lambda: exit_stack.enter_context(contextlib.contextmanager(get_llama_proxy)()) + ) if isinstance(body.prompt, list): assert len(body.prompt) <= 1 body.prompt = body.prompt[0] if len(body.prompt) > 0 else "" @@ -312,6 +320,7 @@ async def create_completion( def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: yield first_response yield from iterator_or_completion + exit_stack.close() send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( @@ -321,6 +330,7 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: request=request, inner_send_chan=send_chan, iterator=iterator(), + on_complete=exit_stack.close, ), sep="\n", ping_message_factory=_ping_message_factory, @@ -449,8 +459,15 @@ async def create_chat_completion( }, } ), - llama_proxy: LlamaProxy = Depends(get_llama_proxy), ) -> llama_cpp.ChatCompletion: + # This is a workaround for an issue in FastAPI dependencies + # where the dependency is cleaned up before a StreamingResponse + # is complete. + # https://github.com/tiangolo/fastapi/issues/11143 + exit_stack = contextlib.ExitStack() + llama_proxy = await run_in_threadpool( + lambda: exit_stack.enter_context(contextlib.contextmanager(get_llama_proxy)()) + ) exclude = { "n", "logit_bias_type", @@ -491,6 +508,7 @@ async def create_chat_completion( def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: yield first_response yield from iterator_or_completion + exit_stack.close() send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( @@ -500,11 +518,13 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: request=request, inner_send_chan=send_chan, iterator=iterator(), + on_complete=exit_stack.close, ), sep="\n", ping_message_factory=_ping_message_factory, ) else: + exit_stack.close() return iterator_or_completion From bd5d17bcf0c61fd2cd43ef48c2f616266f1e29c8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 02:55:19 -0400 Subject: [PATCH 447/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 5fac350b9..023b8807e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 5fac350b9cc49d0446fc291b9c4ad53666c77591 +Subproject commit 023b8807e10bc3ade24a255f01c1ad2a01bb4228 From b4cc9234748cb4bb96db1175a143139d6b7d8557 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 02:55:57 -0400 Subject: [PATCH 448/624] chore: Bump version --- CHANGELOG.md | 8 ++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 42b14c862..a89725676 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.80] + +- feat: Update llama.cpp to ggerganov/llama.cpp@023b8807e10bc3ade24a255f01c1ad2a01bb4228 +- fix(server): Fix bug in FastAPI streaming response where dependency was released before request completes causing SEGFAULT by @abetlen in 296304b60bb83689659883c9cc24f4c074dd88ff +- fix(server): Update default config value for embeddings to False to fix error in text generation where logits were not allocated by llama.cpp by @abetlen in bf5e0bb4b151f4ca2f5a21af68eb832a96a79d75 +- fix(ci): Fix the CUDA workflow by @oobabooga in #1551 +- docs: Update readme examples to use newer Qwen2 model by @jncraton in #1544 + ## [0.2.79] - feat: Update llama.cpp to ggerganov/llama.cpp@9c77ec1d74874ee22bdef8f110e8e8d41389abf2 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 71371b3b0..a83d8ad74 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.79" \ No newline at end of file +__version__ = "0.2.80" \ No newline at end of file From 4fb6fc12a02a68884c25dd9f6a421cacec7604c6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 03:09:19 -0400 Subject: [PATCH 449/624] fix(ci): Use LLAMA_CUDA for cuda wheels --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 31c861905..78d0cba81 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -114,7 +114,7 @@ jobs: $env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH } $env:VERBOSE = '1' - $env:CMAKE_ARGS = '-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all' + $env:CMAKE_ARGS = '-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=all' $env:CMAKE_ARGS = "-DLLAMA_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS" # if ($env:AVXVER -eq 'AVX') { $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' From 387d01d94581dbcb1288304094bb1acee5ca24ae Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 11:06:57 -0400 Subject: [PATCH 450/624] fix(misc): Fix type errors --- llama_cpp/server/app.py | 43 ++++++++++++++++++++++++++--------------- 1 file changed, 27 insertions(+), 16 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index f4d356b23..8fd41a3fc 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -2,6 +2,7 @@ import os import json +import typing import contextlib from threading import Lock @@ -88,11 +89,12 @@ def get_llama_proxy(): llama_outer_lock.release() -_ping_message_factory = None +_ping_message_factory: typing.Optional[typing.Callable[[], bytes]] = None -def set_ping_message_factory(factory): - global _ping_message_factory - _ping_message_factory = factory + +def set_ping_message_factory(factory: typing.Callable[[], bytes]): + global _ping_message_factory + _ping_message_factory = factory def create_app( @@ -155,20 +157,19 @@ def create_app( async def get_event_publisher( request: Request, - inner_send_chan: MemoryObjectSendStream, - iterator: Iterator, - on_complete=None, + inner_send_chan: MemoryObjectSendStream[typing.Any], + iterator: Iterator[typing.Any], + on_complete: typing.Optional[typing.Callable[[], None]] = None, ): + server_settings = next(get_server_settings()) + interrupt_requests = server_settings.interrupt_requests if server_settings else False async with inner_send_chan: try: async for chunk in iterate_in_threadpool(iterator): await inner_send_chan.send(dict(data=json.dumps(chunk))) if await request.is_disconnected(): raise anyio.get_cancelled_exc_class()() - if ( - next(get_server_settings()).interrupt_requests - and llama_outer_lock.locked() - ): + if interrupt_requests and llama_outer_lock.locked(): await inner_send_chan.send(dict(data="[DONE]")) raise anyio.get_cancelled_exc_class()() await inner_send_chan.send(dict(data="[DONE]")) @@ -268,6 +269,11 @@ async def create_completion( llama_proxy = await run_in_threadpool( lambda: exit_stack.enter_context(contextlib.contextmanager(get_llama_proxy)()) ) + if llama_proxy is None: + raise HTTPException( + status_code=status.HTTP_503_SERVICE_UNAVAILABLE, + detail="Service is not available", + ) if isinstance(body.prompt, list): assert len(body.prompt) <= 1 body.prompt = body.prompt[0] if len(body.prompt) > 0 else "" @@ -409,7 +415,7 @@ async def create_chat_completion( {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Who won the world series in 2020"}, ], - "response_format": { "type": "json_object" } + "response_format": {"type": "json_object"}, }, }, "tool_calling": { @@ -434,15 +440,15 @@ async def create_chat_completion( }, "required": ["name", "age"], }, - } + }, } ], "tool_choice": { "type": "function", "function": { "name": "User", - } - } + }, + }, }, }, "logprobs": { @@ -454,7 +460,7 @@ async def create_chat_completion( {"role": "user", "content": "What is the capital of France?"}, ], "logprobs": True, - "top_logprobs": 10 + "top_logprobs": 10, }, }, } @@ -468,6 +474,11 @@ async def create_chat_completion( llama_proxy = await run_in_threadpool( lambda: exit_stack.enter_context(contextlib.contextmanager(get_llama_proxy)()) ) + if llama_proxy is None: + raise HTTPException( + status_code=status.HTTP_503_SERVICE_UNAVAILABLE, + detail="Service is not available", + ) exclude = { "n", "logit_bias_type", From 8992a1a6467322bc8be28b74efa9e6a38d4c33b1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 11:09:31 -0400 Subject: [PATCH 451/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 8da50af92..ee63de7cf 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -311,6 +311,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, # LLAMA_VOCAB_PRE_TYPE_PORO = 15, # LLAMA_VOCAB_PRE_TYPE_VIKING = 16, +# LLAMA_VOCAB_PRE_TYPE_JAIS = 17, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -329,6 +330,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 LLAMA_VOCAB_PRE_TYPE_PORO = 15 LLAMA_VOCAB_PRE_TYPE_VIKING = 16 +LLAMA_VOCAB_PRE_TYPE_JAIS = 17 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 023b8807e..968967376 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 023b8807e10bc3ade24a255f01c1ad2a01bb4228 +Subproject commit 968967376dc2c018d29f897c4883d335bbf384fb From 3a551eb5263fdbd24b36d7770856374c04e92788 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 11:58:26 -0400 Subject: [PATCH 452/624] fix(ci): Update macos image (macos-11 is removed) --- .github/workflows/build-and-release.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 34e9dc139..e736cf964 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -11,7 +11,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-20.04, windows-2019, macos-11] + os: [ubuntu-20.04, windows-2019, macos-12] steps: - uses: actions/checkout@v4 From 01bddd669ca1208f1844ce8d0ba9872532641c9d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 2 Jul 2024 12:02:45 -0400 Subject: [PATCH 453/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index a89725676..a953af7e7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.81] + +- feat: Update llama.cpp to ggerganov/llama.cpp@968967376dc2c018d29f897c4883d335bbf384fb +- fix(ci): Fix CUDA wheels, use LLAMA_CUDA instead of removed LLAMA_CUBLAS by @abetlen in 4fb6fc12a02a68884c25dd9f6a421cacec7604c6 +- fix(ci): Fix MacOS release, use macos-12 image instead of removed macos-11 by @abetlen in 3a551eb5263fdbd24b36d7770856374c04e92788 + ## [0.2.80] - feat: Update llama.cpp to ggerganov/llama.cpp@023b8807e10bc3ade24a255f01c1ad2a01bb4228 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index a83d8ad74..f87e2a388 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.80" \ No newline at end of file +__version__ = "0.2.81" \ No newline at end of file From 7e20e346bd49cc8f0031eb053fe879a38c777b6f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 4 Jul 2024 12:23:47 -0400 Subject: [PATCH 454/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 34 ++++++++++++++++++++++++++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 35 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index ee63de7cf..1fc1968ee 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1444,6 +1444,24 @@ def llama_get_model_tensor( ... +# // Returns true if the model contains an encoder that requires llama_encode() call +# LLAMA_API bool llama_model_has_encoder(const struct llama_model * model); +@ctypes_function("llama_model_has_encoder", [llama_model_p_ctypes], ctypes.c_bool) +def llama_model_has_encoder(model: llama_model_p, /) -> bool: + """Returns true if the model contains an encoder that requires llama_encode() call""" + ... + + +# // For encoder-decoder models, this function returns id of the token that must be provided +# // to the decoder to start generating output sequence. For other models, it returns -1. +# LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); +@ctypes_function("llama_model_decoder_start_token", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_decoder_start_token(model: llama_model_p, /) -> int: + """For encoder-decoder models, this function returns id of the token that must be provided + to the decoder to start generating output sequence. For other models, it returns -1.""" + ... + + # // Returns 0 on success # LLAMA_API uint32_t llama_model_quantize( # const char * fname_inp, @@ -2271,6 +2289,22 @@ def llama_batch_free(batch: llama_batch, /): ... +# // Processes a batch of tokens with the ecoder part of the encoder-decoder model. +# // Stores the encoder output internally for later use by the decoder cross-attention layers. +# // 0 - success +# // < 0 - error +# LLAMA_API int32_t llama_encode( +# struct llama_context * ctx, +# struct llama_batch batch); +@ctypes_function("llama_encode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) +def llama_encode(ctx: llama_context_p, batch: llama_batch, /) -> int: + """Processes a batch of tokens with the ecoder part of the encoder-decoder model. + Stores the encoder output internally for later use by the decoder cross-attention layers. + 0 - success + < 0 - error""" + ... + + # // Positive return values does not mean a fatal error, but rather a warning. # // 0 - success # // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 968967376..51d2ebadb 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 968967376dc2c018d29f897c4883d335bbf384fb +Subproject commit 51d2ebadbbf365b894f3888361df42dbacb12b7a From 62804eeb12322ab167c22bc1d2dea10175e90029 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Jul 2024 20:44:46 -0400 Subject: [PATCH 455/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 72 ++++++++++++++++++++++++++++++++++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 71 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1fc1968ee..5ee3476c9 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -488,6 +488,15 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_POOLING_TYPE_CLS = 2 LLAMA_POOLING_TYPE_LAST = 3 +# enum llama_attention_type { +# LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1, +# LLAMA_ATTENTION_TYPE_CAUSAL = 0, +# LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1, +# }; +LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1 +LLAMA_ATTENTION_TYPE_CAUSAL = 0 +LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1 + # enum llama_split_mode { # LLAMA_SPLIT_MODE_NONE = 0, // single GPU # LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs @@ -775,6 +784,7 @@ class llama_model_params(ctypes.Structure): # enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` # enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id +# enum llama_attention_type attention_type; // attention type to use for embeddings # // ref: https://github.com/ggerganov/llama.cpp/pull/2054 # float rope_freq_base; // RoPE base frequency, 0 = from model @@ -817,6 +827,7 @@ class llama_context_params(ctypes.Structure): n_threads_batch (int): number of threads to use for batch processing rope_scaling_type (int): RoPE scaling type, from `enum llama_rope_scaling_type` pooling_type (int): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) + attention_type (int): attention type to use for embeddings rope_freq_base (float): RoPE base frequency, 0 = from model rope_freq_scale (float): RoPE frequency scaling factor, 0 = from model yarn_ext_factor (float): YaRN extrapolation mix factor, negative = from model @@ -847,6 +858,7 @@ class llama_context_params(ctypes.Structure): n_threads_batch: int rope_scaling_type: int pooling_type: int + attention_type: int rope_freq_base: float rope_freq_scale: float yarn_ext_factor: float @@ -876,6 +888,7 @@ class llama_context_params(ctypes.Structure): ("n_threads_batch", ctypes.c_uint32), ("rope_scaling_type", ctypes.c_int), ("pooling_type", ctypes.c_int), + ("attention_type", ctypes.c_int), ("rope_freq_base", ctypes.c_float), ("rope_freq_scale", ctypes.c_float), ("yarn_ext_factor", ctypes.c_float), @@ -2642,6 +2655,7 @@ def llama_token_eot(model: llama_model_p, /) -> int: ... # /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. # /// @return Returns the number of tokens on success, no more than n_tokens_max # /// @return Returns a negative number on failure - the number of tokens that would have been returned +# /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so. # /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated # /// as plaintext. Does not insert a leading space. # LLAMA_API int32_t llama_tokenize( @@ -2683,7 +2697,7 @@ def llama_tokenize( text_len: The length of the text. tokens: The tokens pointer must be large enough to hold the resulting tokens. n_max_tokens: The maximum number of tokens to return. - add_special: Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. Does not insert a leading space. + add_special: Allow adding special tokenns if the model is configured to do so. parse_special: Allow parsing special tokens. Returns: @@ -2696,13 +2710,14 @@ def llama_tokenize( # // Token Id -> Piece. # // Uses the vocabulary in the provided context. # // Does not write null terminator to the buffer. -# // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. +# // User can skip up to 'lstrip' leading spaces before copying (useful when encoding/decoding multiple tokens with 'add_space_prefix') # // @param special If true, special tokens are rendered in the output. # LLAMA_API int32_t llama_token_to_piece( # const struct llama_model * model, # llama_token token, # char * buf, # int32_t length, +# int32_t lstrip, # bool special); @ctypes_function( "llama_token_to_piece", @@ -2711,6 +2726,7 @@ def llama_tokenize( llama_token, ctypes.c_char_p, ctypes.c_int32, + ctypes.c_int32, ctypes.c_bool, ], ctypes.c_int32, @@ -2720,6 +2736,7 @@ def llama_token_to_piece( token: Union[llama_token, int], buf: Union[ctypes.c_char_p, bytes, CtypesArray[ctypes.c_char]], length: Union[ctypes.c_int, int], + lstrip: Union[ctypes.c_int, int], special: Union[ctypes.c_bool, bool], /, ) -> int: @@ -2733,10 +2750,61 @@ def llama_token_to_piece( token: The token to convert. buf: The buffer to write the token to. length: The length of the buffer. + lstrip: The number of leading spaces to skip. special: If true, special tokens are rendered in the output.""" ... +# /// @details Convert the provided tokens into text (inverse of llama_tokenize()). +# /// @param text The char pointer must be large enough to hold the resulting text. +# /// @return Returns the number of chars/bytes on success, no more than text_len_max. +# /// @return Returns a negative number on failure - the number of chars/bytes that would have been returned. +# /// @param remove_special Allow to remove BOS and EOS tokens if model is configured to do so. +# /// @param unparse_special If true, special tokens are rendered in the output. +# LLAMA_API int32_t llama_detokenize( +# const struct llama_model * model, +# const llama_token * tokens, +# int32_t n_tokens, +# char * text, +# int32_t text_len_max, +# bool remove_special, +# bool unparse_special); +@ctypes_function( + "llama_detokenize", + [ + llama_model_p_ctypes, + ctypes.POINTER(llama_token), + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_int32, + ctypes.c_bool, + ctypes.c_bool, + ], + ctypes.c_int32, +) +def llama_detokenize( + model: llama_model_p, + tokens: CtypesArray[llama_token], + n_tokens: Union[ctypes.c_int, int], + text: bytes, + text_len_max: Union[ctypes.c_int, int], + remove_special: Union[ctypes.c_bool, bool], + unparse_special: Union[ctypes.c_bool, bool], + /, +) -> int: + """Convert the provided tokens into text (inverse of llama_tokenize()). + + Args: + model: The model to use for tokenization. + tokens: The tokens to convert. + n_tokens: The number of tokens. + text: The buffer to write the text to. + text_len_max: The length of the buffer. + remove_special: Allow to remove BOS and EOS tokens if model is configured to do so. + unparse_special: If true, special tokens are rendered in the output.""" + ... + + # /// Apply chat template. Inspired by hf apply_chat_template() on python. # /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" # /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 51d2ebadb..213701b51 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 51d2ebadbbf365b894f3888361df42dbacb12b7a +Subproject commit 213701b51a17175d0d326b566efc03f30ec7fbe6 From 157d91359d5d51cee5a2f44fd9f54734603d74bc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 5 Jul 2024 21:42:02 -0400 Subject: [PATCH 456/624] fix: update token_to_piece --- llama_cpp/_internals.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 423c1b5bc..125d72ded 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -208,7 +208,7 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): def token_to_piece(self, token: int, special: bool = False) -> bytes: assert self.model is not None buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32, special) + llama_cpp.llama_token_to_piece(self.model, token, buf, 32, 0, special) return bytes(buf) def detokenize(self, tokens: List[int], special: bool = False) -> bytes: @@ -218,7 +218,7 @@ def detokenize(self, tokens: List[int], special: bool = False) -> bytes: buffer = (ctypes.c_char * size)() for token in tokens: n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token), buffer, size, special + self.model, llama_cpp.llama_token(token), buffer, size, 0, special ) assert n <= size output += bytes(buffer[:n]) @@ -635,10 +635,10 @@ def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> li def _token_to_piece(model: _LlamaModel, token: int, special: bool = False) -> str: assert model.model is not None result = (ctypes.c_char * 8)(0) - n_tokens = llama_cpp.llama_token_to_piece(model.model, token, result, len(result), special) + n_tokens = llama_cpp.llama_token_to_piece(model.model, token, result, 0, len(result), special) if n_tokens < 0: result = (ctypes.c_char * -n_tokens)(0) - check = llama_cpp.llama_token_to_piece(model.model, token, result, len(result), special) + check = llama_cpp.llama_token_to_piece(model.model, token, result, 0, len(result), special) if check != -n_tokens: raise RuntimeError(f"Failed to get piece: token={token}") else: From 218d3610ae6e8f4849eb7b7661a0239ef3351aea Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:15:22 -0400 Subject: [PATCH 457/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 12 ++++++++---- vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 5ee3476c9..a0f5832c6 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -310,8 +310,10 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_DBRX = 13, # LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, # LLAMA_VOCAB_PRE_TYPE_PORO = 15, -# LLAMA_VOCAB_PRE_TYPE_VIKING = 16, -# LLAMA_VOCAB_PRE_TYPE_JAIS = 17, +# LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16, +# LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17, +# LLAMA_VOCAB_PRE_TYPE_VIKING = 18, +# LLAMA_VOCAB_PRE_TYPE_JAIS = 19, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -329,8 +331,10 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_DBRX = 13 LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 LLAMA_VOCAB_PRE_TYPE_PORO = 15 -LLAMA_VOCAB_PRE_TYPE_VIKING = 16 -LLAMA_VOCAB_PRE_TYPE_JAIS = 17 +LLAMA_VOCAV_PRE_TYPE_CHATGLM3 = 16 +LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17 +LLAMA_VOCAB_PRE_TYPE_VIKING = 18 +LLAMA_VOCAB_PRE_TYPE_JAIS = 19 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 213701b51..7fdb6f73e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 213701b51a17175d0d326b566efc03f30ec7fbe6 +Subproject commit 7fdb6f73e35605c8dbc39e9f19cd9ed84dbc87f2 From 1a55417455660cd626e8e8a37a19659f97c28ef1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:29:29 -0400 Subject: [PATCH 458/624] fix: Update LLAMA_ flags to GGML_ flags --- CMakeLists.txt | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 0c4fa026f..c6b35ed6c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -59,14 +59,14 @@ if (LLAMA_BUILD) if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") # Need to disable these llama.cpp flags on Apple x86_64, # otherwise users may encounter invalid instruction errors - set(LLAMA_AVX "Off" CACHE BOOL "llama: enable AVX" FORCE) - set(LLAMA_AVX2 "Off" CACHE BOOL "llama: enable AVX2" FORCE) - set(LLAMA_FMA "Off" CACHE BOOL "llama: enable FMA" FORCE) - set(LLAMA_F16C "Off" CACHE BOOL "llama: enable F16C" FORCE) + set(GGML_AVX "Off" CACHE BOOL "ggml: enable AVX" FORCE) + set(GGML_AVX2 "Off" CACHE BOOL "ggml: enable AVX2" FORCE) + set(GGML_FMA "Off" CACHE BOOL "gml: enable FMA" FORCE) + set(GGML_F16C "Off" CACHE BOOL "gml: enable F16C" FORCE) endif() if (APPLE) - set(LLAMA_METAL_EMBED_LIBRARY "On" CACHE BOOL "llama: embed metal library" FORCE) + set(GGML_METAL_EMBED_LIBRARY "On" CACHE BOOL "llama: embed metal library" FORCE) endif() add_subdirectory(vendor/llama.cpp) @@ -122,4 +122,4 @@ if (LLAMA_BUILD) ) endif() endif() -endif() \ No newline at end of file +endif() From 09a4f78e05f43fcd93efc82cfddca250af1bad4d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:37:03 -0400 Subject: [PATCH 459/624] fix(ci): Update LLAMA_ flags to GGML_ --- .github/workflows/build-wheels-cuda.yaml | 10 +++++----- .github/workflows/build-wheels-metal.yaml | 2 +- .github/workflows/test.yaml | 2 +- 3 files changed, 7 insertions(+), 7 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 78d0cba81..5d445b5de 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -114,16 +114,16 @@ jobs: $env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH } $env:VERBOSE = '1' - $env:CMAKE_ARGS = '-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=all' - $env:CMAKE_ARGS = "-DLLAMA_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS" + $env:CMAKE_ARGS = '-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=all' + $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS" # if ($env:AVXVER -eq 'AVX') { - $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' + $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off' # } # if ($env:AVXVER -eq 'AVX512') { - # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on' + # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX512=on' # } # if ($env:AVXVER -eq 'basic') { - # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off' + # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX=off -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off' # } python -m build --wheel # write the build tag to the output diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index 11ab795c8..b56fb48aa 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -35,7 +35,7 @@ jobs: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" CIBW_ARCHS: "arm64" - CIBW_ENVIRONMENT: CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on" + CIBW_ENVIRONMENT: CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DGGML_METAL=on" CIBW_BUILD: "cp39-* cp310-* cp311-* cp312-*" with: package-dir: . diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 9a938ae11..78f0b4983 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -97,7 +97,7 @@ jobs: - name: Install dependencies run: | python -m pip install --upgrade pip - CMAKE_ARGS="-DLLAMA_METAL=on" python -m pip install .[all] --verbose + CMAKE_ARGS="-DGGML_METAL=on" python -m pip install .[all] --verbose - name: Test with pytest run: | python -m pytest From 0481a3a8e124b9c7eb5413930cb0597215889f3a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:39:32 -0400 Subject: [PATCH 460/624] fix(docs): Update LLAMA_ flags to GGML_ flags --- README.md | 46 +++++++++++++++++++++---------------------- docs/install/macos.md | 2 +- 2 files changed, 24 insertions(+), 24 deletions(-) diff --git a/README.md b/README.md index 4a710d4e5..8363ef685 100644 --- a/README.md +++ b/README.md @@ -64,13 +64,13 @@ All `llama.cpp` cmake build options can be set via the `CMAKE_ARGS` environment ```bash # Linux and Mac -CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" \ +CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" \ pip install llama-cpp-python ``` ```powershell # Windows -$env:CMAKE_ARGS = "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" +$env:CMAKE_ARGS = "-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python ```
@@ -83,13 +83,13 @@ They can also be set via `pip install -C / --config-settings` command and saved ```bash pip install --upgrade pip # ensure pip is up to date pip install llama-cpp-python \ - -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS" + -C cmake.args="-DGGML_BLAS=ON;-DGGML_BLAS_VENDOR=OpenBLAS" ``` ```txt # requirements.txt -llama-cpp-python -C cmake.args="-DLLAMA_BLAS=ON;-DLLAMA_BLAS_VENDOR=OpenBLAS" +llama-cpp-python -C cmake.args="-DGGML_BLAS=ON;-DGGML_BLAS_VENDOR=OpenBLAS" ``` @@ -101,20 +101,20 @@ Below are some common backends, their build commands and any additional environm
OpenBLAS (CPU) -To install with OpenBLAS, set the `LLAMA_BLAS` and `LLAMA_BLAS_VENDOR` environment variables before installing: +To install with OpenBLAS, set the `GGML_BLAS` and `GGML_BLAS_VENDOR` environment variables before installing: ```bash -CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python ```
CUDA -To install with CUDA support, set the `LLAMA_CUDA=on` environment variable before installing: +To install with CUDA support, set the `GGML_CUDA=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python ``` **Pre-built Wheel (New)** @@ -147,10 +147,10 @@ pip install llama-cpp-python \
Metal -To install with Metal (MPS), set the `LLAMA_METAL=on` environment variable before installing: +To install with Metal (MPS), set the `GGML_METAL=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python ``` **Pre-built Wheel (New)** @@ -170,10 +170,10 @@ pip install llama-cpp-python \
hipBLAS (ROCm) -To install with hipBLAS / ROCm support for AMD cards, set the `LLAMA_HIPBLAS=on` environment variable before installing: +To install with hipBLAS / ROCm support for AMD cards, set the `GGML_HIPBLAS=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_HIPBLAS=on" pip install llama-cpp-python ```
@@ -181,10 +181,10 @@ CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python
Vulkan -To install with Vulkan support, set the `LLAMA_VULKAN=on` environment variable before installing: +To install with Vulkan support, set the `GGML_VULKAN=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_VULKAN=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_VULKAN=on" pip install llama-cpp-python ```
@@ -192,32 +192,32 @@ CMAKE_ARGS="-DLLAMA_VULKAN=on" pip install llama-cpp-python
Kompute -To install with Kompute support, set the `LLAMA_KOMPUTE=on` environment variable before installing: +To install with Kompute support, set the `GGML_KOMPUTE=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_KOMPUTE=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_KOMPUTE=on" pip install llama-cpp-python ```
SYCL -To install with SYCL support, set the `LLAMA_SYCL=on` environment variable before installing: +To install with SYCL support, set the `GGML_SYCL=on` environment variable before installing: ```bash source /opt/intel/oneapi/setvars.sh -CMAKE_ARGS="-DLLAMA_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install llama-cpp-python ```
RPC -To install with RPC support, set the `LLAMA_RPC=on` environment variable before installing: +To install with RPC support, set the `GGML_RPC=on` environment variable before installing: ```bash source /opt/intel/oneapi/setvars.sh -CMAKE_ARGS="-DLLAMA_RPC=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_RPC=on" pip install llama-cpp-python ```
@@ -231,7 +231,7 @@ If you run into issues where it complains it can't find `'nmake'` `'?'` or CMAKE ```ps $env:CMAKE_GENERATOR = "MinGW Makefiles" -$env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" +$env:CMAKE_ARGS = "-DGGML_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" ``` See the above instructions and set `CMAKE_ARGS` to the BLAS backend you want to use. @@ -260,7 +260,7 @@ Otherwise, while installing it will build the llama.cpp x86 version which will b Try installing with ```bash -CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on" pip install --upgrade --verbose --force-reinstall --no-cache-dir llama-cpp-python +CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DGGML_METAL=on" pip install --upgrade --verbose --force-reinstall --no-cache-dir llama-cpp-python ```
@@ -667,7 +667,7 @@ python3 -m llama_cpp.server --model models/7B/llama-model.gguf Similar to Hardware Acceleration section above, you can also install with GPU (cuBLAS) support like this: ```bash -CMAKE_ARGS="-DLLAMA_CUDA=on" FORCE_CMAKE=1 pip install 'llama-cpp-python[server]' +CMAKE_ARGS="-DGGML_CUDA=on" FORCE_CMAKE=1 pip install 'llama-cpp-python[server]' python3 -m llama_cpp.server --model models/7B/llama-model.gguf --n_gpu_layers 35 ``` diff --git a/docs/install/macos.md b/docs/install/macos.md index 240422832..e006fc0a3 100644 --- a/docs/install/macos.md +++ b/docs/install/macos.md @@ -30,7 +30,7 @@ conda activate llama *(you needed xcode installed in order pip to build/compile the C++ code)* ``` pip uninstall llama-cpp-python -y -CMAKE_ARGS="-DLLAMA_METAL=on" pip install -U llama-cpp-python --no-cache-dir +CMAKE_ARGS="-DGGML_METAL=on" pip install -U llama-cpp-python --no-cache-dir pip install 'llama-cpp-python[server]' # you should now have llama-cpp-python v0.1.62 or higher installed From fccff8015eb8880e826d407321f5fbf7bd8dd4c3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:41:36 -0400 Subject: [PATCH 461/624] fix(docs): Remove kompute backend references --- README.md | 10 ---------- 1 file changed, 10 deletions(-) diff --git a/README.md b/README.md index 8363ef685..5641ccaa8 100644 --- a/README.md +++ b/README.md @@ -189,16 +189,6 @@ CMAKE_ARGS="-DGGML_VULKAN=on" pip install llama-cpp-python
-
-Kompute - -To install with Kompute support, set the `GGML_KOMPUTE=on` environment variable before installing: - -```bash -CMAKE_ARGS="-DGGML_KOMPUTE=on" pip install llama-cpp-python -``` -
-
SYCL From 276ea28a3f665ac856aa5ac5ee0af8bcbc333ed2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:46:20 -0400 Subject: [PATCH 462/624] fix(misc): Update LLAMA_ flags to GGML_ --- Makefile | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/Makefile b/Makefile index 90f562fa1..be9b55a3e 100644 --- a/Makefile +++ b/Makefile @@ -22,28 +22,28 @@ build.debug: --editable . build.cuda: - CMAKE_ARGS="-DLLAMA_CUDA=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_CUDA=on" python3 -m pip install --verbose -e . build.openblas: - CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" python3 -m pip install --verbose -e . build.blis: - CMAKE_ARGS="-DLLAMA_BLAS=on -DLLAMA_BLAS_VENDOR=FLAME" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_BLAS=on -DGGML_BLAS_VENDOR=FLAME" python3 -m pip install --verbose -e . build.metal: - CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_METAL=on" python3 -m pip install --verbose -e . build.vulkan: - CMAKE_ARGS="-DLLAMA_VULKAN=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_VULKAN=on" python3 -m pip install --verbose -e . build.kompute: - CMAKE_ARGS="-DLLAMA_KOMPUTE=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_KOMPUTE=on" python3 -m pip install --verbose -e . build.sycl: - CMAKE_ARGS="-DLLAMA_SYCL=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_SYCL=on" python3 -m pip install --verbose -e . build.rpc: - CMAKE_ARGS="-DLLAMA_RPC=on" python3 -m pip install --verbose -e . + CMAKE_ARGS="-DGGML_RPC=on" python3 -m pip install --verbose -e . build.sdist: python3 -m build --sdist @@ -85,4 +85,4 @@ clean: deploy.pypi \ deploy.gh-docs \ docker \ - clean \ No newline at end of file + clean From aaf4cbe81c4387a6afe264843649545be9ead669 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 00:47:51 -0400 Subject: [PATCH 463/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index a953af7e7..47077d8af 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.82] + +- feat: Update llama.cpp to ggerganov/llama.cpp@7fdb6f73e35605c8dbc39e9f19cd9ed84dbc87f2 + ## [0.2.81] - feat: Update llama.cpp to ggerganov/llama.cpp@968967376dc2c018d29f897c4883d335bbf384fb diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index f87e2a388..09ffb594e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.81" \ No newline at end of file +__version__ = "0.2.82" From 14760c63ed6dd7e518efd87142810911ea73b6e2 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue, 9 Jul 2024 11:51:35 -0400 Subject: [PATCH 464/624] chore(deps): bump pypa/cibuildwheel from 2.19.1 to 2.19.2 (#1568) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.19.1 to 2.19.2. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.19.1...v2.19.2) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Andrei --- .github/workflows/build-and-release.yaml | 4 ++-- .github/workflows/build-wheels-metal.yaml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index e736cf964..eeeda803d 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.19.1 + uses: pypa/cibuildwheel@v2.19.2 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -56,7 +56,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.19.1 + uses: pypa/cibuildwheel@v2.19.2 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index b56fb48aa..47c1c3cb5 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -30,7 +30,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.19.1 + uses: pypa/cibuildwheel@v2.19.2 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" From e31f0968ffe456778500fe7c75cbf7844a65238f Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue, 9 Jul 2024 11:51:46 -0400 Subject: [PATCH 465/624] chore(deps): bump microsoft/setup-msbuild from 1.1 to 1.3 (#1569) Bumps [microsoft/setup-msbuild](https://github.com/microsoft/setup-msbuild) from 1.1 to 1.3. - [Release notes](https://github.com/microsoft/setup-msbuild/releases) - [Changelog](https://github.com/microsoft/setup-msbuild/blob/main/building-release.md) - [Commits](https://github.com/microsoft/setup-msbuild/compare/v1.1...v1.3) --- updated-dependencies: - dependency-name: microsoft/setup-msbuild dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Andrei --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 5d445b5de..3c0267095 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -45,7 +45,7 @@ jobs: steps: - name: Add MSBuild to PATH if: runner.os == 'Windows' - uses: microsoft/setup-msbuild@v1.1 + uses: microsoft/setup-msbuild@v1.3 with: vs-version: '[16.11,16.12)' From b77e5070291c205034ba461b05c13f5a271e686f Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Tue, 9 Jul 2024 17:55:21 +0200 Subject: [PATCH 466/624] feat(ci): Dockerfile update base images and post-install cleanup (#1530) * Update Dockerfile * Update Dockerfile * Update Dockerfile * Update Dockerfile * Update Dockerfile * Update Dockerfile * Update Dockerfile * Update build-docker.yaml * Update Dockerfile * Update Dockerfile * Remove docker login --------- Co-authored-by: Andrei Betlen --- docker/cuda_simple/Dockerfile | 2 +- docker/open_llama/Dockerfile | 6 ++++-- docker/openblas_simple/Dockerfile | 7 +++++-- docker/simple/Dockerfile | 6 ++++-- 4 files changed, 14 insertions(+), 7 deletions(-) diff --git a/docker/cuda_simple/Dockerfile b/docker/cuda_simple/Dockerfile index a9e51cdc1..79b6a5bac 100644 --- a/docker/cuda_simple/Dockerfile +++ b/docker/cuda_simple/Dockerfile @@ -1,4 +1,4 @@ -ARG CUDA_IMAGE="12.1.1-devel-ubuntu22.04" +ARG CUDA_IMAGE="12.5.0-devel-ubuntu22.04" FROM nvidia/cuda:${CUDA_IMAGE} # We need to set the host to 0.0.0.0 to allow outside access diff --git a/docker/open_llama/Dockerfile b/docker/open_llama/Dockerfile index 23e37d4d4..70b16dffc 100644 --- a/docker/open_llama/Dockerfile +++ b/docker/open_llama/Dockerfile @@ -1,5 +1,5 @@ # Define the image argument and provide a default value -ARG IMAGE=python:3-slim-bullseye +ARG IMAGE=python:3-slim-bookworm # Use the image as specified FROM ${IMAGE} @@ -12,7 +12,9 @@ RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-reco python3 \ python3-pip \ ninja-build \ - build-essential + build-essential \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context diff --git a/docker/openblas_simple/Dockerfile b/docker/openblas_simple/Dockerfile index e213518b9..804c1b902 100644 --- a/docker/openblas_simple/Dockerfile +++ b/docker/openblas_simple/Dockerfile @@ -1,4 +1,4 @@ -FROM python:3-slim-bullseye +FROM python:3-slim-bookworm # We need to set the host to 0.0.0.0 to allow outside access ENV HOST 0.0.0.0 @@ -6,7 +6,10 @@ ENV HOST 0.0.0.0 COPY . . # Install the package -RUN apt update && apt install -y libopenblas-dev ninja-build build-essential pkg-config +RUN apt update && apt install -y libopenblas-dev ninja-build build-essential pkg-config \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* /tmp/* + RUN python -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context RUN CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama_cpp_python --verbose diff --git a/docker/simple/Dockerfile b/docker/simple/Dockerfile index 41cc68ea2..97a6c19df 100644 --- a/docker/simple/Dockerfile +++ b/docker/simple/Dockerfile @@ -1,5 +1,5 @@ # Define the image argument and provide a default value -ARG IMAGE=python:3-slim-bullseye +ARG IMAGE=python:3-slim-bookworm # Use the image as specified FROM ${IMAGE} @@ -13,7 +13,9 @@ RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-reco python3-pip \ ninja-build \ libopenblas-dev \ - build-essential + build-essential \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* /tmp/* RUN mkdir /app WORKDIR /app From c1ae81554f52f81b6027e36ae0a056dca9819c3c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 12:20:17 -0400 Subject: [PATCH 467/624] fix(misc): Format --- examples/batch-processing/server.py | 1 + examples/gradio_chat/local.py | 22 +- examples/gradio_chat/server.py | 21 +- examples/hf_pull/main.py | 21 +- examples/high_level_api/fastapi_server.py | 1 + .../high_level_api/high_level_api_infill.py | 20 +- examples/low_level_api/Chat.py | 52 +- examples/low_level_api/Miku.py | 46 +- examples/low_level_api/ReasonAct.py | 40 +- examples/low_level_api/common.py | 323 ++++- .../low_level_api/low_level_api_chat_cpp.py | 1197 ++++++++++------- .../low_level_api/low_level_api_llama_cpp.py | 48 +- examples/low_level_api/quantize.py | 7 +- examples/low_level_api/util.py | 156 +-- examples/notebooks/Batching.ipynb | 23 +- examples/notebooks/Clients.ipynb | 8 +- examples/notebooks/Functions.ipynb | 48 +- examples/notebooks/Guidance.ipynb | 12 +- examples/notebooks/Multimodal.ipynb | 19 +- .../notebooks/OpenHermesFunctionCalling.ipynb | 67 +- examples/notebooks/PerformanceTuning.ipynb | 19 +- examples/ray/llm.py | 1 + llama_cpp/_internals.py | 53 +- llama_cpp/_utils.py | 1 + llama_cpp/llama.py | 276 ++-- llama_cpp/llama_cache.py | 11 +- llama_cpp/llama_chat_format.py | 485 ++++--- llama_cpp/llama_cpp.py | 17 +- llama_cpp/llama_grammar.py | 568 +++++--- llama_cpp/llama_tokenizer.py | 11 +- llama_cpp/llama_types.py | 5 +- llama_cpp/server/__main__.py | 5 +- llama_cpp/server/app.py | 4 +- llama_cpp/server/settings.py | 12 +- llama_cpp/server/types.py | 2 +- 35 files changed, 2270 insertions(+), 1332 deletions(-) diff --git a/examples/batch-processing/server.py b/examples/batch-processing/server.py index d3536697a..0b36746f9 100644 --- a/examples/batch-processing/server.py +++ b/examples/batch-processing/server.py @@ -25,6 +25,7 @@ import openai.types.chat as types + @app.post("/v1/chat/completions") def create_chat_completions(): return {"message": "Hello World"} diff --git a/examples/gradio_chat/local.py b/examples/gradio_chat/local.py index a7de8e842..e16bf234a 100644 --- a/examples/gradio_chat/local.py +++ b/examples/gradio_chat/local.py @@ -6,25 +6,26 @@ llama = llama_cpp.Llama.from_pretrained( repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", filename="*q8_0.gguf", - tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B"), - verbose=False + tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained( + "Qwen/Qwen1.5-0.5B" + ), + verbose=False, ) model = "gpt-3.5-turbo" + def predict(message, history): messages = [] for user_message, assistant_message in history: messages.append({"role": "user", "content": user_message}) messages.append({"role": "assistant", "content": assistant_message}) - + messages.append({"role": "user", "content": message}) response = llama.create_chat_completion_openai_v1( - model=model, - messages=messages, - stream=True + model=model, messages=messages, stream=True ) text = "" @@ -52,7 +53,14 @@ def predict(message, history): """ with gr.Blocks(theme=gr.themes.Soft(), js=js, css=css, fill_height=True) as demo: - gr.ChatInterface(predict, fill_height=True, examples=["What is the capital of France?", "Who was the first person on the moon?"]) + gr.ChatInterface( + predict, + fill_height=True, + examples=[ + "What is the capital of France?", + "Who was the first person on the moon?", + ], + ) if __name__ == "__main__": diff --git a/examples/gradio_chat/server.py b/examples/gradio_chat/server.py index 36fa43fbd..52061bea6 100644 --- a/examples/gradio_chat/server.py +++ b/examples/gradio_chat/server.py @@ -2,26 +2,22 @@ from openai import OpenAI -client = OpenAI( - base_url="http://localhost:8000/v1", - api_key="llama.cpp" -) +client = OpenAI(base_url="http://localhost:8000/v1", api_key="llama.cpp") model = "gpt-3.5-turbo" + def predict(message, history): messages = [] for user_message, assistant_message in history: messages.append({"role": "user", "content": user_message}) messages.append({"role": "assistant", "content": assistant_message}) - + messages.append({"role": "user", "content": message}) response = client.chat.completions.create( - model=model, - messages=messages, - stream=True + model=model, messages=messages, stream=True ) text = "" @@ -49,7 +45,14 @@ def predict(message, history): """ with gr.Blocks(theme=gr.themes.Soft(), js=js, css=css, fill_height=True) as demo: - gr.ChatInterface(predict, fill_height=True, examples=["What is the capital of France?", "Who was the first person on the moon?"]) + gr.ChatInterface( + predict, + fill_height=True, + examples=[ + "What is the capital of France?", + "Who was the first person on the moon?", + ], + ) if __name__ == "__main__": diff --git a/examples/hf_pull/main.py b/examples/hf_pull/main.py index d3eb11c39..dfed17516 100644 --- a/examples/hf_pull/main.py +++ b/examples/hf_pull/main.py @@ -5,29 +5,26 @@ llama = llama_cpp.Llama.from_pretrained( repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", filename="*q8_0.gguf", - tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B"), - verbose=False + tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained( + "Qwen/Qwen1.5-0.5B" + ), + verbose=False, ) response = llama.create_chat_completion( - messages=[ - { - "role": "user", - "content": "What is the capital of France?" - } - ], + messages=[{"role": "user", "content": "What is the capital of France?"}], response_format={ "type": "json_object", "schema": { "type": "object", "properties": { "country": {"type": "string"}, - "capital": {"type": "string"} + "capital": {"type": "string"}, }, "required": ["country", "capital"], - } + }, }, - stream=True + stream=True, ) for chunk in response: @@ -36,4 +33,4 @@ continue print(delta["content"], end="", flush=True) -print() \ No newline at end of file +print() diff --git a/examples/high_level_api/fastapi_server.py b/examples/high_level_api/fastapi_server.py index 9421db57b..ee59767d6 100644 --- a/examples/high_level_api/fastapi_server.py +++ b/examples/high_level_api/fastapi_server.py @@ -24,6 +24,7 @@ To actually see the implementation of the server, see llama_cpp/server/app.py """ + import os import uvicorn diff --git a/examples/high_level_api/high_level_api_infill.py b/examples/high_level_api/high_level_api_infill.py index 27af6367e..282333e5a 100644 --- a/examples/high_level_api/high_level_api_infill.py +++ b/examples/high_level_api/high_level_api_infill.py @@ -6,16 +6,16 @@ parser.add_argument("-m", "--model", type=str, default="../models/7B/ggml-models.bin") parser.add_argument("-p", "--prompt", type=str, default="def add(") parser.add_argument("-s", "--suffix", type=str, default="\n return sum\n\n") -parser.add_argument("-i", "--spm-infill", action='store_true') +parser.add_argument("-i", "--spm-infill", action="store_true") args = parser.parse_args() llm = Llama(model_path=args.model, n_gpu_layers=-1, spm_infill=args.spm_infill) output = llm.create_completion( - temperature = 0.0, - repeat_penalty = 1.0, - prompt = args.prompt, - suffix = args.suffix, + temperature=0.0, + repeat_penalty=1.0, + prompt=args.prompt, + suffix=args.suffix, ) # Models sometimes repeat suffix in response, attempt to filter that @@ -25,9 +25,13 @@ unwanted_response_length = len(unwanted_response_suffix) filtered = False -if unwanted_response_suffix and response_stripped[-unwanted_response_length:] == unwanted_response_suffix: +if ( + unwanted_response_suffix + and response_stripped[-unwanted_response_length:] == unwanted_response_suffix +): response = response_stripped[:-unwanted_response_length] filtered = True -print(f"Fill-in-Middle completion{' (filtered)' if filtered else ''}:\n\n{args.prompt}\033[32m{response}\033[{'33' if filtered else '0'}m{args.suffix}\033[0m") - +print( + f"Fill-in-Middle completion{' (filtered)' if filtered else ''}:\n\n{args.prompt}\033[32m{response}\033[{'33' if filtered else '0'}m{args.suffix}\033[0m" +) diff --git a/examples/low_level_api/Chat.py b/examples/low_level_api/Chat.py index fcef8cd80..a755089b2 100644 --- a/examples/low_level_api/Chat.py +++ b/examples/low_level_api/Chat.py @@ -3,10 +3,12 @@ from common import GptParams from low_level_api_chat_cpp import LLaMAInteract + def env_or_def(env, default): - if (env in os.environ): - return os.environ[env] - return default + if env in os.environ: + return os.environ[env] + return default + AI_NAME = env_or_def("AI_NAME", "ChatLLaMa") MODEL = env_or_def("MODEL", "./models/llama-13B/ggml-model.bin") @@ -15,10 +17,10 @@ def env_or_def(env, default): N_THREAD = int(env_or_def("N_THREAD", "8")) today = datetime.datetime.today() -DATE_YEAR=today.strftime("%Y") -DATE_TIME=today.strftime("%H:%M") +DATE_YEAR = today.strftime("%Y") +DATE_TIME = today.strftime("%H:%M") -prompt=f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. +prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. {AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer {USER_NAME}'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. @@ -45,27 +47,29 @@ def env_or_def(env, default): {AI_NAME}: Blue. {USER_NAME}: What time is it? {AI_NAME}: It is {DATE_TIME}. -{USER_NAME}:""" + " ".join(sys.argv[1:]) +{USER_NAME}:""" + " ".join( + sys.argv[1:] +) print("Loading model...") params = GptParams( - n_ctx=2048, - temp=0.7, - top_k=40, - top_p=0.5, - repeat_last_n=256, - n_batch=1024, - repeat_penalty=1.17647, - model=MODEL, - n_threads=N_THREAD, - n_predict=N_PREDICTS, - use_color=True, - interactive=True, - antiprompt=[f"{USER_NAME}:"], - input_prefix=" ", - input_suffix=f"{AI_NAME}:", - prompt=prompt, + n_ctx=2048, + temp=0.7, + top_k=40, + top_p=0.5, + repeat_last_n=256, + n_batch=1024, + repeat_penalty=1.17647, + model=MODEL, + n_threads=N_THREAD, + n_predict=N_PREDICTS, + use_color=True, + interactive=True, + antiprompt=[f"{USER_NAME}:"], + input_prefix=" ", + input_suffix=f"{AI_NAME}:", + prompt=prompt, ) with LLaMAInteract(params) as m: - m.interact() + m.interact() diff --git a/examples/low_level_api/Miku.py b/examples/low_level_api/Miku.py index eb9a2cfa9..e072ab1b1 100644 --- a/examples/low_level_api/Miku.py +++ b/examples/low_level_api/Miku.py @@ -3,10 +3,12 @@ from common import GptParams from low_level_api_chat_cpp import LLaMAInteract + def env_or_def(env, default): - if (env in os.environ): - return os.environ[env] - return default + if env in os.environ: + return os.environ[env] + return default + AI_NAME = env_or_def("AI_NAME", "Miku") MODEL = env_or_def("MODEL", "./models/llama-13B/ggml-model.bin") @@ -14,7 +16,7 @@ def env_or_def(env, default): N_PREDICTS = int(env_or_def("N_PREDICTS", "4096")) N_THREAD = int(env_or_def("N_THREAD", "0")) -prompt=f"""This is a transcript of a 1000 page, never ending conversation between {USER_NAME} and the cute and helpful AI assistant {AI_NAME}. {AI_NAME} is a girl who is an AI running on the users computer. +prompt = f"""This is a transcript of a 1000 page, never ending conversation between {USER_NAME} and the cute and helpful AI assistant {AI_NAME}. {AI_NAME} is a girl who is an AI running on the users computer. {AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next. {AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct she will ask the user for help. {AI_NAME} is a very helpful AI and will help the user with anything they need, she is also very friendly and will try to make the user feel better if they are sad. @@ -32,28 +34,30 @@ def env_or_def(env, default): {AI_NAME}: /think It sounds like {USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off! {AI_NAME}: /think I wonder what {USER_NAME} likes to do in his free time? I should ask him about that! {AI_NAME}: What do you like to do in your free time? ^_^ -{USER_NAME}:""" + " ".join(sys.argv[1:]) +{USER_NAME}:""" + " ".join( + sys.argv[1:] +) print("Loading model...") params = GptParams( - n_batch=1024, - n_ctx=2048, - n_keep=-1, - repeat_last_n=256, - repeat_penalty=1.17647, - temp=0.7, - top_k=40, - top_p=0.5, - model=MODEL, - n_predict=N_PREDICTS, - use_color=True, - interactive=True, - antiprompt=[f"{USER_NAME}:"], - prompt=prompt, + n_batch=1024, + n_ctx=2048, + n_keep=-1, + repeat_last_n=256, + repeat_penalty=1.17647, + temp=0.7, + top_k=40, + top_p=0.5, + model=MODEL, + n_predict=N_PREDICTS, + use_color=True, + interactive=True, + antiprompt=[f"{USER_NAME}:"], + prompt=prompt, ) if N_THREAD > 0: - params.n_threads = N_THREAD + params.n_threads = N_THREAD with LLaMAInteract(params) as m: - m.interact() + m.interact() diff --git a/examples/low_level_api/ReasonAct.py b/examples/low_level_api/ReasonAct.py index 82e5c4487..1f2c59017 100644 --- a/examples/low_level_api/ReasonAct.py +++ b/examples/low_level_api/ReasonAct.py @@ -3,14 +3,16 @@ from common import GptParams from low_level_api_chat_cpp import LLaMAInteract + def env_or_def(env, default): - if (env in os.environ): - return os.environ[env] - return default + if env in os.environ: + return os.environ[env] + return default + MODEL = env_or_def("MODEL", "./models/llama-13B/ggml-model.bin") -prompt=f"""You run in a loop of Thought, Action, Observation. +prompt = f"""You run in a loop of Thought, Action, Observation. At the end of the loop either Answer or restate your Thought and Action. Use Thought to describe your thoughts about the question you have been asked. Use Action to run one of these actions available to you: @@ -27,23 +29,25 @@ def env_or_def(env, default): Question: What is capital of france? Thought: Do I need to use an action? No, I know the answer Answer: Paris is the capital of France -Question:""" + " ".join(sys.argv[1:]) +Question:""" + " ".join( + sys.argv[1:] +) print("Loading model...") params = GptParams( - interactive=True, - interactive_start=True, - top_k=10000, - temp=0.2, - repeat_penalty=1, - n_threads=7, - n_ctx=2048, - antiprompt=["Question:","Observation:"], - model=MODEL, - input_prefix=" ", - n_predict=-1, - prompt=prompt, + interactive=True, + interactive_start=True, + top_k=10000, + temp=0.2, + repeat_penalty=1, + n_threads=7, + n_ctx=2048, + antiprompt=["Question:", "Observation:"], + model=MODEL, + input_prefix=" ", + n_predict=-1, + prompt=prompt, ) with LLaMAInteract(params) as m: - m.interact() + m.interact() diff --git a/examples/low_level_api/common.py b/examples/low_level_api/common.py index 1a5152530..a0212ff0d 100644 --- a/examples/low_level_api/common.py +++ b/examples/low_level_api/common.py @@ -65,82 +65,242 @@ class GptParams: # Set to "\nUser:" etc. # This is an alternative to input_prefix which always adds it, so it potentially duplicates "User:"" fix_prefix: str = "" - input_echo: bool = True, + input_echo: bool = (True,) # Default instructions for Alpaca # switch to "Human" and "Assistant" for Vicuna. # TODO: TBD how they are gonna handle this upstream - instruct_inp_prefix: str="\n\n### Instruction:\n\n" - instruct_inp_suffix: str="\n\n### Response:\n\n" + instruct_inp_prefix: str = "\n\n### Instruction:\n\n" + instruct_inp_suffix: str = "\n\n### Response:\n\n" -def gpt_params_parse(argv = None): - parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) - parser.add_argument("-s", "--seed", type=int, default=-1, help="RNG seed (use random seed for <= 0)",dest="seed") - parser.add_argument("-t", "--threads", type=int, default=min(4, os.cpu_count() or 1), help="number of threads to use during computation",dest="n_threads") - parser.add_argument("-n", "--n_predict", type=int, default=128, help="number of tokens to predict (-1 = infinity)",dest="n_predict") - parser.add_argument("--n_parts", type=int, default=-1, help="number of model parts", dest="n_parts") - parser.add_argument("-c", "--ctx_size", type=int, default=512, help="size of the prompt context",dest="n_ctx") - parser.add_argument("-b", "--batch_size", type=int, default=8, help="batch size for prompt processing",dest="n_batch") - parser.add_argument("--keep", type=int, default=0, help="number of tokens to keep from the initial prompt",dest="n_keep") +def gpt_params_parse(argv=None): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + "-s", + "--seed", + type=int, + default=-1, + help="RNG seed (use random seed for <= 0)", + dest="seed", + ) + parser.add_argument( + "-t", + "--threads", + type=int, + default=min(4, os.cpu_count() or 1), + help="number of threads to use during computation", + dest="n_threads", + ) + parser.add_argument( + "-n", + "--n_predict", + type=int, + default=128, + help="number of tokens to predict (-1 = infinity)", + dest="n_predict", + ) + parser.add_argument( + "--n_parts", type=int, default=-1, help="number of model parts", dest="n_parts" + ) + parser.add_argument( + "-c", + "--ctx_size", + type=int, + default=512, + help="size of the prompt context", + dest="n_ctx", + ) + parser.add_argument( + "-b", + "--batch_size", + type=int, + default=8, + help="batch size for prompt processing", + dest="n_batch", + ) + parser.add_argument( + "--keep", + type=int, + default=0, + help="number of tokens to keep from the initial prompt", + dest="n_keep", + ) parser.add_argument( "-l", "--logit-bias", type=str, - action='append', + action="append", help="--logit-bias TOKEN_ID(+/-)BIAS", - dest="logit_bias_str" - ) - parser.add_argument("--ignore-eos", action="store_true", help="ignore end of stream token and continue generating", dest="ignore_eos") - parser.add_argument("--top_k", type=int, default=40, help="top-k sampling",dest="top_k") - parser.add_argument("--top_p", type=float, default=0.95, help="top-p samplin",dest="top_p") - parser.add_argument("--tfs", type=float, default=1.0, help="tail free sampling, parameter z (1.0 = disabled)",dest="tfs_z") - parser.add_argument("--temp", type=float, default=0.80, help="temperature",dest="temp") - parser.add_argument("--repeat_penalty", type=float, default=1.10, help="penalize repeat sequence of tokens",dest="repeat_penalty") - parser.add_argument("--repeat_last_n", type=int, default=64, help="last n tokens to consider for penalize ",dest="repeat_last_n") - parser.add_argument("--frequency_penalty", type=float, default=0.0, help="repeat alpha frequency penalty (0.0 = disabled)",dest="tfs_z") - parser.add_argument("--presence_penalty", type=float, default=0.0, help="repeat alpha presence penalty (0.0 = disabled)",dest="presence_penalty") - parser.add_argument("--mirostat", type=float, default=1.0, help="use Mirostat sampling.",dest="mirostat") - parser.add_argument("--mirostat_ent", type=float, default=5.0, help="Mirostat target entropy, parameter tau represents the average surprise value",dest="mirostat_tau") - parser.add_argument("--mirostat_lr", type=float, default=0.1, help="Mirostat learning rate, parameter eta",dest="mirostat_eta") - - parser.add_argument("-m", "--model", type=str, default="./models/llama-7B/ggml-model.bin", help="model path",dest="model") - parser.add_argument("-p", "--prompt", type=str, default=None, help="initial prompt",dest="prompt") - parser.add_argument("-f", "--file", type=str, default=None, help="file containing initial prompt to load",dest="file") - parser.add_argument("--session", type=str, default=None, help="file to cache model state in (may be large!)",dest="path_session") - parser.add_argument("--in-prefix", type=str, default="", help="string to prefix user inputs with", dest="input_prefix") - parser.add_argument("--in-suffix", type=str, default="", help="append to input", dest="input_suffix") + dest="logit_bias_str", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="ignore end of stream token and continue generating", + dest="ignore_eos", + ) + parser.add_argument( + "--top_k", type=int, default=40, help="top-k sampling", dest="top_k" + ) + parser.add_argument( + "--top_p", type=float, default=0.95, help="top-p samplin", dest="top_p" + ) + parser.add_argument( + "--tfs", + type=float, + default=1.0, + help="tail free sampling, parameter z (1.0 = disabled)", + dest="tfs_z", + ) + parser.add_argument( + "--temp", type=float, default=0.80, help="temperature", dest="temp" + ) + parser.add_argument( + "--repeat_penalty", + type=float, + default=1.10, + help="penalize repeat sequence of tokens", + dest="repeat_penalty", + ) + parser.add_argument( + "--repeat_last_n", + type=int, + default=64, + help="last n tokens to consider for penalize ", + dest="repeat_last_n", + ) + parser.add_argument( + "--frequency_penalty", + type=float, + default=0.0, + help="repeat alpha frequency penalty (0.0 = disabled)", + dest="tfs_z", + ) + parser.add_argument( + "--presence_penalty", + type=float, + default=0.0, + help="repeat alpha presence penalty (0.0 = disabled)", + dest="presence_penalty", + ) + parser.add_argument( + "--mirostat", + type=float, + default=1.0, + help="use Mirostat sampling.", + dest="mirostat", + ) + parser.add_argument( + "--mirostat_ent", + type=float, + default=5.0, + help="Mirostat target entropy, parameter tau represents the average surprise value", + dest="mirostat_tau", + ) + parser.add_argument( + "--mirostat_lr", + type=float, + default=0.1, + help="Mirostat learning rate, parameter eta", + dest="mirostat_eta", + ) + + parser.add_argument( + "-m", + "--model", + type=str, + default="./models/llama-7B/ggml-model.bin", + help="model path", + dest="model", + ) + parser.add_argument( + "-p", "--prompt", type=str, default=None, help="initial prompt", dest="prompt" + ) + parser.add_argument( + "-f", + "--file", + type=str, + default=None, + help="file containing initial prompt to load", + dest="file", + ) + parser.add_argument( + "--session", + type=str, + default=None, + help="file to cache model state in (may be large!)", + dest="path_session", + ) + parser.add_argument( + "--in-prefix", + type=str, + default="", + help="string to prefix user inputs with", + dest="input_prefix", + ) + parser.add_argument( + "--in-suffix", type=str, default="", help="append to input", dest="input_suffix" + ) parser.add_argument( "-r", "--reverse-prompt", type=str, - action='append', + action="append", help="poll user input upon seeing PROMPT (can be\nspecified more than once for multiple prompts).", - dest="antiprompt" + dest="antiprompt", + ) + + parser.add_argument( + "--lora", + type=str, + default="", + help="apply LoRA adapter (implies --no-mmap)", + dest="lora_adapter", + ) + parser.add_argument( + "--lora-base", + type=str, + default="", + help="optional model to use as a base for the layers modified by the LoRA adapter", + dest="lora_base", ) - - parser.add_argument("--lora", type=str, default="", help="apply LoRA adapter (implies --no-mmap)", dest="lora_adapter") - parser.add_argument("--lora-base", type=str, default="", help="optional model to use as a base for the layers modified by the LoRA adapter", dest="lora_base") - parser.add_argument("--memory_f32", action="store_false", help="use f32 instead of f16 for memory key+value",dest="memory_f16") - parser.add_argument("--random-prompt", action="store_true", help="start with a randomized prompt.", dest="random_prompt") + parser.add_argument( + "--memory_f32", + action="store_false", + help="use f32 instead of f16 for memory key+value", + dest="memory_f16", + ) + parser.add_argument( + "--random-prompt", + action="store_true", + help="start with a randomized prompt.", + dest="random_prompt", + ) parser.add_argument( "--color", action="store_true", help="colorise output to distinguish prompt and user input from generations", - dest="use_color" + dest="use_color", ) parser.add_argument( - "-i", "--interactive", action="store_true", help="run in interactive mode", dest="interactive" + "-i", + "--interactive", + action="store_true", + help="run in interactive mode", + dest="interactive", ) - + parser.add_argument("--embedding", action="store_true", help="", dest="embedding") parser.add_argument( "--interactive-first", action="store_true", help="run in interactive mode and wait for input right away", - dest="interactive_start" + dest="interactive_start", ) parser.add_argument( @@ -148,42 +308,84 @@ def gpt_params_parse(argv = None): "--instruct", action="store_true", help="run in instruction mode (use with Alpaca or Vicuna models)", - dest="instruct" + dest="instruct", + ) + parser.add_argument( + "--no-penalize-nl", + action="store_false", + help="do not penalize newline token", + dest="penalize_nl", + ) + parser.add_argument( + "--perplexity", + action="store_true", + help="compute perplexity over the prompt", + dest="perplexity", + ) + parser.add_argument( + "--no-mmap", + action="store_false", + help="do not memory-map model (slower load but may reduce pageouts if not using mlock)", + dest="use_mmap", + ) + parser.add_argument( + "--mlock", + action="store_true", + help="force system to keep model in RAM rather than swapping or compressing", + dest="use_mlock", + ) + parser.add_argument( + "--mtest", + action="store_true", + help="compute maximum memory usage", + dest="mem_test", + ) + parser.add_argument( + "--verbose-prompt", + action="store_true", + help="print prompt before generation", + dest="verbose_prompt", ) - parser.add_argument("--no-penalize-nl", action="store_false", help="do not penalize newline token", dest="penalize_nl") - parser.add_argument("--perplexity", action="store_true", help="compute perplexity over the prompt", dest="perplexity") - parser.add_argument("--no-mmap", action="store_false",help="do not memory-map model (slower load but may reduce pageouts if not using mlock)",dest="use_mmap") - parser.add_argument("--mlock", action="store_true",help="force system to keep model in RAM rather than swapping or compressing",dest="use_mlock") - parser.add_argument("--mtest", action="store_true",help="compute maximum memory usage",dest="mem_test") - parser.add_argument("--verbose-prompt", action="store_true",help="print prompt before generation",dest="verbose_prompt") - #Custom args - parser.add_argument("--fix-prefix", type=str, default="", help="append to input when generated n_predict tokens", dest="fix_prefix") - parser.add_argument("--input-noecho", action="store_false", help="dont output the input", dest="input_echo") + # Custom args + parser.add_argument( + "--fix-prefix", + type=str, + default="", + help="append to input when generated n_predict tokens", + dest="fix_prefix", + ) + parser.add_argument( + "--input-noecho", + action="store_false", + help="dont output the input", + dest="input_echo", + ) parser.add_argument( "--interactive-start", action="store_true", help="run in interactive mode", - dest="interactive" + dest="interactive", ) args = parser.parse_args(argv) - + logit_bias_str = args.logit_bias_str delattr(args, "logit_bias_str") params = GptParams(**vars(args)) - if (params.lora_adapter): + if params.lora_adapter: params.use_mmap = False - if (logit_bias_str != None): + if logit_bias_str != None: for i in logit_bias_str: - if (m := re.match(r"(\d+)([-+]\d+)", i)): + if m := re.match(r"(\d+)([-+]\d+)", i): params.logit_bias[int(m.group(1))] = float(m.group(2)) return params + def gpt_random_prompt(rng): return [ "So", @@ -198,5 +400,6 @@ def gpt_random_prompt(rng): "They", ][rng % 10] + if __name__ == "__main__": print(gpt_params_parse()) diff --git a/examples/low_level_api/low_level_api_chat_cpp.py b/examples/low_level_api/low_level_api_chat_cpp.py index 02c09afb0..39081be17 100644 --- a/examples/low_level_api/low_level_api_chat_cpp.py +++ b/examples/low_level_api/low_level_api_chat_cpp.py @@ -10,6 +10,7 @@ You should also still be feeding the model with a "primer" prompt that shows it the expected format. """ + import ctypes import sys from time import time @@ -19,198 +20,257 @@ from common import GptParams, gpt_params_parse, gpt_random_prompt import util + # A LLaMA interactive session class LLaMAInteract: - def __init__(self, params: GptParams) -> None: - # input args - self.params = params - if self.params.path_session is None: - self.params.path_session = "" - if self.params.antiprompt is None: - self.params.antiprompt = "" - - if (self.params.perplexity): - raise NotImplementedError("""************ + def __init__(self, params: GptParams) -> None: + # input args + self.params = params + if self.params.path_session is None: + self.params.path_session = "" + if self.params.antiprompt is None: + self.params.antiprompt = "" + + if self.params.perplexity: + raise NotImplementedError( + """************ please use the 'perplexity' tool for perplexity calculations -************""") +************""" + ) - if (self.params.embedding): - raise NotImplementedError("""************ + if self.params.embedding: + raise NotImplementedError( + """************ please use the 'embedding' tool for embedding calculations -************""") +************""" + ) - if (self.params.n_ctx > 2048): - print(f"""warning: model does not support \ + if self.params.n_ctx > 2048: + print( + f"""warning: model does not support \ context sizes greater than 2048 tokens ({self.params.n_ctx} \ -specified) expect poor results""", file=sys.stderr) - - if (self.params.seed <= 0): - self.params.seed = int(time()) - - print(f"seed = {self.params.seed}", file=sys.stderr) - - if (self.params.random_prompt): - self.params.prompt = gpt_random_prompt(self.params.seed) - - # runtime args - self.input_consumed = 0 - self.n_past = 0 - self.n_session_consumed = 0 - self.first_antiprompt = [] - self.remaining_tokens = self.params.n_predict - self.output_echo = self.params.input_echo - self.multibyte_fix = [] - - # model load - self.lparams = llama_cpp.llama_model_default_params() - self.lparams.n_ctx = self.params.n_ctx - self.lparams.n_parts = self.params.n_parts - self.lparams.seed = self.params.seed - self.lparams.memory_f16 = self.params.memory_f16 - self.lparams.use_mlock = self.params.use_mlock - self.lparams.use_mmap = self.params.use_mmap - - self.model = llama_cpp.llama_load_model_from_file( - self.params.model.encode("utf8"), self.lparams) - - # Context Params. - self.cparams = llama_cpp.llama_context_default_params() - - self.ctx = llama_cpp.llama_new_context_with_model(self.model, self.cparams) - if (not self.ctx): - raise RuntimeError(f"error: failed to load model '{self.params.model}'") - - if (self.params.ignore_eos): - self.params.logit_bias[llama_cpp.llama_token_eos()] = -float("inf") - - if (len(self.params.lora_adapter) > 0): - if (llama_cpp.llama_apply_lora_from_file( - self.ctx, - self.params.lora_adapter.encode("utf8"), - self.params.lora_base.encode("utf8") if len(self.params.lora_base) > 0 else None, - self.params.n_threads - ) != 0): - print("error: failed to apply lora adapter") - return - - print(file=sys.stderr) - print(f"system_info: n_threads = {self.params.n_threads} / {cpu_count()} \ -| {llama_cpp.llama_print_system_info().decode('utf8')}", file=sys.stderr) - - # determine the required inference memory per token: - if (self.params.mem_test): - tmp = [0, 1, 2, 3] - llama_cpp.llama_eval(self.ctx, (llama_cpp.c_int * len(tmp))(*tmp), len(tmp), 0, self.n_threads) - llama_cpp.llama_print_timings(self.ctx) - self.exit() - return - - # create internal context - self.n_ctx = llama_cpp.llama_n_ctx(self.ctx) - - # Add a space in front of the first character to match OG llama tokenizer behavior - self.params.prompt = " " + self.params.prompt - - # Load prompt file - if (self.params.file): - with open(self.params.file) as f: - self.params.prompt = f.read() - - self.session_tokens: list[llama_cpp.llama_token] = [] - if (len(self.params.path_session) > 0): - print(f"attempting to load saved session from '{self.params.path_session}'", file=sys.stderr) - - if (path.exists(self.params.path_session)): - _session_tokens = (llama_cpp.llama_token * (self.params.n_ctx))() - _n_token_count_out = llama_cpp.c_size_t() - if (llama_cpp.llama_load_session_file( - self.ctx, - self.params.path_session.encode("utf8"), - _session_tokens, - self.params.n_ctx, - ctypes.byref(_n_token_count_out) - ) != 1): - print(f"error: failed to load session file '{self.params.path_session}'", file=sys.stderr) - return - _n_token_count_out = _n_token_count_out.value - self.session_tokens = _session_tokens[:_n_token_count_out] - print(f"loaded a session with prompt size of {_n_token_count_out} tokens", file=sys.stderr) - else: - print(f"session file does not exist, will create", file=sys.stderr) - - # tokenize the prompt - self.embd = [] - self.embd_inp = self._tokenize(self.params.prompt) - - if (len(self.embd_inp) > self.n_ctx - 4): - raise RuntimeError(f"error: prompt is too long ({len(self.embd_inp)} tokens, max {self.params.n_ctx - 4})") - - # debug message about similarity of saved session, if applicable - self.n_matching_session_tokens = 0 - if len(self.session_tokens) > 0: - for id in self.session_tokens: - if self.n_matching_session_tokens >= len(self.embd_inp) or id != self.embd_inp[self.n_matching_session_tokens]: - break - self.n_matching_session_tokens += 1 - - if self.n_matching_session_tokens >= len(self.embd_inp): - print(f"session file has exact match for prompt!") - elif self.n_matching_session_tokens < (len(self.embd_inp) / 2): - print(f"warning: session file has low similarity to prompt ({self.n_matching_session_tokens} / {len(self.embd_inp)} tokens); will mostly be reevaluated") - else: - print(f"session file matches {self.n_matching_session_tokens} / {len(self.embd_inp)} tokens of prompt") - - self.need_to_save_session = len(self.params.path_session) > 0 and self.n_matching_session_tokens < (len(self.embd_inp) * 3 / 4) - - # number of tokens to keep when resetting context - if (self.params.n_keep < 0 or self.params.n_keep > len(self.embd_inp) or self.params.instruct): - self.params.n_keep = len(self.embd_inp) - - self.inp_prefix = self._tokenize(self.params.instruct_inp_prefix) - self.inp_suffix = self._tokenize(self.params.instruct_inp_suffix, False) - - # in instruct mode, we inject a prefix and a suffix to each input by the user - self.antiecho = None - if (self.params.instruct): - self.params.interactive_start = True - _ptn = self._tokenize(self.params.instruct_inp_prefix.strip(), False) - self.first_antiprompt.append(_ptn) - self.antiecho = util.IterSearch(_ptn) - - # enable interactive mode if reverse prompt or interactive start is specified - if (len(self.params.antiprompt) != 0 or self.params.interactive_start): - self.params.interactive = True - - # determine newline token - self.llama_token_newline = self._tokenize("\n", False) - self.llama_token_eot = self._tokenize(" [end of text]\n", False) - - if (self.params.verbose_prompt): - print(f""" +specified) expect poor results""", + file=sys.stderr, + ) + + if self.params.seed <= 0: + self.params.seed = int(time()) + + print(f"seed = {self.params.seed}", file=sys.stderr) + + if self.params.random_prompt: + self.params.prompt = gpt_random_prompt(self.params.seed) + + # runtime args + self.input_consumed = 0 + self.n_past = 0 + self.n_session_consumed = 0 + self.first_antiprompt = [] + self.remaining_tokens = self.params.n_predict + self.output_echo = self.params.input_echo + self.multibyte_fix = [] + + # model load + self.lparams = llama_cpp.llama_model_default_params() + self.lparams.n_ctx = self.params.n_ctx + self.lparams.n_parts = self.params.n_parts + self.lparams.seed = self.params.seed + self.lparams.memory_f16 = self.params.memory_f16 + self.lparams.use_mlock = self.params.use_mlock + self.lparams.use_mmap = self.params.use_mmap + + self.model = llama_cpp.llama_load_model_from_file( + self.params.model.encode("utf8"), self.lparams + ) + + # Context Params. + self.cparams = llama_cpp.llama_context_default_params() + + self.ctx = llama_cpp.llama_new_context_with_model(self.model, self.cparams) + if not self.ctx: + raise RuntimeError(f"error: failed to load model '{self.params.model}'") + + if self.params.ignore_eos: + self.params.logit_bias[llama_cpp.llama_token_eos()] = -float("inf") + + if len(self.params.lora_adapter) > 0: + if ( + llama_cpp.llama_apply_lora_from_file( + self.ctx, + self.params.lora_adapter.encode("utf8"), + ( + self.params.lora_base.encode("utf8") + if len(self.params.lora_base) > 0 + else None + ), + self.params.n_threads, + ) + != 0 + ): + print("error: failed to apply lora adapter") + return + + print(file=sys.stderr) + print( + f"system_info: n_threads = {self.params.n_threads} / {cpu_count()} \ +| {llama_cpp.llama_print_system_info().decode('utf8')}", + file=sys.stderr, + ) + + # determine the required inference memory per token: + if self.params.mem_test: + tmp = [0, 1, 2, 3] + llama_cpp.llama_eval( + self.ctx, + (llama_cpp.c_int * len(tmp))(*tmp), + len(tmp), + 0, + self.n_threads, + ) + llama_cpp.llama_print_timings(self.ctx) + self.exit() + return + + # create internal context + self.n_ctx = llama_cpp.llama_n_ctx(self.ctx) + + # Add a space in front of the first character to match OG llama tokenizer behavior + self.params.prompt = " " + self.params.prompt + + # Load prompt file + if self.params.file: + with open(self.params.file) as f: + self.params.prompt = f.read() + + self.session_tokens: list[llama_cpp.llama_token] = [] + if len(self.params.path_session) > 0: + print( + f"attempting to load saved session from '{self.params.path_session}'", + file=sys.stderr, + ) + + if path.exists(self.params.path_session): + _session_tokens = (llama_cpp.llama_token * (self.params.n_ctx))() + _n_token_count_out = llama_cpp.c_size_t() + if ( + llama_cpp.llama_load_session_file( + self.ctx, + self.params.path_session.encode("utf8"), + _session_tokens, + self.params.n_ctx, + ctypes.byref(_n_token_count_out), + ) + != 1 + ): + print( + f"error: failed to load session file '{self.params.path_session}'", + file=sys.stderr, + ) + return + _n_token_count_out = _n_token_count_out.value + self.session_tokens = _session_tokens[:_n_token_count_out] + print( + f"loaded a session with prompt size of {_n_token_count_out} tokens", + file=sys.stderr, + ) + else: + print(f"session file does not exist, will create", file=sys.stderr) + + # tokenize the prompt + self.embd = [] + self.embd_inp = self._tokenize(self.params.prompt) + + if len(self.embd_inp) > self.n_ctx - 4: + raise RuntimeError( + f"error: prompt is too long ({len(self.embd_inp)} tokens, max {self.params.n_ctx - 4})" + ) + + # debug message about similarity of saved session, if applicable + self.n_matching_session_tokens = 0 + if len(self.session_tokens) > 0: + for id in self.session_tokens: + if ( + self.n_matching_session_tokens >= len(self.embd_inp) + or id != self.embd_inp[self.n_matching_session_tokens] + ): + break + self.n_matching_session_tokens += 1 + + if self.n_matching_session_tokens >= len(self.embd_inp): + print(f"session file has exact match for prompt!") + elif self.n_matching_session_tokens < (len(self.embd_inp) / 2): + print( + f"warning: session file has low similarity to prompt ({self.n_matching_session_tokens} / {len(self.embd_inp)} tokens); will mostly be reevaluated" + ) + else: + print( + f"session file matches {self.n_matching_session_tokens} / {len(self.embd_inp)} tokens of prompt" + ) + + self.need_to_save_session = len( + self.params.path_session + ) > 0 and self.n_matching_session_tokens < (len(self.embd_inp) * 3 / 4) + + # number of tokens to keep when resetting context + if ( + self.params.n_keep < 0 + or self.params.n_keep > len(self.embd_inp) + or self.params.instruct + ): + self.params.n_keep = len(self.embd_inp) + + self.inp_prefix = self._tokenize(self.params.instruct_inp_prefix) + self.inp_suffix = self._tokenize(self.params.instruct_inp_suffix, False) + + # in instruct mode, we inject a prefix and a suffix to each input by the user + self.antiecho = None + if self.params.instruct: + self.params.interactive_start = True + _ptn = self._tokenize(self.params.instruct_inp_prefix.strip(), False) + self.first_antiprompt.append(_ptn) + self.antiecho = util.IterSearch(_ptn) + + # enable interactive mode if reverse prompt or interactive start is specified + if len(self.params.antiprompt) != 0 or self.params.interactive_start: + self.params.interactive = True + + # determine newline token + self.llama_token_newline = self._tokenize("\n", False) + self.llama_token_eot = self._tokenize(" [end of text]\n", False) + + if self.params.verbose_prompt: + print( + f""" prompt: '{self.params.prompt}' -number of tokens in prompt = {len(self.embd_inp)}""", file=sys.stderr) - - for i in range(len(self.embd_inp)): - print(f"{self.embd_inp[i]} -> '{self.token_to_str(self.embd_inp[i])}'", file=sys.stderr) - - if (self.params.n_keep > 0): - print("static prompt based on n_keep: '") - for i in range(self.params.n_keep): - print(self.token_to_str(self.embd_inp[i]), file=sys.stderr) - print("'", file=sys.stderr) - print(file=sys.stderr) - - if (self.params.interactive): - print("interactive mode on.", file=sys.stderr) - - if (len(self.params.antiprompt) > 0): - for antiprompt in self.params.antiprompt: - print(f"Reverse prompt: '{antiprompt}'", file=sys.stderr) - - if len(self.params.input_prefix) > 0: - print(f"Input prefix: '{self.params.input_prefix}'", file=sys.stderr) - - print(f"""sampling: repeat_last_n = {self.params.repeat_last_n},\ +number of tokens in prompt = {len(self.embd_inp)}""", + file=sys.stderr, + ) + + for i in range(len(self.embd_inp)): + print( + f"{self.embd_inp[i]} -> '{self.token_to_str(self.embd_inp[i])}'", + file=sys.stderr, + ) + + if self.params.n_keep > 0: + print("static prompt based on n_keep: '") + for i in range(self.params.n_keep): + print(self.token_to_str(self.embd_inp[i]), file=sys.stderr) + print("'", file=sys.stderr) + print(file=sys.stderr) + + if self.params.interactive: + print("interactive mode on.", file=sys.stderr) + + if len(self.params.antiprompt) > 0: + for antiprompt in self.params.antiprompt: + print(f"Reverse prompt: '{antiprompt}'", file=sys.stderr) + + if len(self.params.input_prefix) > 0: + print(f"Input prefix: '{self.params.input_prefix}'", file=sys.stderr) + + print( + f"""sampling: repeat_last_n = {self.params.repeat_last_n},\ repeat_penalty = {self.params.repeat_penalty},\ presence_penalty = {self.params.presence_penalty},\ frequency_penalty = {self.params.frequency_penalty},\ @@ -228,77 +288,96 @@ def __init__(self, params: GptParams) -> None: n_predict = {self.params.n_predict},\ n_keep = {self.params.n_keep} -""", file=sys.stderr) +""", + file=sys.stderr, + ) - # determine antiprompt tokens - for i in self.params.antiprompt: - self.first_antiprompt.append(self._tokenize(i, False)) + # determine antiprompt tokens + for i in self.params.antiprompt: + self.first_antiprompt.append(self._tokenize(i, False)) - self.last_n_tokens = [0]*self.n_ctx #TODO: deque doesnt support slices + self.last_n_tokens = [0] * self.n_ctx # TODO: deque doesnt support slices - if (params.interactive): - print("""== Running in interactive mode. == + if params.interactive: + print( + """== Running in interactive mode. == - Press Ctrl+C to interject at any time. - Press Return to return control to LLaMa. - If you want to submit another line, end your input in '\\'. -""", file=sys.stderr) - self.set_color(util.CONSOLE_COLOR_PROMPT) - - # tokenize a prompt - def _tokenize(self, prompt, bos=True): - _arr = (llama_cpp.llama_token * ((len(prompt) + 1) * 4))() - _n = llama_cpp.llama_tokenize(self.model, prompt.encode("utf8", errors="ignore"), len(prompt), _arr, len(_arr), bos, False) - return _arr[:_n] - - def set_color(self, c): - if (self.params.use_color): - print(c, end="") - - def use_antiprompt(self): - return len(self.first_antiprompt) > 0 - - # generate tokens - def generate(self): - while self.remaining_tokens > 0 or self.params.interactive or self.params.n_predict == -1: - # predict - if len(self.embd) > 0: - # infinite text generation via context swapping - # if we run out of context: - # - take the n_keep first tokens from the original prompt (via n_past) - # - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch - if (self.n_past + len(self.embd) > self.n_ctx): - n_left = self.n_past - self.params.n_keep - self.n_past = self.params.n_keep - - # insert n_left/2 tokens at the start of embd from last_n_tokens - _insert = self.last_n_tokens[ - self.n_ctx - int(n_left/2) - len(self.embd):-len(self.embd) - ] - self.embd = _insert + self.embd - self.params.path_session = "" - - # try to reuse a matching prefix from the loaded session instead of re-eval (via n_past) - if self.n_session_consumed < len(self.session_tokens): - for i in range(len(self.embd)): - if self.embd[i] != self.session_tokens[self.n_session_consumed]: - self.session_tokens = self.session_tokens[:self.n_session_consumed] - break - - self.n_past += 1 - self.n_session_consumed += 1 - - if self.n_session_consumed >= len(self.session_tokens): - i += 1 - break - - if i > 0: - self.embd = self.embd[i:] - - # evaluate tokens in batches - # embd is typically prepared beforehand to fit within a batch, but not always - #TODO BUG: The batching code causes nonsensical generation - """for i in range(0, len(self.embd), self.params.n_batch): +""", + file=sys.stderr, + ) + self.set_color(util.CONSOLE_COLOR_PROMPT) + + # tokenize a prompt + def _tokenize(self, prompt, bos=True): + _arr = (llama_cpp.llama_token * ((len(prompt) + 1) * 4))() + _n = llama_cpp.llama_tokenize( + self.model, + prompt.encode("utf8", errors="ignore"), + len(prompt), + _arr, + len(_arr), + bos, + False, + ) + return _arr[:_n] + + def set_color(self, c): + if self.params.use_color: + print(c, end="") + + def use_antiprompt(self): + return len(self.first_antiprompt) > 0 + + # generate tokens + def generate(self): + while ( + self.remaining_tokens > 0 + or self.params.interactive + or self.params.n_predict == -1 + ): + # predict + if len(self.embd) > 0: + # infinite text generation via context swapping + # if we run out of context: + # - take the n_keep first tokens from the original prompt (via n_past) + # - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch + if self.n_past + len(self.embd) > self.n_ctx: + n_left = self.n_past - self.params.n_keep + self.n_past = self.params.n_keep + + # insert n_left/2 tokens at the start of embd from last_n_tokens + _insert = self.last_n_tokens[ + self.n_ctx - int(n_left / 2) - len(self.embd) : -len(self.embd) + ] + self.embd = _insert + self.embd + self.params.path_session = "" + + # try to reuse a matching prefix from the loaded session instead of re-eval (via n_past) + if self.n_session_consumed < len(self.session_tokens): + for i in range(len(self.embd)): + if self.embd[i] != self.session_tokens[self.n_session_consumed]: + self.session_tokens = self.session_tokens[ + : self.n_session_consumed + ] + break + + self.n_past += 1 + self.n_session_consumed += 1 + + if self.n_session_consumed >= len(self.session_tokens): + i += 1 + break + + if i > 0: + self.embd = self.embd[i:] + + # evaluate tokens in batches + # embd is typically prepared beforehand to fit within a batch, but not always + # TODO BUG: The batching code causes nonsensical generation + """for i in range(0, len(self.embd), self.params.n_batch): n_eval = self.params.n_batch _arr = (llama_cpp.llama_token * n_eval)(*self.embd[i:i + n_eval]) if llama_cpp.llama_eval(self.ctx, _arr, n_eval, self.n_past, self.params.n_threads) != 0: @@ -307,278 +386,356 @@ def generate(self): self.n_past += n_eval""" - if (llama_cpp.llama_eval( - self.ctx, (llama_cpp.llama_token * len(self.embd))(*self.embd), len(self.embd), self.n_past - ) != 0): - raise Exception("Failed to llama_eval!") - - if len(self.embd) > 0 and len(self.params.path_session) > 0: - self.session_tokens.extend(self.embd) - self.n_session_consumed = len(self.session_tokens) - - self.n_past += len(self.embd) - self.embd = [] - if len(self.embd_inp) <= self.input_consumed: #&& !is_interacting - # out of user input, sample next token - top_k = llama_cpp.llama_n_vocab(self.ctx) if self.params.top_k <= 0 else self.params.top_k - repeat_last_n = self.n_ctx if self.params.repeat_last_n < 0 else self.params.repeat_last_n - - # optionally save the session on first sample (for faster prompt loading next time) - if len(self.params.path_session) > 0 and self.need_to_save_session: - self.need_to_save_session = False - llama_cpp.llama_save_session_file( - self.ctx, - self.params.path_session.encode("utf8"), - (llama_cpp.llama_token * len(self.session_tokens))(*self.session_tokens), - len(self.session_tokens) - ) - - id = 0 - - logits = llama_cpp.llama_get_logits(self.ctx) - n_vocab = llama_cpp.llama_n_vocab(self.model) - - # Apply params.logit_bias map - for key, value in self.params.logit_bias.items(): - logits[key] += value - - _arr = (llama_cpp.llama_token_data * n_vocab)(*[ - llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) - for token_id in range(n_vocab) - ]) - candidates_p = llama_cpp.ctypes.pointer(llama_cpp.llama_token_data_array(_arr, len(_arr), False)) - - # Apply penalties - nl_logit = logits[llama_cpp.llama_token_nl(self.ctx)] - last_n_repeat = min(len(self.last_n_tokens), repeat_last_n, self.n_ctx) - - _arr = (llama_cpp.llama_token * last_n_repeat)(*self.last_n_tokens[len(self.last_n_tokens) - last_n_repeat:]) - llama_cpp.llama_sample_repetition_penalties( - ctx=self.ctx, - candidates=candidates_p, - last_tokens_data = _arr, - penalty_last_n = last_n_repeat, - penalty_repeat = llama_cpp.c_float(self.params.repeat_penalty), - penalty_freq = llama_cpp.c_float(self.params.frequency_penalty), - penalty_present = llama_cpp.c_float(self.params.presence_penalty), - ) - - # NOT PRESENT IN CURRENT VERSION ? - # llama_cpp.llama_sample_frequency_and_presence_penalti(self.ctx, candidates_p, - # _arr, - # last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty)) - - if not self.params.penalize_nl: - logits[llama_cpp.llama_token_nl()] = nl_logit - - if self.params.temp <= 0: - # Greedy sampling - id = llama_cpp.llama_sample_token_greedy(self.ctx, candidates_p) - else: - if self.params.mirostat == 1: - mirostat_mu = 2.0 * self.params.mirostat_tau - mirostat_m = 100 - llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) - id = llama_cpp.llama_sample_token_mirostat(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_int(mirostat_m), llama_cpp.c_float(mirostat_mu)) - elif self.params.mirostat == 2: - mirostat_mu = 2.0 * self.params.mirostat_tau - llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) - id = llama_cpp.llama_sample_token_mirostat_v2(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_float(mirostat_mu)) - else: - # Temperature sampling - llama_cpp.llama_sample_top_k(self.ctx, candidates_p, top_k, min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_tail_free(self.ctx, candidates_p, llama_cpp.c_float(self.params.tfs_z), min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_typical(self.ctx, candidates_p, llama_cpp.c_float(self.params.typical_p), min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_top_p(self.ctx, candidates_p, llama_cpp.c_float(self.params.top_p), min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) - id = llama_cpp.llama_sample_token(self.ctx, candidates_p) - # print("`{}`".format(candidates_p.size)) - - self.last_n_tokens.pop(0) - self.last_n_tokens.append(id) - - # replace end of text token with newline token when in interactive mode - if (id == llama_cpp.llama_token_eos(self.ctx) and self.params.interactive and not self.params.instruct): - id = self.llama_token_newline[0] - self.embd.append(id) - if (self.use_antiprompt()): - # tokenize and inject first reverse prompt - self.embd_inp += self.first_antiprompt[0] - for id in self.first_antiprompt[0]: - self.embd.append(id) - else: - # add it to the context - self.embd.append(id) - - # echo this to console - self.output_echo = True - - # decrement remaining sampling budget - self.remaining_tokens -= 1 - else: - # output to console if input echo is on - self.output_echo = self.params.input_echo - - # some user input remains from prompt or interaction, forward it to processing - while len(self.embd_inp) > self.input_consumed: - self.embd.append(self.embd_inp[self.input_consumed]) - self.last_n_tokens.pop(0) - self.last_n_tokens.append(self.embd_inp[self.input_consumed]) - self.input_consumed += 1 - if len(self.embd) >= self.params.n_batch: - break - - # display tokens - if self.output_echo: - for id in self.embd: - if self.antiecho != None: - for r in self.antiecho(id): - yield r - else: - yield id - - # reset color to default if we there is no pending user input - if (self.params.input_echo and len(self.embd_inp) == self.input_consumed): - self.set_color(util.CONSOLE_COLOR_DEFAULT) - - if (self.params.interactive and len(self.embd_inp) <= self.input_consumed): - # if antiprompt is present, stop - if (self.use_antiprompt()): - if True in [ - i == self.last_n_tokens[-len(i):] - for i in self.first_antiprompt - ]: - break - - # if we are using instruction mode, and we have processed the initial prompt - if (self.params.interactive_start): - break - - # end of text token - if len(self.embd) > 0 and self.embd[-1] == llama_cpp.llama_token_eos(self.ctx): - if (not self.params.instruct): - for i in self.llama_token_eot: - yield i - break - - # respect n_predict even if antiprompt is present - if (self.params.interactive and self.remaining_tokens <= 0 and self.params.n_predict != -1): - # If we arent in instruction mode, fix the current generation by appending the antiprompt. - # Makes it so if chat ends prematurely you dont append the AI's text etc. - if not self.params.instruct: - self.embd_inp += self.first_antiprompt[0] - self.n_remain = self.params.n_predict - break - - self.params.interactive_start = False - - def __enter__(self): - return self - - def __exit__(self, type, value, tb): - self.exit() - - def exit(self): - llama_cpp.llama_free(self.ctx) - self.set_color(util.CONSOLE_COLOR_DEFAULT) - - def token_to_str(self, token_id: int) -> bytes: - size = 32 - buffer = (ctypes.c_char * size)() - n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token_id), buffer, size) - assert n <= size - return bytes(buffer[:n]) - - # return past text - def past(self): - for id in self.last_n_tokens[-self.n_past:]: - yield self.token_to_str(id).decode("utf8", errors="ignore") - - # write input - def input(self, prompt: str): - if (self.params.instruct and self.last_n_tokens[-len(self.inp_prefix):] != self.inp_prefix): - self.embd_inp += self.inp_prefix - self.embd_inp += self._tokenize(prompt) - if (self.params.instruct): - self.embd_inp += self.inp_suffix - - # write output - def output(self): - self.remaining_tokens = self.params.n_predict - for id in self.generate(): - cur_char = self.token_to_str(id) - - # Add remainder of missing bytes - if None in self.multibyte_fix: - self.multibyte_fix[self.multibyte_fix.index(None)] = cur_char - - # Return completed utf char - if len(self.multibyte_fix) > 0 and not None in self.multibyte_fix: - yield (b"".join(self.multibyte_fix)).decode("utf8") - self.multibyte_fix = [] - continue - - # Contains multi-byte UTF8 - for num, pattern in [(2, 192), (3, 224), (4, 240)]: - # Bitwise AND check - if pattern & int.from_bytes(cur_char, 'little') == pattern: - self.multibyte_fix = [cur_char] + ([None] * (num-1)) - - # Stop incomplete bytes from passing - if len(self.multibyte_fix) > 0: - continue - - yield cur_char.decode("utf8") - - # read user input - def read_input(self): - out = "" - while (t := input()).endswith("\\"): - out += t[:-1] + "\n" - return out + t + "\n" - - # interactive mode - def interact(self): - for i in self.output(): - print(i,end="",flush=True) - self.params.input_echo = False + if ( + llama_cpp.llama_eval( + self.ctx, + (llama_cpp.llama_token * len(self.embd))(*self.embd), + len(self.embd), + self.n_past, + ) + != 0 + ): + raise Exception("Failed to llama_eval!") + + if len(self.embd) > 0 and len(self.params.path_session) > 0: + self.session_tokens.extend(self.embd) + self.n_session_consumed = len(self.session_tokens) + + self.n_past += len(self.embd) + self.embd = [] + if len(self.embd_inp) <= self.input_consumed: # && !is_interacting + # out of user input, sample next token + top_k = ( + llama_cpp.llama_n_vocab(self.ctx) + if self.params.top_k <= 0 + else self.params.top_k + ) + repeat_last_n = ( + self.n_ctx + if self.params.repeat_last_n < 0 + else self.params.repeat_last_n + ) + + # optionally save the session on first sample (for faster prompt loading next time) + if len(self.params.path_session) > 0 and self.need_to_save_session: + self.need_to_save_session = False + llama_cpp.llama_save_session_file( + self.ctx, + self.params.path_session.encode("utf8"), + (llama_cpp.llama_token * len(self.session_tokens))( + *self.session_tokens + ), + len(self.session_tokens), + ) + + id = 0 + + logits = llama_cpp.llama_get_logits(self.ctx) + n_vocab = llama_cpp.llama_n_vocab(self.model) + + # Apply params.logit_bias map + for key, value in self.params.logit_bias.items(): + logits[key] += value + + _arr = (llama_cpp.llama_token_data * n_vocab)( + *[ + llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) + for token_id in range(n_vocab) + ] + ) + candidates_p = llama_cpp.ctypes.pointer( + llama_cpp.llama_token_data_array(_arr, len(_arr), False) + ) + + # Apply penalties + nl_logit = logits[llama_cpp.llama_token_nl(self.ctx)] + last_n_repeat = min(len(self.last_n_tokens), repeat_last_n, self.n_ctx) + + _arr = (llama_cpp.llama_token * last_n_repeat)( + *self.last_n_tokens[len(self.last_n_tokens) - last_n_repeat :] + ) + llama_cpp.llama_sample_repetition_penalties( + ctx=self.ctx, + candidates=candidates_p, + last_tokens_data=_arr, + penalty_last_n=last_n_repeat, + penalty_repeat=llama_cpp.c_float(self.params.repeat_penalty), + penalty_freq=llama_cpp.c_float(self.params.frequency_penalty), + penalty_present=llama_cpp.c_float(self.params.presence_penalty), + ) + + # NOT PRESENT IN CURRENT VERSION ? + # llama_cpp.llama_sample_frequency_and_presence_penalti(self.ctx, candidates_p, + # _arr, + # last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty)) + + if not self.params.penalize_nl: + logits[llama_cpp.llama_token_nl()] = nl_logit + + if self.params.temp <= 0: + # Greedy sampling + id = llama_cpp.llama_sample_token_greedy(self.ctx, candidates_p) + else: + if self.params.mirostat == 1: + mirostat_mu = 2.0 * self.params.mirostat_tau + mirostat_m = 100 + llama_cpp.llama_sample_temperature( + self.ctx, candidates_p, llama_cpp.c_float(self.params.temp) + ) + id = llama_cpp.llama_sample_token_mirostat( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.mirostat_tau), + llama_cpp.c_float(self.params.mirostat_eta), + llama_cpp.c_int(mirostat_m), + llama_cpp.c_float(mirostat_mu), + ) + elif self.params.mirostat == 2: + mirostat_mu = 2.0 * self.params.mirostat_tau + llama_cpp.llama_sample_temperature( + self.ctx, candidates_p, llama_cpp.c_float(self.params.temp) + ) + id = llama_cpp.llama_sample_token_mirostat_v2( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.mirostat_tau), + llama_cpp.c_float(self.params.mirostat_eta), + llama_cpp.c_float(mirostat_mu), + ) + else: + # Temperature sampling + llama_cpp.llama_sample_top_k( + self.ctx, + candidates_p, + top_k, + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_tail_free( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.tfs_z), + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_typical( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.typical_p), + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_top_p( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.top_p), + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_temperature( + self.ctx, candidates_p, llama_cpp.c_float(self.params.temp) + ) + id = llama_cpp.llama_sample_token(self.ctx, candidates_p) + # print("`{}`".format(candidates_p.size)) + + self.last_n_tokens.pop(0) + self.last_n_tokens.append(id) + + # replace end of text token with newline token when in interactive mode + if ( + id == llama_cpp.llama_token_eos(self.ctx) + and self.params.interactive + and not self.params.instruct + ): + id = self.llama_token_newline[0] + self.embd.append(id) + if self.use_antiprompt(): + # tokenize and inject first reverse prompt + self.embd_inp += self.first_antiprompt[0] + for id in self.first_antiprompt[0]: + self.embd.append(id) + else: + # add it to the context + self.embd.append(id) + + # echo this to console + self.output_echo = True + + # decrement remaining sampling budget + self.remaining_tokens -= 1 + else: + # output to console if input echo is on + self.output_echo = self.params.input_echo + + # some user input remains from prompt or interaction, forward it to processing + while len(self.embd_inp) > self.input_consumed: + self.embd.append(self.embd_inp[self.input_consumed]) + self.last_n_tokens.pop(0) + self.last_n_tokens.append(self.embd_inp[self.input_consumed]) + self.input_consumed += 1 + if len(self.embd) >= self.params.n_batch: + break + + # display tokens + if self.output_echo: + for id in self.embd: + if self.antiecho != None: + for r in self.antiecho(id): + yield r + else: + yield id + + # reset color to default if we there is no pending user input + if self.params.input_echo and len(self.embd_inp) == self.input_consumed: + self.set_color(util.CONSOLE_COLOR_DEFAULT) + + if self.params.interactive and len(self.embd_inp) <= self.input_consumed: + # if antiprompt is present, stop + if self.use_antiprompt(): + if True in [ + i == self.last_n_tokens[-len(i) :] + for i in self.first_antiprompt + ]: + break + + # if we are using instruction mode, and we have processed the initial prompt + if self.params.interactive_start: + break + + # end of text token + if len(self.embd) > 0 and self.embd[-1] == llama_cpp.llama_token_eos( + self.ctx + ): + if not self.params.instruct: + for i in self.llama_token_eot: + yield i + break + + # respect n_predict even if antiprompt is present + if ( + self.params.interactive + and self.remaining_tokens <= 0 + and self.params.n_predict != -1 + ): + # If we arent in instruction mode, fix the current generation by appending the antiprompt. + # Makes it so if chat ends prematurely you dont append the AI's text etc. + if not self.params.instruct: + self.embd_inp += self.first_antiprompt[0] + self.n_remain = self.params.n_predict + break + + self.params.interactive_start = False + + def __enter__(self): + return self + + def __exit__(self, type, value, tb): + self.exit() + + def exit(self): + llama_cpp.llama_free(self.ctx) + self.set_color(util.CONSOLE_COLOR_DEFAULT) + + def token_to_str(self, token_id: int) -> bytes: + size = 32 + buffer = (ctypes.c_char * size)() + n = llama_cpp.llama_token_to_piece( + self.model, llama_cpp.llama_token(token_id), buffer, size + ) + assert n <= size + return bytes(buffer[:n]) + + # return past text + def past(self): + for id in self.last_n_tokens[-self.n_past :]: + yield self.token_to_str(id).decode("utf8", errors="ignore") + + # write input + def input(self, prompt: str): + if ( + self.params.instruct + and self.last_n_tokens[-len(self.inp_prefix) :] != self.inp_prefix + ): + self.embd_inp += self.inp_prefix + self.embd_inp += self._tokenize(prompt) + if self.params.instruct: + self.embd_inp += self.inp_suffix + + # write output + def output(self): + self.remaining_tokens = self.params.n_predict + for id in self.generate(): + cur_char = self.token_to_str(id) + + # Add remainder of missing bytes + if None in self.multibyte_fix: + self.multibyte_fix[self.multibyte_fix.index(None)] = cur_char + + # Return completed utf char + if len(self.multibyte_fix) > 0 and not None in self.multibyte_fix: + yield (b"".join(self.multibyte_fix)).decode("utf8") + self.multibyte_fix = [] + continue + + # Contains multi-byte UTF8 + for num, pattern in [(2, 192), (3, 224), (4, 240)]: + # Bitwise AND check + if pattern & int.from_bytes(cur_char, "little") == pattern: + self.multibyte_fix = [cur_char] + ([None] * (num - 1)) + + # Stop incomplete bytes from passing + if len(self.multibyte_fix) > 0: + continue + + yield cur_char.decode("utf8") + + # read user input + def read_input(self): + out = "" + while (t := input()).endswith("\\"): + out += t[:-1] + "\n" + return out + t + "\n" + + # interactive mode + def interact(self): + for i in self.output(): + print(i, end="", flush=True) + self.params.input_echo = False # Using string instead of tokens to check for antiprompt, - # It is more reliable than tokens for interactive mode. - generated_str = "" - while self.params.interactive: - self.set_color(util.CONSOLE_COLOR_USER_INPUT) - if (self.params.instruct): - print('\n> ', end="") - self.input(self.read_input()) - else: - print(self.params.input_prefix, end="") - self.input(f"{self.params.input_prefix}{self.read_input()}{self.params.input_suffix}") - print(self.params.input_suffix,end="") - self.set_color(util.CONSOLE_COLOR_DEFAULT) - - try: - for i in self.output(): - print(i,end="",flush=True) - generated_str += i - for ap in self.params.antiprompt: - if generated_str.endswith(ap): - raise KeyboardInterrupt - except KeyboardInterrupt: - self.set_color(util.CONSOLE_COLOR_DEFAULT) - if not self.params.instruct: - print(self.params.fix_prefix,end="") - self.input(self.params.fix_prefix) + # It is more reliable than tokens for interactive mode. + generated_str = "" + while self.params.interactive: + self.set_color(util.CONSOLE_COLOR_USER_INPUT) + if self.params.instruct: + print("\n> ", end="") + self.input(self.read_input()) + else: + print(self.params.input_prefix, end="") + self.input( + f"{self.params.input_prefix}{self.read_input()}{self.params.input_suffix}" + ) + print(self.params.input_suffix, end="") + self.set_color(util.CONSOLE_COLOR_DEFAULT) + + try: + for i in self.output(): + print(i, end="", flush=True) + generated_str += i + for ap in self.params.antiprompt: + if generated_str.endswith(ap): + raise KeyboardInterrupt + except KeyboardInterrupt: + self.set_color(util.CONSOLE_COLOR_DEFAULT) + if not self.params.instruct: + print(self.params.fix_prefix, end="") + self.input(self.params.fix_prefix) + if __name__ == "__main__": - from datetime import datetime + from datetime import datetime - USER_NAME="User" - AI_NAME="ChatLLaMa" + USER_NAME = "User" + AI_NAME = "ChatLLaMa" - time_now = datetime.now() - prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. + time_now = datetime.now() + prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. {AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer {USER_NAME}’s requests immediately and with details and precision. Transcript below contains only the recorded dialog between two, without any annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. @@ -595,10 +752,10 @@ def interact(self): {USER_NAME}: Name a color. {AI_NAME}: Blue {USER_NAME}: """ - - params = gpt_params_parse() - if params.prompt is None and params.file is None: - params.prompt = prompt - with LLaMAInteract(params) as m: - m.interact() + params = gpt_params_parse() + if params.prompt is None and params.file is None: + params.prompt = prompt + + with LLaMAInteract(params) as m: + m.interact() diff --git a/examples/low_level_api/low_level_api_llama_cpp.py b/examples/low_level_api/low_level_api_llama_cpp.py index ef1b2c016..ba3545771 100644 --- a/examples/low_level_api/low_level_api_llama_cpp.py +++ b/examples/low_level_api/low_level_api_llama_cpp.py @@ -7,23 +7,20 @@ llama_cpp.llama_backend_init(numa=False) N_THREADS = multiprocessing.cpu_count() -MODEL_PATH = os.environ.get('MODEL', "../models/7B/ggml-model.bin") +MODEL_PATH = os.environ.get("MODEL", "../models/7B/ggml-model.bin") prompt = b"\n\n### Instruction:\nWhat is the capital of France?\n\n### Response:\n" lparams = llama_cpp.llama_model_default_params() cparams = llama_cpp.llama_context_default_params() -model = llama_cpp.llama_load_model_from_file(MODEL_PATH.encode('utf-8'), lparams) +model = llama_cpp.llama_load_model_from_file(MODEL_PATH.encode("utf-8"), lparams) ctx = llama_cpp.llama_new_context_with_model(model, cparams) # determine the required inference memory per token: tmp = [0, 1, 2, 3] llama_cpp.llama_eval( - ctx = ctx, - tokens=(llama_cpp.c_int * len(tmp))(*tmp), - n_tokens=len(tmp), - n_past=0 - )# Deprecated + ctx=ctx, tokens=(llama_cpp.c_int * len(tmp))(*tmp), n_tokens=len(tmp), n_past=0 +) # Deprecated n_past = 0 @@ -32,12 +29,12 @@ embd_inp = (llama_cpp.llama_token * (len(prompt) + 1))() n_of_tok = llama_cpp.llama_tokenize( model=model, - text=bytes(str(prompt),'utf-8'), - text_len=len(embd_inp), + text=bytes(str(prompt), "utf-8"), + text_len=len(embd_inp), tokens=embd_inp, n_max_tokens=len(embd_inp), add_bos=False, - special=False + special=False, ) embd_inp = embd_inp[:n_of_tok] @@ -63,11 +60,11 @@ while remaining_tokens > 0: if len(embd) > 0: llama_cpp.llama_eval( - ctx = ctx, + ctx=ctx, tokens=(llama_cpp.c_int * len(embd))(*embd), n_tokens=len(embd), - n_past=n_past - )# Deprecated + n_past=n_past, + ) # Deprecated n_past += len(embd) embd = [] @@ -75,20 +72,26 @@ logits = llama_cpp.llama_get_logits(ctx) n_vocab = llama_cpp.llama_n_vocab(model) - _arr = (llama_cpp.llama_token_data * n_vocab)(*[ - llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) - for token_id in range(n_vocab) - ]) + _arr = (llama_cpp.llama_token_data * n_vocab)( + *[ + llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) + for token_id in range(n_vocab) + ] + ) candidates_p = llama_cpp.ctypes.pointer( - llama_cpp.llama_token_data_array(_arr, len(_arr), False)) + llama_cpp.llama_token_data_array(_arr, len(_arr), False) + ) _arr = (llama_cpp.c_int * len(last_n_tokens_data))(*last_n_tokens_data) - llama_cpp.llama_sample_repetition_penalties(ctx, candidates_p, + llama_cpp.llama_sample_repetition_penalties( + ctx, + candidates_p, _arr, penalty_last_n=last_n_repeat, penalty_repeat=repeat_penalty, penalty_freq=frequency_penalty, - penalty_present=presence_penalty) + penalty_present=presence_penalty, + ) llama_cpp.llama_sample_top_k(ctx, candidates_p, k=40, min_keep=1) llama_cpp.llama_sample_top_p(ctx, candidates_p, p=0.8, min_keep=1) @@ -111,10 +114,11 @@ size = 32 buffer = (ctypes.c_char * size)() n = llama_cpp.llama_token_to_piece( - model, llama_cpp.llama_token(id), buffer, size) + model, llama_cpp.llama_token(id), buffer, size + ) assert n <= size print( - buffer[:n].decode('utf-8'), + buffer[:n].decode("utf-8"), end="", flush=True, ) diff --git a/examples/low_level_api/quantize.py b/examples/low_level_api/quantize.py index cdb038a71..057ac389e 100644 --- a/examples/low_level_api/quantize.py +++ b/examples/low_level_api/quantize.py @@ -22,7 +22,10 @@ def main(args): parser = argparse.ArgumentParser() parser.add_argument("fname_inp", type=str, help="Path to input model") parser.add_argument("fname_out", type=str, help="Path to output model") - parser.add_argument("type", type=int, help="Type of quantization (2: q4_0, 3: q4_1), see llama_cpp.py for enum") + parser.add_argument( + "type", + type=int, + help="Type of quantization (2: q4_0, 3: q4_1), see llama_cpp.py for enum", + ) args = parser.parse_args() main(args) - diff --git a/examples/low_level_api/util.py b/examples/low_level_api/util.py index 9d0ec2f70..ef8b1c1ee 100644 --- a/examples/low_level_api/util.py +++ b/examples/low_level_api/util.py @@ -1,4 +1,3 @@ - ANSI_COLOR_RESET = "\x1b[0m" ANSI_COLOR_YELLOW = "\x1b[33m" ANSI_BOLD = "\x1b[1m" @@ -8,88 +7,95 @@ CONSOLE_COLOR_PROMPT = ANSI_COLOR_YELLOW CONSOLE_COLOR_USER_INPUT = ANSI_BOLD + ANSI_COLOR_GREEN + # Iterative search # Actively searches and prevents a pattern from being returned class IterSearch: - def __init__(self, pattern): - self.pattern = list(pattern) - self.buffer = [] - - def __call__(self, char): - self.buffer += [char] + def __init__(self, pattern): + self.pattern = list(pattern) + self.buffer = [] - if (self.pattern[:len(self.buffer)] == self.buffer): - if (len(self.buffer) >= len(self.pattern)): - self.buffer.clear() - return [] + def __call__(self, char): + self.buffer += [char] - _tmp = self.buffer[:] - self.buffer.clear() - return _tmp + if self.pattern[: len(self.buffer)] == self.buffer: + if len(self.buffer) >= len(self.pattern): + self.buffer.clear() + return [] -class Circle: - def __init__(self, size, default=0): - self.list = [default] * size - self.maxsize = size - self.size = 0 - self.offset = 0 - - def append(self, elem): - if self.size < self.maxsize: - self.list[self.size] = elem - self.size += 1 - else: - self.list[self.offset] = elem - self.offset = (self.offset + 1) % self.maxsize - - def __getitem__(self, val): - if isinstance(val, int): - if 0 > val or val >= self.size: - raise IndexError('Index out of range') - return self.list[val] if self.size < self.maxsize else self.list[(self.offset + val) % self.maxsize] - elif isinstance(val, slice): - start, stop, step = val.start, val.stop, val.step - if step is None: - step = 1 - if start is None: - start = 0 - if stop is None: - stop = self.size - if start < 0: - start = self.size + start - if stop < 0: - stop = self.size + stop - - indices = range(start, stop, step) - return [self.list[(self.offset + i) % self.maxsize] for i in indices if i < self.size] - else: - raise TypeError('Invalid argument type') + _tmp = self.buffer[:] + self.buffer.clear() + return _tmp +class Circle: + def __init__(self, size, default=0): + self.list = [default] * size + self.maxsize = size + self.size = 0 + self.offset = 0 + + def append(self, elem): + if self.size < self.maxsize: + self.list[self.size] = elem + self.size += 1 + else: + self.list[self.offset] = elem + self.offset = (self.offset + 1) % self.maxsize + + def __getitem__(self, val): + if isinstance(val, int): + if 0 > val or val >= self.size: + raise IndexError("Index out of range") + return ( + self.list[val] + if self.size < self.maxsize + else self.list[(self.offset + val) % self.maxsize] + ) + elif isinstance(val, slice): + start, stop, step = val.start, val.stop, val.step + if step is None: + step = 1 + if start is None: + start = 0 + if stop is None: + stop = self.size + if start < 0: + start = self.size + start + if stop < 0: + stop = self.size + stop + + indices = range(start, stop, step) + return [ + self.list[(self.offset + i) % self.maxsize] + for i in indices + if i < self.size + ] + else: + raise TypeError("Invalid argument type") if __name__ == "__main__": - c = Circle(5) - - c.append(1) - print(c.list) - print(c[:]) - assert c[0] == 1 - assert c[:5] == [1] - - for i in range(2,5+1): - c.append(i) - print(c.list) - print(c[:]) - assert c[0] == 1 - assert c[:5] == [1,2,3,4,5] - - for i in range(5+1,9+1): - c.append(i) - print(c.list) - print(c[:]) - assert c[0] == 5 - assert c[:5] == [5,6,7,8,9] - #assert c[:-5] == [5,6,7,8,9] - assert c[:10] == [5,6,7,8,9] - + c = Circle(5) + + c.append(1) + print(c.list) + print(c[:]) + assert c[0] == 1 + assert c[:5] == [1] + + for i in range(2, 5 + 1): + c.append(i) + print(c.list) + print(c[:]) + assert c[0] == 1 + assert c[:5] == [1, 2, 3, 4, 5] + + for i in range(5 + 1, 9 + 1): + c.append(i) + print(c.list) + print(c[:]) + assert c[0] == 5 + assert c[:5] == [5, 6, 7, 8, 9] + # assert c[:-5] == [5,6,7,8,9] + assert c[:10] == [5, 6, 7, 8, 9] diff --git a/examples/notebooks/Batching.ipynb b/examples/notebooks/Batching.ipynb index 687316b32..73b28c744 100644 --- a/examples/notebooks/Batching.ipynb +++ b/examples/notebooks/Batching.ipynb @@ -392,7 +392,9 @@ "source": [ "params = llama_cpp.llama_model_default_params()\n", "params.n_gpu_layers = 35\n", - "model = llama_cpp.llama_load_model_from_file(b\"../../models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf\", params=params) # Update this to whatever" + "model = llama_cpp.llama_load_model_from_file(\n", + " b\"../../models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf\", params=params\n", + ") # Update this to whatever" ] }, { @@ -416,7 +418,9 @@ "prompt = b\"The quick brown fox\"\n", "\n", "tokens = (llama_cpp.llama_token * n_ctx)()\n", - "tokens_len = llama_cpp.llama_tokenize(model, prompt, len(prompt), tokens, len(tokens), True, True)\n", + "tokens_len = llama_cpp.llama_tokenize(\n", + " model, prompt, len(prompt), tokens, len(tokens), True, True\n", + ")\n", "print(tokens[:tokens_len])\n", "\n", "n_kv_req = tokens_len + (n_len - tokens_len) * n_parallel\n", @@ -447,7 +451,6 @@ } ], "source": [ - "\n", "ctx_params = llama_cpp.llama_context_default_params()\n", "ctx_params.seed = 1234\n", "ctx_params.n_ctx = n_kv_req\n", @@ -577,7 +580,7 @@ " for i in range(n_parallel):\n", " if i_batch[i] < 0:\n", " continue\n", - " \n", + "\n", " n_vocab = llama_cpp.llama_n_vocab(model)\n", " logits = llama_cpp.llama_get_logits_ith(ctx, i_batch[i])\n", "\n", @@ -588,7 +591,9 @@ " candidates[token_id].logit = logits[token_id]\n", " candidates[token_id].p = 0.0\n", "\n", - " candidates_p = llama_cpp.llama_token_data_array(candidates, len(candidates), False)\n", + " candidates_p = llama_cpp.llama_token_data_array(\n", + " candidates, len(candidates), False\n", + " )\n", "\n", " top_k = 40\n", " top_p = 0.9\n", @@ -596,8 +601,8 @@ "\n", " llama_cpp.llama_sample_top_k(ctx, ctypes.byref(candidates_p), top_k, 1)\n", " llama_cpp.llama_sample_top_p(ctx, ctypes.byref(candidates_p), top_p, 1)\n", - " llama_cpp.llama_sample_temp (ctx, ctypes.byref(candidates_p), temp)\n", - " \n", + " llama_cpp.llama_sample_temp(ctx, ctypes.byref(candidates_p), temp)\n", + "\n", " new_token_id = llama_cpp.llama_sample_token(ctx, ctypes.byref(candidates_p))\n", "\n", " if new_token_id == llama_cpp.llama_token_eos(ctx) or n_cur == n_len:\n", @@ -617,7 +622,7 @@ " i_batch[i] = batch.n_tokens\n", " batch.n_tokens += 1\n", " n_decode += 1\n", - " \n", + "\n", " if batch.n_tokens == 0:\n", " break\n", "\n", @@ -627,7 +632,7 @@ " print(\"Error decoding\", flush=True)\n", " break\n", " print(n_cur)\n", - " print(streams)\n" + " print(streams)" ] }, { diff --git a/examples/notebooks/Clients.ipynb b/examples/notebooks/Clients.ipynb index caebbb67f..fab82673e 100644 --- a/examples/notebooks/Clients.ipynb +++ b/examples/notebooks/Clients.ipynb @@ -37,11 +37,11 @@ "source": [ "import openai\n", "\n", - "openai.api_key = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + "openai.api_key = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", "openai.api_base = \"http://100.64.159.73:8000/v1\"\n", "\n", "openai.Completion.create(\n", - " model=\"text-davinci-003\", # currently can be anything\n", + " model=\"text-davinci-003\", # currently can be anything\n", " prompt=\"The quick brown fox jumps\",\n", " max_tokens=5,\n", ")" @@ -66,7 +66,9 @@ "source": [ "import os\n", "\n", - "os.environ[\"OPENAI_API_KEY\"] = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + "os.environ[\"OPENAI_API_KEY\"] = (\n", + " \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + ")\n", "os.environ[\"OPENAI_API_BASE\"] = \"http://100.64.159.73:8000/v1\"\n", "\n", "from langchain.llms import OpenAI\n", diff --git a/examples/notebooks/Functions.ipynb b/examples/notebooks/Functions.ipynb index f1e5e9a1d..1f4138165 100644 --- a/examples/notebooks/Functions.ipynb +++ b/examples/notebooks/Functions.ipynb @@ -45,10 +45,11 @@ "\n", "\n", "client = openai.OpenAI(\n", - " api_key = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\", # can be anything\n", - " base_url = \"http://100.64.159.73:8000/v1\" # NOTE: Replace with IP address and port of your llama-cpp-python server\n", + " api_key=\"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\", # can be anything\n", + " base_url=\"http://100.64.159.73:8000/v1\", # NOTE: Replace with IP address and port of your llama-cpp-python server\n", ")\n", "\n", + "\n", "# Example dummy function hard coded to return the same weather\n", "# In production, this could be your backend API or an external API\n", "def get_current_weather(location, unit=\"fahrenheit\"):\n", @@ -56,15 +57,23 @@ " if \"tokyo\" in location.lower():\n", " return json.dumps({\"location\": \"Tokyo\", \"temperature\": \"10\", \"unit\": \"celsius\"})\n", " elif \"san francisco\" in location.lower():\n", - " return json.dumps({\"location\": \"San Francisco\", \"temperature\": \"72\", \"unit\": \"fahrenheit\"})\n", + " return json.dumps(\n", + " {\"location\": \"San Francisco\", \"temperature\": \"72\", \"unit\": \"fahrenheit\"}\n", + " )\n", " elif \"paris\" in location.lower():\n", " return json.dumps({\"location\": \"Paris\", \"temperature\": \"22\", \"unit\": \"celsius\"})\n", " else:\n", " return json.dumps({\"location\": location, \"temperature\": \"unknown\"})\n", "\n", + "\n", "def run_conversation():\n", " # Step 1: send the conversation and available functions to the model\n", - " messages = [{\"role\": \"user\", \"content\": \"What's the weather like in San Francisco, Tokyo, and Paris?\"}]\n", + " messages = [\n", + " {\n", + " \"role\": \"user\",\n", + " \"content\": \"What's the weather like in San Francisco, Tokyo, and Paris?\",\n", + " }\n", + " ]\n", " tools = [\n", " {\n", " \"type\": \"function\",\n", @@ -123,6 +132,8 @@ " messages=messages,\n", " ) # get a new response from the model where it can see the function response\n", " return second_response\n", + "\n", + "\n", "print(run_conversation())" ] }, @@ -169,16 +180,18 @@ "# Enables `response_model`\n", "client = instructor.patch(client=client)\n", "\n", + "\n", "class UserDetail(BaseModel):\n", " name: str\n", " age: int\n", "\n", + "\n", "user = client.chat.completions.create(\n", " model=\"gpt-3.5-turbo\",\n", " response_model=UserDetail,\n", " messages=[\n", " {\"role\": \"user\", \"content\": \"Extract Jason is 25 years old\"},\n", - " ]\n", + " ],\n", ")\n", "\n", "assert isinstance(user, UserDetail)\n", @@ -213,17 +226,22 @@ "source": [ "import enum\n", "\n", + "\n", "class Labels(str, enum.Enum):\n", " \"\"\"Enumeration for single-label text classification.\"\"\"\n", + "\n", " SPAM = \"spam\"\n", " NOT_SPAM = \"not_spam\"\n", "\n", + "\n", "class SinglePrediction(BaseModel):\n", " \"\"\"\n", " Class for a single class label prediction.\n", " \"\"\"\n", + "\n", " class_label: Labels\n", "\n", + "\n", "def classify(data: str) -> SinglePrediction:\n", " \"\"\"Perform single-label classification on the input text.\"\"\"\n", " return client.chat.completions.create(\n", @@ -237,6 +255,7 @@ " ],\n", " ) # type: ignore\n", "\n", + "\n", "prediction = classify(\"Hello there I'm a Nigerian prince and I want to give you money\")\n", "assert prediction.class_label == Labels.SPAM\n", "print(prediction)" @@ -265,19 +284,23 @@ "source": [ "from typing import List\n", "\n", + "\n", "# Define Enum class for multiple labels\n", "class MultiLabels(str, enum.Enum):\n", " TECH_ISSUE = \"tech_issue\"\n", " BILLING = \"billing\"\n", " GENERAL_QUERY = \"general_query\"\n", "\n", + "\n", "# Define the multi-class prediction model\n", "class MultiClassPrediction(BaseModel):\n", " \"\"\"\n", " Class for a multi-class label prediction.\n", " \"\"\"\n", + "\n", " class_labels: List[MultiLabels]\n", "\n", + "\n", "def multi_classify(data: str) -> MultiClassPrediction:\n", " \"\"\"Perform multi-label classification on the input text.\"\"\"\n", " return client.chat.completions.create(\n", @@ -291,6 +314,7 @@ " ],\n", " ) # type: ignore\n", "\n", + "\n", "# Test multi-label classification\n", "ticket = \"My account is locked and I can't access my billing info.\"\n", "prediction = multi_classify(ticket)\n", @@ -331,10 +355,12 @@ "question = \"What is the meaning of life?\"\n", "context = \"The according to the devil the meaning of live is to live a life of sin and debauchery.\"\n", "\n", + "\n", "class QuestionAnswer(BaseModel):\n", " question: str\n", " answer: str\n", "\n", + "\n", "qa: QuestionAnswer = client.chat.completions.create(\n", " model=\"gpt-3.5-turbo\",\n", " response_model=QuestionAnswer,\n", @@ -351,6 +377,7 @@ ")\n", "print(qa)\n", "\n", + "\n", "class QuestionAnswerNoEvil(BaseModel):\n", " question: str\n", " answer: Annotated[\n", @@ -360,6 +387,7 @@ " ),\n", " ]\n", "\n", + "\n", "try:\n", " qa: QuestionAnswerNoEvil = client.chat.completions.create(\n", " model=\"gpt-3.5-turbo\",\n", @@ -405,6 +433,7 @@ "\n", "from pydantic import Field, BaseModel, model_validator, FieldValidationInfo\n", "\n", + "\n", "class Fact(BaseModel):\n", " fact: str = Field(...)\n", " substring_quote: List[str] = Field(...)\n", @@ -424,6 +453,7 @@ " for match in re.finditer(re.escape(quote), context):\n", " yield match.span()\n", "\n", + "\n", "class QuestionAnswer(BaseModel):\n", " question: str = Field(...)\n", " answer: List[Fact] = Field(...)\n", @@ -440,13 +470,17 @@ " temperature=0.0,\n", " response_model=QuestionAnswer,\n", " messages=[\n", - " {\"role\": \"system\", \"content\": \"You are a world class algorithm to answer questions with correct and exact citations.\"},\n", + " {\n", + " \"role\": \"system\",\n", + " \"content\": \"You are a world class algorithm to answer questions with correct and exact citations.\",\n", + " },\n", " {\"role\": \"user\", \"content\": f\"{context}\"},\n", - " {\"role\": \"user\", \"content\": f\"Question: {question}\"}\n", + " {\"role\": \"user\", \"content\": f\"Question: {question}\"},\n", " ],\n", " validation_context={\"text_chunk\": context},\n", " )\n", "\n", + "\n", "question = \"What did the author do during college?\"\n", "context = \"\"\"\n", "My name is Jason Liu, and I grew up in Toronto Canada but I was born in China.\n", diff --git a/examples/notebooks/Guidance.ipynb b/examples/notebooks/Guidance.ipynb index 045856ea2..c52559853 100644 --- a/examples/notebooks/Guidance.ipynb +++ b/examples/notebooks/Guidance.ipynb @@ -28,7 +28,9 @@ "source": [ "import os\n", "\n", - "os.environ[\"OPENAI_API_KEY\"] = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + "os.environ[\"OPENAI_API_KEY\"] = (\n", + " \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + ")\n", "os.environ[\"OPENAI_API_BASE\"] = \"http://100.64.159.73:8000/v1\"\n", "os.environ[\"OPENAI_API_HOST\"] = \"http://100.64.159.73:8000\"\n", "\n", @@ -38,21 +40,23 @@ "guidance.llm = guidance.llms.OpenAI(\"text-davinci-003\", caching=False)\n", "\n", "# define a guidance program that adapts a proverb\n", - "program = guidance(\"\"\"Tweak this proverb to apply to model instructions instead.\n", + "program = guidance(\n", + " \"\"\"Tweak this proverb to apply to model instructions instead.\n", "\n", "{{proverb}}\n", "- {{book}} {{chapter}}:{{verse}}\n", "\n", "UPDATED\n", "Where there is no guidance{{gen 'rewrite' stop=\"\\\\n-\"}}\n", - "- GPT {{gen 'chapter'}}:{{gen 'verse'}}\"\"\")\n", + "- GPT {{gen 'chapter'}}:{{gen 'verse'}}\"\"\"\n", + ")\n", "\n", "# execute the program on a specific proverb\n", "executed_program = program(\n", " proverb=\"Where there is no guidance, a people falls,\\nbut in an abundance of counselors there is safety.\",\n", " book=\"Proverbs\",\n", " chapter=11,\n", - " verse=14\n", + " verse=14,\n", ")" ] }, diff --git a/examples/notebooks/Multimodal.ipynb b/examples/notebooks/Multimodal.ipynb index def68df00..8448ac1f7 100644 --- a/examples/notebooks/Multimodal.ipynb +++ b/examples/notebooks/Multimodal.ipynb @@ -38,23 +38,20 @@ " \"url\": \"https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png\",\n", " },\n", " },\n", - " {\"type\": \"text\", \"text\": \"What does the image say. Format your response as a json object with a single 'text' key.\"},\n", + " {\n", + " \"type\": \"text\",\n", + " \"text\": \"What does the image say. Format your response as a json object with a single 'text' key.\",\n", + " },\n", " ],\n", " }\n", " ],\n", - " response_format={ \n", + " response_format={\n", " \"type\": \"json_object\",\n", - " \"schema\": {\n", - " \"type\": \"object\",\n", - " \"properties\": {\n", - " \"text\": {\n", - " \"type\": \"string\"\n", - " }\n", - " }\n", - " }\n", - " }\n", + " \"schema\": {\"type\": \"object\", \"properties\": {\"text\": {\"type\": \"string\"}}},\n", + " },\n", ")\n", "import json\n", + "\n", "print(json.loads(response.choices[0].message.content))" ] }, diff --git a/examples/notebooks/OpenHermesFunctionCalling.ipynb b/examples/notebooks/OpenHermesFunctionCalling.ipynb index c0de3fdc2..13128be04 100644 --- a/examples/notebooks/OpenHermesFunctionCalling.ipynb +++ b/examples/notebooks/OpenHermesFunctionCalling.ipynb @@ -42,42 +42,58 @@ "import inspect\n", "from typing import get_type_hints\n", "\n", + "\n", "class Article:\n", " pass\n", "\n", + "\n", "class Weather:\n", " pass\n", "\n", + "\n", "class Directions:\n", " pass\n", "\n", - "def calculate_mortgage_payment(loan_amount: int, interest_rate: float, loan_term: int) -> float:\n", + "\n", + "def calculate_mortgage_payment(\n", + " loan_amount: int, interest_rate: float, loan_term: int\n", + ") -> float:\n", " \"\"\"Get the monthly mortgage payment given an interest rate percentage.\"\"\"\n", - " \n", + "\n", " # TODO: you must implement this to actually call it later\n", " pass\n", "\n", - "def get_article_details(title: str, authors: list[str], short_summary: str, date_published: str, tags: list[str]) -> Article:\n", + "\n", + "def get_article_details(\n", + " title: str,\n", + " authors: list[str],\n", + " short_summary: str,\n", + " date_published: str,\n", + " tags: list[str],\n", + ") -> Article:\n", " '''Get article details from unstructured article text.\n", - "date_published: formatted as \"MM/DD/YYYY\"'''\n", - " \n", + " date_published: formatted as \"MM/DD/YYYY\"'''\n", + "\n", " # TODO: you must implement this to actually call it later\n", " pass\n", "\n", + "\n", "def get_weather(zip_code: str) -> Weather:\n", " \"\"\"Get the current weather given a zip code.\"\"\"\n", - " \n", + "\n", " # TODO: you must implement this to actually call it later\n", " pass\n", "\n", + "\n", "def get_directions(start: str, destination: str) -> Directions:\n", " \"\"\"Get directions from Google Directions API.\n", - "start: start address as a string including zipcode (if any)\n", - "destination: end address as a string including zipcode (if any)\"\"\"\n", - " \n", + " start: start address as a string including zipcode (if any)\n", + " destination: end address as a string including zipcode (if any)\"\"\"\n", + "\n", " # TODO: you must implement this to actually call it later\n", " pass\n", "\n", + "\n", "def get_type_name(t):\n", " name = str(t)\n", " if \"list\" in name or \"dict\" in name:\n", @@ -85,6 +101,7 @@ " else:\n", " return t.__name__\n", "\n", + "\n", "def serialize_function_to_json(func):\n", " signature = inspect.signature(func)\n", " type_hints = get_type_hints(func)\n", @@ -92,11 +109,8 @@ " function_info = {\n", " \"name\": func.__name__,\n", " \"description\": func.__doc__,\n", - " \"parameters\": {\n", - " \"type\": \"object\",\n", - " \"properties\": {}\n", - " },\n", - " \"returns\": type_hints.get('return', 'void').__name__\n", + " \"parameters\": {\"type\": \"object\", \"properties\": {}},\n", + " \"returns\": type_hints.get(\"return\", \"void\").__name__,\n", " }\n", "\n", " for name, _ in signature.parameters.items():\n", @@ -105,6 +119,7 @@ "\n", " return json.dumps(function_info, indent=2)\n", "\n", + "\n", "print(serialize_function_to_json(get_article_details))" ] }, @@ -117,13 +132,14 @@ "import xml.etree.ElementTree as ET\n", "import re\n", "\n", + "\n", "def extract_function_calls(completion):\n", " completion = completion.strip()\n", " pattern = r\"((.*?))\"\n", " match = re.search(pattern, completion, re.DOTALL)\n", " if not match:\n", " return None\n", - " \n", + "\n", " multiplefn = match.group(1)\n", " root = ET.fromstring(multiplefn)\n", " functions = root.findall(\"functioncall\")\n", @@ -399,7 +415,7 @@ "prompts = [\n", " \"What's the weather in 10001?\",\n", " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", - " \"What's the current exchange rate for USD to EUR?\"\n", + " \"What's the current exchange rate for USD to EUR?\",\n", "]\n", "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", "\n", @@ -793,7 +809,12 @@ "source": [ "import llama_cpp\n", "\n", - "llama = llama_cpp.Llama(model_path=\"../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf\", n_gpu_layers=-1, n_ctx=2048, verbose=False)" + "llama = llama_cpp.Llama(\n", + " model_path=\"../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf\",\n", + " n_gpu_layers=-1,\n", + " n_ctx=2048,\n", + " verbose=False,\n", + ")" ] }, { @@ -818,7 +839,7 @@ "prompts = [\n", " \"What's the weather in 10001?\",\n", " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", - " \"What's the current exchange rate for USD to EUR?\"\n", + " \"What's the current exchange rate for USD to EUR?\",\n", "]\n", "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", "\n", @@ -830,7 +851,7 @@ " print(function_calls)\n", " else:\n", " print(completion.strip())\n", - " print(\"=\"*100)" + " print(\"=\" * 100)" ] }, { @@ -861,11 +882,13 @@ "prompts = [\n", " \"What's the weather in 05751?\",\n", " \"I'm planning a trip to Killington, Vermont (05751) from Hoboken, NJ (07030). Can you get me weather for both locations and directions?\",\n", - " \"What's the current exchange rate for USD to EUR?\"\n", + " \"What's the current exchange rate for USD to EUR?\",\n", "]\n", "\n", "for prompt in prompts:\n", - " completion = llama.create_completion(generate_hermes_prompt(prompt, functions), max_tokens=-1)[\"choices\"][0][\"text\"]\n", + " completion = llama.create_completion(\n", + " generate_hermes_prompt(prompt, functions), max_tokens=-1\n", + " )[\"choices\"][0][\"text\"]\n", " function_calls = extract_function_calls(completion)\n", "\n", " if function_calls:\n", @@ -875,7 +898,7 @@ " else:\n", " print(completion.strip())\n", "\n", - " print(\"=\"*100)" + " print(\"=\" * 100)" ] }, { diff --git a/examples/notebooks/PerformanceTuning.ipynb b/examples/notebooks/PerformanceTuning.ipynb index 76e26fbd1..ba74e4a41 100644 --- a/examples/notebooks/PerformanceTuning.ipynb +++ b/examples/notebooks/PerformanceTuning.ipynb @@ -13,6 +13,7 @@ "import llama_cpp\n", "\n", "import numpy as np\n", + "\n", "np.int = int\n", "\n", "from skopt.space import Integer, Categorical\n", @@ -25,7 +26,7 @@ " Categorical([True, False], name=\"f16_kv\"),\n", " Categorical([True, False], name=\"use_mlock\"),\n", " Integer(1, multiprocessing.cpu_count(), name=\"n_threads\"),\n", - " Integer(1, 2048, name=\"n_batch\")\n", + " Integer(1, 2048, name=\"n_batch\"),\n", "]\n", "\n", "# TODO: Make this a random prompt to avoid any cache related inconsistencies\n", @@ -41,18 +42,25 @@ "\n", "from skopt.utils import use_named_args\n", "\n", + "\n", "@use_named_args(space)\n", "def objective(**params):\n", " f16_kv = params[\"f16_kv\"]\n", " use_mlock = params[\"use_mlock\"]\n", " n_threads = params[\"n_threads\"]\n", " n_batch = params[\"n_batch\"]\n", - " llm = llama_cpp.Llama(model_path=MODEL_PATH, f16_kv=f16_kv, use_mlock=use_mlock, n_threads=n_threads, n_batch=n_batch)\n", + " llm = llama_cpp.Llama(\n", + " model_path=MODEL_PATH,\n", + " f16_kv=f16_kv,\n", + " use_mlock=use_mlock,\n", + " n_threads=n_threads,\n", + " n_batch=n_batch,\n", + " )\n", "\n", " t1 = time.time()\n", " output = llm(\n", " PROMPT,\n", - " max_tokens=1, # Only optimize prompt processing\n", + " max_tokens=1, # Only optimize prompt processing\n", " stop=[\"###\", \"\\n\"],\n", " echo=True,\n", " )\n", @@ -5240,10 +5248,7 @@ "source": [ "from skopt import gp_minimize\n", "\n", - "res = gp_minimize(\n", - " objective,\n", - " space\n", - ")" + "res = gp_minimize(objective, space)" ] }, { diff --git a/examples/ray/llm.py b/examples/ray/llm.py index 27b15d21d..2325dd303 100755 --- a/examples/ray/llm.py +++ b/examples/ray/llm.py @@ -4,6 +4,7 @@ from ray.serve import Application from llama_cpp import Llama + @serve.deployment class LlamaDeployment: def __init__(self, model_path: str): diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 125d72ded..ff378739f 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -117,9 +117,11 @@ def apply_lora_from_file( self.model, lora_path.encode("utf-8"), scale, - path_base_model.encode("utf-8") - if path_base_model is not None - else ctypes.c_char_p(0), + ( + path_base_model.encode("utf-8") + if path_base_model is not None + else ctypes.c_char_p(0) + ), n_threads, ) @@ -225,7 +227,9 @@ def detokenize(self, tokens: List[int], special: bool = False) -> bytes: # NOTE: Llama1 models automatically added a space at the start of the prompt # this line removes a leading space if the first token is a beginning of sentence token return ( - output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() and output[0:1] == b' ' else output + output[1:] + if len(tokens) > 0 and tokens[0] == self.token_bos() and output[0:1] == b" " + else output ) # Extra @@ -235,20 +239,28 @@ def metadata(self) -> Dict[str, str]: buffer_size = 1024 buffer = ctypes.create_string_buffer(buffer_size) # zero the buffer - buffer.value = b'\0' * buffer_size + buffer.value = b"\0" * buffer_size # iterate over model keys for i in range(llama_cpp.llama_model_meta_count(self.model)): - nbytes = llama_cpp.llama_model_meta_key_by_index(self.model, i, buffer, buffer_size) + nbytes = llama_cpp.llama_model_meta_key_by_index( + self.model, i, buffer, buffer_size + ) if nbytes > buffer_size: buffer_size = nbytes + 1 buffer = ctypes.create_string_buffer(buffer_size) - nbytes = llama_cpp.llama_model_meta_key_by_index(self.model, i, buffer, buffer_size) + nbytes = llama_cpp.llama_model_meta_key_by_index( + self.model, i, buffer, buffer_size + ) key = buffer.value.decode("utf-8") - nbytes = llama_cpp.llama_model_meta_val_str_by_index(self.model, i, buffer, buffer_size) + nbytes = llama_cpp.llama_model_meta_val_str_by_index( + self.model, i, buffer, buffer_size + ) if nbytes > buffer_size: buffer_size = nbytes + 1 buffer = ctypes.create_string_buffer(buffer_size) - nbytes = llama_cpp.llama_model_meta_val_str_by_index(self.model, i, buffer, buffer_size) + nbytes = llama_cpp.llama_model_meta_val_str_by_index( + self.model, i, buffer, buffer_size + ) value = buffer.value.decode("utf-8") metadata[key] = value return metadata @@ -465,7 +477,11 @@ def sample_token_mirostat( ) def sample_token_mirostat_v2( - self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float] + self, + candidates: "_LlamaTokenDataArray", + tau: float, + eta: float, + mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: assert self.ctx is not None return llama_cpp.llama_sample_token_mirostat_v2( @@ -589,7 +605,7 @@ def __init__(self, *, n_vocab: int): size=self.n_vocab, sorted=False, ) - self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) # type: ignore + self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) # type: ignore self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) def copy_logits(self, logits: npt.NDArray[np.single]): @@ -635,10 +651,14 @@ def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> li def _token_to_piece(model: _LlamaModel, token: int, special: bool = False) -> str: assert model.model is not None result = (ctypes.c_char * 8)(0) - n_tokens = llama_cpp.llama_token_to_piece(model.model, token, result, 0, len(result), special) + n_tokens = llama_cpp.llama_token_to_piece( + model.model, token, result, 0, len(result), special + ) if n_tokens < 0: result = (ctypes.c_char * -n_tokens)(0) - check = llama_cpp.llama_token_to_piece(model.model, token, result, 0, len(result), special) + check = llama_cpp.llama_token_to_piece( + model.model, token, result, 0, len(result), special + ) if check != -n_tokens: raise RuntimeError(f"Failed to get piece: token={token}") else: @@ -750,7 +770,10 @@ def prev_str(self, ctx_main: _LlamaContext, n: int) -> str: return ctx_main.model.detokenize(self.prev[-n:]).decode("utf-8") def sample( - self, ctx_main: _LlamaContext, idx: int = 0, logits_array: Optional[npt.NDArray[np.single]] = None + self, + ctx_main: _LlamaContext, + idx: int = 0, + logits_array: Optional[npt.NDArray[np.single]] = None, ): n_vocab = ctx_main.model.n_vocab() id: int = 0 @@ -775,7 +798,7 @@ def sample( if len(self.prev) > 0: nl_token = ctx_main.model.token_nl() nl_logit = logits_array[nl_token] - last_tokens = self.prev[-self.params.penalty_last_n:] + last_tokens = self.prev[-self.params.penalty_last_n :] last_tokens_size = min(len(last_tokens), self.params.penalty_last_n) if last_tokens_size > 0: last_tokens_p = (llama_cpp.llama_token * len(last_tokens))(*last_tokens) diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index 781b26501..29628193b 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -10,6 +10,7 @@ STDOUT_FILENO = 1 STDERR_FILENO = 2 + class suppress_stdout_stderr(object): # NOTE: these must be "saved" here to avoid exceptions when using # this context manager inside of a __del__ method diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 5f5396683..3669efdec 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -86,7 +86,9 @@ def __init__( n_batch: int = 512, n_threads: Optional[int] = None, n_threads_batch: Optional[int] = None, - rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, + rope_scaling_type: Optional[ + int + ] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_UNSPECIFIED, rope_freq_base: float = 0.0, rope_freq_scale: float = 0.0, @@ -229,7 +231,7 @@ def __init__( self.model_params.split_mode = split_mode self.model_params.main_gpu = main_gpu if rpc_servers is not None: - self.model_params.rpc_servers = rpc_servers.encode('utf-8') + self.model_params.rpc_servers = rpc_servers.encode("utf-8") self._rpc_servers = rpc_servers else: self._rpc_servers = None @@ -262,22 +264,34 @@ def __init__( for i, (k, v) in enumerate(kv_overrides.items()): self._kv_overrides_array[i].key = k.encode("utf-8") if isinstance(v, bool): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL + self._kv_overrides_array[i].tag = ( + llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL + ) self._kv_overrides_array[i].value.val_bool = v elif isinstance(v, int): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT + self._kv_overrides_array[i].tag = ( + llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT + ) self._kv_overrides_array[i].value.val_i64 = v elif isinstance(v, float): - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT + self._kv_overrides_array[i].tag = ( + llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT + ) self._kv_overrides_array[i].value.val_f64 = v - elif isinstance(v, str): # type: ignore + elif isinstance(v, str): # type: ignore v_bytes = v.encode("utf-8") - if len(v_bytes) > 128: # TODO: Make this a constant + if len(v_bytes) > 128: # TODO: Make this a constant raise ValueError(f"Value for {k} is too long: {v}") v_bytes = v_bytes.ljust(128, b"\0") - self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR + self._kv_overrides_array[i].tag = ( + llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR + ) # copy min(v_bytes, 128) to str_value - address = typing.cast(int, ctypes.addressof(self._kv_overrides_array[i].value) + llama_cpp.llama_model_kv_override_value.val_str.offset) + address = typing.cast( + int, + ctypes.addressof(self._kv_overrides_array[i].value) + + llama_cpp.llama_model_kv_override_value.val_str.offset, + ) buffer_start = ctypes.cast(address, ctypes.POINTER(ctypes.c_char)) ctypes.memmove( buffer_start, @@ -331,7 +345,7 @@ def __init__( self.context_params.logits_all = ( logits_all if draft_model is None else True ) # Must be set to True for speculative decoding - self.context_params.embeddings = embedding # TODO: Rename to embeddings + self.context_params.embeddings = embedding # TODO: Rename to embeddings self.context_params.offload_kqv = offload_kqv self.context_params.flash_attn = flash_attn # KV cache quantization @@ -355,9 +369,15 @@ def __init__( self._stack = contextlib.ExitStack() - self._model = self._stack.enter_context(contextlib.closing(_LlamaModel( - path_model=self.model_path, params=self.model_params, verbose=self.verbose - ))) + self._model = self._stack.enter_context( + contextlib.closing( + _LlamaModel( + path_model=self.model_path, + params=self.model_params, + verbose=self.verbose, + ) + ) + ) # Override tokenizer self.tokenizer_ = tokenizer or LlamaTokenizer(self) @@ -369,18 +389,26 @@ def __init__( self.context_params.n_ctx = self._model.n_ctx_train() self.context_params.n_batch = self.n_batch - self._ctx = self._stack.enter_context(contextlib.closing(_LlamaContext( - model=self._model, - params=self.context_params, - verbose=self.verbose, - ))) + self._ctx = self._stack.enter_context( + contextlib.closing( + _LlamaContext( + model=self._model, + params=self.context_params, + verbose=self.verbose, + ) + ) + ) - self._batch = self._stack.enter_context(contextlib.closing(_LlamaBatch( - n_tokens=self.n_batch, - embd=0, - n_seq_max=self.context_params.n_ctx, - verbose=self.verbose, - ))) + self._batch = self._stack.enter_context( + contextlib.closing( + _LlamaBatch( + n_tokens=self.n_batch, + embd=0, + n_seq_max=self.context_params.n_ctx, + verbose=self.verbose, + ) + ) + ) if self.lora_path: if self._model.apply_lora_from_file( @@ -398,7 +426,9 @@ def __init__( self.chat_format = chat_format self.chat_handler = chat_handler - self._chat_handlers: Dict[str, llama_chat_format.LlamaChatCompletionHandler] = {} + self._chat_handlers: Dict[str, llama_chat_format.LlamaChatCompletionHandler] = ( + {} + ) self.draft_model = draft_model @@ -433,17 +463,30 @@ def __init__( eos_token_id = self.token_eos() bos_token_id = self.token_bos() - eos_token = self._model.token_get_text(eos_token_id) if eos_token_id != -1 else "" - bos_token = self._model.token_get_text(bos_token_id) if bos_token_id != -1 else "" + eos_token = ( + self._model.token_get_text(eos_token_id) if eos_token_id != -1 else "" + ) + bos_token = ( + self._model.token_get_text(bos_token_id) if bos_token_id != -1 else "" + ) # Unfortunately the llama.cpp API does not return metadata arrays, so we can't get template names from tokenizer.chat_templates - template_choices = dict((name[10:], template) for name, template in self.metadata.items() if name.startswith("tokenizer.chat_template.")) + template_choices = dict( + (name[10:], template) + for name, template in self.metadata.items() + if name.startswith("tokenizer.chat_template.") + ) if "tokenizer.chat_template" in self.metadata: - template_choices["chat_template.default"] = self.metadata["tokenizer.chat_template"] + template_choices["chat_template.default"] = self.metadata[ + "tokenizer.chat_template" + ] if self.verbose and template_choices: - print(f"Available chat formats from metadata: {', '.join(template_choices.keys())}", file=sys.stderr) + print( + f"Available chat formats from metadata: {', '.join(template_choices.keys())}", + file=sys.stderr, + ) for name, template in template_choices.items(): self._chat_handlers[name] = llama_chat_format.Jinja2ChatFormatter( @@ -468,7 +511,10 @@ def __init__( print(f"Guessed chat format: {chat_format}", file=sys.stderr) else: if self.verbose: - print(f"Using gguf chat template: {template_choices['chat_template.default']}", file=sys.stderr) + print( + f"Using gguf chat template: {template_choices['chat_template.default']}", + file=sys.stderr, + ) print(f"Using chat eos_token: {eos_token}", file=sys.stderr) print(f"Using chat bos_token: {bos_token}", file=sys.stderr) @@ -477,7 +523,9 @@ def __init__( if self.chat_format is None and self.chat_handler is None: self.chat_format = "llama-2" if self.verbose: - print(f"Using fallback chat format: {self.chat_format}", file=sys.stderr) + print( + f"Using fallback chat format: {self.chat_format}", file=sys.stderr + ) @property def ctx(self) -> llama_cpp.llama_context_p: @@ -582,13 +630,17 @@ def eval(self, tokens: Sequence[int]): if self.context_params.logits_all: rows = n_tokens cols = self._n_vocab - logits = np.ctypeslib.as_array(self._ctx.get_logits(), shape=(rows * cols, )) - self.scores[n_past : n_past + n_tokens, :].reshape(-1)[: :] = logits + logits = np.ctypeslib.as_array( + self._ctx.get_logits(), shape=(rows * cols,) + ) + self.scores[n_past : n_past + n_tokens, :].reshape(-1)[::] = logits else: rows = 1 cols = self._n_vocab - logits = np.ctypeslib.as_array(self._ctx.get_logits(), shape=(rows * cols, )) - self.scores[n_past + n_tokens - 1, :].reshape(-1)[: :] = logits + logits = np.ctypeslib.as_array( + self._ctx.get_logits(), shape=(rows * cols,) + ) + self.scores[n_past + n_tokens - 1, :].reshape(-1)[::] = logits # Update n_tokens self.n_tokens += n_tokens @@ -880,7 +932,8 @@ def decode_batch(seq_sizes: List[int]): for i, size in enumerate(seq_sizes): ptr = llama_cpp.llama_get_embeddings(self._ctx.ctx) embedding: List[List[float]] = [ - ptr[pos + j * n_embd : pos + (j + 1) * n_embd] for j in range(size) + ptr[pos + j * n_embd : pos + (j + 1) * n_embd] + for j in range(size) ] if normalize: embedding = [_normalize_embedding(e) for e in embedding] @@ -987,14 +1040,24 @@ def _create_completion( prefix_token_id: int = self._model.token_prefix() middle_token_id: int = self._model.token_middle() suffix_token_id: int = self._model.token_suffix() - add_space_prefix: bool = self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true" + add_space_prefix: bool = ( + self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true" + ) bos_tokens: List[int] = [cls_token_id if cls_token_id != -1 else bos_token_id] - eos_tokens: List[int] = [sep_token_id if sep_token_id != -1 else self.token_eos()] + eos_tokens: List[int] = [ + sep_token_id if sep_token_id != -1 else self.token_eos() + ] - if (isinstance(prompt, list) and suffix is None) or self._model.add_bos_token() == 0 or bos_tokens[:1] == [-1]: + if ( + (isinstance(prompt, list) and suffix is None) + or self._model.add_bos_token() == 0 + or bos_tokens[:1] == [-1] + ): bos_tokens = [] - if (isinstance(prompt, list) and suffix is None) or (self._model.add_eos_token() != 1 and sep_token_id == -1): + if (isinstance(prompt, list) and suffix is None) or ( + self._model.add_eos_token() != 1 and sep_token_id == -1 + ): eos_tokens = [] suffix_space_prefix: int = 0 @@ -1008,28 +1071,27 @@ def _create_completion( completion_tokens: List[int] = [] if len(prompt) > 0 else [bos_token_id] # Add blank space to start of prompt to match OG llama tokenizer prefix_tokens: List[int] = ( + [prefix_token_id] if prefix_token_id >= 0 and suffix is not None else [] + ) + ( ( - [prefix_token_id] - if prefix_token_id >= 0 and suffix is not None - else [] - ) - + - ( - ( - self.tokenize(prompt.encode("utf-8"), add_bos=False, special=(prefix_token_id < 0 or suffix is None)) - if prompt != "" - else [] + self.tokenize( + prompt.encode("utf-8"), + add_bos=False, + special=(prefix_token_id < 0 or suffix is None), ) - if isinstance(prompt, str) - else prompt + if prompt != "" + else [] ) + if isinstance(prompt, str) + else prompt ) suffix_tokens: List[int] = ( ( [suffix_token_id] - + - ( - self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)[suffix_space_prefix:] + + ( + self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)[ + suffix_space_prefix: + ] if suffix else [] ) @@ -1038,11 +1100,17 @@ def _create_completion( else [] ) middle_tokens: List[int] = ( - [middle_token_id] - if middle_token_id >= 0 and suffix is not None - else [] + [middle_token_id] if middle_token_id >= 0 and suffix is not None else [] + ) + prompt_tokens: List[int] = ( + bos_tokens + + ( + (suffix_tokens + prefix_tokens + middle_tokens) + if self.spm_infill + else (prefix_tokens + suffix_tokens + middle_tokens) + ) + + eos_tokens ) - prompt_tokens: List[int] = bos_tokens + ((suffix_tokens + prefix_tokens + middle_tokens) if self.spm_infill else (prefix_tokens + suffix_tokens + middle_tokens)) + eos_tokens text: bytes = b"" returned_tokens: int = 0 stop = ( @@ -1179,7 +1247,10 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - remaining_text = self.detokenize(remaining_tokens, prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) + remaining_text = self.detokenize( + remaining_tokens, + prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], + ) remaining_length = len(remaining_text) # We want to avoid yielding any characters from @@ -1201,19 +1272,29 @@ def logit_bias_processor( for token in remaining_tokens: if token == bos_token_id: continue - token_end_position += len(self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens])) + token_end_position += len( + self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ) + ) # Check if stop sequence is in the token if token_end_position > ( remaining_length - first_stop_position ): break - token_str = self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( - "utf-8", errors="ignore" - ) + token_str = self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore") text_offset = len(prompt) + len( - self.detokenize(completion_tokens[:returned_tokens], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( - "utf-8", errors="ignore" - ) + self.detokenize( + completion_tokens[:returned_tokens], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore") ) token_offset = len(prompt_tokens) + returned_tokens logits = self._scores[token_offset - 1, :] @@ -1233,9 +1314,11 @@ def logit_bias_processor( top_logprob.update({token_str: current_logprobs[int(token)]}) logprobs_or_none = { "tokens": [ - self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( - "utf-8", errors="ignore" - ) + self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore") ], "text_offset": [text_offset], "token_logprobs": [current_logprobs[int(token)]], @@ -1249,9 +1332,11 @@ def logit_bias_processor( "model": model_name, "choices": [ { - "text": self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]).decode( - "utf-8", errors="ignore" - ), + "text": self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore"), "index": 0, "logprobs": logprobs_or_none, "finish_reason": None, @@ -1263,7 +1348,11 @@ def logit_bias_processor( decode_success = False for i in range(1, len(remaining_tokens) + 1): try: - bs = self.detokenize(remaining_tokens[:i], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) + bs = self.detokenize( + remaining_tokens[:i], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ) ts = bs.decode("utf-8") decode_success = True break @@ -1313,7 +1402,10 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - all_text = self.detokenize(remaining_tokens, prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) + all_text = self.detokenize( + remaining_tokens, + prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], + ) any_stop = [s for s in stop_sequences if s in all_text] if len(any_stop) > 0: end = min(all_text.index(stop) for stop in any_stop) @@ -1322,7 +1414,12 @@ def logit_bias_processor( token_end_position = 0 for token in remaining_tokens: - token_end_position += len(self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens])) + token_end_position += len( + self.detokenize( + [token], + prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], + ) + ) logprobs_or_none: Optional[CompletionLogprobs] = None if logprobs is not None: @@ -1332,7 +1429,11 @@ def logit_bias_processor( "utf-8", errors="ignore" ) text_offset = len(prompt) + len( - self.detokenize(completion_tokens[:returned_tokens], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]) + self.detokenize( + completion_tokens[:returned_tokens], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ) ) token_offset = len(prompt_tokens) + returned_tokens - 1 logits = self._scores[token_offset, :] @@ -1441,12 +1542,17 @@ def logit_bias_processor( if echo: # Remove leading BOS token if exists - all_tokens = prompt_tokens[1 if prompt_tokens[0] == self.token_bos() else 0:] + completion_tokens + all_tokens = ( + prompt_tokens[1 if prompt_tokens[0] == self.token_bos() else 0 :] + + completion_tokens + ) else: all_tokens = completion_tokens all_token_strs = [ - self.detokenize([token], prev_tokens=all_tokens[:i]).decode("utf-8", errors="ignore") + self.detokenize([token], prev_tokens=all_tokens[:i]).decode( + "utf-8", errors="ignore" + ) for i, token in enumerate(all_tokens) ] all_logprobs = Llama.logits_to_logprobs(self._scores)[token_offset:] @@ -1472,7 +1578,9 @@ def logit_bias_processor( ) token_logprobs.append(logprobs_token[int(token)]) top_logprob: Optional[Dict[str, float]] = { - self.detokenize([i], prev_tokens=all_tokens[:idx]).decode("utf-8", errors="ignore"): logprob + self.detokenize([i], prev_tokens=all_tokens[:idx]).decode( + "utf-8", errors="ignore" + ): logprob for logprob, i in sorted_logprobs[:logprobs] } top_logprob.update({token_str: logprobs_token[int(token)]}) @@ -1765,8 +1873,10 @@ def create_chat_completion( Returns: Generated chat completion or a stream of chat completion chunks. """ - handler = self.chat_handler or self._chat_handlers.get(self.chat_format) or llama_chat_format.get_chat_completion_handler( - self.chat_format + handler = ( + self.chat_handler + or self._chat_handlers.get(self.chat_format) + or llama_chat_format.get_chat_completion_handler(self.chat_format) ) return handler( llama=self, diff --git a/llama_cpp/llama_cache.py b/llama_cpp/llama_cache.py index 9e9870a52..5220c7933 100644 --- a/llama_cpp/llama_cache.py +++ b/llama_cpp/llama_cache.py @@ -40,7 +40,9 @@ def __contains__(self, key: Sequence[int]) -> bool: raise NotImplementedError @abstractmethod - def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState") -> None: + def __setitem__( + self, key: Sequence[int], value: "llama_cpp.llama.LlamaState" + ) -> None: raise NotImplementedError @@ -50,7 +52,9 @@ class LlamaRAMCache(BaseLlamaCache): def __init__(self, capacity_bytes: int = (2 << 30)): super().__init__(capacity_bytes) self.capacity_bytes = capacity_bytes - self.cache_state: OrderedDict[Tuple[int, ...], "llama_cpp.llama.LlamaState"] = OrderedDict() + self.cache_state: OrderedDict[Tuple[int, ...], "llama_cpp.llama.LlamaState"] = ( + OrderedDict() + ) @property def cache_size(self): @@ -63,7 +67,8 @@ def _find_longest_prefix_key( min_len = 0 min_key = None keys = ( - (k, llama_cpp.llama.Llama.longest_token_prefix(k, key)) for k in self.cache_state.keys() + (k, llama_cpp.llama.Llama.longest_token_prefix(k, key)) + for k in self.cache_state.keys() ) for k, prefix_len in keys: if prefix_len > min_len: diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 975f74762..78d83ba86 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -8,7 +8,18 @@ import string from contextlib import ExitStack -from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol, cast +from typing import ( + Any, + Dict, + Iterator, + List, + Literal, + Optional, + Tuple, + Union, + Protocol, + cast, +) import jinja2 from jinja2.sandbox import ImmutableSandboxedEnvironment @@ -191,7 +202,9 @@ def __init__( self.eos_token = eos_token self.bos_token = bos_token self.add_generation_prompt = add_generation_prompt - self.stop_token_ids = set(stop_token_ids) if stop_token_ids is not None else None + self.stop_token_ids = ( + set(stop_token_ids) if stop_token_ids is not None else None + ) self._environment = ImmutableSandboxedEnvironment( loader=jinja2.BaseLoader(), @@ -226,14 +239,20 @@ def raise_exception(message: str): stopping_criteria = None if self.stop_token_ids is not None: + def stop_on_last_token( - tokens: npt.NDArray[np.intc], - logits: npt.NDArray[np.single] + tokens: npt.NDArray[np.intc], logits: npt.NDArray[np.single] ) -> bool: return tokens[-1] in self.stop_token_ids + stopping_criteria = llama.StoppingCriteriaList([stop_on_last_token]) - return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token], stopping_criteria=stopping_criteria, added_special=True) + return ChatFormatterResponse( + prompt=prompt, + stop=[self.eos_token], + stopping_criteria=stopping_criteria, + added_special=True, + ) def to_chat_handler(self) -> LlamaChatCompletionHandler: return chat_formatter_to_chat_completion_handler(self) @@ -430,7 +449,9 @@ def _stream_response_to_function_stream( "type": "function", "function": { "name": tool_name, - "arguments": chunk["choices"][0]["text"], + "arguments": chunk["choices"][0][ + "text" + ], }, } ], @@ -465,9 +486,7 @@ def _stream_response_to_function_stream( "type": "function", "function": { "name": tool_name, - "arguments": chunk["choices"][0][ - "text" - ], + "arguments": chunk["choices"][0]["text"], }, } ], @@ -500,7 +519,6 @@ def _stream_response_to_function_stream( return _stream_response_to_function_stream(chunks) - def chat_formatter_to_chat_completion_handler( chat_formatter: ChatFormatter, ) -> LlamaChatCompletionHandler: @@ -549,7 +567,11 @@ def chat_completion_handler( tools=tools, tool_choice=tool_choice, ) - prompt = llama.tokenize(result.prompt.encode("utf-8"), add_bos=not result.added_special, special=True) + prompt = llama.tokenize( + result.prompt.encode("utf-8"), + add_bos=not result.added_special, + special=True, + ) if result.stop is not None: stop = [] if stop is None else [stop] if isinstance(stop, str) else stop rstop = result.stop if isinstance(result.stop, list) else [result.stop] @@ -560,7 +582,9 @@ def chat_completion_handler( stopping_criteria = result.stopping_criteria if response_format is not None and response_format["type"] == "json_object": - grammar = _grammar_for_response_format(response_format, verbose=llama.verbose) + grammar = _grammar_for_response_format( + response_format, verbose=llama.verbose + ) # Convert legacy functions to tools if functions is not None: @@ -587,7 +611,11 @@ def chat_completion_handler( } tool = None - if tool_choice is not None and isinstance(tool_choice, dict) and tools is not None: + if ( + tool_choice is not None + and isinstance(tool_choice, dict) + and tools is not None + ): name = tool_choice["function"]["name"] tool = next((t for t in tools if t["function"]["name"] == name), None) if tool is None: @@ -656,7 +684,9 @@ def format_autotokenizer( prompt: str = tokenizer.apply_chat_template(messages, tokenize=False) # type: ignore assert isinstance(prompt, str) # Return formatted prompt and eos token by default - return ChatFormatterResponse(prompt=prompt, stop=tokenizer.eos_token, added_special=True) + return ChatFormatterResponse( + prompt=prompt, stop=tokenizer.eos_token, added_special=True + ) return format_autotokenizer @@ -709,7 +739,9 @@ def format_tokenizer_config( bos_token=bos_token, eos_token=eos_token, ) - return ChatFormatterResponse(prompt=prompt, stop=[eos_token, bos_token], added_special=True) + return ChatFormatterResponse( + prompt=prompt, stop=[eos_token, bos_token], added_special=True + ) return format_tokenizer_config @@ -731,8 +763,10 @@ def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[s if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE: return "chatml" - if (metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE or - metadata["tokenizer.chat_template"] == MIXTRAL_INSTRUCT_CHAT_TEMPLATE): + if ( + metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE + or metadata["tokenizer.chat_template"] == MIXTRAL_INSTRUCT_CHAT_TEMPLATE + ): return "mistral-instruct" if metadata["tokenizer.chat_template"] == LLAMA3_INSTRUCT_CHAT_TEMPLATE: @@ -868,13 +902,15 @@ def _format_chatglm3( ret += role return ret -def _grammar_for_json(verbose:bool=False): - return llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF, verbose=verbose) + +def _grammar_for_json(verbose: bool = False): + return llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=verbose + ) + def _grammar_for_json_schema( - schema: str, - verbose: bool = False, - fallback_to_json: bool = True + schema: str, verbose: bool = False, fallback_to_json: bool = True ): try: return llama_grammar.LlamaGrammar.from_json_schema(schema, verbose=verbose) @@ -884,9 +920,10 @@ def _grammar_for_json_schema( else: raise e + def _grammar_for_response_format( - response_format: llama_types.ChatCompletionRequestResponseFormat, - verbose: bool = False + response_format: llama_types.ChatCompletionRequestResponseFormat, + verbose: bool = False, ): if response_format["type"] != "json_object": return None @@ -898,6 +935,7 @@ def _grammar_for_response_format( else: return _grammar_for_json(verbose=verbose) + ### Chat Formats ### @@ -1239,10 +1277,7 @@ def format_mistral_instruct( and isinstance(message["content"], str) ): prompt += "[INST] " + message["content"] - elif ( - message["role"] == "assistant" - and message["content"] is not None - ): + elif message["role"] == "assistant" and message["content"] is not None: prompt += " [/INST]" + message["content"] + eos prompt += " [/INST]" return ChatFormatterResponse(prompt=prompt, stop=stop) @@ -1932,7 +1967,9 @@ def prepare_messages_for_inference( grammar=grammar, ) if stream is False: - completion_or_completion_chunks["choices"][0]["text"] = completion_or_completion_chunks["choices"][0]["text"].lstrip() + completion_or_completion_chunks["choices"][0]["text"] = ( + completion_or_completion_chunks["choices"][0]["text"].lstrip() + ) return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore def get_grammar(function_call): @@ -1969,45 +2006,50 @@ def get_grammar(function_call): return grammar def create_completion(prompt, stop, grammar): - completion = cast(llama_types.Completion, llama.create_completion( - prompt=prompt, - temperature=temperature, - top_p=top_p, - top_k=top_k, - min_p=min_p, - typical_p=typical_p, - stream=stream, - stop=stop, - max_tokens=max_tokens, - presence_penalty=presence_penalty, - frequency_penalty=frequency_penalty, - repeat_penalty=repeat_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - model=model, - logits_processor=logits_processor, - grammar=grammar, - )) + completion = cast( + llama_types.Completion, + llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ), + ) return completion content = "" function_calls, function_bodies = [], [] completion_tokens = 0 - + def generate_streaming(tools, functions, function_call, prompt): assert version == "v2", "Streaming for v1 is not supported" - + chunk_id, chunk_created = None, None - + # If tool_choice/function_call is provided if isinstance(function_call, dict): prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" grammar = get_grammar(function_call["name"]) stops = [STOP_TOKEN, FROM_TOKEN] - tool_id = "".join([random.choice(string.ascii_letters + string.digits) for _ in range(24)]) + tool_id = "".join( + [random.choice(string.ascii_letters + string.digits) for _ in range(24)] + ) completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_text = "" first = True @@ -2021,19 +2063,35 @@ def generate_streaming(tools, functions, function_call, prompt): "index": 0, "id": "call_" + tool_id, "type": "function", - "function": {"name": function_call["name"], "arguments": ""}, + "function": { + "name": function_call["name"], + "arguments": "", + }, } ] } else: - func_call_dict = {"function_call": {"name": function_call["name"], "arguments": ""}} + func_call_dict = { + "function_call": { + "name": function_call["name"], + "arguments": "", + } + } yield llama_types.CreateChatCompletionStreamResponse( id="chat" + chunk["id"], object="chat.completion.chunk", created=chunk["created"], model=chunk["model"], choices=[ - {"index": 0, "logprobs": None, "delta": {"role": None, "content": None, **func_call_dict}} + { + "index": 0, + "logprobs": None, + "delta": { + "role": None, + "content": None, + **func_call_dict, + }, + } ], ) first = False @@ -2052,7 +2110,12 @@ def generate_streaming(tools, functions, function_call, prompt): ] } else: - func_call_dict = {"function_call": {"name": None, "arguments": chunk["choices"][0]["text"].rstrip()}} + func_call_dict = { + "function_call": { + "name": None, + "arguments": chunk["choices"][0]["text"].rstrip(), + } + } if len(chunk["choices"][0]["text"].rstrip()) > 0: yield llama_types.CreateChatCompletionStreamResponse( id="chat" + chunk["id"], @@ -2080,10 +2143,15 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "finish_reason": "tool_calls" if tools is not None else "function_call", + "finish_reason": ( + "tool_calls" if tools is not None else "function_call" + ), "logprobs": None, "delta": { - "role": None, "content": None, "function_call": None, "tool_calls": None + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, }, } ], @@ -2095,7 +2163,9 @@ def generate_streaming(tools, functions, function_call, prompt): # Generate function name first grammar = None stops = CONTENT_TOKEN - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) completion_text = "" for chunk in completion: completion_text += chunk["choices"][0]["text"] @@ -2124,7 +2194,12 @@ def generate_streaming(tools, functions, function_call, prompt): else: prompt += f"{function_name}\n<|content|>" grammar = get_grammar(function_name) - tool_id = "".join([random.choice(string.ascii_letters + string.digits) for _ in range(24)]) + tool_id = "".join( + [ + random.choice(string.ascii_letters + string.digits) + for _ in range(24) + ] + ) if tools is not None: func_call_dict = { "tool_calls": [ @@ -2132,12 +2207,17 @@ def generate_streaming(tools, functions, function_call, prompt): "index": tool_index, "id": "call_" + tool_id, "type": "function", - "function": {"name": function_name, "arguments": ""}, + "function": { + "name": function_name, + "arguments": "", + }, } ] } else: - func_call_dict = {"function_call": {"name": function_name, "arguments": ""}} + func_call_dict = { + "function_call": {"name": function_name, "arguments": ""} + } # Stream function name yield llama_types.CreateChatCompletionStreamResponse( id="chat" + chunk_id, @@ -2158,10 +2238,16 @@ def generate_streaming(tools, functions, function_call, prompt): ) # Generate content stops = [RECIPIENT_TOKEN, STOP_TOKEN] - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) if function_name == "all": completion_text = "" - stop_sequence, buffer, is_end = "\n<|from|>assistant\n<|recipient|>", [], False + stop_sequence, buffer, is_end = ( + "\n<|from|>assistant\n<|recipient|>", + [], + False, + ) for i, chunk in enumerate(completion): completion_text += chunk["choices"][0]["text"] if is_end: @@ -2179,9 +2265,12 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": chunk["choices"][0][ + "logprobs" + ], "delta": { - "role": "assistant", "content": buffer.pop(0) + "role": "assistant", + "content": buffer.pop(0), }, } ], @@ -2204,17 +2293,30 @@ def generate_streaming(tools, functions, function_call, prompt): "logprobs": chunk["choices"][0]["logprobs"], "delta": { "role": "assistant", - "content": chunk["choices"][0]["text"] if i > 0 else chunk["choices"][0]["text"].lstrip() + "content": ( + chunk["choices"][0]["text"] + if i > 0 + else chunk["choices"][0][ + "text" + ].lstrip() + ), }, } ], ) # Check whether the model wants to generate another turn - if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text: + if ( + "<|from|> assistant" in completion_text + or "<|from|>assistant" in completion_text + ): if completion_text.endswith("\n<|from|>assistant\n"): - cleaned_completion_text = completion_text[:-len("\n<|from|>assistant\n")].strip() + cleaned_completion_text = completion_text[ + : -len("\n<|from|>assistant\n") + ].strip() elif completion_text.endswith("\n<|from|> assistant\n"): - cleaned_completion_text = completion_text[:-len("\n<|from|> assistant\n")].strip() + cleaned_completion_text = completion_text[ + : -len("\n<|from|> assistant\n") + ].strip() else: cleaned_completion_text = completion_text.strip() prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" @@ -2250,13 +2352,22 @@ def generate_streaming(tools, functions, function_call, prompt): "type": "function", "function": { "name": None, - "arguments": chunk["choices"][0]["text"].rstrip(), + "arguments": chunk["choices"][0][ + "text" + ].rstrip(), }, } ] } else: - func_call_dict = {"function_call": {"name": None, "arguments": chunk["choices"][0]["text"].rstrip()}} + func_call_dict = { + "function_call": { + "name": None, + "arguments": chunk["choices"][0][ + "text" + ].rstrip(), + } + } yield llama_types.CreateChatCompletionStreamResponse( id="chat" + chunk_id, object="chat.completion.chunk", @@ -2276,9 +2387,16 @@ def generate_streaming(tools, functions, function_call, prompt): ) prompt += completion_text.strip() grammar = None - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) - completion_text += "".join([chunk["choices"][0]["text"] for chunk in completion]) - if ("<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text) and tools is not None: + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_text += "".join( + [chunk["choices"][0]["text"] for chunk in completion] + ) + if ( + "<|from|> assistant" in completion_text + or "<|from|>assistant" in completion_text + ) and tools is not None: prompt += "\n<|from|>assistant\n<|recipient|>" tool_index += 1 else: @@ -2291,16 +2409,23 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "finish_reason": "tool_calls" if tools is not None else "function_call", + "finish_reason": ( + "tool_calls" + if tools is not None + else "function_call" + ), "logprobs": None, "delta": { - "role": None, "content": None, "function_call": None, "tool_calls": None + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, }, } ], ) break - + if stream is not False: return generate_streaming( tools=tools, functions=functions, function_call=function_call, prompt=prompt @@ -2324,7 +2449,6 @@ def generate_streaming(tools, functions, function_call, prompt): completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] - # If the generation does not involve a function call if ( @@ -2348,7 +2472,9 @@ def generate_streaming(tools, functions, function_call, prompt): completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() ) grammar = get_grammar(function_calls[-1]) - completion = create_completion(prompt=prompt, stop=END_FUNCTION_CALL_TOKEN, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=END_FUNCTION_CALL_TOKEN, grammar=grammar + ) completion_tokens += completion["usage"]["completion_tokens"] function_bodies.append(completion["choices"][0]["text"].strip()) # If the prompt involves a function call, just append generated parameters to function_bodies @@ -2362,7 +2488,9 @@ def generate_streaming(tools, functions, function_call, prompt): function_calls.append(function_call) grammar = get_grammar(function_call) stops = [STOP_TOKEN, FROM_TOKEN] - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] function_bodies.append(completion_text.strip()) @@ -2372,7 +2500,9 @@ def generate_streaming(tools, functions, function_call, prompt): # Generate function name first grammar = None stops = CONTENT_TOKEN - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] function_name = completion_text.strip() @@ -2385,23 +2515,32 @@ def generate_streaming(tools, functions, function_call, prompt): grammar = get_grammar(function_call) # Generate content stops = [RECIPIENT_TOKEN, STOP_TOKEN] - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) completion_text = completion["choices"][0]["text"] completion_tokens += completion["usage"]["completion_tokens"] if function_name == "all": if completion_text.endswith("\n<|from|>assistant\n"): - content += completion_text[:-len("\n<|from|>assistant\n")] + content += completion_text[: -len("\n<|from|>assistant\n")] if completion_text.endswith("\n<|from|> assistant\n"): content += completion_text[-len("\n<|from|> assistant\n")] else: content += completion_text content = content.lstrip() # Check whether the model wants to generate another turn - if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text: + if ( + "<|from|> assistant" in completion_text + or "<|from|>assistant" in completion_text + ): if completion_text.endswith("\n<|from|>assistant\n"): - cleaned_completion_text = completion_text[:-len("\n<|from|>assistant\n")].strip() + cleaned_completion_text = completion_text[ + : -len("\n<|from|>assistant\n") + ].strip() elif completion_text.endswith("\n<|from|> assistant\n"): - cleaned_completion_text = completion_text[-len("\n<|from|> assistant\n")].strip() + cleaned_completion_text = completion_text[ + -len("\n<|from|> assistant\n") + ].strip() else: cleaned_completion_text = completion_text.strip() prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" @@ -2412,9 +2551,14 @@ def generate_streaming(tools, functions, function_call, prompt): # Check whether the model wants to generate another turn prompt += completion_text.strip() grammar = None - completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) completion_tokens += completion["usage"]["completion_tokens"] - if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]: + if ( + "<|from|> assistant" in completion["choices"][0]["text"] + or "<|from|>assistant" in completion["choices"][0]["text"] + ): prompt += "\n<|from|>assistant\n<|recipient|>" else: break @@ -2442,7 +2586,13 @@ def generate_streaming(tools, functions, function_call, prompt): ) # TODO: support stream mode - function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = {} + function_call_dict: Union[ + Dict[str, str], + Dict[ + Literal["function_call"], + llama_types.ChatCompletionRequestAssistantMessageFunctionCall, + ], + ] = {} if len(tool_calls) > 0: if tools is not None: function_call_dict["tool_calls"] = tool_calls @@ -2474,7 +2624,9 @@ def generate_streaming(tools, functions, function_call, prompt): class Llava15ChatHandler: - DEFAULT_SYSTEM_MESSAGE: Optional[str] = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." + DEFAULT_SYSTEM_MESSAGE: Optional[str] = ( + "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." + ) CHAT_FORMAT = ( "{% for message in messages %}" @@ -2487,7 +2639,6 @@ class Llava15ChatHandler: "{% endif %}" "{% if message.content is iterable %}" "\nUSER: " - "{% for content in message.content %}" "{% if content.type == 'image_url' and content.image_url is string %}" "{{ content.image_url }}" @@ -2496,13 +2647,11 @@ class Llava15ChatHandler: "{{ content.image_url.url }}" "{% endif %}" "{% endfor %}" - "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" "{% endfor %}" - "{% endif %}" "{% endif %}" "{% if message.role == 'assistant' and message.content is not none %}" @@ -2520,30 +2669,30 @@ def __init__(self, clip_model_path: str, verbose: bool = True): self.clip_model_path = clip_model_path self.verbose = verbose - self._llava_cpp = llava_cpp # TODO: Fix + self._llava_cpp = llava_cpp # TODO: Fix self._exit_stack = ExitStack() - self._last_image_embed: Optional[llava_cpp.CtypesPointer[llava_cpp.llava_image_embed]] = None + self._last_image_embed: Optional[ + llava_cpp.CtypesPointer[llava_cpp.llava_image_embed] + ] = None self._last_image_hash: Optional[int] = None if not os.path.exists(clip_model_path): raise ValueError(f"Clip model path does not exist: {clip_model_path}") with suppress_stdout_stderr(disable=self.verbose): - clip_ctx = self._llava_cpp.clip_model_load( - self.clip_model_path.encode(), 0 - ) + clip_ctx = self._llava_cpp.clip_model_load(self.clip_model_path.encode(), 0) if clip_ctx is None: raise ValueError(f"Failed to load clip model: {clip_model_path}") - + self.clip_ctx = clip_ctx def clip_free(): with suppress_stdout_stderr(disable=self.verbose): self._llava_cpp.clip_free(self.clip_ctx) - + self._exit_stack.callback(clip_free) - + def last_image_embed_free(): with suppress_stdout_stderr(disable=self.verbose): if self._last_image_embed is not None: @@ -2598,7 +2747,11 @@ def __call__( system_prompt = _get_system_message(messages) if system_prompt == "" and self.DEFAULT_SYSTEM_MESSAGE is not None: - messages = [llama_types.ChatCompletionRequestSystemMessage(role="system", content=self.DEFAULT_SYSTEM_MESSAGE)] + messages + messages = [ + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=self.DEFAULT_SYSTEM_MESSAGE + ) + ] + messages image_urls = self.get_image_urls(messages) template = ImmutableSandboxedEnvironment( @@ -2614,7 +2767,11 @@ def __call__( split_text = self.split_text_on_image_urls(text, image_urls) def embed_image_bytes(image_bytes: bytes): - if self._last_image_embed is not None and self._last_image_hash is not None and hash(image_bytes) == self._last_image_hash: + if ( + self._last_image_embed is not None + and self._last_image_hash is not None + and hash(image_bytes) == self._last_image_hash + ): return self._last_image_embed with suppress_stdout_stderr(disable=self.verbose): # Free the previous image embed @@ -2622,13 +2779,13 @@ def embed_image_bytes(image_bytes: bytes): self._llava_cpp.llava_image_embed_free(self._last_image_embed) self._last_image_embed = None self._last_image_hash = None - embed = ( - self._llava_cpp.llava_image_embed_make_with_bytes( - self.clip_ctx, - llama.context_params.n_threads_batch, - (ctypes.c_uint8 * len(image_bytes)).from_buffer(bytearray(image_bytes)), - len(image_bytes), - ) + embed = self._llava_cpp.llava_image_embed_make_with_bytes( + self.clip_ctx, + llama.context_params.n_threads_batch, + (ctypes.c_uint8 * len(image_bytes)).from_buffer( + bytearray(image_bytes) + ), + len(image_bytes), ) self._last_image_embed = embed self._last_image_hash = hash(image_bytes) @@ -2639,15 +2796,21 @@ def embed_image_bytes(image_bytes: bytes): llama._ctx.kv_cache_clear() for type_, value in split_text: if type_ == "text": - tokens = llama.tokenize(value.encode("utf8"), add_bos=False, special=True) + tokens = llama.tokenize( + value.encode("utf8"), add_bos=False, special=True + ) if llama.n_tokens + len(tokens) > llama.n_ctx(): - raise ValueError(f"Prompt exceeds n_ctx: {llama.n_tokens + len(tokens)} > {llama.n_ctx()}") + raise ValueError( + f"Prompt exceeds n_ctx: {llama.n_tokens + len(tokens)} > {llama.n_ctx()}" + ) llama.eval(tokens) else: image_bytes = self.load_image(value) embed = embed_image_bytes(image_bytes) if llama.n_tokens + embed.contents.n_image_pos > llama.n_ctx(): - raise ValueError(f"Prompt exceeds n_ctx: {llama.n_tokens + embed.contents.n_image_pos} > {llama.n_ctx()}") + raise ValueError( + f"Prompt exceeds n_ctx: {llama.n_tokens + embed.contents.n_image_pos} > {llama.n_ctx()}" + ) n_past = ctypes.c_int(llama.n_tokens) n_past_p = ctypes.pointer(n_past) with suppress_stdout_stderr(disable=self.verbose): @@ -2692,7 +2855,11 @@ def embed_image_bytes(image_bytes: bytes): } tool = None - if tool_choice is not None and isinstance(tool_choice, dict) and tools is not None: + if ( + tool_choice is not None + and isinstance(tool_choice, dict) + and tools is not None + ): name = tool_choice["function"]["name"] tool = next((t for t in tools if t["function"]["name"] == name), None) if tool is None: @@ -2809,9 +2976,10 @@ def from_pretrained( ) -> "Llava15ChatHandler": import fnmatch from pathlib import Path + try: - from huggingface_hub import hf_hub_download, HfFileSystem # type: ignore - from huggingface_hub.utils import validate_repo_id # type: ignore + from huggingface_hub import hf_hub_download, HfFileSystem # type: ignore + from huggingface_hub.utils import validate_repo_id # type: ignore except ImportError: raise ImportError( "Llama.from_pretrained requires the huggingface-hub package. " @@ -2824,7 +2992,7 @@ def from_pretrained( files = [ file["name"] if isinstance(file, dict) else file - for file in hffs.ls(repo_id) # type: ignore + for file in hffs.ls(repo_id) # type: ignore ] # split each file into repo_id, subfolder, filename @@ -2880,6 +3048,7 @@ def from_pretrained( **kwargs, ) + class ObsidianChatHandler(Llava15ChatHandler): # Prompt Format # The model followed ChatML format. However, with ### as the seperator @@ -2906,7 +3075,6 @@ class ObsidianChatHandler(Llava15ChatHandler): "{{ message.content }}" "{% endif %}" "{% if message.content is iterable %}" - "{% for content in message.content %}" "{% if content.type == 'image_url' and content.image_url is string %}" "{{ content.image_url }}" @@ -2915,13 +3083,11 @@ class ObsidianChatHandler(Llava15ChatHandler): "{{ content.image_url.url }}" "{% endif %}" "{% endfor %}" - "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" "{% endfor %}" - "{% endif %}" "###\n" "{% endif %}" @@ -2938,6 +3104,7 @@ class ObsidianChatHandler(Llava15ChatHandler): "{% endif %}" ) + class MoondreamChatHandler(Llava15ChatHandler): # Chat Format: # f"\n\n{chat_history}Question: {question}\n\nAnswer:" @@ -2945,7 +3112,6 @@ class MoondreamChatHandler(Llava15ChatHandler): "{% for message in messages %}" "{% if message.role == 'user' %}" "{% if message.content is iterable %}" - # "{% for content in message.content %}" "{% if content.type == 'image_url' %}" @@ -2957,35 +3123,30 @@ class MoondreamChatHandler(Llava15ChatHandler): "{% endif %}" "{% endif %}" "{% endfor %}" - # Question: "{% for content in message.content %}" "{% if content.type == 'text' %}" "Question: {{ content.text }}\n\n" "{% endif %}" "{% endfor %}" - "{% endif %}" - # Question: "{% if message.content is string %}" "Question: {{ message.content }}\n\n" "{% endif %}" - "{% endif %}" - # Answer: "{% if message.role == 'assistant' %}" "Answer:{{ message.content }}\n\n" "{% endif %}" "{% endfor %}" - # Generation prompt "{% if add_generation_prompt %}" "Answer:" "{% endif %}" ) + class Llava16ChatHandler(Llava15ChatHandler): DEFAULT_SYSTEM_MESSAGE = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. " @@ -2999,7 +3160,6 @@ class Llava16ChatHandler(Llava15ChatHandler): "{% endif %}" "{% if message.role == 'user' %}" "{% if message.content is iterable %}" - # "{% for content in message.content %}" "{% if content.type == 'image_url' %}" @@ -3011,35 +3171,30 @@ class Llava16ChatHandler(Llava15ChatHandler): "{% endif %}" "{% endif %}" "{% endfor %}" - # Question: "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" "{% endfor %}" - "{% endif %}" - # Question: "{% if message.content is string %}" "{{ message.content }}" "{% endif %}" - "{% endif %}" - # Answer: "{% if message.role == 'assistant' %}" "{{ message.content }}" "{% endif %}" "{% endfor %}" - # Generation prompt "{% if add_generation_prompt %}" "Answer:" "{% endif %}" ) + class NanoLlavaChatHandler(Llava15ChatHandler): # Prompt Format # The model follow the ChatML standard, however, without \n at the end of <|im_end|>: @@ -3065,7 +3220,6 @@ class NanoLlavaChatHandler(Llava15ChatHandler): "{{ message.content }}" "{% endif %}" "{% if message.content is iterable %}" - "{% for content in message.content %}" "{% if content.type == 'image_url' and content.image_url is string %}" "{{ content.image_url }}" @@ -3074,13 +3228,11 @@ class NanoLlavaChatHandler(Llava15ChatHandler): "{{ content.image_url.url }}" "{% endif %}" "{% endfor %}" - "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" "{% endfor %}" - "{% endif %}" "<|im_end|>" "{% endif %}" @@ -3097,6 +3249,7 @@ class NanoLlavaChatHandler(Llava15ChatHandler): "{% endif %}" ) + class Llama3VisionAlphaChatHandler(Llava15ChatHandler): # question = "" + q @@ -3105,15 +3258,10 @@ class Llama3VisionAlphaChatHandler(Llava15ChatHandler): CHAT_FORMAT = ( "{% for message in messages %}" - "<|start_header_id|>" - "{% if message.role == 'user' %}" - "user<|end_header_id|>\n\n" - "{% if message.content is iterable %}" - # "{% for content in message.content %}" "{% if content.type == 'image_url' %}" @@ -3125,39 +3273,32 @@ class Llama3VisionAlphaChatHandler(Llava15ChatHandler): "{% endif %}" "{% endif %}" "{% endfor %}" - # Question: "{% for content in message.content %}" "{% if content.type == 'text' %}" "{{ content.text }}" "{% endif %}" "{% endfor %}" - "{% endif %}" - # Question: "{% if message.content is string %}" "{{ message.content }}" "{% endif %}" - "{% endif %}" - # Answer: "{% if message.role == 'assistant' %}" "assistant<|end_header_id|>\n\n" "{{ message.content }}" "{% endif %}" - "<|eot_id|>" - "{% endfor %}" - # Generation prompt "{% if add_generation_prompt %}" "<|start_header_id|>assistant<|end_header_id|>\n\n" "{% endif %}" ) + # alias Llama3VisionAlpha = Llama3VisionAlphaChatHandler @@ -3276,7 +3417,11 @@ def chatml_function_calling( }, } - stop = [stop, "<|im_end|>"] if isinstance(stop, str) else stop + ["<|im_end|>"] if stop else ["<|im_end|>"] + stop = ( + [stop, "<|im_end|>"] + if isinstance(stop, str) + else stop + ["<|im_end|>"] if stop else ["<|im_end|>"] + ) # Case 1: No tool choice by user if ( @@ -3489,7 +3634,9 @@ def chatml_function_calling( logits_processor=logits_processor, grammar=grammar, ) - completion_or_chunks = cast(llama_types.CreateCompletionResponse, completion_or_chunks) + completion_or_chunks = cast( + llama_types.CreateCompletionResponse, completion_or_chunks + ) completions.append(completion_or_chunks) completions_tool_name.append(tool_name) prompt += completion_or_chunks["choices"][0]["text"] @@ -3526,12 +3673,22 @@ def chatml_function_calling( ) # Merge completions - function_call_dict: Union[Dict[str, str], Dict[Literal["function_call"], llama_types.ChatCompletionRequestAssistantMessageFunctionCall]] = { - "function_call": { - "name": tool_name, - "arguments": completions[0]["choices"][0]["text"], + function_call_dict: Union[ + Dict[str, str], + Dict[ + Literal["function_call"], + llama_types.ChatCompletionRequestAssistantMessageFunctionCall, + ], + ] = ( + { + "function_call": { + "name": tool_name, + "arguments": completions[0]["choices"][0]["text"], + } } - } if len(completions) == 1 else {} + if len(completions) == 1 + else {} + ) return { "id": "chat" + completion["id"], "object": "chat.completion", @@ -3562,13 +3719,17 @@ def chatml_function_calling( zip(completions_tool_name, completions) ) ], - **function_call_dict + **function_call_dict, }, } ], "usage": { "completion_tokens": sum( - completion["usage"]["completion_tokens"] if "usage" in completion else 0 + ( + completion["usage"]["completion_tokens"] + if "usage" in completion + else 0 + ) for completion in completions ), "prompt_tokens": sum( diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index a0f5832c6..decc08e65 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -56,7 +56,7 @@ def _load_shared_library(lib_base_name: str): # Add the library directory to the DLL search path on Windows (if needed) if sys.platform == "win32": os.add_dll_directory(str(_base_path)) - os.environ['PATH'] = str(_base_path) + os.pathsep + os.environ['PATH'] + os.environ["PATH"] = str(_base_path) + os.pathsep + os.environ["PATH"] if sys.platform == "win32" and sys.version_info >= (3, 8): os.add_dll_directory(str(_base_path)) @@ -490,7 +490,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_POOLING_TYPE_NONE = 0 LLAMA_POOLING_TYPE_MEAN = 1 LLAMA_POOLING_TYPE_CLS = 2 -LLAMA_POOLING_TYPE_LAST = 3 +LLAMA_POOLING_TYPE_LAST = 3 # enum llama_attention_type { # LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1, @@ -812,6 +812,7 @@ class llama_model_params(ctypes.Structure): # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU # bool flash_attn; // whether to use flash attention [EXPERIMENTAL] + # // Abort callback # // if it returns true, execution of llama_decode() will be aborted # // currently works only with CPU execution @@ -1472,10 +1473,13 @@ def llama_model_has_encoder(model: llama_model_p, /) -> bool: # // For encoder-decoder models, this function returns id of the token that must be provided # // to the decoder to start generating output sequence. For other models, it returns -1. # LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); -@ctypes_function("llama_model_decoder_start_token", [llama_model_p_ctypes], ctypes.c_int32) +@ctypes_function( + "llama_model_decoder_start_token", [llama_model_p_ctypes], ctypes.c_int32 +) def llama_model_decoder_start_token(model: llama_model_p, /) -> int: """For encoder-decoder models, this function returns id of the token that must be provided - to the decoder to start generating output sequence. For other models, it returns -1.""" + to the decoder to start generating output sequence. For other models, it returns -1. + """ ... @@ -2568,7 +2572,9 @@ def llama_token_is_eog(model: llama_model_p, token: Union[llama_token, int], /) @ctypes_function( "llama_token_is_control", [llama_model_p_ctypes, llama_token], ctypes.c_bool ) -def llama_token_is_control(model: llama_model_p, token: Union[llama_token, int], /) -> bool: +def llama_token_is_control( + model: llama_model_p, token: Union[llama_token, int], / +) -> bool: """Identify if Token Id is a control token or a render-able token""" ... @@ -3364,6 +3370,7 @@ def llama_grammar_accept_token( # // Model split # // + # /// @details Build a split GGUF final path for this chunk. # /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" # // Returns the split_path length. diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index d9a3823b4..0ac7354bb 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -167,12 +167,10 @@ def __getitem__(self, index: int) -> str: return value[index] if index < len(value) else "" @overload - def __add__(self: Ptr, other: int) -> Ptr: - ... + def __add__(self: Ptr, other: int) -> Ptr: ... @overload - def __add__(self: Ptr, other: Ptr) -> int: - ... + def __add__(self: Ptr, other: Ptr) -> int: ... def __add__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: return ( @@ -182,12 +180,10 @@ def __add__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: ) @overload - def __sub__(self: Ptr, other: int) -> Ptr: - ... + def __sub__(self: Ptr, other: int) -> Ptr: ... @overload - def __sub__(self: Ptr, other: Ptr) -> int: - ... + def __sub__(self: Ptr, other: Ptr) -> int: ... def __sub__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: return ( @@ -1396,40 +1392,48 @@ def print_grammar(file: TextIO, state: parse_state) -> None: SPACE_RULE = '" "?' -def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False): +def _build_repetition( + item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False +): if not separator_rule: if min_items == 0 and max_items == 1: - return f'{item_rule}?' + return f"{item_rule}?" elif min_items == 1 and max_items is None: - return f'{item_rule}+' + return f"{item_rule}+" - result = '' + result = "" if min_items > 0: if item_rule_is_literal and separator_rule is None: result = '"' + (item_rule[1:-1] * min_items) + '"' else: - result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items) + result = (f" {separator_rule} " if separator_rule else " ").join( + [item_rule] * min_items + ) def opt_repetitions(up_to_n, prefix_with_sep=False): - ''' - - n=4, no sep: '(a (a (a (a)?)?)?)?' - - n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?' - - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?' - ''' - - content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule + """ + - n=4, no sep: '(a (a (a (a)?)?)?)?' + - n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?' + - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?' + """ + + content = ( + f"{separator_rule} {item_rule}" + if prefix_with_sep and separator_rule + else item_rule + ) if up_to_n == 0: - return '' + return "" elif up_to_n == 1: - return f'({content})?' + return f"({content})?" elif separator_rule and not prefix_with_sep: - return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?' + return f"({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?" else: - return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n) + return (f"({content} " * up_to_n).rstrip() + (")?" * up_to_n) if min_items > 0 and max_items != min_items: - result += ' ' + result += " " if max_items is not None: result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0) @@ -1437,56 +1441,81 @@ def opt_repetitions(up_to_n, prefix_with_sep=False): item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})' if min_items == 0 and separator_rule: - result = f'({item_rule} {item_operator}*)?' + result = f"({item_rule} {item_operator}*)?" else: - result += f'{item_operator}*' + result += f"{item_operator}*" return result - class BuiltinRule: def __init__(self, content: str, deps: list = None): self.content = content self.deps = deps or [] -_up_to_15_digits = _build_repetition('[0-9]', 0, 15) + +_up_to_15_digits = _build_repetition("[0-9]", 0, 15) PRIMITIVE_RULES = { - 'boolean' : BuiltinRule('("true" | "false") space', []), - 'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []), - 'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []), - 'number' : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']), - 'integer' : BuiltinRule('("-"? integral-part) space', ['integral-part']), - 'value' : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']), - 'object' : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']), - 'array' : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']), - 'uuid' : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []), - 'char' : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []), - 'string' : BuiltinRule(r'"\"" char* "\"" space', ['char']), - 'null' : BuiltinRule('"null" space', []), + "boolean": BuiltinRule('("true" | "false") space', []), + "decimal-part": BuiltinRule("[0-9] " + _up_to_15_digits, []), + "integral-part": BuiltinRule("[0-9] | [1-9] " + _up_to_15_digits, []), + "number": BuiltinRule( + '("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', + ["integral-part", "decimal-part"], + ), + "integer": BuiltinRule('("-"? integral-part) space', ["integral-part"]), + "value": BuiltinRule( + "object | array | string | number | boolean | null", + ["object", "array", "string", "number", "boolean", "null"], + ), + "object": BuiltinRule( + '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', + ["string", "value"], + ), + "array": BuiltinRule( + '"[" space ( value ("," space value)* )? "]" space', ["value"] + ), + "uuid": BuiltinRule( + r'"\"" ' + + ' "-" '.join("[0-9a-fA-F]" * n for n in [8, 4, 4, 4, 12]) + + r' "\"" space', + [], + ), + "char": BuiltinRule( + r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', + [], + ), + "string": BuiltinRule(r'"\"" char* "\"" space', ["char"]), + "null": BuiltinRule('"null" space', []), } # TODO: support "uri", "email" string formats STRING_FORMAT_RULES = { - 'date' : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []), - 'time' : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []), - 'date-time' : BuiltinRule('date "T" time', ['date', 'time']), - 'date-string' : BuiltinRule('"\\"" date "\\"" space', ['date']), - 'time-string' : BuiltinRule('"\\"" time "\\"" space', ['time']), - 'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']), + "date": BuiltinRule( + '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( "0" [1-9] | [1-2] [0-9] | "3" [0-1] )', + [], + ), + "time": BuiltinRule( + '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', + [], + ), + "date-time": BuiltinRule('date "T" time', ["date", "time"]), + "date-string": BuiltinRule('"\\"" date "\\"" space', ["date"]), + "time-string": BuiltinRule('"\\"" time "\\"" space', ["time"]), + "date-time-string": BuiltinRule('"\\"" date-time "\\"" space', ["date-time"]), } -DOTALL = '[\\U00000000-\\U0010FFFF]' -DOT = '[^\\x0A\\x0D]' - -RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()]) +DOTALL = "[\\U00000000-\\U0010FFFF]" +DOT = "[^\\x0A\\x0D]" - -NON_LITERAL_SET = set('|.()[]{}*+?') -ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?') +RESERVED_NAMES = set( + ["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()] +) +NON_LITERAL_SET = set("|.()[]{}*+?") +ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set("[]()|{}*+?") class SchemaConverter: @@ -1496,7 +1525,7 @@ def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern): self._dotall = dotall self._raw_pattern = raw_pattern self._rules = { - 'space': SPACE_RULE, + "space": SPACE_RULE, } self._refs = {} self._refs_being_resolved = set() @@ -1507,76 +1536,89 @@ def _format_literal(self, literal): ) return f'"{escaped}"' - def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str: - ''' - not_literal('a') -> '[^a]' - not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?' - ''' - assert len(literal) > 0, 'Empty literal not supported' + def not_literal( + self, literal: str, dotall: bool = True, maybe_escaped_underscores=False + ) -> str: + """ + not_literal('a') -> '[^a]' + not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?' + """ + assert len(literal) > 0, "Empty literal not supported" + def recurse(i: int): c = literal[i] - if maybe_escaped_underscores and c == '_': - yield f'[^{c}\\\\]' - yield ' | ' + if maybe_escaped_underscores and c == "_": + yield f"[^{c}\\\\]" + yield " | " yield f'"\\\\"? "{c}"' else: - yield f'[^{c}]' + yield f"[^{c}]" if i < len(literal) - 1: - yield ' | ' + yield " | " yield self._format_literal(c) - yield ' (' + yield " (" yield from recurse(i + 1) - yield ')?' + yield ")?" - return ''.join(('(', *recurse(0), ')')) + return "".join(("(", *recurse(0), ")")) def _add_rule(self, name, rule): - esc_name = INVALID_RULE_CHARS_RE.sub('-', name) + esc_name = INVALID_RULE_CHARS_RE.sub("-", name) if esc_name not in self._rules or self._rules[esc_name] == rule: key = esc_name else: i = 0 - while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule: + while ( + f"{esc_name}{i}" in self._rules + and self._rules[f"{esc_name}{i}"] != rule + ): i += 1 - key = f'{esc_name}{i}' + key = f"{esc_name}{i}" self._rules[key] = rule return key def resolve_refs(self, schema: dict, url: str): - ''' - Resolves all $ref fields in the given schema, fetching any remote schemas, - replacing $ref with absolute reference URL and populating self._refs with the - respective referenced (sub)schema dictionaries. - ''' + """ + Resolves all $ref fields in the given schema, fetching any remote schemas, + replacing $ref with absolute reference URL and populating self._refs with the + respective referenced (sub)schema dictionaries. + """ + def visit(n: dict): if isinstance(n, list): return [visit(x) for x in n] elif isinstance(n, dict): - ref = n.get('$ref') + ref = n.get("$ref") if ref is not None and ref not in self._refs: - if ref.startswith('https://'): - assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)' + if ref.startswith("https://"): + assert ( + self._allow_fetch + ), "Fetching remote schemas is not allowed (use --allow-fetch for force)" import requests - frag_split = ref.split('#') + frag_split = ref.split("#") base_url = frag_split[0] target = self._refs.get(base_url) if target is None: - target = self.resolve_refs(requests.get(ref).json(), base_url) + target = self.resolve_refs( + requests.get(ref).json(), base_url + ) self._refs[base_url] = target - if len(frag_split) == 1 or frag_split[-1] == '': + if len(frag_split) == 1 or frag_split[-1] == "": return target - elif ref.startswith('#/'): + elif ref.startswith("#/"): target = schema - ref = f'{url}{ref}' - n['$ref'] = ref + ref = f"{url}{ref}" + n["$ref"] = ref else: - raise ValueError(f'Unsupported ref {ref}') + raise ValueError(f"Unsupported ref {ref}") - for sel in ref.split('#')[-1].split('/')[1:]: - assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}' + for sel in ref.split("#")[-1].split("/")[1:]: + assert ( + target is not None and sel in target + ), f"Error resolving ref {ref}: {sel} not in {target}" target = target[sel] self._refs[ref] = target @@ -1585,28 +1627,33 @@ def visit(n: dict): visit(v) return n + return visit(schema) def _generate_union_rule(self, name, alt_schemas): - return ' | '.join(( - self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') - for i, alt_schema in enumerate(alt_schemas) - )) + return " | ".join( + ( + self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') + for i, alt_schema in enumerate(alt_schemas) + ) + ) def _visit_pattern(self, pattern, name): - ''' - Transforms a regular expression pattern into a GBNF rule. + """ + Transforms a regular expression pattern into a GBNF rule. - Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions - Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md + Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions + Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md - Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. + Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. - Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which - we define sub-rules to keep the output lean. - ''' + Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which + we define sub-rules to keep the output lean. + """ - assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"' + assert pattern.startswith("^") and pattern.endswith( + "$" + ), 'Pattern must start with "^" and end with "$"' pattern = pattern[1:-1] sub_rule_ids = {} @@ -1615,12 +1662,12 @@ def _visit_pattern(self, pattern, name): def to_rule(s: Tuple[str, bool]) -> str: (txt, is_literal) = s - return "\"" + txt + "\"" if is_literal else txt + return '"' + txt + '"' if is_literal else txt def transform() -> Tuple[str, bool]: - ''' - Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. - ''' + """ + Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. + """ nonlocal i nonlocal pattern nonlocal sub_rule_ids @@ -1638,64 +1685,72 @@ def get_dot(): else: # Accept any character... except \n and \r line break chars (\x0A and \xOD) rule = DOT - return self._add_rule(f'dot', rule) + return self._add_rule(f"dot", rule) def join_seq(): nonlocal seq ret = [] for is_literal, g in groupby(seq, lambda x: x[1]): if is_literal: - ret.append((''.join(x[0] for x in g), True)) + ret.append(("".join(x[0] for x in g), True)) else: ret.extend(g) if len(ret) == 1: return ret[0] - return (' '.join(to_rule(x) for x in seq), False) + return (" ".join(to_rule(x) for x in seq), False) while i < length: c = pattern[i] - if c == '.': + if c == ".": seq.append((get_dot(), False)) i += 1 - elif c == '(': + elif c == "(": i += 1 if i < length: - assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' - seq.append((f'({to_rule(transform())})', False)) - elif c == ')': + assert ( + pattern[i] != "?" + ), f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' + seq.append((f"({to_rule(transform())})", False)) + elif c == ")": i += 1 - assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}' + assert ( + start > 0 and pattern[start - 1] == "(" + ), f"Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}" return join_seq() - elif c == '[': + elif c == "[": square_brackets = c i += 1 - while i < length and pattern[i] != ']': - if pattern[i] == '\\': - square_brackets += pattern[i:i+2] + while i < length and pattern[i] != "]": + if pattern[i] == "\\": + square_brackets += pattern[i : i + 2] i += 2 else: square_brackets += pattern[i] i += 1 - assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}' - square_brackets += ']' + assert ( + i < length + ), f"Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}" + square_brackets += "]" i += 1 seq.append((square_brackets, False)) - elif c == '|': - seq.append(('|', False)) + elif c == "|": + seq.append(("|", False)) i += 1 - elif c in ('*', '+', '?'): + elif c in ("*", "+", "?"): seq[-1] = (to_rule(seq[-1]) + c, False) i += 1 - elif c == '{': + elif c == "{": curly_brackets = c i += 1 - while i < length and pattern[i] != '}': + while i < length and pattern[i] != "}": curly_brackets += pattern[i] i += 1 - assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}' - curly_brackets += '}' + assert ( + i < length + ), f"Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}" + curly_brackets += "}" i += 1 - nums = [s.strip() for s in curly_brackets[1:-1].split(',')] + nums = [s.strip() for s in curly_brackets[1:-1].split(",")] min_times = 0 max_times = None try: @@ -1707,35 +1762,49 @@ def join_seq(): min_times = int(nums[0]) if nums[0] else 0 max_times = int(nums[1]) if nums[1] else None except ValueError: - raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/') + raise ValueError( + f"Invalid quantifier {curly_brackets} in /{pattern}/" + ) (sub, sub_is_literal) = seq[-1] if not sub_is_literal: id = sub_rule_ids.get(sub) if id is None: - id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub) + id = self._add_rule(f"{name}-{len(sub_rule_ids) + 1}", sub) sub_rule_ids[sub] = id sub = id - seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False) + seq[-1] = ( + _build_repetition( + f'"{sub}"' if sub_is_literal else sub, + min_times, + max_times, + item_rule_is_literal=sub_is_literal, + ), + False, + ) else: - literal = '' + literal = "" while i < length: - if pattern[i] == '\\' and i < length - 1: + if pattern[i] == "\\" and i < length - 1: next = pattern[i + 1] if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS: i += 1 literal += pattern[i] i += 1 else: - literal += pattern[i:i+2] + literal += pattern[i : i + 2] i += 2 elif pattern[i] == '"' and not self._raw_pattern: literal += '\\"' i += 1 - elif pattern[i] not in NON_LITERAL_SET and \ - (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET): + elif pattern[i] not in NON_LITERAL_SET and ( + i == length - 1 + or literal == "" + or pattern[i + 1] == "." + or pattern[i + 1] not in NON_LITERAL_SET + ): literal += pattern[i] i += 1 else: @@ -1747,12 +1816,15 @@ def join_seq(): return self._add_rule( name, - to_rule(transform()) if self._raw_pattern \ - else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space") - + ( + to_rule(transform()) + if self._raw_pattern + else '"\\"" ' + to_rule(transform()) + ' "\\"" space' + ), + ) def _resolve_ref(self, ref): - ref_name = ref.split('/')[-1] + ref_name = ref.split("/")[-1] if ref_name not in self._rules and ref not in self._refs_being_resolved: self._refs_being_resolved.add(ref) resolved = self._refs[ref] @@ -1764,131 +1836,203 @@ def _generate_constant_rule(self, value): return self._format_literal(json.dumps(value)) def visit(self, schema, name): - schema_type = schema.get('type') - schema_format = schema.get('format') - rule_name = name + '-' if name in RESERVED_NAMES else name or 'root' + schema_type = schema.get("type") + schema_format = schema.get("format") + rule_name = name + "-" if name in RESERVED_NAMES else name or "root" - if (ref := schema.get('$ref')) is not None: + if (ref := schema.get("$ref")) is not None: return self._add_rule(rule_name, self._resolve_ref(ref)) - elif 'oneOf' in schema or 'anyOf' in schema: - return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf'])) + elif "oneOf" in schema or "anyOf" in schema: + return self._add_rule( + rule_name, + self._generate_union_rule(name, schema.get("oneOf") or schema["anyOf"]), + ) elif isinstance(schema_type, list): - return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type])) + return self._add_rule( + rule_name, + self._generate_union_rule(name, [{"type": t} for t in schema_type]), + ) - elif 'const' in schema: - return self._add_rule(rule_name, self._generate_constant_rule(schema['const'])) + elif "const" in schema: + return self._add_rule( + rule_name, self._generate_constant_rule(schema["const"]) + ) - elif 'enum' in schema: - rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) + elif "enum" in schema: + rule = " | ".join((self._generate_constant_rule(v) for v in schema["enum"])) return self._add_rule(rule_name, rule) - elif schema_type in (None, 'object') and \ - ('properties' in schema or \ - ('additionalProperties' in schema and schema['additionalProperties'] is not True)): - required = set(schema.get('required', [])) - properties = list(schema.get('properties', {}).items()) - return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties'))) + elif schema_type in (None, "object") and ( + "properties" in schema + or ( + "additionalProperties" in schema + and schema["additionalProperties"] is not True + ) + ): + required = set(schema.get("required", [])) + properties = list(schema.get("properties", {}).items()) + return self._add_rule( + rule_name, + self._build_object_rule( + properties, required, name, schema.get("additionalProperties") + ), + ) - elif schema_type in (None, 'object') and 'allOf' in schema: + elif schema_type in (None, "object") and "allOf" in schema: required = set() properties = [] hybrid_name = name + def add_component(comp_schema, is_required): - if (ref := comp_schema.get('$ref')) is not None: + if (ref := comp_schema.get("$ref")) is not None: comp_schema = self._refs[ref] - if 'properties' in comp_schema: - for prop_name, prop_schema in comp_schema['properties'].items(): + if "properties" in comp_schema: + for prop_name, prop_schema in comp_schema["properties"].items(): properties.append((prop_name, prop_schema)) if is_required: required.add(prop_name) - for t in schema['allOf']: - if 'anyOf' in t: - for tt in t['anyOf']: + for t in schema["allOf"]: + if "anyOf" in t: + for tt in t["anyOf"]: add_component(tt, is_required=False) else: add_component(t, is_required=True) - return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[])) + return self._add_rule( + rule_name, + self._build_object_rule( + properties, required, hybrid_name, additional_properties=[] + ), + ) - elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema): - items = schema.get('items') or schema['prefixItems'] + elif schema_type in (None, "array") and ( + "items" in schema or "prefixItems" in schema + ): + items = schema.get("items") or schema["prefixItems"] if isinstance(items, list): return self._add_rule( rule_name, - '"[" space ' + - ' "," space '.join( + '"[" space ' + + ' "," space '.join( self.visit(item, f'{name}{"-" if name else ""}tuple-{i}') - for i, item in enumerate(items)) + - ' "]" space') + for i, item in enumerate(items) + ) + + ' "]" space', + ) else: item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item') min_items = schema.get("minItems", 0) max_items = schema.get("maxItems") - return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space') + return self._add_rule( + rule_name, + '"[" space ' + + _build_repetition( + item_rule_name, min_items, max_items, separator_rule='"," space' + ) + + ' "]" space', + ) - elif schema_type in (None, 'string') and 'pattern' in schema: - return self._visit_pattern(schema['pattern'], rule_name) + elif schema_type in (None, "string") and "pattern" in schema: + return self._visit_pattern(schema["pattern"], rule_name) - elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''): + elif schema_type in (None, "string") and re.match( + r"^uuid[1-5]?$", schema_format or "" + ): return self._add_primitive( - 'root' if rule_name == 'root' else schema_format, - PRIMITIVE_RULES['uuid'] + "root" if rule_name == "root" else schema_format, + PRIMITIVE_RULES["uuid"], ) - elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES: - prim_name = f'{schema_format}-string' - return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name])) - - elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema): - char_rule = self._add_primitive('char', PRIMITIVE_RULES['char']) - min_len = schema.get('minLength', 0) - max_len = schema.get('maxLength') + elif ( + schema_type in (None, "string") + and f"{schema_format}-string" in STRING_FORMAT_RULES + ): + prim_name = f"{schema_format}-string" + return self._add_rule( + rule_name, + self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]), + ) - return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space') + elif schema_type == "string" and ( + "minLength" in schema or "maxLength" in schema + ): + char_rule = self._add_primitive("char", PRIMITIVE_RULES["char"]) + min_len = schema.get("minLength", 0) + max_len = schema.get("maxLength") + + return self._add_rule( + rule_name, + r'"\"" ' + + _build_repetition(char_rule, min_len, max_len) + + r' "\"" space', + ) - elif (schema_type == 'object') or (len(schema) == 0): - return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object'])) + elif (schema_type == "object") or (len(schema) == 0): + return self._add_rule( + rule_name, self._add_primitive("object", PRIMITIVE_RULES["object"]) + ) else: - assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}' + assert schema_type in PRIMITIVE_RULES, f"Unrecognized schema: {schema}" # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero - return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type]) + return self._add_primitive( + "root" if rule_name == "root" else schema_type, + PRIMITIVE_RULES[schema_type], + ) def _add_primitive(self, name: str, rule: BuiltinRule): n = self._add_rule(name, rule.content) for dep in rule.deps: dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep) - assert dep_rule, f'Rule {dep} not known' + assert dep_rule, f"Rule {dep} not known" if dep not in self._rules: self._add_primitive(dep, dep_rule) return n - def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]): + def _build_object_rule( + self, + properties: List[Tuple[str, Any]], + required: Set[str], + name: str, + additional_properties: Union[bool, Any], + ): prop_order = self._prop_order # sort by position in prop_order (if specified) then by original order - sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))] + sorted_props = [ + kv[0] + for _, kv in sorted( + enumerate(properties), + key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]), + ) + ] prop_kv_rule_names = {} for prop_name, prop_schema in properties: - prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}') + prop_rule_name = self.visit( + prop_schema, f'{name}{"-" if name else ""}{prop_name}' + ) prop_kv_rule_names[prop_name] = self._add_rule( f'{name}{"-" if name else ""}{prop_name}-kv', - fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}' + rf'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}', ) required_props = [k for k in sorted_props if k in required] optional_props = [k for k in sorted_props if k not in required] if additional_properties == True or isinstance(additional_properties, dict): sub_name = f'{name}{"-" if name else ""}additional' - value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value') + value_rule = self.visit( + {} if additional_properties == True else additional_properties, + f"{sub_name}-value", + ) prop_kv_rule_names["*"] = self._add_rule( - f'{sub_name}-kv', - self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}' + f"{sub_name}-kv", + self._add_primitive("string", PRIMITIVE_RULES["string"]) + + f' ":" space {value_rule}', ) optional_props.append("*") @@ -1896,51 +2040,55 @@ def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[st rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props) if optional_props: - rule += ' (' + rule += " (" if required_props: rule += ' "," space ( ' def get_recursive_refs(ks, first_is_optional): [k, *rest] = ks kv_rule_name = prop_kv_rule_names[k] - if k == '*': + if k == "*": res = self._add_rule( f'{name}{"-" if name else ""}additional-kvs', - f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*' + f'{kv_rule_name} ( "," space ' + kv_rule_name + " )*", ) elif first_is_optional: res = f'( "," space {kv_rule_name} )?' else: res = kv_rule_name if len(rest) > 0: - res += ' ' + self._add_rule( + res += " " + self._add_rule( f'{name}{"-" if name else ""}{k}-rest', - get_recursive_refs(rest, first_is_optional=True) + get_recursive_refs(rest, first_is_optional=True), ) return res - rule += ' | '.join( + rule += " | ".join( get_recursive_refs(optional_props[i:], first_is_optional=False) for i in range(len(optional_props)) ) if required_props: - rule += ' )' - rule += ' )?' + rule += " )" + rule += " )?" rule += ' "}" space' return rule def format_grammar(self): - return '\n'.join( - f'{name} ::= {rule}' + return "\n".join( + f"{name} ::= {rule}" for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0]) ) + + def json_schema_to_gbnf(schema: str, prop_order: Optional[List[str]] = None): prop_order = prop_order or [] schema = json.loads(schema) prop_order = {name: idx for idx, name in enumerate(prop_order)} - converter = SchemaConverter(prop_order=prop_order, allow_fetch=False, dotall=False, raw_pattern=False) + converter = SchemaConverter( + prop_order=prop_order, allow_fetch=False, dotall=False, raw_pattern=False + ) schema = converter.resolve_refs(schema, "stdin") converter.visit(schema, "") return converter.format_grammar() diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py index 5a8b13b7a..029bf2acc 100644 --- a/llama_cpp/llama_tokenizer.py +++ b/llama_cpp/llama_tokenizer.py @@ -17,7 +17,7 @@ def tokenize( self, text: bytes, add_bos: bool = True, special: bool = True ) -> List[int]: """Tokenize the text into tokens. - + Args: text: The text to tokenize. add_bos: Whether to add a beginning of sequence token. @@ -29,10 +29,11 @@ def detokenize( self, tokens: List[int], prev_tokens: Optional[List[int]] = None ) -> bytes: """Detokenize the tokens into text. - + Args: tokens: The tokens to detokenize. - prev_tokens: If tokens is a continuation of a previous sequence, the previous tokens.""" + prev_tokens: If tokens is a continuation of a previous sequence, the previous tokens. + """ raise NotImplementedError @@ -80,7 +81,9 @@ def detokenize( self, tokens: List[int], prev_tokens: Optional[List[int]] = None ) -> bytes: if prev_tokens is not None: - text = self.hf_tokenizer.decode(prev_tokens + tokens).encode("utf-8", errors="ignore") + text = self.hf_tokenizer.decode(prev_tokens + tokens).encode( + "utf-8", errors="ignore" + ) prev_text = self.hf_tokenizer.decode(prev_tokens).encode( "utf-8", errors="ignore" ) diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 4677785ae..573d7d651 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -6,6 +6,7 @@ https://github.com/openai/openai-openapi/blob/master/openapi.yaml """ + from typing import Any, List, Optional, Dict, Union from typing_extensions import TypedDict, NotRequired, Literal @@ -156,7 +157,9 @@ class ChatCompletionFunctionCallOption(TypedDict): class ChatCompletionRequestResponseFormat(TypedDict): type: Literal["text", "json_object"] - schema: NotRequired[JsonType] # https://docs.endpoints.anyscale.com/guides/json_mode/ + schema: NotRequired[ + JsonType + ] # https://docs.endpoints.anyscale.com/guides/json_mode/ class ChatCompletionRequestMessageContentPartText(TypedDict): diff --git a/llama_cpp/server/__main__.py b/llama_cpp/server/__main__.py index a6f1f4e9c..bbac4957e 100644 --- a/llama_cpp/server/__main__.py +++ b/llama_cpp/server/__main__.py @@ -21,6 +21,7 @@ Then visit http://localhost:8000/docs to see the interactive API docs. """ + from __future__ import annotations import os @@ -68,7 +69,9 @@ def main(): json.dumps(yaml.safe_load(f)) ) else: - config_file_settings = ConfigFileSettings.model_validate_json(f.read()) + config_file_settings = ConfigFileSettings.model_validate_json( + f.read() + ) server_settings = ServerSettings.model_validate(config_file_settings) model_settings = config_file_settings.models else: diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 8fd41a3fc..cd3255176 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -162,7 +162,9 @@ async def get_event_publisher( on_complete: typing.Optional[typing.Callable[[], None]] = None, ): server_settings = next(get_server_settings()) - interrupt_requests = server_settings.interrupt_requests if server_settings else False + interrupt_requests = ( + server_settings.interrupt_requests if server_settings else False + ) async with inner_send_chan: try: async for chunk in iterate_in_threadpool(iterator): diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index 9d9b533d5..b20655813 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -182,15 +182,17 @@ class ModelSettings(BaseSettings): default=True, description="Whether to print debug information." ) - @model_validator(mode="before") # pre=True to ensure this runs before any other validation + @model_validator( + mode="before" + ) # pre=True to ensure this runs before any other validation def set_dynamic_defaults(self) -> Self: # If n_threads or n_threads_batch is -1, set it to multiprocessing.cpu_count() cpu_count = multiprocessing.cpu_count() values = cast(Dict[str, int], self) - if values.get('n_threads', 0) == -1: - values['n_threads'] = cpu_count - if values.get('n_threads_batch', 0) == -1: - values['n_threads_batch'] = cpu_count + if values.get("n_threads", 0) == -1: + values["n_threads"] = cpu_count + if values.get("n_threads_batch", 0) == -1: + values["n_threads_batch"] = cpu_count return self diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py index a75f9e55b..fdd164456 100644 --- a/llama_cpp/server/types.py +++ b/llama_cpp/server/types.py @@ -216,7 +216,7 @@ class CreateChatCompletionRequest(BaseModel): min_tokens: int = min_tokens_field logprobs: Optional[bool] = Field( default=False, - description="Whether to output the logprobs or not. Default is True" + description="Whether to output the logprobs or not. Default is True", ) top_logprobs: Optional[int] = Field( default=None, From 08f2bb36fefdc63559da12707ee830097447f29a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 9 Jul 2024 14:06:46 -0400 Subject: [PATCH 468/624] fix(minor): Minor ruff fixes --- llama_cpp/_internals.py | 1 + llama_cpp/llama.py | 8 +++----- llama_cpp/llama_chat_format.py | 9 +++++++-- 3 files changed, 11 insertions(+), 7 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index ff378739f..d5d3b2179 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -4,6 +4,7 @@ import ctypes from typing import ( + Dict, List, Optional, Sequence, diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3669efdec..d8c2e0cdd 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -11,10 +11,11 @@ import warnings import contextlib import multiprocessing -from types import TracebackType from typing import ( + Any, List, + Literal, Optional, Union, Generator, @@ -23,14 +24,11 @@ Deque, Callable, Dict, - Type, ) from collections import deque from pathlib import Path -from llama_cpp.llama_types import List - from .llama_types import * from .llama_grammar import LlamaGrammar from .llama_cache import ( @@ -901,7 +899,7 @@ def embed( pooling_type = self.pooling_type() logits_all = pooling_type == llama_cpp.LLAMA_POOLING_TYPE_NONE - if self.context_params.embeddings == False: + if self.context_params.embeddings is False: raise RuntimeError( "Llama model must be created with embedding=True to call this method" ) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 78d83ba86..ea8d07feb 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1,6 +1,7 @@ from __future__ import annotations import os +import sys import json import ctypes import dataclasses @@ -627,6 +628,8 @@ def chat_completion_handler( json.dumps(schema), verbose=llama.verbose ) except Exception as e: + if llama.verbose: + print(str(e), file=sys.stderr) grammar = llama_grammar.LlamaGrammar.from_string( llama_grammar.JSON_GBNF, verbose=llama.verbose ) @@ -1611,12 +1614,12 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): function_call = completion_text.split(".")[-1][:-1] new_prompt = prompt + completion_text + stop elif isinstance(function_call, str) and function_call != "none": - new_prompt = prompt + f":\n" + new_prompt = prompt + ":\n" elif isinstance(function_call, dict): new_prompt = prompt + f" to=functions.{function_call['name']}:\n" function_call = function_call["name"] else: - new_prompt = prompt + f":\n" + new_prompt = prompt + ":\n" function_body = None for function in functions or []: @@ -2871,6 +2874,8 @@ def embed_image_bytes(image_bytes: bytes): json.dumps(schema), verbose=llama.verbose ) except Exception as e: + if llama.verbose: + print(str(e), file=sys.stderr) grammar = llama_grammar.LlamaGrammar.from_string( llama_grammar.JSON_GBNF, verbose=llama.verbose ) From f7f4fa8782ddb965ba394bfa7b7b42d1fabc7482 Mon Sep 17 00:00:00 2001 From: yentur <63314411+yentur@users.noreply.github.com> Date: Tue, 9 Jul 2024 21:24:06 +0300 Subject: [PATCH 469/624] feat(ci): Update simple Dockerfile (#1459) Co-authored-by: omer.yentur Co-authored-by: Andrei --- docker/simple/Dockerfile | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/docker/simple/Dockerfile b/docker/simple/Dockerfile index 97a6c19df..2838fd1ff 100644 --- a/docker/simple/Dockerfile +++ b/docker/simple/Dockerfile @@ -23,7 +23,9 @@ COPY . /app RUN python3 -m pip install --upgrade pip -RUN make deps && make build && make clean +RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context + +RUN pip install llama-cpp-python --verbose; # Set environment variable for the host ENV HOST=0.0.0.0 @@ -33,4 +35,4 @@ ENV PORT=8000 EXPOSE 8000 # Run the server start script -CMD ["/bin/sh", "/app/docker/simple/run.sh"] +CMD ["/bin/sh", "/app/run.sh"] From 7613d23db7bc3cfce7d4123f220dcfa6657236ee Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 17 Jul 2024 18:06:58 -0400 Subject: [PATCH 470/624] feat: Update llama.cpp --- llama_cpp/llama.py | 23 ++++++--- llama_cpp/llama_cpp.py | 114 ++++++++++++++++++++++++++++------------- vendor/llama.cpp | 2 +- 3 files changed, 95 insertions(+), 44 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index d8c2e0cdd..4b6f6b8f4 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -408,15 +408,24 @@ def __init__( ) ) + self._lora_adapter: Optional[llama_cpp.llama_lora_adapter_p] = None + if self.lora_path: - if self._model.apply_lora_from_file( - self.lora_path, - self.lora_scale, - self.lora_base, - self.n_threads, + assert self._model.model is not None + self._lora_adapter = llama_cpp.llama_lora_adapter_init( + self._model.model, + self.lora_path.encode("utf-8"), + ) + if self._lora_adapter is None: + raise RuntimeError( + f"Failed to initialize LoRA adapter from lora path: {self.lora_path}" + ) + assert self._ctx.ctx is not None + if llama_cpp.llama_lora_adapter_set( + self._ctx.ctx, self._lora_adapter, self.lora_scale ): raise RuntimeError( - f"Failed to apply LoRA from lora path: {self.lora_path} to base path: {self.lora_base}" + f"Failed to set LoRA adapter from lora path: {self.lora_path}" ) if self.verbose: @@ -2077,6 +2086,8 @@ def close(self) -> None: self._stack.close() def __del__(self) -> None: + if self._lora_adapter is not None: + llama_cpp.llama_lora_adapter_free(self._lora_adapter) self.close() @staticmethod diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index decc08e65..35dc030b1 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -401,7 +401,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors -# LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 +# // LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 # // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed # // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed # LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors @@ -430,14 +430,16 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors # LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors - +# LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors +# # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; LLAMA_FTYPE_ALL_F32 = 0 LLAMA_FTYPE_MOSTLY_F16 = 1 LLAMA_FTYPE_MOSTLY_Q4_0 = 2 LLAMA_FTYPE_MOSTLY_Q4_1 = 3 -LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4 LLAMA_FTYPE_MOSTLY_Q8_0 = 7 LLAMA_FTYPE_MOSTLY_Q5_0 = 8 LLAMA_FTYPE_MOSTLY_Q5_1 = 9 @@ -464,6 +466,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_MOSTLY_IQ4_XS = 30 LLAMA_FTYPE_MOSTLY_IQ1_M = 31 LLAMA_FTYPE_MOSTLY_BF16 = 32 +LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33 +LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34 +LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { @@ -1100,6 +1105,12 @@ class llama_chat_message(ctypes.Structure): ] +# // lora adapter +# struct llama_lora_adapter; +llama_lora_adapter_p = ctypes.c_void_p +llama_lora_adapter_p_ctypes = ctypes.POINTER(ctypes.c_void_p) + + # // Helpers for getting default parameters # LLAMA_API struct llama_model_params llama_model_default_params(void); @ctypes_function( @@ -1507,43 +1518,72 @@ def llama_model_quantize( ... -# // Apply a LoRA adapter to a loaded model -# // path_base_model is the path to a higher quality model to use as a base for -# // the layers modified by the adapter. Can be NULL to use the current loaded model. -# // The model needs to be reloaded before applying a new adapter, otherwise the adapter -# // will be applied on top of the previous one -# // Returns 0 on success -# LLAMA_API int32_t llama_model_apply_lora_from_file( -# const struct llama_model * model, -# const char * path_lora, -# float scale, -# const char * path_base_model, -# int32_t n_threads); -@ctypes_function( - "llama_model_apply_lora_from_file", - [ - llama_model_p_ctypes, - ctypes.c_char_p, - ctypes.c_float, - ctypes.c_char_p, - ctypes.c_int32, - ], +# // Load a LoRA adapter from file +# // The loaded adapter will be associated to the given model, and will be free when the model is deleted +# LLAMA_API struct llama_lora_adapter * llama_lora_adapter_init( +# struct llama_model * model, +# const char * path_lora); +@ctypes_function( + "llama_lora_adapter_init", + [llama_model_p_ctypes, ctypes.c_char_p], + llama_lora_adapter_p_ctypes, +) +def llama_lora_adapter_init( + model: llama_model_p, path_lora: bytes, / +) -> Optional[llama_lora_adapter_p]: + """Load a LoRA adapter from file + The loaded adapter will be associated to the given model, and will be free when the model is deleted""" + ... + + +# // Add a loaded LoRA adapter to given context +# // This will not modify model's weight +# LLAMA_API int32_t llama_lora_adapter_set( +# struct llama_context * ctx, +# struct llama_lora_adapter * adapter, +# float scale); +@ctypes_function( + "llama_lora_adapter_set", + [llama_context_p_ctypes, llama_lora_adapter_p_ctypes, ctypes.c_float], ctypes.c_int32, ) -def llama_model_apply_lora_from_file( - model: llama_model_p, - path_lora: Union[ctypes.c_char_p, bytes], - scale: Union[ctypes.c_float, float], - path_base_model: Union[ctypes.c_char_p, bytes, None], - n_threads: Union[ctypes.c_int32, int], - /, +def llama_lora_adapter_set( + ctx: llama_context_p, adapter: llama_lora_adapter_p, scale: float, / +) -> int: + """Add a loaded LoRA adapter to given context + This will not modify model's weight""" + ... + + +# // Remove a LoRA adapter from given context +# // Return -1 if the adapter is not present in the context +# LLAMA_API int32_t llama_lora_adapter_remove( +# struct llama_context * ctx, +# struct llama_lora_adapter * adapter); +@ctypes_function( + "llama_lora_adapter_remove", + [llama_context_p_ctypes, llama_lora_adapter_p_ctypes], + ctypes.c_int32, +) +def llama_lora_adapter_remove( + ctx: llama_context_p, adapter: llama_lora_adapter_p, / ) -> int: - """Apply a LoRA adapter to a loaded model - path_base_model is the path to a higher quality model to use as a base for - the layers modified by the adapter. Can be NULL to use the current loaded model. - The model needs to be reloaded before applying a new adapter, otherwise the adapter - will be applied on top of the previous one - Returns 0 on success""" + """Remove a LoRA adapter from given context + Return -1 if the adapter is not present in the context""" + ... + + +# // Manually free a LoRA adapter +# // Note: loaded adapters will be free when the associated model is deleted +# LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter); +@ctypes_function( + "llama_lora_adapter_free", + [llama_lora_adapter_p_ctypes], + None, +) +def llama_lora_adapter_free(adapter: llama_lora_adapter_p, /): + """Manually free a LoRA adapter + Note: loaded adapters will be free when the associated model is deleted""" ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 7fdb6f73e..30f80ca0b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 7fdb6f73e35605c8dbc39e9f19cd9ed84dbc87f2 +Subproject commit 30f80ca0bcee58669ada7a94244eeccc8c4807cc From 66d5cddbc31976636cf560e5b0b86e4333e85f2b Mon Sep 17 00:00:00 2001 From: Grider Date: Wed, 17 Jul 2024 23:09:28 +0100 Subject: [PATCH 471/624] fix(server): Use split_mode from model settings (#1594) Co-authored-by: Andrei --- llama_cpp/server/model.py | 1 + 1 file changed, 1 insertion(+) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index ad39c1004..c486f8885 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -223,6 +223,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: **kwargs, # Model Params n_gpu_layers=settings.n_gpu_layers, + split_mode=settings.split_mode, main_gpu=settings.main_gpu, tensor_split=settings.tensor_split, vocab_only=settings.vocab_only, From 797f54c3171a8aa56b83459651ca8ecc7b3bdf26 Mon Sep 17 00:00:00 2001 From: Eric Curtin Date: Wed, 17 Jul 2024 23:09:55 +0100 Subject: [PATCH 472/624] fix(docs): Update README.md typo (#1589) Missing ll --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 5641ccaa8..b0dfdd5b5 100644 --- a/README.md +++ b/README.md @@ -755,7 +755,7 @@ pip install -e .[all] make clean ``` -You can also test out specific commits of `lama.cpp` by checking out the desired commit in the `vendor/llama.cpp` submodule and then running `make clean` and `pip install -e .` again. Any changes in the `llama.h` API will require +You can also test out specific commits of `llama.cpp` by checking out the desired commit in the `vendor/llama.cpp` submodule and then running `make clean` and `pip install -e .` again. Any changes in the `llama.h` API will require changes to the `llama_cpp/llama_cpp.py` file to match the new API (additional changes may be required elsewhere). ## FAQ From 070047662fe6261a32eaae0b380b52119d9af54e Mon Sep 17 00:00:00 2001 From: ddh0 Date: Wed, 17 Jul 2024 20:51:16 -0500 Subject: [PATCH 473/624] fix: Change repeat_penalty to 1.0 to match llama.cpp defaults (#1590) * reapply changes after sync with main branch * change repeat_penalty to 1.0 to match llama.cpp defaults --------- Co-authored-by: Andrei --- llama_cpp/_internals.py | 2 +- llama_cpp/llama.py | 14 +++++++------- 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index d5d3b2179..dcd4e17ff 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -721,7 +721,7 @@ class _LlamaSamplingParams: typical_p: float = 1.00 temp: float = 0.80 penalty_last_n: int = 64 - penalty_repeat: float = 1.10 + penalty_repeat: float = 1.0 penalty_freq: float = 0.00 penalty_present: float = 0.00 mirostat: int = 0 diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 4b6f6b8f4..005045f5c 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -658,7 +658,7 @@ def sample( min_p: float = 0.05, typical_p: float = 1.0, temp: float = 0.80, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, frequency_penalty: float = 0.0, presence_penalty: float = 0.0, tfs_z: float = 1.0, @@ -733,7 +733,7 @@ def generate( min_p: float = 0.05, typical_p: float = 1.0, temp: float = 0.80, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, reset: bool = True, frequency_penalty: float = 0.0, presence_penalty: float = 0.0, @@ -751,7 +751,7 @@ def generate( Examples: >>> llama = Llama("models/ggml-7b.bin") >>> tokens = llama.tokenize(b"Hello, world!") - >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.1): + >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.0): ... print(llama.detokenize([token])) Args: @@ -1020,7 +1020,7 @@ def _create_completion( stop: Optional[Union[str, List[str]]] = [], frequency_penalty: float = 0.0, presence_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, top_k: int = 40, stream: bool = False, seed: Optional[int] = None, @@ -1639,7 +1639,7 @@ def create_completion( stop: Optional[Union[str, List[str]]] = [], frequency_penalty: float = 0.0, presence_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, top_k: int = 40, stream: bool = False, seed: Optional[int] = None, @@ -1736,7 +1736,7 @@ def __call__( stop: Optional[Union[str, List[str]]] = [], frequency_penalty: float = 0.0, presence_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, top_k: int = 40, stream: bool = False, seed: Optional[int] = None, @@ -1833,7 +1833,7 @@ def create_chat_completion( max_tokens: Optional[int] = None, presence_penalty: float = 0.0, frequency_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, tfs_z: float = 1.0, mirostat_mode: int = 0, mirostat_tau: float = 5.0, From 3638f7333b8ece31c61f83b0cbbfc6c1dfe06814 Mon Sep 17 00:00:00 2001 From: Michael Schock Date: Wed, 17 Jul 2024 18:52:04 -0700 Subject: [PATCH 474/624] feat: Add 'required' literal to ChatCompletionToolChoiceOption (#1597) --- llama_cpp/llama_types.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 573d7d651..bbb58afc3 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -275,7 +275,7 @@ class ChatCompletionNamedToolChoice(TypedDict): ChatCompletionToolChoiceOption = Union[ - Literal["none", "auto"], ChatCompletionNamedToolChoice + Literal["none", "auto", "required"], ChatCompletionNamedToolChoice ] From f95057a4750a14de28bcf05b7dd380475f9f9c20 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat, 20 Jul 2024 17:16:13 -0400 Subject: [PATCH 475/624] chore(deps): bump microsoft/setup-msbuild from 1.3 to 2 (#1585) Bumps [microsoft/setup-msbuild](https://github.com/microsoft/setup-msbuild) from 1.3 to 2. - [Release notes](https://github.com/microsoft/setup-msbuild/releases) - [Changelog](https://github.com/microsoft/setup-msbuild/blob/main/building-release.md) - [Commits](https://github.com/microsoft/setup-msbuild/compare/v1.3...v2) --- updated-dependencies: - dependency-name: microsoft/setup-msbuild dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 3c0267095..0733a68c5 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -45,7 +45,7 @@ jobs: steps: - name: Add MSBuild to PATH if: runner.os == 'Windows' - uses: microsoft/setup-msbuild@v1.3 + uses: microsoft/setup-msbuild@v2 with: vs-version: '[16.11,16.12)' From 5105f400f32d48b2eb4d8bac563072fb38912115 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 22 Jul 2024 15:20:12 -0400 Subject: [PATCH 476/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 10 ++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 35dc030b1..881b41b51 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -247,8 +247,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 6 -LLAMA_SESSION_VERSION = 6 +# define LLAMA_SESSION_VERSION 7 +LLAMA_SESSION_VERSION = 7 # define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ @@ -314,6 +314,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17, # LLAMA_VOCAB_PRE_TYPE_VIKING = 18, # LLAMA_VOCAB_PRE_TYPE_JAIS = 19, +# LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20, +# LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21, +# LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -335,6 +338,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17 LLAMA_VOCAB_PRE_TYPE_VIKING = 18 LLAMA_VOCAB_PRE_TYPE_JAIS = 19 +LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20 +LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21 +LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 30f80ca0b..081fe431a 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 30f80ca0bcee58669ada7a94244eeccc8c4807cc +Subproject commit 081fe431aa8fb6307145c4feb3eed4f48cab19f8 From 816d4912d9d2971198d2300a840ce4c100152502 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 22 Jul 2024 15:26:00 -0400 Subject: [PATCH 477/624] chore: Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 47077d8af..d71e1609e 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.83] + +- feat: Update llama.cpp to ggerganov/llama.cpp@081fe431aa8fb6307145c4feb3eed4f48cab19f8 +- feat: Add 'required' literal to ChatCompletionToolChoiceOption by @mjschock in #1597 +- fix: Change repeat_penalty to 1.0 to match llama.cpp defaults by @ddh0 in #1590 +- fix(docs): Update README.md typo by @ericcurtin in #1589 +- fix(server): Use split_mode from model settings by @grider-withourai in #1594 +- feat(ci): Dockerfile update base images and post-install cleanup by @Smartappli in #1530 + ## [0.2.82] - feat: Update llama.cpp to ggerganov/llama.cpp@7fdb6f73e35605c8dbc39e9f19cd9ed84dbc87f2 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 09ffb594e..65364d878 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.82" +__version__ = "0.2.83" From a14b49d64e760183504e1a5b37f4c9beb36a20ff Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 23 Jul 2024 21:26:20 -0400 Subject: [PATCH 478/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 234 ++++++++++++++++++++++++++--------------- vendor/llama.cpp | 2 +- 2 files changed, 153 insertions(+), 83 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 881b41b51..e70f2832e 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1192,7 +1192,8 @@ def llama_backend_init(): [ctypes.c_int], None, ) -def llama_numa_init(numa: int, /): ... +def llama_numa_init(numa: int, /): + ... # // Call once at the end of the program - currently only used for MPI @@ -1217,7 +1218,8 @@ def llama_backend_free(): ) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / -) -> Optional[llama_model_p]: ... +) -> Optional[llama_model_p]: + ... # LLAMA_API void llama_free_model(struct llama_model * model); @@ -1226,7 +1228,8 @@ def llama_load_model_from_file( [llama_model_p_ctypes], None, ) -def llama_free_model(model: llama_model_p, /): ... +def llama_free_model(model: llama_model_p, /): + ... # LLAMA_API struct llama_context * llama_new_context_with_model( @@ -1239,7 +1242,8 @@ def llama_free_model(model: llama_model_p, /): ... ) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / -) -> Optional[llama_context_p]: ... +) -> Optional[llama_context_p]: + ... # // Frees all allocated memory @@ -1260,87 +1264,104 @@ def llama_free(ctx: llama_context_p, /): [], ctypes.c_int64, ) -def llama_time_us() -> int: ... +def llama_time_us() -> int: + ... # LLAMA_API size_t llama_max_devices(void); @ctypes_function("llama_max_devices", [], ctypes.c_size_t) -def llama_max_devices() -> int: ... +def llama_max_devices() -> int: + ... # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) -def llama_supports_mmap() -> bool: ... +def llama_supports_mmap() -> bool: + ... # LLAMA_API bool llama_supports_mlock (void); @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) -def llama_supports_mlock() -> bool: ... +def llama_supports_mlock() -> bool: + ... # LLAMA_API bool llama_supports_gpu_offload(void); @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) -def llama_supports_gpu_offload() -> bool: ... +def llama_supports_gpu_offload() -> bool: + ... # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) -def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: + ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ctx(ctx: llama_context_p, /) -> int: ... +def llama_n_ctx(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_batch(ctx: llama_context_p, /) -> int: ... +def llama_n_batch(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); @ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ubatch(ctx: llama_context_p, /) -> int: ... +def llama_n_ubatch(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); @ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... +def llama_n_seq_max(ctx: llama_context_p, /) -> int: + ... # LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) -def llama_pooling_type(ctx: llama_context_p, /) -> int: ... +def llama_pooling_type(ctx: llama_context_p, /) -> int: + ... # LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: ... +def llama_vocab_type(model: llama_model_p, /) -> int: + ... # LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: ... +def llama_rope_type(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: ... +def llama_n_vocab(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_ctx_train(model: llama_model_p, /) -> int: ... +def llama_n_ctx_train(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_embd(model: llama_model_p, /) -> int: ... +def llama_n_embd(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_layer (const struct llama_model * model); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_layer(model: llama_model_p, /) -> int: ... +def llama_n_layer(model: llama_model_p, /) -> int: + ... # // Get the model's RoPE frequency scaling factor @@ -1538,7 +1559,8 @@ def llama_lora_adapter_init( model: llama_model_p, path_lora: bytes, / ) -> Optional[llama_lora_adapter_p]: """Load a LoRA adapter from file - The loaded adapter will be associated to the given model, and will be free when the model is deleted""" + The loaded adapter will be associated to the given model, and will be free when the model is deleted + """ ... @@ -2092,7 +2114,8 @@ def llama_state_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> bool: ... +) -> bool: + ... # LLAMA_API DEPRECATED(bool llama_load_session_file( @@ -2120,7 +2143,8 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: ... +) -> int: + ... # LLAMA_API bool llama_state_save_file( @@ -2144,7 +2168,8 @@ def llama_state_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> bool: ... +) -> bool: + ... # LLAMA_API DEPRECATED(bool llama_save_session_file( @@ -2169,7 +2194,8 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: ... +) -> int: + ... # // Get the exact size needed to copy the KV cache of a single sequence @@ -2247,7 +2273,8 @@ def llama_state_seq_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: ... +) -> int: + ... # LLAMA_API size_t llama_state_seq_load_file( @@ -2277,7 +2304,8 @@ def llama_state_seq_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: ... +) -> int: + ... # // @@ -2582,7 +2610,8 @@ def llama_get_embeddings_seq( ) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / -) -> bytes: ... +) -> bytes: + ... # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); @@ -2591,7 +2620,8 @@ def llama_token_get_text( ) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / -) -> float: ... +) -> float: + ... # LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token); @@ -2600,7 +2630,8 @@ def llama_token_get_score( ) def llama_token_get_attr( model: llama_model_p, token: Union[llama_token, int], / -) -> int: ... +) -> int: + ... # // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) @@ -2664,7 +2695,7 @@ def llama_token_nl(model: llama_model_p, /) -> int: # // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); +# LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); @ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_int32) def llama_add_bos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" @@ -2672,7 +2703,7 @@ def llama_add_bos_token(model: llama_model_p, /) -> int: # // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); +# LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); @ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_int32) def llama_add_eos_token(model: llama_model_p, /) -> int: """Returns -1 if unknown, 1 for true or 0 for false.""" @@ -2689,17 +2720,20 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: ... +def llama_token_middle(model: llama_model_p, /) -> int: + ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: ... +def llama_token_suffix(model: llama_model_p, /) -> int: + ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: ... +def llama_token_eot(model: llama_model_p, /) -> int: + ... # // @@ -2861,6 +2895,11 @@ def llama_detokenize( ... +# // +# // Chat templates +# // + + # /// Apply chat template. Inspired by hf apply_chat_template() on python. # /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" # /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template @@ -2895,7 +2934,8 @@ def llama_chat_apply_template( chat: CtypesArray[llama_chat_message], n_msg: int, /, -) -> int: ... +) -> int: + ... # // @@ -2950,6 +2990,79 @@ def llama_grammar_copy(grammar: llama_grammar_p, /) -> llama_grammar_p: ... +# /// @details Apply constraints from grammar +# LLAMA_API void llama_grammar_sample( +# const struct llama_grammar * grammar, +# const struct llama_context * ctx, +# llama_token_data_array * candidates); +@ctypes_function( + "llama_grammar_sample", + [ + llama_grammar_p, + llama_context_p_ctypes, + llama_token_data_array_p, + ], + None, +) +def llama_grammar_sample( + grammar: llama_grammar_p, + ctx: llama_context_p, + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], + /, +): + """Apply constraints from grammar""" + ... + + +# LLAMA_API DEPRECATED(void llama_sample_grammar( +# struct llama_context * ctx, +# llama_token_data_array * candidates, +# const struct llama_grammar * grammar), +# "use llama_grammar_sample instead"); +@ctypes_function( + "llama_sample_grammar", + [llama_context_p_ctypes, llama_token_data_array_p, llama_grammar_p], + None, +) +def llama_sample_grammar( + ctx: llama_context_p, + candidates: Union[ + CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ], + grammar, # type: llama_grammar_p + /, +): + """Apply constraints from grammar + + Parameters: + candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. + grammar: A grammar object containing the rules and constraints to apply to the generated text. + """ + ... + + +# /// @details Accepts the sampled token into the grammar +# LLAMA_API void llama_grammar_accept_token( +# struct llama_grammar * grammar, +# struct llama_context * ctx, +# llama_token token); +@ctypes_function( + "llama_grammar_accept_token", + [llama_grammar_p, llama_context_p_ctypes, llama_token], + None, +) +def llama_grammar_accept_token( + grammar: llama_grammar_p, + ctx: llama_context_p, + token: Union[llama_token, int], + /, +): + """Accepts the sampled token into the grammar""" + ... + + # // # // Sampling functions # // @@ -3236,33 +3349,6 @@ def llama_sample_temp( ... -# /// @details Apply constraints from grammar -# LLAMA_API void llama_sample_grammar( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# const struct llama_grammar * grammar); -@ctypes_function( - "llama_sample_grammar", - [llama_context_p_ctypes, llama_token_data_array_p, llama_grammar_p], - None, -) -def llama_sample_grammar( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - grammar, # type: llama_grammar_p - /, -): - """Apply constraints from grammar - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - grammar: A grammar object containing the rules and constraints to apply to the generated text. - """ - ... - - # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. # /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -3395,23 +3481,6 @@ def llama_sample_token( ... -# /// @details Accepts the sampled token into the grammar -# LLAMA_API void llama_grammar_accept_token( -# struct llama_context * ctx, -# struct llama_grammar * grammar, -# llama_token token); -@ctypes_function( - "llama_grammar_accept_token", - [llama_context_p_ctypes, llama_grammar_p, llama_token], - None, -) -def llama_grammar_accept_token( - ctx: llama_context_p, grammar: llama_grammar_p, token: Union[llama_token, int], / -) -> None: - """Accepts the sampled token into the grammar""" - ... - - # // # // Model split # // @@ -3533,4 +3602,5 @@ def llama_log_set( [ctypes.c_void_p, llama_context_p_ctypes], None, ) -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): ... +def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): + ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 081fe431a..b841d0740 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 081fe431aa8fb6307145c4feb3eed4f48cab19f8 +Subproject commit b841d0740855c5af1344a81f261139a45a2b39ee From dccb148580e4794879cd66ff6c5bea6af95a66e2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 28 Jul 2024 11:58:22 -0400 Subject: [PATCH 479/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 91 ++++++++++++++++++++++++++++++------------ vendor/llama.cpp | 2 +- 2 files changed, 66 insertions(+), 27 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index e70f2832e..d563d77bc 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -233,9 +233,6 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_DEFAULT_SEED 0xFFFFFFFF LLAMA_DEFAULT_SEED = 0xFFFFFFFF -# define LLAMA_MAX_RNG_STATE (64*1024) -LLAMA_MAX_RNG_STATE = 64 * 1024 - # define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla' LLAMA_FILE_MAGIC_GGLA = 0x67676C61 @@ -247,13 +244,13 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 7 -LLAMA_SESSION_VERSION = 7 +# define LLAMA_SESSION_VERSION 8 +LLAMA_SESSION_VERSION = 8 # define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ -# define LLAMA_STATE_SEQ_VERSION 1 -LLAMA_STATE_SEQ_VERSION = 1 +# define LLAMA_STATE_SEQ_VERSION 2 +LLAMA_STATE_SEQ_VERSION = 2 # struct llama_model; llama_model_p = NewType("llama_model_p", int) @@ -1583,7 +1580,7 @@ def llama_lora_adapter_set( ... -# // Remove a LoRA adapter from given context +# // Remove a specific LoRA adapter from given context # // Return -1 if the adapter is not present in the context # LLAMA_API int32_t llama_lora_adapter_remove( # struct llama_context * ctx, @@ -1601,6 +1598,19 @@ def llama_lora_adapter_remove( ... +# // Remove all LoRA adapters from given context +# LLAMA_API void llama_lora_adapter_clear( +# struct llama_context * ctx); +@ctypes_function( + "llama_lora_adapter_clear", + [llama_context_p_ctypes], + None, +) +def llama_lora_adapter_clear(ctx: llama_context_p, /): + """Remove all LoRA adapters from given context""" + ... + + # // Manually free a LoRA adapter # // Note: loaded adapters will be free when the associated model is deleted # LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter); @@ -1992,17 +2002,17 @@ def llama_kv_cache_update(ctx: llama_context_p, /): # // -# Returns the maximum size in bytes of the state (rng, logits, embedding -# and kv_cache) - will often be smaller after compacting tokens -# LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx); +# // Returns the *actual* size in bytes of the state +# // (rng, logits, embedding and kv_cache) +# // Only use when saving the state, not when restoring it, otherwise the size may be too small. +# LLAMA_API size_t llama_state_get_size(struct llama_context * ctx); @ctypes_function("llama_state_get_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_state_get_size(ctx: llama_context_p, /) -> int: - """Returns the maximum size in bytes of the state (rng, logits, embedding - and kv_cache) - will often be smaller after compacting tokens""" + """Returns the *actual* size in bytes of the state (rng, logits, embedding and kv_cache) - will often be smaller after compacting tokens""" ... -# LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx), +# LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx), # "use llama_state_get_size instead"); @ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_get_state_size(ctx: llama_context_p, /) -> int: @@ -2011,22 +2021,27 @@ def llama_get_state_size(ctx: llama_context_p, /) -> int: ... -# Copies the state to the specified destination address. -# Destination needs to have allocated enough memory. -# Returns the number of bytes copied +# // Copies the state to the specified destination address. +# // Destination needs to have allocated enough memory. +# // Returns the number of bytes copied # LLAMA_API size_t llama_state_get_data( # struct llama_context * ctx, -# uint8_t * dst); +# uint8_t * dst, +# size_t size); @ctypes_function( "llama_state_get_data", [ llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), + ctypes.c_size_t, ], ctypes.c_size_t, ) def llama_state_get_data( - ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], / + ctx: llama_context_p, + dst: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + /, ) -> int: """Copies the state to the specified destination address. Destination needs to have allocated enough memory. @@ -2059,14 +2074,18 @@ def llama_copy_state_data( # // Returns the number of bytes read # LLAMA_API size_t llama_state_set_data( # struct llama_context * ctx, -# const uint8_t * src); +# const uint8_t * src, +# size_t size); @ctypes_function( "llama_state_set_data", - [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), ctypes.c_size_t], ctypes.c_size_t, ) def llama_state_set_data( - ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], / + ctx: llama_context_p, + src: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + /, ) -> int: """Set the state reading from the specified address Returns the number of bytes read""" @@ -2216,14 +2235,24 @@ def llama_state_seq_get_size(ctx: llama_context_p, seq_id: llama_seq_id, /) -> i # LLAMA_API size_t llama_state_seq_get_data( # struct llama_context * ctx, # uint8_t * dst, +# size_t size, # llama_seq_id seq_id); @ctypes_function( "llama_state_seq_get_data", - [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), llama_seq_id], + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ctypes.c_size_t, + llama_seq_id, + ], ctypes.c_size_t, ) def llama_state_seq_get_data( - ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], seq_id: llama_seq_id, / + ctx: llama_context_p, + dst: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + seq_id: llama_seq_id, + /, ) -> int: """Copy the KV cache of a single sequence into the specified buffer""" ... @@ -2236,14 +2265,24 @@ def llama_state_seq_get_data( # LLAMA_API size_t llama_state_seq_set_data( # struct llama_context * ctx, # const uint8_t * src, +# size_t size, # llama_seq_id dest_seq_id); @ctypes_function( "llama_state_seq_set_data", - [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), llama_seq_id], + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ctypes.c_size_t, + llama_seq_id, + ], ctypes.c_size_t, ) def llama_state_seq_set_data( - ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], dest_seq_id: llama_seq_id, / + ctx: llama_context_p, + src: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + dest_seq_id: llama_seq_id, + /, ) -> int: """Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence""" ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index b841d0740..4730faca6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit b841d0740855c5af1344a81f261139a45a2b39ee +Subproject commit 4730faca618ff9cee0780580145e3cbe86f24876 From 9ed6b2727c0790a106d4be8c4d25b1d6eceff050 Mon Sep 17 00:00:00 2001 From: Abhay Nigam <36666689+mashuk999@users.noreply.github.com> Date: Sun, 28 Jul 2024 21:29:01 +0530 Subject: [PATCH 480/624] fix: Correcting run.sh filepath in Simple Docker implementation (#1626) --- docker/simple/Dockerfile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docker/simple/Dockerfile b/docker/simple/Dockerfile index 2838fd1ff..3594df1a5 100644 --- a/docker/simple/Dockerfile +++ b/docker/simple/Dockerfile @@ -35,4 +35,4 @@ ENV PORT=8000 EXPOSE 8000 # Run the server start script -CMD ["/bin/sh", "/app/run.sh"] +CMD ["/bin/sh", "/app/docker/simple/run.sh"] From 4bf3b43df95e62ddf74413a22a1df3411d206acc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 28 Jul 2024 12:06:36 -0400 Subject: [PATCH 481/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index d71e1609e..4d74c3032 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.84] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4730faca618ff9cee0780580145e3cbe86f24876 +- fix: fix: Correcting run.sh filepath in Simple Docker implementation by @mashuk999 in #1626 + ## [0.2.83] - feat: Update llama.cpp to ggerganov/llama.cpp@081fe431aa8fb6307145c4feb3eed4f48cab19f8 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 65364d878..d94a23db5 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.83" +__version__ = "0.2.84" From cffb4ec8c40e556e10b4e1168a43ba7f18bd957d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 30 Jul 2024 22:11:49 -0400 Subject: [PATCH 482/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4730faca6..268c56600 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4730faca618ff9cee0780580145e3cbe86f24876 +Subproject commit 268c5660062270a2c19a36fc655168aa287aaec2 From 53c6f329595bc3d8d242f58e0f1341d57bedfd92 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jul 2024 12:10:19 -0400 Subject: [PATCH 483/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 268c56600..398ede5ef 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 268c5660062270a2c19a36fc655168aa287aaec2 +Subproject commit 398ede5efeb07b9adf9fbda7ea63f630d476a792 From 0b1a8d8eff24595554513f0586929cba3e583bc2 Mon Sep 17 00:00:00 2001 From: "yuri@FreeBSD" Date: Wed, 31 Jul 2024 09:12:19 -0700 Subject: [PATCH 484/624] feat: FreeBSD compatibility (#1635) --- llama_cpp/llama_cpp.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d563d77bc..727195cf5 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -28,7 +28,7 @@ def _load_shared_library(lib_base_name: str): # for llamacpp) and "llama" (default name for this repo) _lib_paths: List[pathlib.Path] = [] # Determine the file extension based on the platform - if sys.platform.startswith("linux"): + if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"): _lib_paths += [ _base_path / f"lib{lib_base_name}.so", ] From 8297a0d32ea1b5d88451a1805323a0a4bde466b2 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Wed, 31 Jul 2024 18:13:00 +0200 Subject: [PATCH 485/624] fix(docker): Update Dockerfile build options from `LLAMA_` to `GGML_` (#1634) --- docker/openblas_simple/Dockerfile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docker/openblas_simple/Dockerfile b/docker/openblas_simple/Dockerfile index 804c1b902..5ee667dc0 100644 --- a/docker/openblas_simple/Dockerfile +++ b/docker/openblas_simple/Dockerfile @@ -12,7 +12,7 @@ RUN apt update && apt install -y libopenblas-dev ninja-build build-essential pkg RUN python -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context -RUN CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama_cpp_python --verbose +RUN CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama_cpp_python --verbose # Run the server CMD python3 -m llama_cpp.server From ac0217404c72a66bd47a84fa741a09cf712b8975 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Wed, 31 Jul 2024 18:13:24 +0200 Subject: [PATCH 486/624] fix(docker): Fix GGML_CUDA param (#1633) --- docker/cuda_simple/Dockerfile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docker/cuda_simple/Dockerfile b/docker/cuda_simple/Dockerfile index 79b6a5bac..0bbf20ffe 100644 --- a/docker/cuda_simple/Dockerfile +++ b/docker/cuda_simple/Dockerfile @@ -15,13 +15,13 @@ COPY . . # setting build related env vars ENV CUDA_DOCKER_ARCH=all -ENV LLAMA_CUBLAS=1 +ENV GGML_CUDA=1 # Install depencencies RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context # Install llama-cpp-python (build with cuda) -RUN CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python +RUN CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python # Run the server CMD python3 -m llama_cpp.server From 8a12c9f34680c4bd3e6c84557f386a7253b8fbbe Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Wed, 31 Jul 2024 18:13:53 +0200 Subject: [PATCH 487/624] fix(docker): Update Dockerfile BLAS options (#1632) --- docker/open_llama/Dockerfile | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/docker/open_llama/Dockerfile b/docker/open_llama/Dockerfile index 70b16dffc..f05e66ef2 100644 --- a/docker/open_llama/Dockerfile +++ b/docker/open_llama/Dockerfile @@ -20,13 +20,13 @@ RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fa # Perform the conditional installations based on the image RUN echo "Image: ${IMAGE}" && \ - if [ "${IMAGE}" = "python:3-slim-bullseye" ] ; then \ + if [ "${IMAGE}" = "python:3-slim-bookworm" ] ; then \ echo "OpenBLAS install:" && \ apt-get install -y --no-install-recommends libopenblas-dev && \ - LLAMA_OPENBLAS=1 pip install llama-cpp-python --verbose; \ + CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python --verbose; \ else \ echo "CuBLAS install:" && \ - LLAMA_CUBLAS=1 pip install llama-cpp-python --verbose; \ + CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --verbose; \ fi # Clean up apt cache From 1f0b9a2872065a89c16a4567b561ddd10f2533fe Mon Sep 17 00:00:00 2001 From: Shamit Verma Date: Wed, 31 Jul 2024 21:45:51 +0530 Subject: [PATCH 488/624] fix : Missing LoRA adapter after API change (#1630) --- llama_cpp/llama.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 005045f5c..0cb5ca2fc 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2083,11 +2083,14 @@ def pooling_type(self) -> str: def close(self) -> None: """Explicitly free the model from memory.""" - self._stack.close() + if hasattr(self,'_stack'): + if self._stack is not None: + self._stack.close() def __del__(self) -> None: - if self._lora_adapter is not None: - llama_cpp.llama_lora_adapter_free(self._lora_adapter) + if hasattr(self,'_lora_adapter'): + if self._lora_adapter is not None: + llama_cpp.llama_lora_adapter_free(self._lora_adapter) self.close() @staticmethod From f7b9e6d42981bb75106699e0712a6378c39ef92c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 31 Jul 2024 12:24:21 -0400 Subject: [PATCH 489/624] chore: Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 4d74c3032..6dc25442b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.85] + +- feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792 +- fix: Missing LoRA adapter after API change by @shamitv in #1630 +- fix(docker): Update Dockerfile BLAS options by @olivierdebauche in #1632 +- fix(docker): Fix GGML_CUDA param by @olivierdebauche in #1633 +- fix(docker): Update Dockerfile build options from `LLAMA_` to `GGML_` by @olivierdebauche in #1634 +- feat: FreeBSD compatibility by @yurivict in #1635 + ## [0.2.84] - feat: Update llama.cpp to ggerganov/llama.cpp@4730faca618ff9cee0780580145e3cbe86f24876 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index d94a23db5..97c8d8174 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.84" +__version__ = "0.2.85" From 5575fedd33141ffdeb87304da3b1225da58db76a Mon Sep 17 00:00:00 2001 From: tc-wolf <50339167+tc-wolf@users.noreply.github.com> Date: Sun, 4 Aug 2024 16:07:54 -0500 Subject: [PATCH 490/624] fix: llama_grammar_accept_token arg order (#1649) Old was: llama_grammar_accept_token(ctx, grammar, token) Now this is: llama_grammar_accept_token(grammar, ctx, token) --- llama_cpp/_internals.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index dcd4e17ff..357023e3c 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -511,7 +511,7 @@ def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: def grammar_accept_token(self, grammar: LlamaGrammar, token: int): assert self.ctx is not None assert grammar.grammar is not None - llama_cpp.llama_grammar_accept_token(self.ctx, grammar.grammar, token) + llama_cpp.llama_grammar_accept_token(grammar.grammar, self.ctx, token) def reset_timings(self): assert self.ctx is not None From dff186cc5911a03c1c2de2505ab467c419aa6821 Mon Sep 17 00:00:00 2001 From: ExtReMLapin <3909752+ExtReMLapin@users.noreply.github.com> Date: Wed, 7 Aug 2024 02:19:36 +0200 Subject: [PATCH 491/624] feat: Ported back new grammar changes from C++ to Python implementation (#1637) * Backported . (any chat) from llama.cpp * unfinished {count,optionalmax) * implemented slice function in std:vector * fixed mistake done while reading * ported https://github.com/ggerganov/llama.cpp/pull/7194 * multiple fixes, var copy * Rewrite LlamaGrammar internals in python style * bugfix --------- Co-authored-by: Andrei --- llama_cpp/llama_cpp.py | 2 +- llama_cpp/llama_grammar.py | 1652 +++++++++++++++--------------------- 2 files changed, 682 insertions(+), 972 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 727195cf5..d598cf9f1 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -3002,7 +3002,7 @@ def llama_grammar_init( n_rules: Union[ctypes.c_size_t, int], start_rule_index: Union[ctypes.c_size_t, int], /, -) -> llama_grammar_p: +) -> Optional[llama_grammar_p]: """Initialize a grammar from a set of rules.""" ... diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 0ac7354bb..e60817e1d 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1,520 +1,101 @@ """Python implementation of llama grammar parser directly translated from C++ source file in vendor/llama.cpp/common/grammar-parser.cpp.""" # flake8: noqa -from pathlib import Path import sys -from ctypes import * # type: ignore -from enum import Enum -from itertools import islice, groupby +import ctypes +import enum +import typing +import dataclasses + +from itertools import groupby from typing import ( Any, - Callable, - Dict, Set, - Generic, List, Optional, - OrderedDict, - TextIO, Tuple, - TypeVar, Union, - overload, ) import llama_cpp.llama_cpp as llama_cpp -# Type aliases -llama_grammar_element = llama_cpp.llama_grammar_element -llama_grammar_element_p = llama_cpp.llama_grammar_element_p -llama_grammar_p = llama_cpp.llama_grammar_p - -# Type variables -Ptr = TypeVar("Ptr", bound="const_char_p") -T = TypeVar("T") -U = TypeVar("U") -V = TypeVar("V") -W = TypeVar("W") - - -class Sentinel: - """Used to mark the end of a iterator of std::vector & std::map.""" - - -class LlamaGrammar: - """Keeps reference counts of all the arguments, so that they are not - garbage collected by Python.""" - - def __del__(self) -> None: - """Free the grammar pointer when the object is deleted.""" - if self.grammar is not None: - llama_cpp.llama_grammar_free(self.grammar) - self.grammar = None - - def __init__( - self, - parsed_grammar: "parse_state", - ) -> None: - """Initialize the grammar pointer from the parsed state.""" - self._grammar_rules = ( - parsed_grammar.c_rules() - ) # type: std.vector[std.vector[LlamaGrammarElement]] - self._n_rules = self._grammar_rules.size() # type: int - self._start_rule_index = parsed_grammar.symbol_ids.at("root") # type: int - self.init() - - @classmethod - def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": - """Convert a GBNF grammar to a Llama grammar.""" - parsed_grammar = parse(const_char_p(grammar)) # type: parse_state - if parsed_grammar.rules.empty(): - raise ValueError( - f"{cls.from_string.__name__}: error parsing grammar file: parsed_grammar.rules is empty" - ) - if verbose: - print(f"{cls.from_string.__name__} grammar:", file=sys.stderr) - print_grammar(sys.stderr, parsed_grammar) - print(file=sys.stderr) - return cls(parsed_grammar) - - @classmethod - def from_json_schema( - cls, - json_schema: str, - verbose: bool = True, - ) -> "LlamaGrammar": - """Convert a JSON schema to a Llama grammar.""" - return cls.from_string(json_schema_to_gbnf(json_schema), verbose=verbose) - - @classmethod - def from_file(cls, file: Union[str, Path], verbose: bool = True) -> "LlamaGrammar": - try: - with open(file) as f: - grammar = f.read() - except Exception as err: - raise Exception( - f"{cls.from_file.__name__}: error reading grammar file: {err}" - ) - - if grammar: - return cls.from_string(grammar, verbose=verbose) - - raise ValueError( - f"{cls.from_file.__name__}: error parsing grammar file: params_grammer is empty" - ) - - def init(self) -> None: - # Step 1: Convert LlamaGrammarElement to llama_grammar_element - self._element_lists = [ - [ - llama_grammar_element(c_int(elem.type.value), c_uint32(elem.value)) - for elem in subvector - ] - for subvector in self._grammar_rules - ] # type: List[List[llama_grammar_element]] - - # Step 2: Convert each list to llama_grammar_element array and get pointer - self._element_arrays = [ - (llama_grammar_element * len(sublist))(*sublist) - for sublist in self._element_lists - ] # type: List[Array[llama_grammar_element]] - - # Step 3: Get pointer of each array - self._element_array_pointers = [ - cast(subarray, llama_grammar_element_p) for subarray in self._element_arrays - ] # type: List[llama_grammar_element_p] - - # Step 4: Make array of these pointers and get its pointer - self._rules = (llama_grammar_element_p * len(self._element_array_pointers))( - *self._element_array_pointers - ) - self.grammar = llama_cpp.llama_grammar_init( - self._rules, c_size_t(self._n_rules), c_size_t(self._start_rule_index) - ) +class GrammarElementType(enum.IntEnum): + END = llama_cpp.LLAMA_GRETYPE_END + ALT = llama_cpp.LLAMA_GRETYPE_ALT + RULE_REF = llama_cpp.LLAMA_GRETYPE_RULE_REF + CHAR = llama_cpp.LLAMA_GRETYPE_CHAR + CHAR_NOT = llama_cpp.LLAMA_GRETYPE_CHAR_NOT + CHAR_RNG_UPPER = llama_cpp.LLAMA_GRETYPE_CHAR_RNG_UPPER + CHAR_ALT = llama_cpp.LLAMA_GRETYPE_CHAR_ALT + CHAR_ANY = llama_cpp.LLAMA_GRETYPE_CHAR_ANY - def reset(self) -> None: - if self.grammar is not None: - llama_cpp.llama_grammar_free(self.grammar) - self.init() +@dataclasses.dataclass +class GrammarElement: + type: GrammarElementType + value: int -class LlamaGrammarElement: - def __init__(self, type: "llama_gretype", value: int): - self.type = type - self.value = value # Unicode code point or rule ID +@dataclasses.dataclass +class ParseState: + symbol_ids: typing.Dict[str, int] = dataclasses.field(default_factory=dict) + rules: typing.List[typing.List[GrammarElement]] = dataclasses.field(default_factory=list) -class const_char_p: - """C++ implementation of const char *.""" - - def __init__(self, value: Union[str, Ptr], move: Optional[int] = None): - if isinstance(value, const_char_p): - # We're copying an existing const_char_p - self.value = value.value - self.pos = value.pos + (move or 0) - return - - # We're creating a new const_char_p - self.value = value - self.pos = move or 0 - - def __str__(self) -> str: - assert self.value is not None, "null pointer" - return self.value[self.pos :] - - def __getitem__(self, index: int) -> str: - value = str(self) - return value[index] if index < len(value) else "" - - @overload - def __add__(self: Ptr, other: int) -> Ptr: ... - - @overload - def __add__(self: Ptr, other: Ptr) -> int: ... - - def __add__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: - return ( - self.__class__(self.value, self.pos + other) - if isinstance(other, int) - else self.pos + other.pos - ) - - @overload - def __sub__(self: Ptr, other: int) -> Ptr: ... - - @overload - def __sub__(self: Ptr, other: Ptr) -> int: ... - - def __sub__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: - return ( - self.__class__(self.value, self.pos - other) - if isinstance(other, int) - else self.pos - other.pos - ) - - def __eq__(self: Ptr, other: Ptr) -> bool: - assert self.value == other.value, "comparing pointers from different strings" - return self.pos == other.pos - - def __lt__(self: Ptr, other: Ptr) -> bool: - assert self.value == other.value, "comparing pointers from different strings" - return self.pos < other.pos - - def __gt__(self: Ptr, other: Ptr) -> bool: - assert self.value == other.value, "comparing pointers from different strings" - return self.pos > other.pos - - -class std: - @staticmethod - def string(ptr: const_char_p, length: Optional[int] = None) -> str: - """C++ implementation of std::string constructor.""" - value = str(ptr) - if length is not None: - value = value[:length] - return value - - class vector(Generic[T], List[T]): - """C++ implementation of std::vector.""" - - class iterator: - def __init__(self, vector: "std.vector[T]", index: int): - self._vector = vector - self._index = index - self._version = vector._version - - def _check_version(self): - if self._version != self._vector._version: - raise RuntimeError("Iterator used after vector was modified.") - - def __iter__(self): - return self - - def __next__(self) -> T: - self._check_version() - if self._index >= self._vector.size(): - raise StopIteration - value = self._vector[self._index] - self._index += 1 - return value - - def __add__(self, value: int) -> "std.vector[T].iterator": - return self.__class__(self._vector, self._index + value) - - def __sub__(self, value: int) -> "std.vector[T].iterator": - return self.__class__(self._vector, self._index - value) - - def __init__(self): - self._version = 0 - - def modify(self): - # This is a bit of a hack to make sure iterators are invalidated - self._version += 1 - - def push_back(self, value: T) -> None: - self.modify() - self.append(value) - - def pop_back(self) -> None: - self.modify() - if not self.empty(): - self.pop() - - def back(self) -> T: - return self[-1] - - def size(self) -> int: - return len(self) - - def clear(self) -> None: - self.modify() - super().clear() - - def empty(self) -> bool: - return self.size() == 0 - - def data(self) -> "std.vector[T]": - return self - - def resize( - self, - new_size: int, - fill_value_factory: Optional[Callable[[], T]] = None, - ) -> None: - if new_size > self.size(): - if fill_value_factory is None: - raise ValueError("A fill value factory function must be provided.") - self.reserve(new_size, fill_value_factory) - elif new_size < self.size(): - self[:] = self[:new_size] - - def reserve(self, capacity: int, fill_value_factory: Callable[[], T]) -> None: - if capacity > self.size(): - fill_value = fill_value_factory() - self.extend([fill_value] * (capacity - self.size())) - - def front(self) -> T: - if not self.empty(): - return self[0] - else: - raise IndexError("Vector is empty.") - - def assign(self, count: int, value: T) -> None: - self.clear() - self.extend([value] * count) - - def insert( - self, - pos: "std.vector[T].iterator", - first: "std.vector[T].iterator", - last: "std.vector[T].iterator", - ) -> None: - self[pos._index : pos._index] = list( - islice(first._vector, first._index, last._index) - ) - - def begin(self) -> "std.vector[T].iterator": - return self.iterator(self, 0) - - def end(self) -> "std.vector[T].iterator": - return self.iterator(self, self.size()) - - class map(Generic[T, U], OrderedDict[T, U]): - """C++ implementation of std::map.""" - - class iterator(Generic[V, W]): - def __init__(self, _map: "std.map[T, U]", key: Union[T, Sentinel]): - self._map = _map - self.iter = iter(_map) - self.key = key - self._advance() - - def _sanitize_key(self) -> T: - if isinstance(self.key, Sentinel): - raise StopIteration - return self.key - - def _advance(self) -> None: - try: - while next(self.iter) != self.key: - pass - except StopIteration: - self.key = Sentinel() - - def __next__(self) -> Tuple[T, U]: - key = self._sanitize_key() - if key in self._map: - value = self._map[key] - self._advance() - return key, value - else: - raise StopIteration - - def get(self) -> Tuple[T, U]: - key = self._sanitize_key() - return key, self._map[key] - - @property - def first(self) -> T: - return self._sanitize_key() - - @property - def second(self) -> U: - return self._map[self._sanitize_key()] - - def insert( - self, key: T, value: U - ) -> Tuple["std.map[T, U].iterator[T, U]", bool]: - if key in self: - return self.iterator(self, key), False - else: - self[key] = value - return self.iterator(self, key), True - - def find(self, key: T) -> "std.map[T, U].iterator[T, U]": - if key in self: - return self.iterator(self, key) - else: - return self.end() - - def at(self, key: T) -> U: - if key in self: - return self[key] - else: - raise KeyError("The provided key is not found in the map.") - - def erase(self, iterator: "std.map[T, U].iterator[T, U]") -> None: - key = iterator.first - if key in self: - del self[key] - - def size(self) -> int: - return len(self) - - def empty(self) -> bool: - return self.size() == 0 - - def lower_bound(self, key: T) -> "std.map[T, U].iterator[T, U]": - try: - keys = sorted(list(self.keys())) # type: ignore - for k in keys: - if k >= key: - return self.iterator(self, k) - raise ValueError("No key found that is not less than the input key") - except TypeError: - raise TypeError("Keys of type T cannot be sorted.") - - def begin(self) -> "std.map[T, U].iterator[T, U]": - return self.iterator(self, next(iter(self))) - - def end(self) -> "std.map[T, U].iterator[T, U]": - return self.iterator(self, Sentinel()) - - -# // grammar element type -# enum llama_gretype { -# // end of rule definition -# LLAMA_GRETYPE_END = 0, - -# // start of alternate definition for rule -# LLAMA_GRETYPE_ALT = 1, - -# // non-terminal element: reference to rule -# LLAMA_GRETYPE_RULE_REF = 2, - -# // terminal element: character (code point) -# LLAMA_GRETYPE_CHAR = 3, - -# // inverse char(s) ([^a], [^a-b] [^abc]) -# LLAMA_GRETYPE_CHAR_NOT = 4, - -# // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to -# // be an inclusive range ([a-z]) -# LLAMA_GRETYPE_CHAR_RNG_UPPER = 5, - - -# // modifies a preceding LLAMA_GRETYPE_CHAR or -# // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) -# LLAMA_GRETYPE_CHAR_ALT = 6, -# }; -class llama_gretype(Enum): - """grammar element type""" - - LLAMA_GRETYPE_END = 0 # end of rule definition - LLAMA_GRETYPE_ALT = 1 # start of alternate definition for rule - LLAMA_GRETYPE_RULE_REF = 2 # non-terminal element: reference to rule - LLAMA_GRETYPE_CHAR = 3 # terminal element: character (code point) - LLAMA_GRETYPE_CHAR_NOT = 4 # inverse char(s) ([^a], [^a-b] [^abc]) - LLAMA_GRETYPE_CHAR_RNG_UPPER = 5 # modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to be an inclusive range ([a-z]) - LLAMA_GRETYPE_CHAR_ALT = 6 # modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) - - -# struct parse_state { -# std::map symbol_ids; -# std::vector> rules; -# std::vector c_rules(); -# }; -class parse_state: - def __init__(self): - self.symbol_ids: std.map[str, int] = std.map() - self.rules: std.vector[std.vector[LlamaGrammarElement]] = std.vector() - - # std::vector parse_state::c_rules() { - # std::vector ret; - # for (const auto & rule : rules) { - # ret.push_back(rule.data()); - # } - # return ret; - # } - def c_rules(self) -> std.vector[std.vector[LlamaGrammarElement]]: - ret = std.vector() # type: std.vector[std.vector[LlamaGrammarElement]] - for rule in self.rules: - ret.push_back(rule.data()) - return ret - - def __repr__(self) -> str: - return ( - f"parse_state(symbol_ids={len(self.symbol_ids)}, rules={len(self.rules)})" - ) +# static std::pair decode_utf8(const char * src) { +# static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; +# uint8_t first_byte = static_cast(*src); +# uint8_t highbits = first_byte >> 4; +# int len = lookup[highbits]; +# uint8_t mask = (1 << (8 - len)) - 1; +# uint32_t value = first_byte & mask; +# const char * end = src + len; // may overrun! +# const char * pos = src + 1; +# for ( ; pos < end && *pos; pos++) { +# value = (value << 6) + (static_cast(*pos) & 0x3F); +# } +# return std::make_pair(value, pos); +# } +def decode_utf8(src: str) -> typing.Tuple[int, str]: + lookup: list[int] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4] + first_byte: int = ord(src[0]) + highbits: int = first_byte >> 4 + length: int = lookup[highbits] + mask: int = (1 << (8 - length)) - 1 + value: int = first_byte & mask + end: int = min(len(src), length) # Prevent overrun + + pos: int = 1 + for pos in range(1, end): + if not src[pos]: + break + value = (value << 6) + (ord(src[pos]) & 0x3F) -# struct llama_grammar { -# const std::vector> rules; -# std::vector> stacks; -# }; -# class llama_grammar: -# def __init__( -# self, -# rules: std.vector[std.vector[llama_grammar_element]], -# stacks: std.vector[std.vector[llama_grammar_element]], -# ): -# self.rules = rules -# self.stacks = stacks + return value, src[pos:] if pos < len(src) else "" -# uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) { +# static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) { # uint32_t next_id = static_cast(state.symbol_ids.size()); -# auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id)); +# auto result = state.symbol_ids.emplace(std::string(src, len), next_id); # return result.first->second; # } -def get_symbol_id(state: parse_state, src: const_char_p, len: int) -> int: - next_id = state.symbol_ids.size() # type: int - result = state.symbol_ids.insert(std.string(src, len), next_id) - return result[0].second # type: ignore +def get_symbol_id(state: ParseState, name: str) -> int: + next_id = len(state.symbol_ids) + return state.symbol_ids.setdefault(name, next_id) -# uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) { +# static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) { # uint32_t next_id = static_cast(state.symbol_ids.size()); # state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id; # return next_id; # } -def generate_symbol_id(state: parse_state, base_name: str) -> int: - next_id = state.symbol_ids.size() # type: int - state.symbol_ids[base_name + "_" + str(next_id)] = next_id +def generate_symbol_id(state: ParseState, base_name: str) -> int: + next_id = len(state.symbol_ids) + state.symbol_ids[f"{base_name}_{next_id}"] = next_id return next_id -# void add_rule( +# static void add_rule( # parse_state & state, # uint32_t rule_id, # const std::vector & rule) { @@ -523,51 +104,27 @@ def generate_symbol_id(state: parse_state, base_name: str) -> int: # } # state.rules[rule_id] = rule; # } -def add_rule( - state: parse_state, - rule_id: int, - rule: std.vector[LlamaGrammarElement], -) -> None: - if state.rules.size() <= rule_id: - state.rules.resize( - rule_id + 1, - fill_value_factory=std.vector[LlamaGrammarElement], - ) +def add_rule(state: ParseState, rule_id: int, rule: typing.List[GrammarElement]) -> None: + if len(state.rules) <= rule_id: + state.rules.extend([[]] * (rule_id + 1 - len(state.rules))) state.rules[rule_id] = rule -# std::pair decode_utf8(const char * src) { -# static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; -# uint8_t first_byte = static_cast(*src); -# uint8_t highbits = first_byte >> 4; -# int len = lookup[highbits]; -# uint8_t mask = (1 << (8 - len)) - 1; -# uint32_t value = first_byte & mask; -# const char * end = src + len; // may overrun! -# const char * pos = src + 1; -# for ( ; pos < end && *pos; pos++) { -# value = (value << 6) + (static_cast(*pos) & 0x3F); -# } -# return std::make_pair(value, pos); +# static bool is_digit_char(char c) { +# return '0' <= c && c <= '9'; # } -def decode_utf8(src: const_char_p) -> Tuple[int, const_char_p]: - """Decodes a UTF-8 character from the source string.""" - # Get the codepoint of the first character - value = ord(src[0]) - # Move the pointer ahead one character - pos = src + 1 - - return value, pos +def is_digit_char(c: str) -> bool: + return "0" <= c <= "9" -# bool is_word_char(char c) { -# return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9'); +# static bool is_word_char(char c) { +# return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c); # } def is_word_char(c: str) -> bool: - return ("a" <= c <= "z") or ("A" <= c <= "Z") or c == "-" or ("0" <= c <= "9") + return ("a" <= c <= "z") or ("A" <= c <= "Z") or c == "-" or is_digit_char(c) -# std::pair parse_hex(const char * src, int size) { +# static std::pair parse_hex(const char * src, int size) { # const char * pos = src; # const char * end = src + size; # uint32_t value = 0; @@ -589,13 +146,12 @@ def is_word_char(c: str) -> bool: # } # return std::make_pair(value, pos); # } -def parse_hex(src: const_char_p, size: int) -> Tuple[int, const_char_p]: - pos = const_char_p(src) # type: const_char_p - end = src + size # type: const_char_p - value = 0 # type: int - while pos < end and pos[0]: +def parse_hex(src: str, size: int) -> typing.Tuple[int, str]: + pos = 0 + value = 0 + for _ in range(size): value <<= 4 - c = pos[0] # type: str + c = src[pos] if "a" <= c <= "f": value += ord(c) - ord("a") + 10 elif "A" <= c <= "F": @@ -605,12 +161,74 @@ def parse_hex(src: const_char_p, size: int) -> Tuple[int, const_char_p]: else: break pos += 1 - if pos != end: - raise RuntimeError("expecting " + str(size) + " hex chars at " + str(src)) - return (value, pos) + if pos != size: + raise ValueError(f"expecting {size} hex chars at {src}") + return value, src[pos:] + + +# static const char * parse_space(const char * src, bool newline_ok) { +# const char * pos = src; +# while (*pos == ' ' || *pos == '\t' || *pos == '#' || +# (newline_ok && (*pos == '\r' || *pos == '\n'))) { +# if (*pos == '#') { +# while (*pos && *pos != '\r' && *pos != '\n') { +# pos++; +# } +# } else { +# pos++; +# } +# } +# return pos; +# } +def parse_space(src: str, newline_ok: bool) -> str: + pos = src + while pos and (pos[0] in (' ', '\t', '#') or (newline_ok and pos[0] in ('\r', '\n'))): + if pos[0] == "#": + while pos and pos[0] not in ("\r", "\n"): + pos = pos[1:] + else: + pos = pos[1:] + return pos -# std::pair parse_char(const char * src) { +# static const char * parse_name(const char * src) { +# const char * pos = src; +# while (is_word_char(*pos)) { +# pos++; +# } +# if (pos == src) { +# throw std::runtime_error(std::string("expecting name at ") + src); +# } +# return pos; +# } +def parse_name(src: str) -> typing.Tuple[str, str]: + pos = src + while pos and is_word_char(pos[0]): + pos = pos[1:] + if pos == src: + raise ValueError(f"expecting name at {src}") + return src[:len(src) - len(pos)], pos + +# static const char * parse_int(const char * src) { +# const char * pos = src; +# while (is_digit_char(*pos)) { +# pos++; +# } +# if (pos == src) { +# throw std::runtime_error(std::string("expecting integer at ") + src); +# } +# return pos; +# } +def parse_int(src: str) -> typing.Tuple[int, str]: + pos = src + while pos and is_digit_char(pos[0]): + pos = pos[1:] + if pos == src: + raise ValueError(f"expecting integer at {src}") + return int(src[:len(src) - len(pos)]), pos + + +# static std::pair parse_char(const char * src) { # if (*src == '\\') { # switch (src[1]) { # case 'x': return parse_hex(src + 2, 2); @@ -632,273 +250,320 @@ def parse_hex(src: const_char_p, size: int) -> Tuple[int, const_char_p]: # } # throw std::runtime_error("unexpected end of input"); # } -def parse_char(src: const_char_p) -> Tuple[int, const_char_p]: +def parse_char(src: str) -> typing.Tuple[int, str]: + if not src: + raise ValueError("unexpected end of input") if src[0] == "\\": - case = src[1] # type: str - if case == "x": - return parse_hex(src + 2, 2) - elif case == "u": - return parse_hex(src + 2, 4) - elif case == "U": - return parse_hex(src + 2, 8) - elif case == "t": - return (ord("\t"), src + 2) # implicit cast - elif case == "r": - return (ord("\r"), src + 2) # implicit cast - elif case == "n": - return (ord("\n"), src + 2) # implicit cast - elif case in ("\\", '"', "[", "]"): - return (ord(case), src + 2) # implicit cast + if src[1] == "x": + return parse_hex(src[2:], 2) + elif src[1] == "u": + return parse_hex(src[2:], 4) + elif src[1] == "U": + return parse_hex(src[2:], 8) + elif src[1] == "t": + return ord("\t"), src[2:] + elif src[1] == "r": + return ord("\r"), src[2:] + elif src[1] == "n": + return ord("\n"), src[2:] + elif src[1] in ('\\', '"', '[', ']'): + return ord(src[1]), src[2:] else: - raise RuntimeError("unknown escape at " + str(src)) - elif src[0]: - return decode_utf8(src) - else: - raise RuntimeError("unexpected end of input") - - -# const char * parse_name(const char * src) { -# const char * pos = src; -# while (is_word_char(*pos)) { -# pos++; -# } -# if (pos == src) { -# throw std::runtime_error(std::string("expecting name at ") + src); -# } -# return pos; -# } -def parse_name(src: const_char_p) -> const_char_p: - pos = const_char_p(src) # type: const_char_p - while is_word_char(pos[0]): - pos += 1 - if pos == src: - raise RuntimeError("expecting name at " + str(src)) - return pos + raise ValueError(f"unknown escape at {src}") + return decode_utf8(src) - -# const char * parse_space(const char * src, bool newline_ok) { +# static const char * parse_sequence( +# parse_state & state, +# const char * src, +# const std::string & rule_name, +# std::vector & out_elements, +# bool is_nested) { +# size_t last_sym_start = out_elements.size(); # const char * pos = src; -# while (*pos == ' ' || *pos == '\t' || *pos == '#' || -# (newline_ok && (*pos == '\r' || *pos == '\n'))) { -# if (*pos == '#') { -# while (*pos && *pos != '\r' && *pos != '\n') { +# +# auto handle_repetitions = [&](int min_times, int max_times) { +# +# if (last_sym_start == out_elements.size()) { +# throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos); +# } +# +# // apply transformation to previous symbol (last_sym_start to end) according to +# // the following rewrite rules: +# // S{m,n} --> S S S (m times) S'(n-m) +# // S'(x) ::= S S'(x-1) | +# // (... n-m definitions of these S' rules ...) +# // S'(1) ::= S | +# // S{m,} --> S S S (m times) S' +# // S' ::= S S' | +# // S* --> S{0,} +# // --> S' ::= S S' | +# // S+ --> S{1,} +# // --> S S' +# // S' ::= S S' | +# // S? --> S{0,1} +# // --> S' +# // S' ::= S | +# +# std::vector previous_elements(out_elements.begin() + last_sym_start, out_elements.end()); +# if (min_times == 0) { +# out_elements.resize(last_sym_start); +# } else { +# // Repeat the previous elements (min_times - 1) times +# for (int i = 1; i < min_times; i++) { +# out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end()); +# } +# } +# +# uint32_t last_rec_rule_id = 0; +# auto n_opt = max_times < 0 ? 1 : max_times - min_times; +# +# std::vector rec_rule(previous_elements); +# for (int i = 0; i < n_opt; i++) { +# rec_rule.resize(previous_elements.size()); +# uint32_t rec_rule_id = generate_symbol_id(state, rule_name); +# if (i > 0 || max_times < 0) { +# rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id}); +# } +# rec_rule.push_back({LLAMA_GRETYPE_ALT, 0}); +# rec_rule.push_back({LLAMA_GRETYPE_END, 0}); +# add_rule(state, rec_rule_id, rec_rule); +# last_rec_rule_id = rec_rule_id; +# } +# if (n_opt > 0) { +# out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id}); +# } +# }; +# +# while (*pos) { +# if (*pos == '"') { // literal string +# pos++; +# last_sym_start = out_elements.size(); +# while (*pos != '"') { +# if (!*pos) { +# throw std::runtime_error("unexpected end of input"); +# } +# auto char_pair = parse_char(pos); +# pos = char_pair.second; +# out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first}); +# } +# pos = parse_space(pos + 1, is_nested); +# } else if (*pos == '[') { // char range(s) +# pos++; +# enum llama_gretype start_type = LLAMA_GRETYPE_CHAR; +# if (*pos == '^') { # pos++; +# start_type = LLAMA_GRETYPE_CHAR_NOT; +# } +# last_sym_start = out_elements.size(); +# while (*pos != ']') { +# if (!*pos) { +# throw std::runtime_error("unexpected end of input"); +# } +# auto char_pair = parse_char(pos); +# pos = char_pair.second; +# enum llama_gretype type = last_sym_start < out_elements.size() +# ? LLAMA_GRETYPE_CHAR_ALT +# : start_type; +# +# out_elements.push_back({type, char_pair.first}); +# if (pos[0] == '-' && pos[1] != ']') { +# if (!pos[1]) { +# throw std::runtime_error("unexpected end of input"); +# } +# auto endchar_pair = parse_char(pos + 1); +# pos = endchar_pair.second; +# out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first}); +# } +# } +# pos = parse_space(pos + 1, is_nested); +# } else if (is_word_char(*pos)) { // rule reference +# const char * name_end = parse_name(pos); +# uint32_t ref_rule_id = get_symbol_id(state, pos, name_end - pos); +# pos = parse_space(name_end, is_nested); +# last_sym_start = out_elements.size(); +# out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id}); +# } else if (*pos == '(') { // grouping +# // parse nested alternates into synthesized rule +# pos = parse_space(pos + 1, true); +# uint32_t sub_rule_id = generate_symbol_id(state, rule_name); +# pos = parse_alternates(state, pos, rule_name, sub_rule_id, true); +# last_sym_start = out_elements.size(); +# // output reference to synthesized rule +# out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); +# if (*pos != ')') { +# throw std::runtime_error(std::string("expecting ')' at ") + pos); +# } +# pos = parse_space(pos + 1, is_nested); +# } else if (*pos == '.') { // any char +# last_sym_start = out_elements.size(); +# out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0}); +# pos = parse_space(pos + 1, is_nested); +# } else if (*pos == '*') { +# pos = parse_space(pos + 1, is_nested); +# handle_repetitions(0, -1); +# } else if (*pos == '+') { +# pos = parse_space(pos + 1, is_nested); +# handle_repetitions(1, -1); +# } else if (*pos == '?') { +# pos = parse_space(pos + 1, is_nested); +# handle_repetitions(0, 1); +# } else if (*pos == '{') { +# pos = parse_space(pos + 1, is_nested); +# +# if (!is_digit_char(*pos)) { +# throw std::runtime_error(std::string("expecting an int at ") + pos); # } +# const char * int_end = parse_int(pos); +# int min_times = std::stoul(std::string(pos, int_end - pos)); +# pos = parse_space(int_end, is_nested); +# +# int max_times = -1; +# +# if (*pos == '}') { +# max_times = min_times; +# pos = parse_space(pos + 1, is_nested); +# } else if (*pos == ',') { +# pos = parse_space(pos + 1, is_nested); +# +# if (is_digit_char(*pos)) { +# const char * int_end = parse_int(pos); +# max_times = std::stoul(std::string(pos, int_end - pos)); +# pos = parse_space(int_end, is_nested); +# } +# +# if (*pos != '}') { +# throw std::runtime_error(std::string("expecting '}' at ") + pos); +# } +# pos = parse_space(pos + 1, is_nested); +# } else { +# throw std::runtime_error(std::string("expecting ',' at ") + pos); +# } +# handle_repetitions(min_times, max_times); # } else { -# pos++; +# break; # } # } # return pos; # } -def parse_space(src: const_char_p, newline_ok: bool) -> const_char_p: - pos = const_char_p(src) # type: const_char_p - while pos[0] in (" ", "\t", "#") or (newline_ok and pos[0] in ("\r", "\n")): - if pos[0] == "#": - while pos[0] is not None and pos[0] not in ("\r", "\n"): - pos += 1 - else: - pos += 1 - return pos +def parse_sequence(state: ParseState, src: str, rule_name: str, out_elements: typing.List[GrammarElement], is_nested: bool) -> str: + last_sym_start = len(out_elements) + pos = src + def handle_repetitions(min_times: int, max_times: int) -> None: + nonlocal state, src, rule_name, out_elements, is_nested, last_sym_start, pos -# const char * parse_sequence( -# parse_state & state, -# const char * src, -# const std::string & rule_name, -# std::vector & out_elements, -# bool is_nested) { -def parse_sequence( - state: parse_state, - src: const_char_p, - rule_name: str, - out_elements: std.vector[LlamaGrammarElement], - is_nested: bool, -) -> const_char_p: - # size_t last_sym_start = out_elements.size(); - # const char * pos = src; - last_sym_start = out_elements.size() # type: int - pos = const_char_p(src) # type: const_char_p - # while (*pos) { - while pos[0]: - # if (*pos == '"') { // literal string - # pos++; - # last_sym_start = out_elements.size(); - # while (*pos != '"') { - # auto char_pair = parse_char(pos); - # pos = char_pair.second; - # out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first}); - # } - # pos = parse_space(pos + 1, is_nested); - if pos[0] == '"': # literal string - pos += 1 - last_sym_start = out_elements.size() - while pos[0] != '"': - char_pair = parse_char(pos) # type: Tuple[int, const_char_p] - pos = char_pair[1] - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_CHAR, char_pair[0]) - ) - pos = parse_space(pos + 1, is_nested) - # } else if (*pos == '[') { // char range(s) - # pos++; - # enum llama_gretype start_type = LLAMA_GRETYPE_CHAR; - elif pos[0] == "[": # char range(s) - pos += 1 - start_type = llama_gretype.LLAMA_GRETYPE_CHAR # type: llama_gretype - # if (*pos == '^') { - # pos++; - # start_type = LLAMA_GRETYPE_CHAR_NOT; - # } - # last_sym_start = out_elements.size(); + if last_sym_start == len(out_elements): + raise ValueError(f"expecting preceding item to */+/?/{{ at {pos}") + + previous_elements = out_elements[last_sym_start:] + if min_times == 0: + del out_elements[last_sym_start:] + else: + for i in range(1, min_times): + out_elements.extend(previous_elements) + + last_rec_rule_id = 0 + n_opt = 1 if max_times < 0 else max_times - min_times + + rec_rule = previous_elements[:] + for i in range(n_opt): + rec_rule = rec_rule[:len(previous_elements)] + rec_rule_id = generate_symbol_id(state, rule_name) + if i > 0 or max_times < 0: + rec_rule.append(GrammarElement(GrammarElementType.RULE_REF, rec_rule_id if max_times < 0 else last_rec_rule_id)) + rec_rule.append(GrammarElement(GrammarElementType.ALT, 0)) + rec_rule.append(GrammarElement(GrammarElementType.END, 0)) + add_rule(state, rec_rule_id, rec_rule) + last_rec_rule_id = rec_rule_id + if n_opt > 0: + out_elements.append(GrammarElement(GrammarElementType.RULE_REF, last_rec_rule_id)) + + while pos: + if pos[0] == '"': + pos = pos[1:] + last_sym_start = len(out_elements) + while not pos.startswith('"'): + if not pos: + raise ValueError("unexpected end of input") + char, pos = parse_char(pos) + out_elements.append(GrammarElement(GrammarElementType.CHAR, char)) + pos = parse_space(pos[1:], is_nested) + elif pos[0] == "[": + pos = pos[1:] + start_type = GrammarElementType.CHAR if pos[0] == "^": - pos += 1 - start_type = llama_gretype.LLAMA_GRETYPE_CHAR_NOT - last_sym_start = out_elements.size() - # while (*pos != ']') { - # auto char_pair = parse_char(pos); - # pos = char_pair.second; - # enum llama_gretype type = last_sym_start < out_elements.size() - # ? LLAMA_GRETYPE_CHAR_ALT - # : start_type; - # out_elements.push_back({type, char_pair.first}); + pos = pos[1:] + start_type = GrammarElementType.CHAR_NOT + last_sym_start = len(out_elements) while pos[0] != "]": - char_pair = parse_char(pos) # type: Tuple[int, const_char_p] - pos = char_pair[1] - type = ( - llama_gretype.LLAMA_GRETYPE_CHAR_ALT - if last_sym_start < out_elements.size() - else start_type - ) # type: llama_gretype - out_elements.push_back(LlamaGrammarElement(type, char_pair[0])) - # if (pos[0] == '-' && pos[1] != ']') { - # auto endchar_pair = parse_char(pos + 1); - # pos = endchar_pair.second; - # out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first}); - # } - # } + if not pos: + raise ValueError("unexpected end of input") + char, pos = parse_char(pos) + type = GrammarElementType.CHAR_ALT if last_sym_start < len(out_elements) else start_type + out_elements.append(GrammarElement(type, char)) if pos[0] == "-" and pos[1] != "]": - endchar_pair = parse_char(pos + 1) # type: Tuple[int, const_char_p] - pos = endchar_pair[1] - out_elements.push_back( - LlamaGrammarElement( - llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER, - endchar_pair[0], - ) - ) - # pos = parse_space(pos + 1, is_nested); - pos = parse_space(pos + 1, is_nested) - # } else if (is_word_char(*pos)) { // rule reference - # const char * name_end = parse_name(pos); - # uint32_t ref_rule_id = get_symbol_id(state, pos, name_end - pos); - # pos = parse_space(name_end, is_nested); - # last_sym_start = out_elements.size(); - # out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id}); - elif is_word_char(pos[0]): # rule reference - name_end = parse_name(pos) # type: const_char_p - ref_rule_id = get_symbol_id(state, pos, name_end - pos) # type: int - pos = parse_space(name_end, is_nested) - last_sym_start = out_elements.size() - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_RULE_REF, ref_rule_id) - ) - # } else if (*pos == '(') { // grouping - # // parse nested alternates into synthesized rule - # pos = parse_space(pos + 1, true); - # uint32_t sub_rule_id = generate_symbol_id(state, rule_name); - # pos = parse_alternates(state, pos, rule_name, sub_rule_id, true); - # last_sym_start = out_elements.size(); - # // output reference to synthesized rule - # out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - # if (*pos != ')') { - # throw std::runtime_error(std::string("expecting ')' at ") + pos); - # } - # pos = parse_space(pos + 1, is_nested); - elif pos[0] == "(": # grouping - # parse nested alternates into synthesized rule - pos = parse_space(pos + 1, True) - sub_rule_id = generate_symbol_id(state, rule_name) # type: int - pos = parse_alternates(state, pos, rule_name, sub_rule_id, True) - last_sym_start = out_elements.size() - # output reference to synthesized rule - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_RULE_REF, sub_rule_id) - ) + if not pos[1]: + raise ValueError("unexpected end of input") + endchar, pos = parse_char(pos[1:]) + out_elements.append(GrammarElement(GrammarElementType.CHAR_RNG_UPPER, endchar)) + pos = parse_space(pos[1:], is_nested) + elif pos and is_word_char(pos[0]): + name, rest = parse_name(pos) + ref_rule_id = get_symbol_id(state, name) + pos = parse_space(rest, is_nested) + last_sym_start = len(out_elements) + out_elements.append(GrammarElement(GrammarElementType.RULE_REF, ref_rule_id)) + elif pos.startswith("("): + pos = parse_space(pos[1:], newline_ok=True) + sub_rule_id = generate_symbol_id(state, rule_name) + pos = parse_alternates(state, pos, rule_name, sub_rule_id, is_nested=True) + last_sym_start = len(out_elements) + out_elements.append(GrammarElement(GrammarElementType.RULE_REF, sub_rule_id)) if pos[0] != ")": - raise RuntimeError("expecting ')' at " + str(pos)) - pos = parse_space(pos + 1, is_nested) - # } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator - # if (last_sym_start == out_elements.size()) { - # throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos); - # } - elif pos[0] in ("*", "+", "?"): # repetition operator - if last_sym_start == out_elements.size(): - raise RuntimeError("expecting preceding item to */+/? at " + str(pos)) - # // apply transformation to previous symbol (last_sym_start to end) according to - # // rewrite rules: - # // S* --> S' ::= S S' | - # // S+ --> S' ::= S S' | S - # // S? --> S' ::= S | - # uint32_t sub_rule_id = generate_symbol_id(state, rule_name); - # std::vector sub_rule; - # // add preceding symbol to generated rule - # sub_rule.insert( - # sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end()); - sub_rule_id = generate_symbol_id(state, rule_name) # type: int - sub_rule = std.vector[ - LlamaGrammarElement - ]() # type: std.vector[LlamaGrammarElement] - sub_rule.insert( - sub_rule.end(), - out_elements.begin() + last_sym_start, - out_elements.end(), - ) - # if (*pos == '*' || *pos == '+') { - # // cause generated rule to recurse - # sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - # } - # // mark start of alternate def - # sub_rule.push_back({LLAMA_GRETYPE_ALT, 0}); - if pos[0] in ("*", "+"): - sub_rule.push_back( - LlamaGrammarElement( - llama_gretype.LLAMA_GRETYPE_RULE_REF, sub_rule_id - ) - ) - sub_rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_ALT, 0)) - # if (*pos == '+') { - # // add preceding symbol as alternate only for '+' (otherwise empty) - # sub_rule.insert( - # sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end()); - # } - # sub_rule.push_back({LLAMA_GRETYPE_END, 0}); - # add_rule(state, sub_rule_id, sub_rule); - # // in original rule, replace previous symbol with reference to generated rule - # out_elements.resize(last_sym_start); - # out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - # pos = parse_space(pos + 1, is_nested); - if pos[0] == "+": - # add preceding symbol as alternate only for '+' (otherwise empty) - sub_rule.insert( - sub_rule.end(), - out_elements.begin() + last_sym_start, - out_elements.end(), - ) - sub_rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_END, 0)) - add_rule(state, sub_rule_id, sub_rule) - # in original rule, replace previous symbol with reference to generated rule - out_elements.resize(last_sym_start) - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_RULE_REF, sub_rule_id) - ) - pos = parse_space(pos + 1, is_nested) - # } else { - # break; - # } + raise ValueError(f"expecting ')' at {pos}") + pos = parse_space(pos[1:], is_nested) + elif pos.startswith("."): + last_sym_start = len(out_elements) + out_elements.append(GrammarElement(GrammarElementType.CHAR_ANY, 0)) + pos = parse_space(pos[1:], is_nested) + elif pos.startswith("*"): + pos = parse_space(pos[1:], is_nested) + handle_repetitions(0, -1) + elif pos.startswith("+"): + pos = parse_space(pos[1:], is_nested) + handle_repetitions(1, -1) + elif pos.startswith("?"): + pos = parse_space(pos[1:], is_nested) + handle_repetitions(0, 1) + elif pos.startswith("{"): + pos = parse_space(pos[1:], is_nested) + + if not is_digit_char(pos): + raise ValueError(f"expecting an int at {pos}") + min_times, pos = parse_int(pos) + pos = parse_space(pos, is_nested) + + max_times = -1 + + if pos[0] == "}": + max_times = min_times + pos = parse_space(pos[1:], is_nested) + elif pos[0] == ",": + pos = parse_space(pos[1:], is_nested) + + if is_digit_char(pos): + max_times, pos = parse_int(pos) + pos = parse_space(pos, is_nested) + + if pos[0] != "}": + raise ValueError("expecting '}' at {}".format(pos)) + + pos = parse_space(pos[1:], is_nested) + else: + raise ValueError(f"expecting ',' at {pos}") + handle_repetitions(min_times, max_times) else: break - # } - # return pos; - # } return pos @@ -919,39 +584,32 @@ def parse_sequence( # add_rule(state, rule_id, rule); # return pos; # } -def parse_alternates( - state: parse_state, - src: const_char_p, - rule_name: str, - rule_id: int, - is_nested: bool, -) -> const_char_p: - rule = std.vector() # type: std.vector[LlamaGrammarElement] - pos = parse_sequence(state, src, rule_name, rule, is_nested) # type: const_char_p - while pos[0] == "|": - rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_ALT, 0)) - pos = parse_space(pos + 1, True) +def parse_alternates(state: ParseState, src: str, rule_name: str, rule_id: int, is_nested: bool) -> str: + rule = [] + pos = parse_sequence(state, src, rule_name, rule, is_nested) + while pos.startswith("|"): + rule.append(GrammarElement(GrammarElementType.ALT, 0)) + pos = parse_space(pos[1:], newline_ok=True) pos = parse_sequence(state, pos, rule_name, rule, is_nested) - rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_END, 0)) + rule.append(GrammarElement(GrammarElementType.END, 0)) add_rule(state, rule_id, rule) return pos -# const char * parse_rule(parse_state & state, const char * src) { +# static const char * parse_rule(parse_state & state, const char * src) { # const char * name_end = parse_name(src); # const char * pos = parse_space(name_end, false); # size_t name_len = name_end - src; # uint32_t rule_id = get_symbol_id(state, src, name_len); # const std::string name(src, name_len); - +# # if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) { # throw std::runtime_error(std::string("expecting ::= at ") + pos); # } # pos = parse_space(pos + 3, true); - +# # pos = parse_alternates(state, pos, name, rule_id, false); - - +# # if (*pos == '\r') { # pos += pos[1] == '\n' ? 2 : 1; # } else if (*pos == '\n') { @@ -961,26 +619,26 @@ def parse_alternates( # } # return parse_space(pos, true); # } -def parse_rule(state: parse_state, src: const_char_p) -> const_char_p: - name_end = parse_name(src) # type: const_char_p - pos = parse_space(name_end, False) # type: const_char_p - name_len = name_end - src # type: int - rule_id = get_symbol_id(state, src, name_len) # type: int - name = std.string(src, name_len) # type: str - - if not (pos[0] == ":" and pos[1] == ":" and pos[2] == "="): - raise RuntimeError("expecting ::= at " + str(pos)) - - pos = parse_space(pos + 3, True) # type: const_char_p - pos = parse_alternates(state, pos, name, rule_id, False) # type: const_char_p - - if pos[0] == "\r": - pos += 2 if pos[1] == "\n" else 1 - elif pos[0] == "\n": - pos += 1 - elif pos[0]: - raise RuntimeError("expecting newline or end at " + str(pos)) - return parse_space(pos, True) +def parse_rule(state: ParseState, src: str) -> str: + pos = src + name, pos = parse_name(pos) + pos = parse_space(pos, newline_ok=False) + rule_id = get_symbol_id(state, name) + + if not pos.startswith("::="): + raise ValueError(f"expecting ::= at {pos}") + + pos = parse_space(pos[3:], newline_ok=True) + + pos = parse_alternates(state, pos, name, rule_id, is_nested=False) + + if pos.startswith("\r"): + pos = pos[2:] if pos[1] == "\n" else pos[1:] + elif pos.startswith("\n"): + pos = pos[1:] + elif pos: + raise ValueError(f"expecting newline or end at {pos}") + return parse_space(pos, newline_ok=True) # parse_state parse(const char * src) { @@ -990,204 +648,255 @@ def parse_rule(state: parse_state, src: const_char_p) -> const_char_p: # while (*pos) { # pos = parse_rule(state, pos); # } +# // Validate the state to ensure that all rules are defined +# for (const auto & rule : state.rules) { +# for (const auto & elem : rule) { +# if (elem.type == LLAMA_GRETYPE_RULE_REF) { +# // Ensure that the rule at that location exists +# if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) { +# // Get the name of the rule that is missing +# for (const auto & kv : state.symbol_ids) { +# if (kv.second == elem.value) { +# throw std::runtime_error("Undefined rule identifier '" + kv.first + "'"); +# } +# } +# } +# } +# } +# } # return state; # } catch (const std::exception & err) { # fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what()); # return parse_state(); # } # } -def parse(src: const_char_p) -> parse_state: - try: - state = parse_state() # type: parse_state - pos = parse_space(src, True) # type: const_char_p - while pos[0]: - pos = parse_rule(state, pos) - return state - except Exception as err: - print(f"{parse.__name__}: error parsing grammar: {err}") - return parse_state() - - -# void print_grammar_char(FILE * file, uint32_t c) { -# if (0x20 <= c && c <= 0x7f) { -# fprintf(file, "%c", static_cast(c)); -# } else { -# // cop out of encoding UTF-8 -# fprintf(file, "", c); -# } -# } -def print_grammar_char(file: TextIO, c: int) -> None: - if 0x20 <= c and c <= 0x7F: - file.write(chr(c)) - else: - # cop out of encoding UTF-8 - file.write(f"") - - -# bool is_char_element(llama_grammar_element elem) { +def parse(src: str) -> ParseState: + state = ParseState() + pos = src + pos = parse_space(pos, newline_ok=True) + while pos: + pos = parse_rule(state, pos) + # validate + for rule in state.rules: + for elem in rule: + if elem.type == GrammarElementType.RULE_REF: + if elem.value >= len(state.rules) or not state.rules[elem.value]: + for k, v in state.symbol_ids.items(): + if v == elem.value: + raise ValueError(f"Undefined rule identifier '{k}'") + return state + + +# static bool is_char_element(llama_grammar_element elem) { # switch (elem.type) { # case LLAMA_GRETYPE_CHAR: return true; # case LLAMA_GRETYPE_CHAR_NOT: return true; # case LLAMA_GRETYPE_CHAR_ALT: return true; # case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true; +# case LLAMA_GRETYPE_CHAR_ANY: return true; # default: return false; # } # } -def is_char_element(elem: LlamaGrammarElement) -> bool: +def is_char_element(elem: GrammarElement) -> bool: return elem.type in ( - llama_gretype.LLAMA_GRETYPE_CHAR, - llama_gretype.LLAMA_GRETYPE_CHAR_NOT, - llama_gretype.LLAMA_GRETYPE_CHAR_ALT, - llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER, + GrammarElementType.CHAR, + GrammarElementType.CHAR_NOT, + GrammarElementType.CHAR_ALT, + GrammarElementType.CHAR_RNG_UPPER, + GrammarElementType.CHAR_ANY ) -# void print_rule( +def print_grammar_char(file: typing.TextIO, c: int) -> None: + if 0x20 <= c <= 0x7f: + print(chr(c), end="", file=file) + else: + print(f"", end="", file=file) + + +# static void print_rule( # FILE * file, # uint32_t rule_id, # const std::vector & rule, # const std::map & symbol_id_names) { +# if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) { +# throw std::runtime_error( +# "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id)); +# } +# fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str()); +# for (size_t i = 0, end = rule.size() - 1; i < end; i++) { +# llama_grammar_element elem = rule[i]; +# switch (elem.type) { +# case LLAMA_GRETYPE_END: +# throw std::runtime_error( +# "unexpected end of rule: " + std::to_string(rule_id) + "," + +# std::to_string(i)); +# case LLAMA_GRETYPE_ALT: +# fprintf(file, "| "); +# break; +# case LLAMA_GRETYPE_RULE_REF: +# fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str()); +# break; +# case LLAMA_GRETYPE_CHAR: +# fprintf(file, "["); +# print_grammar_char(file, elem.value); +# break; +# case LLAMA_GRETYPE_CHAR_NOT: +# fprintf(file, "[^"); +# print_grammar_char(file, elem.value); +# break; +# case LLAMA_GRETYPE_CHAR_RNG_UPPER: +# if (i == 0 || !is_char_element(rule[i - 1])) { +# throw std::runtime_error( +# "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " + +# std::to_string(rule_id) + "," + std::to_string(i)); +# } +# fprintf(file, "-"); +# print_grammar_char(file, elem.value); +# break; +# case LLAMA_GRETYPE_CHAR_ALT: +# if (i == 0 || !is_char_element(rule[i - 1])) { +# throw std::runtime_error( +# "LLAMA_GRETYPE_CHAR_ALT without preceding char: " + +# std::to_string(rule_id) + "," + std::to_string(i)); +# } +# print_grammar_char(file, elem.value); +# break; +# case LLAMA_GRETYPE_CHAR_ANY: +# fprintf(file, "."); +# break; +# } +# if (is_char_element(elem)) { +# switch (rule[i + 1].type) { +# case LLAMA_GRETYPE_CHAR_ALT: +# case LLAMA_GRETYPE_CHAR_RNG_UPPER: +# case LLAMA_GRETYPE_CHAR_ANY: +# break; +# default: +# fprintf(file, "] "); +# } +# } +# } +# fprintf(file, "\n"); +# } def print_rule( - file: TextIO, + file: typing.TextIO, rule_id: int, - rule: std.vector[LlamaGrammarElement], - symbol_id_names: std.map[int, str], + rule: typing.List[GrammarElement], + symbol_id_names: typing.Dict[int, str], ) -> None: - # if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) { - # throw std::runtime_error( - # "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id)); - # } - # fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str()); - if rule.empty() or rule.back().type != llama_gretype.LLAMA_GRETYPE_END: - raise RuntimeError( - "malformed rule, does not end with LLAMA_GRETYPE_END: " + str(rule_id) - ) - print(f"{symbol_id_names.at(rule_id)} ::=", file=file, end=" ") - # for (size_t i = 0, end = rule.size() - 1; i < end; i++) { - # llama_grammar_element elem = rule[i]; - # switch (elem.type) { - # case LLAMA_GRETYPE_END: - # throw std::runtime_error( - # "unexpected end of rule: " + std::to_string(rule_id) + "," + - # std::to_string(i)); - # case LLAMA_GRETYPE_ALT: - # fprintf(file, "| "); - # break; - # case LLAMA_GRETYPE_RULE_REF: - # fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str()); - # break; - # case LLAMA_GRETYPE_CHAR: - # fprintf(file, "["); - # print_grammar_char(file, elem.value); - # break; - # case LLAMA_GRETYPE_CHAR_NOT: - # fprintf(file, "[^"); - # print_grammar_char(file, elem.value); - # break; - # case LLAMA_GRETYPE_CHAR_RNG_UPPER: - # if (i == 0 || !is_char_element(rule[i - 1])) { - # throw std::runtime_error( - # "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " + - # std::to_string(rule_id) + "," + std::to_string(i)); - # } - # fprintf(file, "-"); - # print_grammar_char(file, elem.value); - # break; - # case LLAMA_GRETYPE_CHAR_ALT: - # if (i == 0 || !is_char_element(rule[i - 1])) { - # throw std::runtime_error( - # "LLAMA_GRETYPE_CHAR_ALT without preceding char: " + - # std::to_string(rule_id) + "," + std::to_string(i)); - # } - # print_grammar_char(file, elem.value); - # break; - # } + if not rule or rule[-1].type != GrammarElementType.END: + raise ValueError(f"malformed rule, does not end with LLAMA_GRETYPE_END: {rule_id}") + + print(f"{symbol_id_names[rule_id]} ::=", end=" ", file=file) + for i, elem in enumerate(rule[:-1]): - case = elem.type # type: llama_gretype - if case is llama_gretype.LLAMA_GRETYPE_END: - raise RuntimeError("unexpected end of rule: " + str(rule_id) + "," + str(i)) - elif case is llama_gretype.LLAMA_GRETYPE_ALT: - print("| ", file=file, end="") - elif case is llama_gretype.LLAMA_GRETYPE_RULE_REF: - print(f"{symbol_id_names.at(elem.value)} ", file=file, end="") - elif case is llama_gretype.LLAMA_GRETYPE_CHAR: - print("[", file=file, end="") + if elem.type == GrammarElementType.END: + raise ValueError(f"unexpected end of rule: {rule_id}, {i}") + if elem.type == GrammarElementType.ALT: + print("| ", end="", file=file) + elif elem.type == GrammarElementType.RULE_REF: + print(f"{symbol_id_names[elem.value]} ", end="", file=file) + elif elem.type == GrammarElementType.CHAR: + print("[", end="", file=file) print_grammar_char(file, elem.value) - elif case is llama_gretype.LLAMA_GRETYPE_CHAR_NOT: - print("[^", file=file, end="") + elif elem.type == GrammarElementType.CHAR_NOT: + print("[^", end="", file=file) print_grammar_char(file, elem.value) - elif case is llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER: + elif elem.type == GrammarElementType.CHAR_RNG_UPPER: if i == 0 or not is_char_element(rule[i - 1]): - raise RuntimeError( - "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " - + str(rule_id) - + "," - + str(i) - ) - print("-", file=file, end="") + raise ValueError(f"LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: {rule_id}, {i}") + print(f"-", end="", file=file) print_grammar_char(file, elem.value) - elif case is llama_gretype.LLAMA_GRETYPE_CHAR_ALT: + elif elem.type == GrammarElementType.CHAR_ALT: if i == 0 or not is_char_element(rule[i - 1]): - raise RuntimeError( - "LLAMA_GRETYPE_CHAR_ALT without preceding char: " - + str(rule_id) - + "," - + str(i) - ) + raise ValueError(f"LLAMA_GRETYPE_CHAR_ALT without preceding char: {rule_id}, {i}") print_grammar_char(file, elem.value) - # if (is_char_element(elem)) { - # switch (rule[i + 1].type) { - # case LLAMA_GRETYPE_CHAR_ALT: - # case LLAMA_GRETYPE_CHAR_RNG_UPPER: - # break; - # default: - # fprintf(file, "] "); + elif elem.type == GrammarElementType.CHAR_ANY: + print(".", end="", file=file) if is_char_element(elem): - if rule[i + 1].type in ( - llama_gretype.LLAMA_GRETYPE_CHAR_ALT, - llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER, - ): - pass - else: - print("] ", file=file, end="") - # } - # } - # } - # fprintf(file, "\n"); - # } + if rule[i + 1].type in (GrammarElementType.CHAR_ALT, GrammarElementType.CHAR_RNG_UPPER, GrammarElementType.CHAR_ANY): + continue + print("] ", end="", file=file) print(file=file) -# void print_grammar(FILE * file, const parse_state & state) { -# try { -# std::map symbol_id_names; -# for (auto kv : state.symbol_ids) { -# symbol_id_names[kv.second] = kv.first; -# } -# for (size_t i = 0, end = state.rules.size(); i < end; i++) { -# // fprintf(file, "%zu: ", i); -# // print_rule_binary(file, state.rules[i]); -# print_rule(file, i, state.rules[i], symbol_id_names); -# // fprintf(file, "\n"); -# } -# } catch (const std::exception & err) { -# fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what()); -# } -# } -def print_grammar(file: TextIO, state: parse_state) -> None: +def print_grammar(file: typing.TextIO, state: ParseState) -> None: try: - symbol_id_names = std.map() # type: std.map[int, str] - for kv in state.symbol_ids.items(): - symbol_id_names[kv[1]] = kv[0] - + symbol_id_names = {v: k for k, v in state.symbol_ids.items()} for i, rule in enumerate(state.rules): print_rule(file, i, rule, symbol_id_names) except Exception as err: - print( - f"{print_grammar.__name__}: error printing grammar: {err}", - file=sys.stderr, + print(f"\nerror printing grammar: {err}", file=file) + raise err + + +class LlamaGrammar: + def __init__(self, parse_state: ParseState): + self.parse_state = parse_state + + self._grammar_rules = parse_state.rules + self._n_rules = len(self._grammar_rules) + self._start_rule_index = parse_state.symbol_ids["root"] + + self._element_lists = [ + [ + llama_cpp.llama_grammar_element(ctypes.c_int(elem.type), ctypes.c_uint32(elem.value)) + for elem in subvector + ] + for subvector in self._grammar_rules + ] + + # Step 2: Convert each list to llama_grammar_element array and get pointer + self._element_arrays = [ + (llama_cpp.llama_grammar_element * len(sublist))(*sublist) + for sublist in self._element_lists + ] + + # Step 3: Get pointer of each array + self._element_array_pointers = [ + ctypes.cast(subarray, llama_cpp.llama_grammar_element_p) for subarray in self._element_arrays + ] + + # Step 4: Make array of these pointers and get its pointer + self._rules = (llama_cpp.llama_grammar_element_p * len(self._element_array_pointers))( + *self._element_array_pointers ) + self.grammar = None + self._init_grammar() + + + def _init_grammar(self): + grammar = llama_cpp.llama_grammar_init( + self._rules, ctypes.c_size_t(self._n_rules), ctypes.c_size_t(self._start_rule_index) + ) + + if grammar is None: + raise ValueError("Failed to create grammar") + + self.grammar = grammar + + def __del__(self): + if self.grammar is not None: + llama_cpp.llama_grammar_free(self.grammar) + self.grammar = None + + def reset(self): + if self.grammar is not None: + llama_cpp.llama_grammar_free(self.grammar) + self._init_grammar() + + @classmethod + def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": + parsed_grammar = parse(grammar) + print_grammar(file=sys.stdout, state=parsed_grammar) + return cls(parsed_grammar) + + @classmethod + def from_json_schema(cls, json_schema: str, verbose: bool = True) -> "LlamaGrammar": + return cls.from_string(json_schema_to_gbnf(json_schema), verbose=verbose) + """llama.cpp gbnf rules from vendor/llama.cpp/grammars""" @@ -1358,12 +1067,13 @@ def print_grammar(file: TextIO, state: parse_state) -> None: string ::= "\"" ( [^"\\\x7F\x00-\x1F] | - "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes + "\\" (["\\bfnrt] | "u" [0-9a-fA-F]{4}) # escapes )* "\"" ws -number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws +number ::= ("-"? ([0-9] | [1-9] [0-9]{0,15})) ("." [0-9]+)? ([eE] [-+]? [0-9] [1-9]{0,15})? ws -ws ::= ([ \t\n] ws)? +# Optional space: by convention, applied in this grammar after literal chars when allowed +ws ::= | " " | "\n" [ \t]{0,20} """ LIST_GBNF = r""" From 18f58fe39fceac821b9cf88ae40355292ade2a8e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 6 Aug 2024 20:21:18 -0400 Subject: [PATCH 492/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++-- vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d598cf9f1..643a321f5 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -944,7 +944,7 @@ class llama_context_params(ctypes.Structure): # int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() # enum llama_ftype ftype; // quantize to this llama_ftype # enum ggml_type output_tensor_type; // output tensor type -# enum ggml_type token_embedding_type; // itoken embeddings tensor type +# enum ggml_type token_embedding_type; // token embeddings tensor type # bool allow_requantize; // allow quantizing non-f32/f16 tensors # bool quantize_output_tensor; // quantize output.weight # bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored @@ -960,7 +960,7 @@ class llama_model_quantize_params(ctypes.Structure): nthread (int): number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() ftype (int): quantize to this llama_ftype output_tensor_type (int): output tensor type - token_embedding_type (int): itoken embeddings tensor type + token_embedding_type (int): token embeddings tensor type allow_requantize (bool): allow quantizing non-f32/f16 tensors quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 398ede5ef..725e3d943 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 398ede5efeb07b9adf9fbda7ea63f630d476a792 +Subproject commit 725e3d94379d5b619c027347308bccf2e0ead89f From ce6466f0eeb5c28d90b3cd26fb5a12959f45138c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 6 Aug 2024 20:23:57 -0400 Subject: [PATCH 493/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 6dc25442b..6630f1958 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.86] + +- feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792 +- feat: Ported back new grammar changes from C++ to Python implementation by @ExtReMLapin in (#1637) +- fix: llama_grammar_accept_token arg order by @tc-wolf in (#1649) + ## [0.2.85] - feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 97c8d8174..7b672c046 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.85" +__version__ = "0.2.86" From 198f47dc1bd202fd2b71b29e041a9f33fe40bfad Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 7 Aug 2024 09:33:05 -0400 Subject: [PATCH 494/624] feat(ci): Re-build wheel index automatically when releases are created --- .github/workflows/generate-index-from-release.yaml | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 500c4613c..662cb6f65 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -1,9 +1,11 @@ name: Wheels Index on: - # Trigger on any new release - release: - types: [published] + # Trigger on new release + workflow_run: + workflows: ["Release", "Build Wheels (CUDA)", "Build Wheels (Metal)"] + types: + - completed # Allows you to run this workflow manually from the Actions tab workflow_dispatch: From a07b337241264ab3b5cadf14b8fd81d1729b86e8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 7 Aug 2024 09:33:45 -0400 Subject: [PATCH 495/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 725e3d943..be55695ef 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 725e3d94379d5b619c027347308bccf2e0ead89f +Subproject commit be55695eff44784a141a863f273661a6bce63dfc From 9cad5714ae6e7c250af8d0bbb179f631368c928b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 7 Aug 2024 09:50:37 -0400 Subject: [PATCH 496/624] fix: Include all llama.cpp source files and subdirectories --- pyproject.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/pyproject.toml b/pyproject.toml index 8345cb1f0..7aebb74f5 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -60,7 +60,7 @@ wheel.packages = ["llama_cpp"] cmake.verbose = true cmake.minimum-version = "3.21" minimum-version = "0.5.1" -sdist.include = [".git", "vendor/llama.cpp/.git"] +sdist.include = [".git", "vendor/llama.cpp/*"] [tool.scikit-build.metadata.version] provider = "scikit_build_core.metadata.regex" From 84321162f8cad3e1caf552d7f0b18bc869e2cebf Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 7 Aug 2024 09:52:11 -0400 Subject: [PATCH 497/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 6630f1958..e5c7f01c1 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.87] + +- feat: Update llama.cpp to ggerganov/llama.cpp@be55695eff44784a141a863f273661a6bce63dfc +- fix: Include all llama.cpp source files and subdirectories by @abetlen in 9cad5714ae6e7c250af8d0bbb179f631368c928b +- feat(ci): Re-build wheel index automatically when releases are created by @abetlen in 198f47dc1bd202fd2b71b29e041a9f33fe40bfad + ## [0.2.86] - feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 7b672c046..37d91cd54 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.86" +__version__ = "0.2.87" From e966f3bcc1296fae7ba0d490eb12bf8ce3ced9fb Mon Sep 17 00:00:00 2001 From: Xu Song Date: Wed, 7 Aug 2024 22:41:39 +0800 Subject: [PATCH 498/624] feat: Add more detailed log for prefix-match (#1659) Co-authored-by: Andrei --- llama_cpp/llama.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 0cb5ca2fc..3fa3b563a 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -777,11 +777,12 @@ def generate( else: break if longest_prefix > 0: - if self.verbose: - print("Llama.generate: prefix-match hit", file=sys.stderr) reset = False tokens = tokens[longest_prefix:] self.n_tokens = longest_prefix + if self.verbose: + print(f"Llama.generate: {longest_prefix} prefix-match hit, " + f"remaining {len(tokens)} prompt tokens to eval", file=sys.stderr) # Reset the model state if reset: From 131db407700df32efd0ced55c75c0bd44def9b1d Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed, 7 Aug 2024 10:42:27 -0400 Subject: [PATCH 499/624] chore(deps): bump pypa/cibuildwheel from 2.19.2 to 2.20.0 (#1657) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.19.2 to 2.20.0. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.19.2...v2.20.0) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 4 ++-- .github/workflows/build-wheels-metal.yaml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index eeeda803d..785f38847 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -29,7 +29,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.19.2 + uses: pypa/cibuildwheel@v2.20.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -56,7 +56,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.19.2 + uses: pypa/cibuildwheel@v2.20.0 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index 47c1c3cb5..dd8b9c8fc 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -30,7 +30,7 @@ jobs: python -m pip install -e .[all] - name: Build wheels - uses: pypa/cibuildwheel@v2.19.2 + uses: pypa/cibuildwheel@v2.20.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" From 5e39a854631c0a8442374d47acb90310fa751501 Mon Sep 17 00:00:00 2001 From: Benedikt Heidrich Date: Wed, 7 Aug 2024 16:52:50 +0200 Subject: [PATCH 500/624] feat: Enable recursive search of HFFS.ls when using `from_pretrained` (#1656) Enable rercursive ls for hffs. Otherwise models in subfolder wouldn't be found. Co-authored-by: Andrei --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3fa3b563a..49608908b 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2160,7 +2160,7 @@ def from_pretrained( files = [ file["name"] if isinstance(file, dict) else file - for file in hffs.ls(repo_id) + for file in hffs.ls(repo_id, recursive=True)) ] # split each file into repo_id, subfolder, filename From c5de5d3e882c6ec5abf346d34cf329b4a8a35916 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Aug 2024 11:32:33 -0400 Subject: [PATCH 501/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index be55695ef..afd27f01f 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit be55695eff44784a141a863f273661a6bce63dfc +Subproject commit afd27f01fe832ece3d07ef03b7d34a9e80c4a895 From 0998ea0deea076a547d54bd598d6b413b588ee2b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Aug 2024 11:34:20 -0400 Subject: [PATCH 502/624] fix: grammar prints on each call. Closes #1666 --- llama_cpp/llama_grammar.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index e60817e1d..bc2a9cc45 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -890,7 +890,8 @@ def reset(self): @classmethod def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": parsed_grammar = parse(grammar) - print_grammar(file=sys.stdout, state=parsed_grammar) + if verbose: + print_grammar(file=sys.stdout, state=parsed_grammar) return cls(parsed_grammar) @classmethod From 7aaf7015800b98fd246b4c747e59d766318dec71 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 Aug 2024 11:35:31 -0400 Subject: [PATCH 503/624] fix: typo --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 49608908b..a37a62aef 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2160,7 +2160,7 @@ def from_pretrained( files = [ file["name"] if isinstance(file, dict) else file - for file in hffs.ls(repo_id, recursive=True)) + for file in hffs.ls(repo_id, recursive=True) ] # split each file into repo_id, subfolder, filename From 45de9d5fa6ca445a6e0ef23ef5e7d15348808ab2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 10 Aug 2024 18:33:33 -0400 Subject: [PATCH 504/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 8 ++++++++ vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 643a321f5..1a8424aa1 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1505,6 +1505,14 @@ def llama_model_has_encoder(model: llama_model_p, /) -> bool: ... +# // Returns true if the model contains a decoder that requires llama_decode() call +# LLAMA_API bool llama_model_has_decoder(const struct llama_model * model); +@ctypes_function("llama_model_has_decoder", [llama_model_p_ctypes], ctypes.c_bool) +def llama_model_has_decoder(model: llama_model_p, /) -> bool: + """Returns true if the model contains a decoder that requires llama_decode() call""" + ... + + # // For encoder-decoder models, this function returns id of the token that must be provided # // to the decoder to start generating output sequence. For other models, it returns -1. # LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); diff --git a/vendor/llama.cpp b/vendor/llama.cpp index afd27f01f..6e02327e8 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit afd27f01fe832ece3d07ef03b7d34a9e80c4a895 +Subproject commit 6e02327e8b7837358e0406bf90a4632e18e27846 From 4244151d7a27a29d8ac56dfbdc7ae10b68defaa5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 12 Aug 2024 02:29:30 -0400 Subject: [PATCH 505/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6e02327e8..4134999e0 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6e02327e8b7837358e0406bf90a4632e18e27846 +Subproject commit 4134999e01f31256b15342b41c4de9e2477c4a6c From 95a1533172b7d31bf2fa9a54b4f2a9776fbd4c45 Mon Sep 17 00:00:00 2001 From: ExtReMLapin <3909752+ExtReMLapin@users.noreply.github.com> Date: Mon, 12 Aug 2024 20:12:54 +0200 Subject: [PATCH 506/624] fix: Added back from_file method to LlamaGrammar (#1673) --- llama_cpp/llama_grammar.py | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index bc2a9cc45..4cd52c2d5 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -1,6 +1,7 @@ """Python implementation of llama grammar parser directly translated from C++ source file in vendor/llama.cpp/common/grammar-parser.cpp.""" # flake8: noqa +from pathlib import Path import sys import ctypes import enum @@ -893,6 +894,23 @@ def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": if verbose: print_grammar(file=sys.stdout, state=parsed_grammar) return cls(parsed_grammar) + + @classmethod + def from_file(cls, file: Union[str, Path], verbose: bool = True) -> "LlamaGrammar": + try: + with open(file) as f: + grammar = f.read() + except Exception as err: + raise Exception( + f"{cls.from_file.__name__}: error reading grammar file: {err}" + ) + + if grammar: + return cls.from_string(grammar, verbose=verbose) + + raise ValueError( + f"{cls.from_file.__name__}: error parsing grammar file: params_grammer is empty" + ) @classmethod def from_json_schema(cls, json_schema: str, verbose: bool = True) -> "LlamaGrammar": From 9bab46f643570a29d356337eb5ab995af8862aea Mon Sep 17 00:00:00 2001 From: Laurent Sorber Date: Mon, 12 Aug 2024 20:19:26 +0200 Subject: [PATCH 507/624] fix: only print 'cache saved' in verbose mode (#1668) Co-authored-by: Andrei --- llama_cpp/llama.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index a37a62aef..b780ef6c6 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1523,7 +1523,8 @@ def logit_bias_processor( if self.verbose: print("Llama._create_completion: cache save", file=sys.stderr) self.cache[prompt_tokens + completion_tokens] = self.save_state() - print("Llama._create_completion: cache saved", file=sys.stderr) + if self.verbose: + print("Llama._create_completion: cache saved", file=sys.stderr) return if self.cache: From 8ed663b867e6ba56dc03096984675dd35667f2b7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 12 Aug 2024 14:19:56 -0400 Subject: [PATCH 508/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4134999e0..fc4ca27b2 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4134999e01f31256b15342b41c4de9e2477c4a6c +Subproject commit fc4ca27b25464a11b3b86c9dbb5b6ed6065965c2 From fc19cc783802ed95eb4c8be358537b12c1fec78f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 13 Aug 2024 05:22:04 -0400 Subject: [PATCH 509/624] chore: Bump version --- CHANGELOG.md | 9 +++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index e5c7f01c1..f7738e1a9 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.88] + +- feat: Update llama.cpp to ggerganov/llama.cpp@fc4ca27b25464a11b3b86c9dbb5b6ed6065965c2 +- fix: only print 'cache saved' in verbose mode by @lsorber in #1668 +- fix: Added back from_file method to LlamaGrammar by @ExtReMLapin in #1673 +- fix: grammar prints on each call by @abetlen in 0998ea0deea076a547d54bd598d6b413b588ee2b +- feat: Enable recursive search of HFFS.ls when using from_pretrained by @benHeidabetlen in #1656 +- feat: Add more detailed log for prefix-match by @xu-song in #1659 + ## [0.2.87] - feat: Update llama.cpp to ggerganov/llama.cpp@be55695eff44784a141a863f273661a6bce63dfc diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 37d91cd54..0030e8f7e 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.87" +__version__ = "0.2.88" From 63d65acabc7d7402cdb85926a70b2ef81b472ac4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 15 Aug 2024 14:46:46 -0400 Subject: [PATCH 510/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 8 ++++---- llama_cpp/llama.py | 4 ++-- llama_cpp/llama_cpp.py | 26 ++++++++++++-------------- vendor/llama.cpp | 2 +- 4 files changed, 19 insertions(+), 21 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 357023e3c..6dae88c8f 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -179,11 +179,11 @@ def token_eot(self) -> int: assert self.model is not None return llama_cpp.llama_token_eot(self.model) - def add_bos_token(self) -> int: + def add_bos_token(self) -> bool: assert self.model is not None return llama_cpp.llama_add_bos_token(self.model) - def add_eos_token(self) -> int: + def add_eos_token(self) -> bool: assert self.model is not None return llama_cpp.llama_add_eos_token(self.model) @@ -691,8 +691,8 @@ def _detokenize_bpe(model: _LlamaModel, tokens: List[int]) -> str: def _should_add_bos(model: _LlamaModel) -> bool: assert model.model is not None add_bos = llama_cpp.llama_add_bos_token(model.model) - if add_bos != -1: - return add_bos != 0 + if add_bos: + return add_bos else: return llama_cpp.llama_vocab_type(model.model) == llama_cpp.LLAMA_VOCAB_TYPE_SPM diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index b780ef6c6..29674c6d8 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1058,13 +1058,13 @@ def _create_completion( if ( (isinstance(prompt, list) and suffix is None) - or self._model.add_bos_token() == 0 + or not self._model.add_bos_token() or bos_tokens[:1] == [-1] ): bos_tokens = [] if (isinstance(prompt, list) and suffix is None) or ( - self._model.add_eos_token() != 1 and sep_token_id == -1 + not self._model.add_eos_token() and sep_token_id == -1 ): eos_tokens = [] diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 1a8424aa1..00064f324 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -314,6 +314,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20, # LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21, # LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, +# LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, +# LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -338,6 +340,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20 LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21 LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22 +LLAMA_VOCAB_PRE_TYPE_BLOOM = 23 +LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24 # // note: these values should be synchronized with ggml_rope @@ -345,13 +349,11 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # enum llama_rope_type { # LLAMA_ROPE_TYPE_NONE = -1, # LLAMA_ROPE_TYPE_NORM = 0, -# LLAMA_ROPE_TYPE_NEOX = 2, -# LLAMA_ROPE_TYPE_GLM = 4, +# LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX, # }; LLAMA_ROPE_TYPE_NONE = -1 LLAMA_ROPE_TYPE_NORM = 0 -LLAMA_ROPE_TYPE_NEOX = 2 -LLAMA_ROPE_TYPE_GLM = 4 +LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX = 2 # enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file @@ -2741,19 +2743,15 @@ def llama_token_nl(model: llama_model_p, /) -> int: ... -# // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); -@ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_int32) -def llama_add_bos_token(model: llama_model_p, /) -> int: - """Returns -1 if unknown, 1 for true or 0 for false.""" +# LLAMA_API bool llama_add_bos_token(const struct llama_model * model); +@ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_bool) +def llama_add_bos_token(model: llama_model_p, /) -> bool: ... -# // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); -@ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_int32) -def llama_add_eos_token(model: llama_model_p, /) -> int: - """Returns -1 if unknown, 1 for true or 0 for false.""" +# LLAMA_API bool llama_add_eos_token(const struct llama_model * model); +@ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_bool) +def llama_add_eos_token(model: llama_model_p, /) -> bool: ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index fc4ca27b2..4b9afbbe9 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit fc4ca27b25464a11b3b86c9dbb5b6ed6065965c2 +Subproject commit 4b9afbbe9037f8a2d659097c0c7d9fce32c6494c From 78e35c4998f3209cb70c188a47cffd17c1f12bce Mon Sep 17 00:00:00 2001 From: Junpei Kawamoto Date: Thu, 15 Aug 2024 15:00:46 -0600 Subject: [PATCH 511/624] fix: missing dependencies for test (#1680) --- pyproject.toml | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/pyproject.toml b/pyproject.toml index 7aebb74f5..ce50c673f 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -41,6 +41,10 @@ test = [ "pytest>=7.4.0", "httpx>=0.24.1", "scipy>=1.10", + "fastapi>=0.100.0", + "sse-starlette>=1.6.1", + "starlette-context>=0.3.6,<0.4", + "pydantic-settings>=2.0.1", ] dev = [ "black>=23.3.0", From 3c7501b35755205d1df70181cd46e769ec2e9790 Mon Sep 17 00:00:00 2001 From: Junpei Kawamoto Date: Thu, 15 Aug 2024 15:01:18 -0600 Subject: [PATCH 512/624] fix: Llama.close didn't free lora adapter (#1679) --- llama_cpp/llama.py | 19 +++++++++++-------- 1 file changed, 11 insertions(+), 8 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 29674c6d8..9b1a3d263 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -198,6 +198,7 @@ def __init__( A Llama instance. """ self.verbose = verbose + self._stack = contextlib.ExitStack() set_verbose(verbose) @@ -365,8 +366,6 @@ def __init__( if not os.path.exists(model_path): raise ValueError(f"Model path does not exist: {model_path}") - self._stack = contextlib.ExitStack() - self._model = self._stack.enter_context( contextlib.closing( _LlamaModel( @@ -420,6 +419,15 @@ def __init__( raise RuntimeError( f"Failed to initialize LoRA adapter from lora path: {self.lora_path}" ) + + def free_lora_adapter(): + if self._lora_adapter is None: + return + llama_cpp.llama_lora_adapter_free(self._lora_adapter) + self._lora_adapter = None + + self._stack.callback(free_lora_adapter) + assert self._ctx.ctx is not None if llama_cpp.llama_lora_adapter_set( self._ctx.ctx, self._lora_adapter, self.lora_scale @@ -2085,14 +2093,9 @@ def pooling_type(self) -> str: def close(self) -> None: """Explicitly free the model from memory.""" - if hasattr(self,'_stack'): - if self._stack is not None: - self._stack.close() + self._stack.close() def __del__(self) -> None: - if hasattr(self,'_lora_adapter'): - if self._lora_adapter is not None: - llama_cpp.llama_lora_adapter_free(self._lora_adapter) self.close() @staticmethod From 7bf07ec51873ab96ac326bac126044039768ef4b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 16 Aug 2024 14:43:25 -0400 Subject: [PATCH 513/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 2 ++ vendor/llama.cpp | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 00064f324..476713aee 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -316,6 +316,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, # LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, # LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, +# LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -342,6 +343,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22 LLAMA_VOCAB_PRE_TYPE_BLOOM = 23 LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24 +LLAMA_VOCAB_PRE_TYPE_EXAONE = 25 # // note: these values should be synchronized with ggml_rope diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 4b9afbbe9..8b3befc0e 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 4b9afbbe9037f8a2d659097c0c7d9fce32c6494c +Subproject commit 8b3befc0e2ed8fb18b903735831496b8b0c80949 From a2ba73165223f8946f3e1180e3cb67d7d9236da2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 19 Aug 2024 13:36:31 -0400 Subject: [PATCH 514/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8b3befc0e..cfac111e2 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8b3befc0e2ed8fb18b903735831496b8b0c80949 +Subproject commit cfac111e2b3953cdb6b0126e67a2487687646971 From d7328efabf841a750259dc42089efece13efc0f5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 19 Aug 2024 13:36:49 -0400 Subject: [PATCH 515/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index f7738e1a9..063fe722b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.89] + +- feat: Update llama.cpp to ggerganov/llama.cpp@cfac111e2b3953cdb6b0126e67a2487687646971 +- fix: Llama.close didn't free lora adapter by @jkawamoto in #1679 +- fix: missing dependencies for test by @jkawamoto in #1680 + ## [0.2.88] - feat: Update llama.cpp to ggerganov/llama.cpp@fc4ca27b25464a11b3b86c9dbb5b6ed6065965c2 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 0030e8f7e..e3cece2f6 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.88" +__version__ = "0.2.89" From a20f13fc3b08e7c2163d84d5da51da78ec9f48c3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 21 Aug 2024 10:11:21 -0400 Subject: [PATCH 516/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index cfac111e2..fc54ef0d1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit cfac111e2b3953cdb6b0126e67a2487687646971 +Subproject commit fc54ef0d1c138133a01933296d50a36a1ab64735 From 259ee151da9a569f58f6d4979e97cfd5d5bc3ecd Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 22 Aug 2024 09:17:45 -0400 Subject: [PATCH 517/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 8 ++++++++ vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 476713aee..b35f56a79 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1530,6 +1530,14 @@ def llama_model_decoder_start_token(model: llama_model_p, /) -> int: ... +# // Returns true if the model is recurrent (like Mamba, RWKV, etc.) +# LLAMA_API bool llama_model_is_recurrent(const struct llama_model * model); +@ctypes_function("llama_model_is_recurrent", [llama_model_p_ctypes], ctypes.c_bool) +def llama_model_is_recurrent(model: llama_model_p, /) -> bool: + """Returns true if the model is recurrent (like Mamba, RWKV, etc.)""" + ... + + # // Returns 0 on success # LLAMA_API uint32_t llama_model_quantize( # const char * fname_inp, diff --git a/vendor/llama.cpp b/vendor/llama.cpp index fc54ef0d1..1731d4238 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit fc54ef0d1c138133a01933296d50a36a1ab64735 +Subproject commit 1731d4238f9e4f925a750810e7f5480827c66dcf From 82ae7f93716c47b4ed0a809f280d5352e2fcdea9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 28 Aug 2024 00:47:48 -0400 Subject: [PATCH 518/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1731d4238..20f1789df 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1731d4238f9e4f925a750810e7f5480827c66dcf +Subproject commit 20f1789dfb4e535d64ba2f523c64929e7891f428 From f70df824985d875226793b94dacc0c302a4256b2 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 01:05:32 -0400 Subject: [PATCH 519/624] feat: Add MiniCPMv26 chat handler. --- llama_cpp/llama_chat_format.py | 92 +++++++++++++++++++++++++--------- 1 file changed, 67 insertions(+), 25 deletions(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index ea8d07feb..3c7330296 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -2707,6 +2707,31 @@ def last_image_embed_free(): def load_image(self, image_url: str) -> bytes: return self._load_image(image_url) + def _embed_image_bytes(self, image_bytes: bytes, n_threads_batch: int = 1): + if ( + self._last_image_embed is not None + and self._last_image_hash is not None + and hash(image_bytes) == self._last_image_hash + ): + return self._last_image_embed + with suppress_stdout_stderr(disable=self.verbose): + # Free the previous image embed + if self._last_image_embed is not None: + self._llava_cpp.llava_image_embed_free(self._last_image_embed) + self._last_image_embed = None + self._last_image_hash = None + embed = self._llava_cpp.llava_image_embed_make_with_bytes( + self.clip_ctx, + n_threads_batch, + (ctypes.c_uint8 * len(image_bytes)).from_buffer( + bytearray(image_bytes) + ), + len(image_bytes), + ) + self._last_image_embed = embed + self._last_image_hash = hash(image_bytes) + return embed + def __call__( self, *, @@ -2769,30 +2794,9 @@ def __call__( ) split_text = self.split_text_on_image_urls(text, image_urls) - def embed_image_bytes(image_bytes: bytes): - if ( - self._last_image_embed is not None - and self._last_image_hash is not None - and hash(image_bytes) == self._last_image_hash - ): - return self._last_image_embed - with suppress_stdout_stderr(disable=self.verbose): - # Free the previous image embed - if self._last_image_embed is not None: - self._llava_cpp.llava_image_embed_free(self._last_image_embed) - self._last_image_embed = None - self._last_image_hash = None - embed = self._llava_cpp.llava_image_embed_make_with_bytes( - self.clip_ctx, - llama.context_params.n_threads_batch, - (ctypes.c_uint8 * len(image_bytes)).from_buffer( - bytearray(image_bytes) - ), - len(image_bytes), - ) - self._last_image_embed = embed - self._last_image_hash = hash(image_bytes) - return embed + if self.verbose: + print(text, file=sys.stderr) + # Evaluate prompt llama.reset() @@ -2809,7 +2813,7 @@ def embed_image_bytes(image_bytes: bytes): llama.eval(tokens) else: image_bytes = self.load_image(value) - embed = embed_image_bytes(image_bytes) + embed = self._embed_image_bytes(image_bytes, llama.context_params.n_threads_batch) if llama.n_tokens + embed.contents.n_image_pos > llama.n_ctx(): raise ValueError( f"Prompt exceeds n_ctx: {llama.n_tokens + embed.contents.n_image_pos} > {llama.n_ctx()}" @@ -3308,6 +3312,44 @@ class Llama3VisionAlphaChatHandler(Llava15ChatHandler): Llama3VisionAlpha = Llama3VisionAlphaChatHandler +class MiniCPMv26(Llava15ChatHandler): + DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant." + + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if loop.first and messages[0]['role'] != 'system' %}" + "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n" + "{% endif %}" + "<|im_start|>{{ message['role'] }}\n" + "{% if message['content'] is iterable %}" + "{% for content in message['content'] %}" + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + + "{% for content in message['content'] %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "{% if message['content'] is string %}" + "{{ message['content'] }}" + "{% endif %}" + "<|im_end|>\n" + "{% endfor %}" + "{% if add_generation_prompt %}" + "<|im_start|>assistant\n" + "{% endif %}" + ) + + @register_chat_completion_handler("chatml-function-calling") def chatml_function_calling( llama: llama.Llama, From e251a0bd7d05f993f19e79e3f8328e4d3b41110a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 01:26:34 -0400 Subject: [PATCH 520/624] fix: Update name to MiniCPMv26ChatHandler --- llama_cpp/llama_chat_format.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 3c7330296..21091bf50 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -3312,7 +3312,7 @@ class Llama3VisionAlphaChatHandler(Llava15ChatHandler): Llama3VisionAlpha = Llama3VisionAlphaChatHandler -class MiniCPMv26(Llava15ChatHandler): +class MiniCPMv26ChatHandler(Llava15ChatHandler): DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant." CHAT_FORMAT = ( From c68e7fbedfff8c385167a3d5232ad6a4c12597f1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 01:51:25 -0400 Subject: [PATCH 521/624] fix: pull all gh releases for self-hosted python index --- .../generate-index-from-release.yaml | 3 + scripts/get-releases.sh | 47 +++++++ scripts/releases-to-pep-503.sh | 122 ++++++++++++------ 3 files changed, 133 insertions(+), 39 deletions(-) create mode 100755 scripts/get-releases.sh diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 662cb6f65..51961088d 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -35,7 +35,10 @@ jobs: - name: Setup Pages uses: actions/configure-pages@v5 - name: Build + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} run: | + ./scripts/get-releases.sh ./scripts/releases-to-pep-503.sh index/whl/cpu '^[v]?[0-9]+\.[0-9]+\.[0-9]+$' ./scripts/releases-to-pep-503.sh index/whl/cu121 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu121$' ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' diff --git a/scripts/get-releases.sh b/scripts/get-releases.sh new file mode 100755 index 000000000..4c904da78 --- /dev/null +++ b/scripts/get-releases.sh @@ -0,0 +1,47 @@ +#!/bin/bash + +# Function to get all releases +get_all_releases() { + local page=1 + local per_page=100 + local releases="" + local new_releases + + # Prepare headers + local headers=(-H "Accept: application/vnd.github.v3+json") + if [ -n "$GITHUB_TOKEN" ]; then + headers+=(-H "Authorization: Bearer $GITHUB_TOKEN") + fi + + while true; do + response=$(curl -s "${headers[@]}" \ + "https://api.github.com/repos/abetlen/llama-cpp-python/releases?page=$page&per_page=$per_page") + + # Check if the response is valid JSON + if ! echo "$response" | jq empty > /dev/null 2>&1; then + echo "Error: Invalid response from GitHub API" >&2 + echo "Response: $response" >&2 + return 1 + fi + + new_releases=$(echo "$response" | jq -r '.[].tag_name') + if [ -z "$new_releases" ]; then + break + fi + releases="$releases $new_releases" + ((page++)) + done + + echo $releases +} + +# Get all releases and save to file +releases=$(get_all_releases) +if [ $? -ne 0 ]; then + echo "Failed to fetch releases. Please check your internet connection and try again later." >&2 + exit 1 +fi + +echo "$releases" | tr ' ' '\n' > all_releases.txt + +echo "All releases have been saved to all_releases.txt" diff --git a/scripts/releases-to-pep-503.sh b/scripts/releases-to-pep-503.sh index 44fbbf3cf..71910efcb 100755 --- a/scripts/releases-to-pep-503.sh +++ b/scripts/releases-to-pep-503.sh @@ -1,60 +1,104 @@ #!/bin/bash +# Enable exit on error +set -e + +# Function for logging +log_error() { + echo "ERROR: $1" >&2 +} + +log_info() { + echo "INFO: $1" +} + # Get output directory or default to index/whl/cpu output_dir=${1:-"index/whl/cpu"} -# Create output directory -mkdir -p $output_dir - -# Change to output directory -pushd $output_dir +# Get pattern from second arg or default to valid python package version pattern +pattern=${2:-"^[v]?[0-9]+\.[0-9]+\.[0-9]+$"} -# Create an index html file -echo "" > index.html -echo "" >> index.html -echo " " >> index.html -echo " " >> index.html -echo " llama-cpp-python" >> index.html -echo "
" >> index.html -echo " " >> index.html -echo "" >> index.html -echo "" >> index.html +# Get the current directory (where the script is run from) +current_dir="$(pwd)" -# Create llama-cpp-python directory -mkdir -p llama-cpp-python +# Check if all_releases.txt exists +if [ ! -f "$current_dir/all_releases.txt" ]; then + log_error "all_releases.txt not found in the current directory." + exit 1 +fi -# Change to llama-cpp-python directory -pushd llama-cpp-python +# Create output directory +mkdir -p "$output_dir" # Create an index html file -echo "" > index.html -echo "" >> index.html -echo " " >> index.html -echo "

Links for llama-cpp-python

" >> index.html +cat << EOF > "$output_dir/index.html" + + + + + llama-cpp-python +
+ + -# Get all releases -releases=$(curl -s https://api.github.com/repos/abetlen/llama-cpp-python/releases | jq -r .[].tag_name) +EOF -# Get pattern from second arg or default to valid python package version pattern -pattern=${2:-"^[v]?[0-9]+\.[0-9]+\.[0-9]+$"} +# Create llama-cpp-python directory +mkdir -p "$output_dir/llama-cpp-python" + +# Create an index html file in llama-cpp-python directory +cat << EOF > "$output_dir/llama-cpp-python/index.html" + + + +

Links for llama-cpp-python

+EOF # Filter releases by pattern -releases=$(echo $releases | tr ' ' '\n' | grep -E $pattern) +releases=$(grep -E "$pattern" "$current_dir/all_releases.txt") + +# Prepare curl headers +headers=('--header' 'Accept: application/vnd.github.v3+json') +if [ -n "$GITHUB_TOKEN" ]; then + headers+=('--header' "authorization: Bearer $GITHUB_TOKEN") +fi +headers+=('--header' 'content-type: application/json') # For each release, get all assets for release in $releases; do - assets=$(curl -s https://api.github.com/repos/abetlen/llama-cpp-python/releases/tags/$release | jq -r .assets) + log_info "Processing release: $release" + response=$(curl -s "${headers[@]}" \ + "https://api.github.com/repos/abetlen/llama-cpp-python/releases/tags/$release") + + if [ -z "$response" ]; then + log_error "Empty response from GitHub API for release $release" + continue + fi + + if ! echo "$response" | jq -e '.assets' > /dev/null 2>&1; then + log_error "Invalid or unexpected response from GitHub API for release $release" + log_error "Response: $response" + continue + fi + # Get release version from release ie v0.1.0-cu121 -> v0.1.0 - release_version=$(echo $release | grep -oE "^[v]?[0-9]+\.[0-9]+\.[0-9]+") - echo "

$release_version

" >> index.html - for asset in $(echo $assets | jq -r .[].browser_download_url); do - if [[ $asset == *".whl" ]]; then - echo " $asset" >> index.html - echo "
" >> index.html - fi + release_version=$(echo "$release" | grep -oE "^[v]?[0-9]+\.[0-9]+\.[0-9]+") + echo "

$release_version

" >> "$output_dir/llama-cpp-python/index.html" + + wheel_urls=$(echo "$response" | jq -r '.assets[] | select(.name | endswith(".whl")) | .browser_download_url') + if [ -z "$wheel_urls" ]; then + log_error "No wheel files found for release $release" + continue + fi + + echo "$wheel_urls" | while read -r asset; do + echo " $asset" >> "$output_dir/llama-cpp-python/index.html" + echo "
" >> "$output_dir/llama-cpp-python/index.html" done done -echo " " >> index.html -echo "" >> index.html -echo "" >> index.html +echo " " >> "$output_dir/llama-cpp-python/index.html" +echo "" >> "$output_dir/llama-cpp-python/index.html" +echo "" >> "$output_dir/llama-cpp-python/index.html" + +log_info "Index generation complete. Output directory: $output_dir" From 97d527eec2b879f38e9eed599f8c7d2ae3a0b23e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 02:19:13 -0400 Subject: [PATCH 522/624] feat: Add server chat_format minicpm-v-2.6 for MiniCPMv26ChatHandler --- llama_cpp/server/model.py | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index c486f8885..901998094 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -157,6 +157,20 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: chat_handler = llama_cpp.llama_chat_format.Llama3VisionAlpha( clip_model_path=settings.clip_model_path, verbose=settings.verbose ) + elif settings.chat_format == "minicpm-v-2.6": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.MiniCPMv26ChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.MiniCPMv26ChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) elif settings.chat_format == "hf-autotokenizer": assert ( settings.hf_pretrained_model_name_or_path is not None From b570fd38c1f5e2b29b801d16e8b556409f25527b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 02:37:08 -0400 Subject: [PATCH 523/624] =?UTF-8?q?docs:=20Add=20project=20icon=20courtesy?= =?UTF-8?q?=20of=20=F0=9F=A4=97?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- README.md | 2 ++ docs/icon.svg | 5 +++++ 2 files changed, 7 insertions(+) create mode 100644 docs/icon.svg diff --git a/README.md b/README.md index b0dfdd5b5..b8065149c 100644 --- a/README.md +++ b/README.md @@ -1,3 +1,5 @@ +![icon.svg](./docs/icon.svg) + # 🦙 Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) [![Documentation Status](https://readthedocs.org/projects/llama-cpp-python/badge/?version=latest)](https://llama-cpp-python.readthedocs.io/en/latest/?badge=latest) diff --git a/docs/icon.svg b/docs/icon.svg new file mode 100644 index 000000000..9f2d08753 --- /dev/null +++ b/docs/icon.svg @@ -0,0 +1,5 @@ + + + + + From cbbfad45929ebbc90d79bfaaa1dbdf2825a2db3f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 02:40:06 -0400 Subject: [PATCH 524/624] docs: center icon and resize --- README.md | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index b8065149c..ea544d4d1 100644 --- a/README.md +++ b/README.md @@ -1,6 +1,8 @@ -![icon.svg](./docs/icon.svg) +

+ +

-# 🦙 Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) +# Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) [![Documentation Status](https://readthedocs.org/projects/llama-cpp-python/badge/?version=latest)](https://llama-cpp-python.readthedocs.io/en/latest/?badge=latest) [![Tests](https://github.com/abetlen/llama-cpp-python/actions/workflows/test.yaml/badge.svg?branch=main)](https://github.com/abetlen/llama-cpp-python/actions/workflows/test.yaml) From ad2deafa8d615e9eaf0f8c3976e465fb1a3ea15f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 02:51:39 -0400 Subject: [PATCH 525/624] docs: Add MiniCPM-V-2.6 to multi-modal model list --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index ea544d4d1..b63efa3c8 100644 --- a/README.md +++ b/README.md @@ -503,6 +503,7 @@ Below are the supported multi-modal models and their respective chat handlers (P | [moondream2](https://huggingface.co/vikhyatk/moondream2) | `MoondreamChatHandler` | `moondream2` | | [nanollava](https://huggingface.co/abetlen/nanollava-gguf) | `NanollavaChatHandler` | `nanollava` | | [llama-3-vision-alpha](https://huggingface.co/abetlen/llama-3-vision-alpha-gguf) | `Llama3VisionAlphaChatHandler` | `llama-3-vision-alpha` | +| [minicpm-v-2.6](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) | `MiniCPMv26ChatHandler` | `minicpm-v-2.6` | Then you'll need to use a custom chat handler to load the clip model and process the chat messages and images. From 332720d00332334f42576bebf3a8a3da1a27bfaf Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 02:56:37 -0400 Subject: [PATCH 526/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 20f1789df..1d1ccce67 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 20f1789dfb4e535d64ba2f523c64929e7891f428 +Subproject commit 1d1ccce67613674c75c9c7e3fa4c1e24e428ba48 From 077ecb6771fbd373d5507444f6e0b6e9bb7cf4e8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 02:57:36 -0400 Subject: [PATCH 527/624] chore: Bump version --- llama_cpp/__init__.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index e3cece2f6..db1f1d0fa 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.89" +__version__ = "0.2.90" From 45001ac550dc6360b4809dd7564d3304d4ad8fa6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 10:12:23 -0400 Subject: [PATCH 528/624] misc(fix): Update CHANGELOG --- CHANGELOG.md | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index 063fe722b..f01b1b9bf 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.2.90] + +- feat: Update llama.cpp to ggerganov/llama.cpp@1d1ccce67613674c75c9c7e3fa4c1e24e428ba48 +- feat: Add support for `MiniCPMv26ChatHandler` and `minicpm-v-26` in server by @abetlen in f70df824985d875226793b94dacc0c302a4256b2 + ## [0.2.89] - feat: Update llama.cpp to ggerganov/llama.cpp@cfac111e2b3953cdb6b0126e67a2487687646971 From 4b1e3642daa06412008ab2e7366aad09281a70bb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 10:25:21 -0400 Subject: [PATCH 529/624] docs: Update README --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index b63efa3c8..29807bbdf 100644 --- a/README.md +++ b/README.md @@ -1,5 +1,5 @@

- +

# Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) From 8b853c0f256c7318938b3894721637174938af17 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 12:41:05 -0400 Subject: [PATCH 530/624] docs: Update README --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 29807bbdf..a8f14e898 100644 --- a/README.md +++ b/README.md @@ -1,5 +1,5 @@

- +

# Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) From 9cba3b8198c5f8d3c3bd12b5ddf5c32838c25040 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 29 Aug 2024 12:41:28 -0400 Subject: [PATCH 531/624] docs: Update README --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index a8f14e898..cbf1c4158 100644 --- a/README.md +++ b/README.md @@ -1,5 +1,5 @@

- +

# Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) From d981d32dac3d35b39cc164a10b4edc11cc5b89db Mon Sep 17 00:00:00 2001 From: benniekiss <63211101+benniekiss@users.noreply.github.com> Date: Thu, 29 Aug 2024 13:04:00 -0400 Subject: [PATCH 532/624] feat: Enable detokenizing special tokens with `special=True` (#1596) * enable detokenizing special tokens * enable skipping_special_tokens in hf_tokenizer detokenize() * process prev_tokens * fix doc strings * Revert changes to LlamaTokenizer prev_tokens and set special to False by default --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 9 ++++++--- llama_cpp/llama_tokenizer.py | 25 ++++++++++++++----------- 2 files changed, 20 insertions(+), 14 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 9b1a3d263..45cbd7bca 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -578,6 +578,8 @@ def tokenize( Args: text: The utf-8 encoded string to tokenize. + add_bos: Whether to add a beginning of sequence token. + special: Whether to tokenize special tokens. Raises: RuntimeError: If the tokenization failed. @@ -588,18 +590,19 @@ def tokenize( return self.tokenizer_.tokenize(text, add_bos, special) def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None + self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False ) -> bytes: """Detokenize a list of tokens. Args: tokens: The list of tokens to detokenize. - prev_tokens: The list of previous tokens. Offset mapping will be performed if provided + prev_tokens: The list of previous tokens. Offset mapping will be performed if provided. + special: Whether to detokenize special tokens. Returns: The detokenized string. """ - return self.tokenizer_.detokenize(tokens, prev_tokens=prev_tokens) + return self.tokenizer_.detokenize(tokens, prev_tokens=prev_tokens, special=special) def set_cache(self, cache: Optional[BaseLlamaCache]): """Set the cache. diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py index 029bf2acc..2e7590d14 100644 --- a/llama_cpp/llama_tokenizer.py +++ b/llama_cpp/llama_tokenizer.py @@ -19,20 +19,22 @@ def tokenize( """Tokenize the text into tokens. Args: - text: The text to tokenize. + text: The utf-8 encoded string to tokenize. add_bos: Whether to add a beginning of sequence token. - special: Whether to tokenize text literally or as special tokens.""" + special: Whether to tokenize special tokens. + """ raise NotImplementedError @abc.abstractmethod def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None + self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False ) -> bytes: """Detokenize the tokens into text. Args: - tokens: The tokens to detokenize. - prev_tokens: If tokens is a continuation of a previous sequence, the previous tokens. + tokens: The list of tokens to detokenize. + prev_tokens: The list of previous tokens. Offset mapping will be performed if provided. + special: Whether to detokenize special tokens. """ raise NotImplementedError @@ -47,9 +49,9 @@ def tokenize( return self._model.tokenize(text, add_bos=add_bos, special=special) def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None + self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False ) -> bytes: - return self._model.detokenize(tokens) + return self._model.detokenize(tokens, special=special) def encode( self, text: str, add_bos: bool = True, special: bool = True @@ -78,18 +80,19 @@ def tokenize( ) def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None + self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False ) -> bytes: + skip_special_tokens = not special if prev_tokens is not None: - text = self.hf_tokenizer.decode(prev_tokens + tokens).encode( + text = self.hf_tokenizer.decode(prev_tokens + tokens, skip_special_tokens=skip_special_tokens).encode( "utf-8", errors="ignore" ) - prev_text = self.hf_tokenizer.decode(prev_tokens).encode( + prev_text = self.hf_tokenizer.decode(prev_tokens, skip_special_tokens=skip_special_tokens).encode( "utf-8", errors="ignore" ) return text[len(prev_text) :] else: - return self.hf_tokenizer.decode(tokens).encode("utf-8", errors="ignore") + return self.hf_tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens).encode("utf-8", errors="ignore") @classmethod def from_pretrained(cls, pretrained_model_name_or_path: str) -> "LlamaHFTokenizer": From 98eb092d3c6e7c142c4ba2faaca6c091718abbb3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 30 Aug 2024 00:14:49 -0400 Subject: [PATCH 533/624] fix: Use system message in og qwen format. Closes #1697 --- llama_cpp/llama_chat_format.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 21091bf50..dfb0af65e 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1009,7 +1009,7 @@ def format_qwen( **kwargs: Any, ) -> ChatFormatterResponse: _roles = dict(user="<|im_start|>user", assistant="<|im_start|>assistant") - system_message = "You are a helpful assistant." + system_message = _get_system_message(messages) or "You are a helpful assistant." system_template = "<|im_start|>system\n{system_message}" system_message = system_template.format(system_message=system_message) _messages = _map_roles(messages, _roles) From dcb0d0cf2aae4b5f4b30da49106f69c6bd5346a8 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 30 Aug 2024 00:33:58 -0400 Subject: [PATCH 534/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 36 +++++++++++++++++++++++------------- vendor/llama.cpp | 2 +- 2 files changed, 24 insertions(+), 14 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index b35f56a79..d2e35a2d1 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -795,8 +795,8 @@ class llama_model_params(ctypes.Structure): # uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode # uint32_t n_ubatch; // physical maximum batch size # uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models) -# uint32_t n_threads; // number of threads to use for generation -# uint32_t n_threads_batch; // number of threads to use for batch processing +# int32_t n_threads; // number of threads to use for generation +# int32_t n_threads_batch; // number of threads to use for batch processing # enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` # enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id @@ -901,8 +901,8 @@ class llama_context_params(ctypes.Structure): ("n_batch", ctypes.c_uint32), ("n_ubatch", ctypes.c_uint32), ("n_seq_max", ctypes.c_uint32), - ("n_threads", ctypes.c_uint32), - ("n_threads_batch", ctypes.c_uint32), + ("n_threads", ctypes.c_int32), + ("n_threads_batch", ctypes.c_int32), ("rope_scaling_type", ctypes.c_int), ("pooling_type", ctypes.c_int), ("attention_type", ctypes.c_int), @@ -1197,6 +1197,16 @@ def llama_numa_init(numa: int, /): ... +# // Optional: an auto threadpool gets created in ggml if not passed explicitly +# LLAMA_API void llama_attach_threadpool( +# struct llama_context * ctx, +# ggml_threadpool_t threadpool, +# ggml_threadpool_t threadpool_batch); + + +# LLAMA_API void llama_detach_threadpool(struct llama_context * ctx); + + # // Call once at the end of the program - currently only used for MPI # LLAMA_API void llama_backend_free(void); @ctypes_function( @@ -2478,20 +2488,20 @@ def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: # // Set the number of threads used for decoding # // n_threads is the number of threads used for generation (single token) # // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) -# LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); +# LLAMA_API void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch); @ctypes_function( "llama_set_n_threads", [ llama_context_p_ctypes, - ctypes.c_uint32, - ctypes.c_uint32, + ctypes.c_int32, + ctypes.c_int32, ], None, ) def llama_set_n_threads( ctx: llama_context_p, - n_threads: Union[ctypes.c_uint32, int], - n_threads_batch: Union[ctypes.c_uint32, int], + n_threads: Union[ctypes.c_int32, int], + n_threads_batch: Union[ctypes.c_int32, int], /, ): """Set the number of threads used for decoding @@ -2502,16 +2512,16 @@ def llama_set_n_threads( # // Get the number of threads used for generation of a single token. -# LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx); -@ctypes_function("llama_n_threads", [llama_context_p_ctypes], ctypes.c_uint32) +# LLAMA_API int32_t llama_n_threads(struct llama_context * ctx); +@ctypes_function("llama_n_threads", [llama_context_p_ctypes], ctypes.c_int32) def llama_n_threads(ctx: llama_context_p, /) -> int: """Get the number of threads used for generation of a single token""" ... # // Get the number of threads used for prompt and batch processing (multiple token). -# LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx); -@ctypes_function("llama_n_threads_batch", [llama_context_p_ctypes], ctypes.c_uint32) +# LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx); +@ctypes_function("llama_n_threads_batch", [llama_context_p_ctypes], ctypes.c_int32) def llama_n_threads_batch(ctx: llama_context_p, /) -> int: """Get the number of threads used for prompt and batch processing (multiple token)""" ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 1d1ccce67..42c76d135 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 1d1ccce67613674c75c9c7e3fa4c1e24e428ba48 +Subproject commit 42c76d1358021ccdbe8ba89109c143dd7ae166df From 9769e5719ad45d4be6bc3ebe01d0f568d3d5001e Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 31 Aug 2024 08:59:41 -0400 Subject: [PATCH 535/624] feat: Update llama.cpp --- llama_cpp/llama.py | 2 +- llama_cpp/llama_cpp.py | 6 +++--- vendor/llama.cpp | 2 +- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 45cbd7bca..7fac936d1 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -153,7 +153,7 @@ def __init__( model_path: Path to the model. n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded. split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options. - main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_LAYER: ignored + main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_MODE_LAYER: ignored tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split. rpc_servers: Comma separated list of RPC servers to use for offloading vocab_only: Only load the vocabulary no weights. diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index d2e35a2d1..c1ba14fe5 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -711,9 +711,9 @@ class llama_model_kv_override(ctypes.Structure): # enum llama_split_mode split_mode; // how to split the model across multiple GPUs # // main_gpu interpretation depends on split_mode: -# // LLAMA_SPLIT_NONE: the GPU that is used for the entire model -# // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results -# // LLAMA_SPLIT_LAYER: ignored +# // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model +# // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results +# // LLAMA_SPLIT_MODE_LAYER: ignored # int32_t main_gpu; # // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 42c76d135..a47667cff 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 42c76d1358021ccdbe8ba89109c143dd7ae166df +Subproject commit a47667cff41f5a198eb791974e0afcc1cddd3229 From c3fc80a2cf5e88360c982a3187751083e881bd16 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 1 Sep 2024 21:06:18 -0400 Subject: [PATCH 536/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 3 +++ vendor/llama.cpp | 2 +- 2 files changed, 4 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index c1ba14fe5..501323774 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -276,6 +276,7 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE # LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece # LLAMA_VOCAB_TYPE_UGM = 4, // T5 tokenizer based on Unigram +# LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization # }; LLAMA_VOCAB_TYPE_NONE = 0 """For models without vocab""" @@ -287,6 +288,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa """BERT tokenizer based on WordPiece""" LLAMA_VOCAB_TYPE_UGM = 4 """T5 tokenizer based on Unigram""" +LLAMA_VOCAB_TYPE_RWKV = 5 +"""RWKV tokenizer based on greedy tokenization""" # // pre-tokenization types diff --git a/vendor/llama.cpp b/vendor/llama.cpp index a47667cff..8f1d81a0b 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit a47667cff41f5a198eb791974e0afcc1cddd3229 +Subproject commit 8f1d81a0b6f50b9bad72db0b6fcd299ad9ecd48c From 9497bcd6da7c1675f1309892fd3206e44c38940f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 4 Sep 2024 23:01:55 -0400 Subject: [PATCH 537/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8f1d81a0b..581c30518 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8f1d81a0b6f50b9bad72db0b6fcd299ad9ecd48c +Subproject commit 581c305186a0ff93f360346c57e21fe16e967bb7 From c032fc65b0873337ed39e5d63e15468a5d797646 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Sep 2024 04:12:29 -0400 Subject: [PATCH 538/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 4 ++++ vendor/llama.cpp | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 501323774..ecaf2369c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -443,6 +443,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors # # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; @@ -479,6 +481,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33 LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34 LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35 +LLAMA_FTYPE_MOSTLY_TQ1_0 = 36 +LLAMA_FTYPE_MOSTLY_TQ2_0 = 37 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 581c30518..8ebe8ddeb 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 581c305186a0ff93f360346c57e21fe16e967bb7 +Subproject commit 8ebe8ddebd68526757c631cd019de009697c63c2 From e529940f45d42ed8aa31334123b8d66bc67b0e78 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Thu, 19 Sep 2024 01:22:05 +0200 Subject: [PATCH 539/624] feat(ci): Speed up CI workflows using `uv`, add support for CUDA 12.5 wheels * Update build-wheels-cuda.yaml * Update build-wheels-cuda.yaml * revert * Bump pyhton from 3.8 to 3.9 * Remove python 3.8 * Remove Python 3.7 and 3.8 deprecated * Bump python from 3.8 to 3.9 * Add python 3.9 * Add python 3.9, remove macos-11 deprecated, add macos-14 * Bump python 3.8 to 3.9 * Add python 3.13 * Add python 3.13 * python 3.13 remove * remove python 3.13 * remove python 3.8 * Bump macos-13 to macos-14 * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update generate-index-from-release.yaml Add avx, avx2 and avx512 * Update test.yaml * Update test-pypi.yaml * Update publish.yaml * Update publish-to-test.yaml * Update build-wheels-cuda.yaml Cuda with AVX2 by default * Update build-wheels-cuda.yaml * remove DEPRECATED 32 bits * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml Upgrade matrix os to latest version * Update build-wheels-metal.yaml * Update build-wheels-cuda.yaml * Update test.yaml * Update test-pypi.yaml * Update test.yaml Add cache: 'pip' * Update publish-to-test.yaml * Update build-wheels-metal.yaml Add cache: 'pip' * Update build-wheels-cuda.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml remove x86_64 * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * revert * Remove cpu variants * Update build-wheels-metal.yaml * Update build-and-release.yaml * Update publish-to-test.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update publish.yaml * Update test-pypi.yaml * Update test.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update publish.yaml * Update test-pypi.yaml * Update publish-to-test.yaml * Update test.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update publish-to-test.yaml * Update publish.yaml * Update test-pypi.yaml * Update test.yaml * Update test.yaml * Update build-and-release.yaml * Update publish-to-test.yaml * Update build-wheels-metal.yaml * Update test-pypi.yaml * Update test.yaml * Update build-and-release.yaml * Update build-wheels-metal.yaml * Update build-wheels-metal.yaml * Update publish.yaml * Update publish-to-test.yaml * Update test-pypi.yaml * Update test.yaml * Update build-wheels-cuda.yaml * Update generate-index-from-release.yaml * Update README.md * Update README.md * Update test.yaml --------- Co-authored-by: Andrei Betlen --- .github/workflows/build-and-release.yaml | 44 +++++++++-- .github/workflows/build-wheels-cuda.yaml | 2 +- .github/workflows/build-wheels-metal.yaml | 17 ++++- .../generate-index-from-release.yaml | 1 + .github/workflows/publish-to-test.yaml | 24 +++++- .github/workflows/publish.yaml | 23 +++++- .github/workflows/test-pypi.yaml | 59 +++++++++++++-- .github/workflows/test.yaml | 75 +++++++++++++++++-- README.md | 3 +- 9 files changed, 217 insertions(+), 31 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 785f38847..ab8ff4e1c 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -21,12 +21,25 @@ jobs: # Used to host cibuildwheel - uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.9" - - name: Install dependencies + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' run: | python -m pip install --upgrade pip - python -m pip install -e .[all] + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd - name: Build wheels uses: pypa/cibuildwheel@v2.20.0 @@ -79,16 +92,33 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - uses: actions/setup-python@v5 with: - python-version: "3.8" - - name: Install dependencies + python-version: "3.9" + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' run: | - python -m pip install --upgrade pip build - python -m pip install -e .[all] + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Build source distribution run: | python -m build --sdist + - uses: actions/upload-artifact@v4 with: name: sdist diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 0733a68c5..b1e5fc1a9 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -22,7 +22,7 @@ jobs: $matrix = @{ 'os' = @('ubuntu-latest', 'windows-2019') 'pyver' = @("3.9", "3.10", "3.11", "3.12") - 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.0") 'releasetag' = @("basic") } diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index dd8b9c8fc..551820822 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -23,11 +23,24 @@ jobs: with: python-version: "3.12" cache: 'pip' + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash - - name: Install dependencies + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install -e .[all] + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd - name: Build wheels uses: pypa/cibuildwheel@v2.20.0 diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 51961088d..8d80e7e47 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -44,6 +44,7 @@ jobs: ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' ./scripts/releases-to-pep-503.sh index/whl/cu123 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu123$' ./scripts/releases-to-pep-503.sh index/whl/cu124 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + ./scripts/releases-to-pep-503.sh index/whl/cu125 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu125$' ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' - name: Upload artifact uses: actions/upload-pages-artifact@v3 diff --git a/.github/workflows/publish-to-test.yaml b/.github/workflows/publish-to-test.yaml index 19613233b..de3ae42aa 100644 --- a/.github/workflows/publish-to-test.yaml +++ b/.github/workflows/publish-to-test.yaml @@ -19,24 +19,42 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - name: Set up Python uses: actions/setup-python@v5 with: python-version: "3.11" cache: 'pip' + - name: Append Dev Version to __version__ run: | DEV_VERSION=${{ github.event.inputs.dev_version }} CURRENT_VERSION=$(awk -F= '/__version__ =/ {print $2}' llama_cpp/__init__.py | tr -d ' "') NEW_VERSION="${CURRENT_VERSION}.dev${DEV_VERSION}" sed -i 's/__version__ = \".*\"/__version__ = \"'"${NEW_VERSION}"'\"/' llama_cpp/__init__.py - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' run: | - python -m pip install --upgrade pip build - python -m pip install -e .[all] + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Build source distribution run: | python -m build --sdist + - name: Publish to Test PyPI uses: pypa/gh-action-pypi-publish@release/v1 with: diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index c6abb43b3..67816ea03 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -13,17 +13,34 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - name: Set up Python uses: actions/setup-python@v5 with: python-version: "3.9" - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' run: | - python -m pip install --upgrade pip build - python -m pip install -e .[all] + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Build source distribution run: | python -m build --sdist + - name: Publish distribution to PyPI # TODO: move to tag based releases # if: startsWith(github.ref, 'refs/tags') diff --git a/.github/workflows/test-pypi.yaml b/.github/workflows/test-pypi.yaml index d7131956d..335033bba 100644 --- a/.github/workflows/test-pypi.yaml +++ b/.github/workflows/test-pypi.yaml @@ -16,10 +16,25 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install llama-cpp-python[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install --verbose llama-cpp-python[all] + python -m pip install uv + python -m uv pip install llama-cpp-python[all] --verbose + shell: cmd + - name: Test with pytest run: | python -c "import llama_cpp" @@ -37,10 +52,25 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install llama-cpp-python[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install --verbose llama-cpp-python[all] + python -m pip install uv + python -m uv pip install llama-cpp-python[all] --verbose + shell: cmd + - name: Test with pytest run: | python -c "import llama_cpp" @@ -57,11 +87,26 @@ jobs: uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - cache: 'pip' - - name: Install dependencies + cache: 'pip' + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install llama-cpp-python[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install --verbose llama-cpp-python[all] + python -m pip install uv + python -m uv pip install llama-cpp-python[all] --verbose + shell: cmd + - name: Test with pytest run: | python -c "import llama_cpp" diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 78f0b4983..9de9c9229 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -20,15 +20,31 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} cache: 'pip' - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install .[all] -v + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Test with pytest run: | python -m pytest @@ -44,15 +60,31 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} cache: 'pip' - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install .[all] -v + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Test with pytest run: | python -m pytest @@ -68,15 +100,31 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - name: Set up Python ${{ matrix.python-version }} uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} cache: 'pip' - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' run: | python -m pip install --upgrade pip - python -m pip install .[all] --verbose + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Test with pytest run: | python -m pytest @@ -90,11 +138,24 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" + - name: Set up Python 3.9 uses: actions/setup-python@v5 with: python-version: "3.9" - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace CMAKE_ARGS="-DLLAMA_METAL=on" python -m uv pip install .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip CMAKE_ARGS="-DGGML_METAL=on" python -m pip install .[all] --verbose diff --git a/README.md b/README.md index cbf1c4158..12b4b127d 100644 --- a/README.md +++ b/README.md @@ -125,7 +125,7 @@ CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python It is also possible to install a pre-built wheel with CUDA support. As long as your system meets some requirements: -- CUDA Version is 12.1, 12.2, 12.3, or 12.4 +- CUDA Version is 12.1, 12.2, 12.3, 12.4 or 12.5 - Python Version is 3.10, 3.11 or 3.12 ```bash @@ -138,6 +138,7 @@ Where `` is one of the following: - `cu122`: CUDA 12.2 - `cu123`: CUDA 12.3 - `cu124`: CUDA 12.4 +- `cu124`: CUDA 12.5 For example, to install the CUDA 12.1 wheel: From a4e1451abe02c02fc9c831abb591d0559a54b849 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed, 18 Sep 2024 19:26:51 -0400 Subject: [PATCH 540/624] chore(deps): bump pypa/cibuildwheel from 2.20.0 to 2.21.1 (#1743) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.20.0 to 2.21.1. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.20.0...v2.21.1) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Andrei --- .github/workflows/build-and-release.yaml | 4 ++-- .github/workflows/build-wheels-metal.yaml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index ab8ff4e1c..3f85a30fe 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -42,7 +42,7 @@ jobs: shell: cmd - name: Build wheels - uses: pypa/cibuildwheel@v2.20.0 + uses: pypa/cibuildwheel@v2.21.1 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -69,7 +69,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.20.0 + uses: pypa/cibuildwheel@v2.21.1 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index 551820822..8e6fb5cb7 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -43,7 +43,7 @@ jobs: shell: cmd - name: Build wheels - uses: pypa/cibuildwheel@v2.20.0 + uses: pypa/cibuildwheel@v2.21.1 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" From f8fcb3ea3424bcfba3a5437626a994771a02324b Mon Sep 17 00:00:00 2001 From: Andrei Date: Wed, 18 Sep 2024 20:00:19 -0400 Subject: [PATCH 541/624] feat: Update sampling API for llama.cpp (#1742) * Initial samplng api update * Fix logger * Update tests * Update * Remove seed * Add sampling chain * Remove unnused test * Use Qwen2 0.5B for ci tests * Fix typo * Fix typo * Update cache version * Use real model for tests * Add huggingface-hub as a test dependency * Remove RUST_LOG=trace * Add actual logit processor test --- .github/workflows/test.yaml | 81 ++- llama_cpp/_internals.py | 361 +++++------ llama_cpp/_logger.py | 19 +- llama_cpp/llama.py | 227 ++++--- llama_cpp/llama_cpp.py | 1178 ++++++++++++++++------------------- llama_cpp/llama_grammar.py | 881 +------------------------- pyproject.toml | 1 + tests/test_llama.py | 358 ++++------- tests/test_llama_grammar.py | 10 +- vendor/llama.cpp | 2 +- 10 files changed, 1035 insertions(+), 2083 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 9de9c9229..f24f5b7e1 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -1,5 +1,4 @@ name: Tests - on: pull_request: branches: @@ -8,14 +7,34 @@ on: branches: - main +env: + REPO_ID: Qwen/Qwen2-0.5B-Instruct-GGUF + MODEL_FILE: qwen2-0_5b-instruct-q8_0.gguf + jobs: - build-linux: + download-model: + runs-on: ubuntu-latest + steps: + - name: Set up Python + uses: actions/setup-python@v5 + with: + python-version: "3.9" + - name: Install huggingface-hub + run: pip install huggingface-hub + - name: Download model + run: huggingface-cli download ${{ env.REPO_ID }} ${{ env.MODEL_FILE }} + - name: Cache model + uses: actions/cache@v4 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + build-linux: + needs: download-model runs-on: ubuntu-latest strategy: matrix: python-version: ["3.9", "3.10", "3.11", "3.12"] - steps: - uses: actions/checkout@v4 with: @@ -26,36 +45,35 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' - + - name: Restore model cache + uses: actions/cache@v3 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - name: Install dependencies (Linux/MacOS) if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv - RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install -e .[all] --verbose shell: bash - - name: Install dependencies (Windows) if: runner.os == 'Windows' - env: - RUST_LOG: trace run: | python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose - shell: cmd - + shell: cmd - name: Test with pytest run: | python -m pytest build-windows: - + needs: download-model runs-on: windows-latest strategy: matrix: python-version: ["3.9", "3.10", "3.11", "3.12"] - steps: - uses: actions/checkout@v4 with: @@ -66,19 +84,23 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' + + - name: Restore model cache + uses: actions/cache@v3 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - name: Install dependencies (Linux/MacOS) if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv - RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install -e .[all] --verbose shell: bash - name: Install dependencies (Windows) if: runner.os == 'Windows' - env: - RUST_LOG: trace run: | python -m pip install --upgrade pip python -m pip install uv @@ -90,12 +112,11 @@ jobs: python -m pytest build-macos: - + needs: download-model runs-on: macos-latest strategy: matrix: python-version: ["3.9", "3.10", "3.11", "3.12"] - steps: - uses: actions/checkout@v4 with: @@ -106,34 +127,36 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' + + - name: Restore model cache + uses: actions/cache@v3 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - name: Install dependencies (Linux/MacOS) if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv - RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install -e .[all] --verbose shell: bash - name: Install dependencies (Windows) if: runner.os == 'Windows' - env: - RUST_LOG: trace run: | python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose - shell: cmd + shell: cmd - name: Test with pytest run: | python -m pytest - build-macos-metal: - + needs: download-model runs-on: macos-latest - steps: - uses: actions/checkout@v4 with: @@ -144,18 +167,22 @@ jobs: with: python-version: "3.9" + - name: Restore model cache + uses: actions/cache@v3 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + - name: Install dependencies (Linux/MacOS) if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv - RUST_LOG=trace CMAKE_ARGS="-DLLAMA_METAL=on" python -m uv pip install .[all] --verbose + CMAKE_ARGS="-DLLAMA_METAL=on" python -m uv pip install .[all] --verbose shell: bash - name: Install dependencies (Windows) if: runner.os == 'Windows' - env: - RUST_LOG: trace run: | python -m pip install --upgrade pip CMAKE_ARGS="-DGGML_METAL=on" python -m pip install .[all] --verbose diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 6dae88c8f..6475695c6 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -6,6 +6,7 @@ from typing import ( Dict, List, + Tuple, Optional, Sequence, ) @@ -25,7 +26,7 @@ # Python wrappers over llama.h structs -class _LlamaModel: +class LlamaModel: """Intermediate Python wrapper for a llama.cpp llama_model. NOTE: For stability it's recommended you use the Llama class instead.""" @@ -41,19 +42,21 @@ def __init__( self.verbose = verbose self._exit_stack = ExitStack() - self.model = None + model = None if not os.path.exists(path_model): raise ValueError(f"Model path does not exist: {path_model}") with suppress_stdout_stderr(disable=verbose): - self.model = llama_cpp.llama_load_model_from_file( + model = llama_cpp.llama_load_model_from_file( self.path_model.encode("utf-8"), self.params ) - if self.model is None: + if model is None: raise ValueError(f"Failed to load model from file: {path_model}") + self.model = model + def free_model(): if self.model is None: return @@ -69,128 +72,84 @@ def __del__(self): self.close() def vocab_type(self) -> int: - assert self.model is not None return llama_cpp.llama_vocab_type(self.model) def n_vocab(self) -> int: - assert self.model is not None return llama_cpp.llama_n_vocab(self.model) def n_ctx_train(self) -> int: - assert self.model is not None return llama_cpp.llama_n_ctx_train(self.model) def n_embd(self) -> int: - assert self.model is not None return llama_cpp.llama_n_embd(self.model) def rope_freq_scale_train(self) -> float: - assert self.model is not None return llama_cpp.llama_rope_freq_scale_train(self.model) def desc(self) -> str: - assert self.model is not None buf = ctypes.create_string_buffer(1024) llama_cpp.llama_model_desc(self.model, buf, 1024) return buf.value.decode("utf-8") def size(self) -> int: - assert self.model is not None return llama_cpp.llama_model_size(self.model) def n_params(self) -> int: - assert self.model is not None return llama_cpp.llama_model_n_params(self.model) def get_tensor(self, name: str) -> ctypes.c_void_p: - assert self.model is not None return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) - def apply_lora_from_file( - self, - lora_path: str, - scale: float, - path_base_model: Optional[str], - n_threads: int, - ): - assert self.model is not None - return llama_cpp.llama_model_apply_lora_from_file( - self.model, - lora_path.encode("utf-8"), - scale, - ( - path_base_model.encode("utf-8") - if path_base_model is not None - else ctypes.c_char_p(0) - ), - n_threads, - ) # Vocab def token_get_text(self, token: int) -> str: - # TODO: Fix - assert self.model is not None return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8") def token_get_score(self, token: int) -> float: - assert self.model is not None return llama_cpp.llama_token_get_score(self.model, token) def token_get_attr(self, token: int) -> int: - assert self.model is not None return llama_cpp.llama_token_get_attr(self.model, token) # Special tokens def token_bos(self) -> int: - assert self.model is not None return llama_cpp.llama_token_bos(self.model) def token_eos(self) -> int: - assert self.model is not None return llama_cpp.llama_token_eos(self.model) def token_cls(self) -> int: - assert self.model is not None return llama_cpp.llama_token_cls(self.model) def token_sep(self) -> int: - assert self.model is not None return llama_cpp.llama_token_sep(self.model) def token_nl(self) -> int: - assert self.model is not None return llama_cpp.llama_token_nl(self.model) def token_prefix(self) -> int: - assert self.model is not None return llama_cpp.llama_token_prefix(self.model) def token_middle(self) -> int: - assert self.model is not None return llama_cpp.llama_token_middle(self.model) def token_suffix(self) -> int: - assert self.model is not None return llama_cpp.llama_token_suffix(self.model) def token_eot(self) -> int: - assert self.model is not None return llama_cpp.llama_token_eot(self.model) def add_bos_token(self) -> bool: - assert self.model is not None return llama_cpp.llama_add_bos_token(self.model) def add_eos_token(self) -> bool: - assert self.model is not None return llama_cpp.llama_add_eos_token(self.model) # Tokenization def tokenize(self, text: bytes, add_bos: bool, special: bool): - assert self.model is not None n_ctx = self.n_ctx_train() tokens = (llama_cpp.llama_token * n_ctx)() n_tokens = llama_cpp.llama_tokenize( @@ -209,13 +168,11 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): return list(tokens[:n_tokens]) def token_to_piece(self, token: int, special: bool = False) -> bytes: - assert self.model is not None buf = ctypes.create_string_buffer(32) llama_cpp.llama_token_to_piece(self.model, token, buf, 32, 0, special) return bytes(buf) def detokenize(self, tokens: List[int], special: bool = False) -> bytes: - assert self.model is not None output = b"" size = 32 buffer = (ctypes.c_char * size)() @@ -235,7 +192,6 @@ def detokenize(self, tokens: List[int], special: bool = False) -> bytes: # Extra def metadata(self) -> Dict[str, str]: - assert self.model is not None metadata: Dict[str, str] = {} buffer_size = 1024 buffer = ctypes.create_string_buffer(buffer_size) @@ -272,14 +228,14 @@ def default_params(): return llama_cpp.llama_model_default_params() -class _LlamaContext: +class LlamaContext: """Intermediate Python wrapper for a llama.cpp llama_context. NOTE: For stability it's recommended you use the Llama class instead.""" def __init__( self, *, - model: _LlamaModel, + model: LlamaModel, params: llama_cpp.llama_context_params, verbose: bool = True, ): @@ -288,15 +244,13 @@ def __init__( self.verbose = verbose self._exit_stack = ExitStack() - self.ctx = None + ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params) - assert self.model.model is not None - - self.ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params) - - if self.ctx is None: + if ctx is None: raise ValueError("Failed to create llama_context") + self.ctx = ctx + def free_ctx(): if self.ctx is None: return @@ -312,35 +266,27 @@ def __del__(self): self.close() def n_ctx(self) -> int: - assert self.ctx is not None return llama_cpp.llama_n_ctx(self.ctx) def pooling_type(self) -> int: - assert self.ctx is not None return llama_cpp.llama_pooling_type(self.ctx) def kv_cache_clear(self): - assert self.ctx is not None llama_cpp.llama_kv_cache_clear(self.ctx) def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int): - assert self.ctx is not None llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1) def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int): - assert self.ctx is not None llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1) def kv_cache_seq_keep(self, seq_id: int): - assert self.ctx is not None llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id) def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): - assert self.ctx is not None llama_cpp.llama_kv_cache_seq_add(self.ctx, seq_id, p0, p1, shift) def get_state_size(self) -> int: - assert self.ctx is not None return llama_cpp.llama_get_state_size(self.ctx) # TODO: copy_state_data @@ -351,9 +297,7 @@ def get_state_size(self) -> int: # TODO: llama_save_session_file - def decode(self, batch: "_LlamaBatch"): - assert self.ctx is not None - assert batch.batch is not None + def decode(self, batch: LlamaBatch): return_code = llama_cpp.llama_decode( self.ctx, batch.batch, @@ -362,25 +306,21 @@ def decode(self, batch: "_LlamaBatch"): raise RuntimeError(f"llama_decode returned {return_code}") def set_n_threads(self, n_threads: int, n_threads_batch: int): - assert self.ctx is not None llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch) def get_logits(self): - assert self.ctx is not None return llama_cpp.llama_get_logits(self.ctx) def get_logits_ith(self, i: int): - assert self.ctx is not None return llama_cpp.llama_get_logits_ith(self.ctx, i) def get_embeddings(self): - assert self.ctx is not None return llama_cpp.llama_get_embeddings(self.ctx) # Sampling functions def set_rng_seed(self, seed: int): - assert self.ctx is not None + # TODO: Fix llama_cpp.llama_set_rng_seed(self.ctx, seed) def sample_repetition_penalties( @@ -392,7 +332,6 @@ def sample_repetition_penalties( penalty_freq: float, penalty_present: float, ): - assert self.ctx is not None llama_cpp.llama_sample_repetition_penalties( self.ctx, llama_cpp.byref(candidates.candidates), @@ -404,26 +343,22 @@ def sample_repetition_penalties( ) def sample_softmax(self, candidates: "_LlamaTokenDataArray"): - assert self.ctx is not None llama_cpp.llama_sample_softmax( self.ctx, llama_cpp.byref(candidates.candidates), ) def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): - assert self.ctx is not None llama_cpp.llama_sample_top_k( self.ctx, llama_cpp.byref(candidates.candidates), k, min_keep ) def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - assert self.ctx is not None llama_cpp.llama_sample_top_p( self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - assert self.ctx is not None llama_cpp.llama_sample_min_p( self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) @@ -431,7 +366,6 @@ def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: i def sample_tail_free( self, candidates: "_LlamaTokenDataArray", z: float, min_keep: int ): - assert self.ctx is not None llama_cpp.llama_sample_tail_free( self.ctx, llama_cpp.byref(candidates.candidates), z, min_keep ) @@ -439,20 +373,16 @@ def sample_tail_free( def sample_typical( self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int ): - assert self.ctx is not None llama_cpp.llama_sample_typical( self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): - assert self.ctx is not None llama_cpp.llama_sample_temp( self.ctx, llama_cpp.byref(candidates.candidates), temp ) def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): - assert self.ctx is not None - assert grammar.grammar is not None llama_cpp.llama_sample_grammar( self.ctx, llama_cpp.byref(candidates.candidates), @@ -467,7 +397,6 @@ def sample_token_mirostat( m: int, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: - assert self.ctx is not None return llama_cpp.llama_sample_token_mirostat( self.ctx, llama_cpp.byref(candidates.candidates), @@ -484,7 +413,6 @@ def sample_token_mirostat_v2( eta: float, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: - assert self.ctx is not None return llama_cpp.llama_sample_token_mirostat_v2( self.ctx, llama_cpp.byref(candidates.candidates), @@ -494,14 +422,12 @@ def sample_token_mirostat_v2( ) def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: - assert self.ctx is not None return llama_cpp.llama_sample_token_greedy( self.ctx, llama_cpp.byref(candidates.candidates), ) def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: - assert self.ctx is not None return llama_cpp.llama_sample_token( self.ctx, llama_cpp.byref(candidates.candidates), @@ -509,17 +435,13 @@ def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: # Grammar def grammar_accept_token(self, grammar: LlamaGrammar, token: int): - assert self.ctx is not None - assert grammar.grammar is not None llama_cpp.llama_grammar_accept_token(grammar.grammar, self.ctx, token) def reset_timings(self): - assert self.ctx is not None - llama_cpp.llama_reset_timings(self.ctx) + llama_cpp.llama_perf_context_reset(self.ctx) def print_timings(self): - assert self.ctx is not None - llama_cpp.llama_print_timings(self.ctx) + llama_cpp.llama_perf_context_print(self.ctx) # Utility functions @staticmethod @@ -528,7 +450,7 @@ def default_params(): return llama_cpp.llama_context_default_params() -class _LlamaBatch: +class LlamaBatch: def __init__( self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True ): @@ -538,11 +460,15 @@ def __init__( self.verbose = verbose self._exit_stack = ExitStack() - self.batch = None - self.batch = llama_cpp.llama_batch_init( + batch = llama_cpp.llama_batch_init( self._n_tokens, self.embd, self.n_seq_max ) + if batch is None: + raise ValueError("Failed to create llama_batch") + + self.batch = batch + def free_batch(): if self.batch is None: return @@ -558,15 +484,12 @@ def __del__(self): self.close() def n_tokens(self) -> int: - assert self.batch is not None return self.batch.n_tokens def reset(self): - assert self.batch is not None self.batch.n_tokens = 0 def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): - assert self.batch is not None n_tokens = len(batch) self.batch.n_tokens = n_tokens for i in range(n_tokens): @@ -578,7 +501,6 @@ def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): self.batch.logits[n_tokens - 1] = True def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool): - assert self.batch is not None n_tokens = len(batch) n_tokens0 = self.batch.n_tokens self.batch.n_tokens += n_tokens @@ -592,7 +514,7 @@ def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool): self.batch.logits[n_tokens - 1] = True -class _LlamaTokenDataArray: +class LlamaTokenDataArray: def __init__(self, *, n_vocab: int): self.n_vocab = n_vocab self.candidates_data = np.recarray( @@ -617,90 +539,9 @@ def copy_logits(self, logits: npt.NDArray[np.single]): self.candidates.size = self.n_vocab -# Python wrappers over common/common -def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> list[int]: - assert model.model is not None - n_tokens = len(text) + 1 if add_bos else len(text) - result = (llama_cpp.llama_token * n_tokens)() - n_tokens = llama_cpp.llama_tokenize( - model.model, - text.encode("utf-8"), - len(text), - result, - n_tokens, - add_bos, - special, - ) - if n_tokens < 0: - result = (llama_cpp.llama_token * -n_tokens)() - check = llama_cpp.llama_tokenize( - model.model, - text.encode("utf-8"), - len(text), - result, - len(result), - add_bos, - special, - ) - if check != -n_tokens: - raise RuntimeError(f'Failed to tokenize: text="{text}" n_tokens={n_tokens}') - else: - result = result[:n_tokens] - return list(result) - - -def _token_to_piece(model: _LlamaModel, token: int, special: bool = False) -> str: - assert model.model is not None - result = (ctypes.c_char * 8)(0) - n_tokens = llama_cpp.llama_token_to_piece( - model.model, token, result, 0, len(result), special - ) - if n_tokens < 0: - result = (ctypes.c_char * -n_tokens)(0) - check = llama_cpp.llama_token_to_piece( - model.model, token, result, 0, len(result), special - ) - if check != -n_tokens: - raise RuntimeError(f"Failed to get piece: token={token}") - else: - result = result[:n_tokens] - return bytes(result).decode("utf-8") - - -def _detokenize_spm(model: _LlamaModel, tokens: List[int]) -> str: - bos_id = model.token_bos() - result = "" - for i, token in enumerate(tokens): - piece = _token_to_piece(model, token) - if ( - (tokens[0] == bos_id and i == 1) or (tokens[0] != bos_id and i == 0) - ) and piece[0] == " ": - piece = piece[1:] - result += piece - return result - - -def _detokenize_bpe(model: _LlamaModel, tokens: List[int]) -> str: - result = "" - for token in tokens: - piece = _token_to_piece(model, token) - result += piece - return result - - -def _should_add_bos(model: _LlamaModel) -> bool: - assert model.model is not None - add_bos = llama_cpp.llama_add_bos_token(model.model) - if add_bos: - return add_bos - else: - return llama_cpp.llama_vocab_type(model.model) == llama_cpp.LLAMA_VOCAB_TYPE_SPM - - # Embedding functions - -def _normalize_embedding(embedding): +def normalize_embedding(embedding): norm = float(np.linalg.norm(embedding)) if norm == 0.0: return embedding @@ -711,7 +552,7 @@ def _normalize_embedding(embedding): @dataclass -class _LlamaSamplingParams: +class LlamaSamplingParams: n_prev: int = 64 n_probs: int = 0 top_k: int = 40 @@ -738,8 +579,8 @@ class _LlamaSamplingParams: @dataclass -class _LlamaSamplingContext: - params: _LlamaSamplingParams = field(default_factory=_LlamaSamplingParams) +class LlamaSamplingContext: + params: LlamaSamplingParams = field(default_factory=LlamaSamplingParams) mirostat_mu: ctypes.c_float = field(default_factory=ctypes.c_float) grammar: Optional[LlamaGrammar] = None # NOTE: Missing parsed_grammar @@ -753,7 +594,7 @@ def reset(self): self.grammar.reset() def cp(self): - return _LlamaSamplingContext( + return LlamaSamplingContext( params=self.params, mirostat_mu=self.mirostat_mu, grammar=self.grammar, @@ -767,12 +608,12 @@ def last(self) -> Optional[int]: else: return None - def prev_str(self, ctx_main: _LlamaContext, n: int) -> str: + def prev_str(self, ctx_main: LlamaContext, n: int) -> str: return ctx_main.model.detokenize(self.prev[-n:]).decode("utf-8") def sample( self, - ctx_main: _LlamaContext, + ctx_main: LlamaContext, idx: int = 0, logits_array: Optional[npt.NDArray[np.single]] = None, ): @@ -790,7 +631,7 @@ def sample( for token, logit_bias in self.params.logit_bias.items(): logits_array[token] += logit_bias - token_data_array = _LlamaTokenDataArray( + token_data_array = LlamaTokenDataArray( n_vocab=n_vocab ) # TODO: Only create this once token_data_array.copy_logits(logits_array) @@ -862,7 +703,141 @@ def sample( id = ctx_main.sample_token(token_data_array) return id - def accept(self, ctx_main: _LlamaContext, id: int, apply_grammar: bool): + def accept(self, ctx_main: LlamaContext, id: int, apply_grammar: bool): if apply_grammar and self.grammar is not None: ctx_main.grammar_accept_token(self.grammar, id) self.prev.append(id) + + +from typing import List, Callable, Optional, Union +import ctypes +import llama_cpp + +class CustomSampler: + def __init__(self, apply_func: typing.Callable[[llama_cpp.llama_token_data_array], None]): + self.apply_func = apply_func + + def apply_wrapper(sampler: llama_cpp.llama_sampler_p, cur_p: llama_cpp.llama_token_data_array_p): + self.apply_func(cur_p) + + def free_wrapper(sampler: llama_cpp.llama_sampler_p): + pass + + sampler_i = llama_cpp.llama_sampler_i() + sampler_i.apply = llama_cpp.llama_sampler_i_apply(apply_wrapper) + self._apply_wrapper_ref = apply_wrapper + + sampler_i.name = llama_cpp.llama_sampler_i_name(0) + sampler_i.accept = llama_cpp.llama_sampler_i_accept(0) + sampler_i.reset = llama_cpp.llama_sampler_i_reset(0) + sampler_i.clone = llama_cpp.llama_sampler_i_clone(0) + sampler_i.free = llama_cpp.llama_sampler_i_free(0) + + self.sampler = llama_cpp.llama_sampler() + self.sampler.iface = ctypes.pointer(sampler_i) + self.sampler.ctx = None + + def get_sampler(self) -> llama_cpp.llama_sampler_p: + return ctypes.pointer(self.sampler) + +class LlamaSampler: + def __init__(self): + params = llama_cpp.llama_sampler_chain_params() + self.sampler = llama_cpp.llama_sampler_chain_init(params) + self.samplers: List[llama_cpp.llama_sampler_p] = [] + self.custom_samplers: List[Tuple[int, CustomSampler]] = [] + + def add_greedy(self): + sampler = llama_cpp.llama_sampler_init_greedy() + self._add_sampler(sampler) + + def add_dist(self, seed: int): + sampler = llama_cpp.llama_sampler_init_dist(seed) + self._add_sampler(sampler) + + def add_softmax(self): + sampler = llama_cpp.llama_sampler_init_softmax() + self._add_sampler(sampler) + + def add_top_k(self, k: int): + sampler = llama_cpp.llama_sampler_init_top_k(k) + self._add_sampler(sampler) + + def add_top_p(self, p: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_top_p(p, min_keep) + self._add_sampler(sampler) + + def add_min_p(self, p: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_min_p(p, min_keep) + self._add_sampler(sampler) + + def add_tail_free(self, z: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_tail_free(z, min_keep) + self._add_sampler(sampler) + + def add_typical(self, p: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_typical(p, min_keep) + self._add_sampler(sampler) + + def add_temp(self, temp: float): + sampler = llama_cpp.llama_sampler_init_temp(temp) + self._add_sampler(sampler) + + def add_temp_ext(self, t: float, delta: float, exponent: float): + sampler = llama_cpp.llama_sampler_init_temp_ext(t, delta, exponent) + self._add_sampler(sampler) + + def add_mirostat(self, n_vocab: int, seed: int, tau: float, eta: float, m: int): + sampler = llama_cpp.llama_sampler_init_mirostat( + n_vocab, seed, tau, eta, m + ) + self._add_sampler(sampler) + + def add_mirostat_v2(self, seed: int, tau: float, eta: float): + sampler = llama_cpp.llama_sampler_init_mirostat_v2(seed, tau, eta) + self._add_sampler(sampler) + + def add_grammar(self, model: LlamaModel, grammar: LlamaGrammar): + sampler = llama_cpp.llama_sampler_init_grammar(model.model, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8")) + self._add_sampler(sampler) + + def add_penalties(self, n_vocab: int, special_eos_id: int, linefeed_id: int, penalty_last_n: int, penalty_repeat: float, penalty_freq: float, penalty_present: float, penalize_nl: bool, ignore_eos: bool): + sampler = llama_cpp.llama_sampler_init_penalties(n_vocab, special_eos_id, linefeed_id, penalty_last_n, penalty_repeat, penalty_freq, penalty_present, penalize_nl, ignore_eos) + self._add_sampler(sampler) + + def init_logit_bias(self, n_vocab: int, n_logit_bias, logit_bias: llama_cpp.llama_logit_bias_p): + sampler = llama_cpp.llama_sampler_init_logit_bias(n_vocab, n_logit_bias, logit_bias) + self._add_sampler(sampler) + + def add_custom(self, apply_func: Callable[[llama_cpp.llama_token_data_array], None]): + custom_sampler = CustomSampler(apply_func) + sampler = custom_sampler.get_sampler() + self._add_sampler(sampler) + # NOTE: Must remove custom samplers before free or llama.cpp will try to free them + self.custom_samplers.append((llama_cpp.llama_sampler_chain_n(self.sampler) - 1, custom_sampler)) + + def _add_sampler(self, sampler: llama_cpp.llama_sampler_p): + assert self.sampler is not None + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + self.samplers.append(sampler) + + def get_seed(self) -> int: + assert self.sampler is not None + return llama_cpp.llama_sampler_get_seed(self.sampler) + + def sample(self, ctx: LlamaContext, idx: int) -> int: + assert self.sampler is not None + return llama_cpp.llama_sampler_sample(self.sampler, ctx.ctx, idx) + + def close(self): + if self.sampler: + # NOTE: Must remove custom samplers before free or llama.cpp will try to free them + for i, _ in reversed(self.custom_samplers): + llama_cpp.llama_sampler_chain_remove(self.sampler, i) + llama_cpp.llama_sampler_free(self.sampler) + self.sampler = None + self.samplers.clear() + self.custom_samplers.clear() + + def __del__(self): + self.close() diff --git a/llama_cpp/_logger.py b/llama_cpp/_logger.py index 7638170a9..157af692f 100644 --- a/llama_cpp/_logger.py +++ b/llama_cpp/_logger.py @@ -5,21 +5,24 @@ import llama_cpp # enum ggml_log_level { -# GGML_LOG_LEVEL_ERROR = 2, -# GGML_LOG_LEVEL_WARN = 3, -# GGML_LOG_LEVEL_INFO = 4, -# GGML_LOG_LEVEL_DEBUG = 5 +# GGML_LOG_LEVEL_NONE = 0, +# GGML_LOG_LEVEL_INFO = 1, +# GGML_LOG_LEVEL_WARN = 2, +# GGML_LOG_LEVEL_ERROR = 3, +# GGML_LOG_LEVEL_DEBUG = 4, # }; GGML_LOG_LEVEL_TO_LOGGING_LEVEL = { - 2: logging.ERROR, - 3: logging.WARNING, - 4: logging.INFO, - 5: logging.DEBUG, + 0: logging.CRITICAL, + 1: logging.INFO, + 2: logging.WARNING, + 3: logging.ERROR, + 4: logging.DEBUG, } logger = logging.getLogger("llama-cpp-python") +# typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data); @llama_cpp.llama_log_callback def llama_log_callback( level: int, diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 7fac936d1..557134a6e 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -46,15 +46,7 @@ import numpy as np import numpy.typing as npt -from ._internals import ( - _LlamaModel, # type: ignore - _LlamaContext, # type: ignore - _LlamaBatch, # type: ignore - _LlamaTokenDataArray, # type: ignore - _LlamaSamplingParams, # type: ignore - _LlamaSamplingContext, # type: ignore - _normalize_embedding, # type: ignore -) +import llama_cpp._internals as internals from ._logger import set_verbose from ._utils import suppress_stdout_stderr @@ -368,7 +360,7 @@ def __init__( self._model = self._stack.enter_context( contextlib.closing( - _LlamaModel( + internals.LlamaModel( path_model=self.model_path, params=self.model_params, verbose=self.verbose, @@ -388,7 +380,7 @@ def __init__( self._ctx = self._stack.enter_context( contextlib.closing( - _LlamaContext( + internals.LlamaContext( model=self._model, params=self.context_params, verbose=self.verbose, @@ -398,7 +390,7 @@ def __init__( self._batch = self._stack.enter_context( contextlib.closing( - _LlamaBatch( + internals.LlamaBatch( n_tokens=self.n_batch, embd=0, n_seq_max=self.context_params.n_ctx, @@ -410,7 +402,6 @@ def __init__( self._lora_adapter: Optional[llama_cpp.llama_lora_adapter_p] = None if self.lora_path: - assert self._model.model is not None self._lora_adapter = llama_cpp.llama_lora_adapter_init( self._model.model, self.lora_path.encode("utf-8"), @@ -428,7 +419,6 @@ def free_lora_adapter(): self._stack.callback(free_lora_adapter) - assert self._ctx.ctx is not None if llama_cpp.llama_lora_adapter_set( self._ctx.ctx, self._lora_adapter, self.lora_scale ): @@ -453,7 +443,7 @@ def free_lora_adapter(): self._token_nl = self.token_nl() self._token_eos = self.token_eos() - self._candidates = _LlamaTokenDataArray(n_vocab=self._n_vocab) + self._candidates = internals.LlamaTokenDataArray(n_vocab=self._n_vocab) self.n_tokens = 0 self.input_ids: npt.NDArray[np.intc] = np.ndarray((n_ctx,), dtype=np.intc) @@ -542,14 +532,14 @@ def free_lora_adapter(): f"Using fallback chat format: {self.chat_format}", file=sys.stderr ) + self._sampler = None + @property def ctx(self) -> llama_cpp.llama_context_p: - assert self._ctx.ctx is not None return self._ctx.ctx @property def model(self) -> llama_cpp.llama_model_p: - assert self._model.model is not None return self._model.model @property @@ -618,8 +608,8 @@ def set_seed(self, seed: int): Args: seed: The random seed. """ - assert self._ctx.ctx is not None - llama_cpp.llama_set_rng_seed(self._ctx.ctx, seed) + # TODO: Fix this + # llama_cpp.llama_set_rng_seed(self._ctx.ctx, seed) def reset(self): """Reset the model state.""" @@ -631,8 +621,6 @@ def eval(self, tokens: Sequence[int]): Args: tokens: The list of tokens to evaluate. """ - assert self._ctx.ctx is not None - assert self._batch.batch is not None self._ctx.kv_cache_seq_rm(-1, self.n_tokens, -1) for i in range(0, len(tokens), self.n_batch): batch = tokens[i : min(len(tokens), i + self.n_batch)] @@ -662,6 +650,93 @@ def eval(self, tokens: Sequence[int]): # Update n_tokens self.n_tokens += n_tokens + def _init_sampler( + self, + top_k: int = 40, + top_p: float = 0.95, + min_p: float = 0.05, + typical_p: float = 1.0, + temp: float = 0.80, + repeat_penalty: float = 1.0, + frequency_penalty: float = 0.0, + presence_penalty: float = 0.0, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_eta: float = 0.1, + mirostat_tau: float = 5.0, + penalize_nl: bool = True, + logits_processor: Optional[LogitsProcessorList] = None, + grammar: Optional[LlamaGrammar] = None, + seed: Optional[int] = None, + ): + sampler = internals.LlamaSampler() + + if logits_processor is not None: + # Create and add a custom sampler + def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): + size = token_data_array.contents.size + data_soa = token_data_array.contents.data + data_soa_address = ctypes.addressof(data_soa.contents) + # NOTE: This is probably broken + recarray = np.recarray( + shape=(size,), + dtype=np.dtype( + [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True + ), + buf=(llama_cpp.llama_token_data * size).from_address(data_soa_address), + ) + for logit_processor in logits_processor: + recarray.logit[:] = logit_processor(self._input_ids, recarray.logit) + sampler.add_custom(apply_func) + + sampler.add_penalties( + n_vocab=self._n_vocab, + special_eos_id=self._token_eos, + linefeed_id=self._token_nl, + penalty_last_n=self.last_n_tokens_size, + penalty_repeat=repeat_penalty, + penalty_freq=frequency_penalty, + penalty_present=presence_penalty, + penalize_nl=penalize_nl, + ignore_eos=False + ) + + if grammar is not None: + sampler.add_grammar(self._model, grammar) + + if temp < 0.0: + sampler.add_softmax() + sampler.add_dist(seed or llama_cpp.LLAMA_DEFAULT_SEED) + elif temp == 0.0: + sampler.add_greedy() + else: + if mirostat_mode == 1: + mirostat_m = 100 + sampler.add_mirostat( + self._n_vocab, + seed or llama_cpp.LLAMA_DEFAULT_SEED, + mirostat_tau, + mirostat_eta, + mirostat_m, + ) + elif mirostat_mode == 2: + sampler.add_mirostat_v2( + seed or llama_cpp.LLAMA_DEFAULT_SEED, + mirostat_tau, + mirostat_eta, + ) + else: + n_probs = 0 + min_keep = max(1, n_probs) + sampler.add_top_k(top_k) + sampler.add_tail_free(tfs_z, min_keep) + sampler.add_typical(typical_p, min_keep) + sampler.add_top_p(top_p, min_keep) + sampler.add_min_p(min_p, min_keep) + sampler.add_temp(temp) + sampler.add_dist(seed or llama_cpp.LLAMA_DEFAULT_SEED) + return sampler + def sample( self, top_k: int = 40, @@ -692,49 +767,35 @@ def sample( Returns: The sampled token. """ - assert self._ctx is not None assert self.n_tokens > 0 - if idx is None: - logits: npt.NDArray[np.single] = self._scores[-1, :] - else: - logits = self._scores[idx, :] - - if logits_processor is not None: - logits[:] = ( - logits_processor(self._input_ids, logits) - if idx is None - else logits_processor(self._input_ids[: idx + 1], logits) + tmp_sampler = False + + if self._sampler is None: + tmp_sampler = True + self._sampler = self._init_sampler( + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + penalize_nl=penalize_nl, + logits_processor=logits_processor, + grammar=grammar, ) - sampling_params = _LlamaSamplingParams( - top_k=top_k, - top_p=top_p, - min_p=min_p, - tfs_z=tfs_z, - typical_p=typical_p, - temp=temp, - penalty_last_n=self.last_n_tokens_size, - penalty_repeat=repeat_penalty, - penalty_freq=frequency_penalty, - penalty_present=presence_penalty, - mirostat=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - penalize_nl=penalize_nl, - ) - sampling_context = _LlamaSamplingContext( - params=sampling_params, - grammar=grammar, - ) - sampling_context.prev = list(self.eval_tokens) - id = sampling_context.sample(ctx_main=self._ctx, logits_array=logits) - sampling_context.accept( - ctx_main=self._ctx, - id=id, - apply_grammar=grammar is not None, - ) - return id + assert self.ctx is not None + token = self._sampler.sample(self._ctx, -1) + if tmp_sampler: + self._sampler = None + return token def generate( self, @@ -756,6 +817,7 @@ def generate( logits_processor: Optional[LogitsProcessorList] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, grammar: Optional[LlamaGrammar] = None, + seed: Optional[int] = None, ) -> Generator[int, Optional[Sequence[int]], None]: """Create a generator of tokens from a prompt. @@ -778,6 +840,24 @@ def generate( """ # Reset mirostat sampling self._mirostat_mu = ctypes.c_float(2.0 * mirostat_tau) + self._sampler = self._init_sampler( + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + penalize_nl=penalize_nl, + logits_processor=logits_processor, + grammar=grammar, + seed=seed, + ) # Check for kv cache prefix match if reset and self.n_tokens > 0: @@ -799,9 +879,9 @@ def generate( if reset: self.reset() - # Reset the grammar - if grammar is not None: - grammar.reset() + # # Reset the grammar + # if grammar is not None: + # grammar.reset() sample_idx = self.n_tokens + len(tokens) - 1 tokens = list(tokens) @@ -867,7 +947,6 @@ def create_embedding( Returns: An embedding object. """ - assert self._model.model is not None model_name: str = model if model is not None else self.model_path input = input if isinstance(input, list) else [input] @@ -912,7 +991,6 @@ def embed( Returns: A list of embeddings """ - assert self._ctx.ctx is not None n_embd = self.n_embd() n_batch = self.n_batch @@ -926,7 +1004,7 @@ def embed( ) if self.verbose: - llama_cpp.llama_reset_timings(self._ctx.ctx) + llama_cpp.llama_perf_context_reset(self._ctx.ctx) if isinstance(input, str): inputs = [input] @@ -940,7 +1018,6 @@ def embed( data: Union[List[List[float]], List[List[List[float]]]] = [] def decode_batch(seq_sizes: List[int]): - assert self._ctx.ctx is not None llama_cpp.llama_kv_cache_clear(self._ctx.ctx) self._ctx.decode(self._batch) self._batch.reset() @@ -955,7 +1032,7 @@ def decode_batch(seq_sizes: List[int]): for j in range(size) ] if normalize: - embedding = [_normalize_embedding(e) for e in embedding] + embedding = [internals.normalize_embedding(e) for e in embedding] data.append(embedding) pos += size else: @@ -963,7 +1040,7 @@ def decode_batch(seq_sizes: List[int]): ptr = llama_cpp.llama_get_embeddings_seq(self._ctx.ctx, i) embedding: List[float] = ptr[:n_embd] if normalize: - embedding = _normalize_embedding(embedding) + embedding = internals.normalize_embedding(embedding) data.append(embedding) # init state @@ -1006,7 +1083,7 @@ def decode_batch(seq_sizes: List[int]): decode_batch(s_batch) if self.verbose: - llama_cpp.llama_print_timings(self._ctx.ctx) + llama_cpp.llama_perf_context_print(self._ctx.ctx) output = data[0] if isinstance(input, str) else data @@ -1048,7 +1125,6 @@ def _create_completion( ) -> Union[ Iterator[CreateCompletionResponse], Iterator[CreateCompletionStreamResponse] ]: - assert self._ctx is not None assert suffix is None or suffix.__class__ is str completion_id: str = f"cmpl-{str(uuid.uuid4())}" @@ -1211,8 +1287,9 @@ def logit_bias_processor( if self.verbose: print("Llama._create_completion: cache miss", file=sys.stderr) - if seed is not None: - self._ctx.set_rng_seed(seed) + # TODO: Fix this + # if seed is not None: + # self._ctx.set_rng_seed(seed) finish_reason = "length" multibyte_fix = 0 @@ -1233,8 +1310,8 @@ def logit_bias_processor( stopping_criteria=stopping_criteria, logits_processor=logits_processor, grammar=grammar, + seed=seed, ): - assert self._model.model is not None if llama_cpp.llama_token_is_eog(self._model.model, token): text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" @@ -2019,7 +2096,6 @@ def __setstate__(self, state): self.__init__(**state) def save_state(self) -> LlamaState: - assert self._ctx.ctx is not None if self.verbose: print("Llama.save_state: saving llama state", file=sys.stderr) state_size = llama_cpp.llama_get_state_size(self._ctx.ctx) @@ -2049,7 +2125,6 @@ def save_state(self) -> LlamaState: ) def load_state(self, state: LlamaState) -> None: - assert self._ctx.ctx is not None # Only filling in up to `n_tokens` and then zero-ing out the rest self.scores[: state.n_tokens, :] = state.scores.copy() self.scores[state.n_tokens :, :] = 0.0 diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index ecaf2369c..261d7c6b4 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -233,6 +233,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_DEFAULT_SEED 0xFFFFFFFF LLAMA_DEFAULT_SEED = 0xFFFFFFFF +# define LLAMA_TOKEN_NULL -1 +LLAMA_TOKEN_NULL = -1 + # define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla' LLAMA_FILE_MAGIC_GGLA = 0x67676C61 @@ -244,8 +247,8 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 8 -LLAMA_SESSION_VERSION = 8 +# define LLAMA_SESSION_VERSION 9 +LLAMA_SESSION_VERSION = 9 # define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ @@ -260,6 +263,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa llama_context_p = NewType("llama_context_p", int) llama_context_p_ctypes = ctypes.c_void_p +# # struct llama_sampler; +# llama_sampler_p = NewType("llama_sampler_p", int) +# llama_sampler_p_ctypes = ctypes.c_void_p # typedef int32_t llama_pos; llama_pos = ctypes.c_int32 @@ -561,6 +567,7 @@ class llama_token_data(ctypes.Structure): # typedef struct llama_token_data_array { # llama_token_data * data; # size_t size; +# int64_t selected; // this is the index in the data array (i.e. not the token id) # bool sorted; # } llama_token_data_array; class llama_token_data_array(ctypes.Structure): @@ -569,16 +576,19 @@ class llama_token_data_array(ctypes.Structure): Attributes: data (ctypes.Array[llama_token_data]): token data size (int): size of the array + selected (int): index in the data array (i.e. not the token id) sorted (bool): whether the array is sorted""" if TYPE_CHECKING: data: CtypesArray[llama_token_data] size: int + selected: int sorted: bool _fields_ = [ ("data", llama_token_data_p), ("size", ctypes.c_size_t), + ("selected", ctypes.c_int64), ("sorted", ctypes.c_bool), ] @@ -797,7 +807,6 @@ class llama_model_params(ctypes.Structure): # // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations # // https://github.com/ggerganov/llama.cpp/pull/7544 # struct llama_context_params { -# uint32_t seed; // RNG seed, -1 for random # uint32_t n_ctx; // text context, 0 = from model # uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode # uint32_t n_ubatch; // physical maximum batch size @@ -830,6 +839,7 @@ class llama_model_params(ctypes.Structure): # bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU # bool flash_attn; // whether to use flash attention [EXPERIMENTAL] +# bool no_perf; // whether to measure performance timings # // Abort callback @@ -842,7 +852,6 @@ class llama_context_params(ctypes.Structure): """Parameters for llama_context Attributes: - seed (int): RNG seed, -1 for random n_ctx (int): text context, 0 = from model n_batch (int): logical maximum batch size that can be submitted to llama_decode n_ubatch (int): physical maximum batch size @@ -873,7 +882,6 @@ class llama_context_params(ctypes.Structure): """ if TYPE_CHECKING: - seed: int n_ctx: int n_batch: int n_ubatch: int @@ -903,7 +911,6 @@ class llama_context_params(ctypes.Structure): abort_callback_data: ctypes.c_void_p _fields_ = [ - ("seed", ctypes.c_uint32), ("n_ctx", ctypes.c_uint32), ("n_batch", ctypes.c_uint32), ("n_ubatch", ctypes.c_uint32), @@ -1009,101 +1016,44 @@ class llama_model_quantize_params(ctypes.Structure): ] -# // grammar types -# struct llama_grammar; -llama_grammar_p = ctypes.c_void_p - -# // grammar element type -# enum llama_gretype { -# // end of rule definition -# LLAMA_GRETYPE_END = 0, - -# // start of alternate definition for rule -# LLAMA_GRETYPE_ALT = 1, - -# // non-terminal element: reference to rule -# LLAMA_GRETYPE_RULE_REF = 2, - -# // terminal element: character (code point) -# LLAMA_GRETYPE_CHAR = 3, - -# // inverse char(s) ([^a], [^a-b] [^abc]) -# LLAMA_GRETYPE_CHAR_NOT = 4, - -# // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to -# // be an inclusive range ([a-z]) -# LLAMA_GRETYPE_CHAR_RNG_UPPER = 5, +# typedef struct llama_logit_bias { +# llama_token token; +# float bias; +# } llama_logit_bias; +class llama_logit_bias(ctypes.Structure): + """Used to store logit bias -# // modifies a preceding LLAMA_GRETYPE_CHAR or -# // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) -# LLAMA_GRETYPE_CHAR_ALT = 6, + Attributes: + token (llama_token): token id + bias (float): bias""" -# // any character (.) -# LLAMA_GRETYPE_CHAR_ANY = 7, -# }; -LLAMA_GRETYPE_END = 0 -LLAMA_GRETYPE_ALT = 1 -LLAMA_GRETYPE_RULE_REF = 2 -LLAMA_GRETYPE_CHAR = 3 -LLAMA_GRETYPE_CHAR_NOT = 4 -LLAMA_GRETYPE_CHAR_RNG_UPPER = 5 -LLAMA_GRETYPE_CHAR_ALT = 6 -LLAMA_GRETYPE_CHAR_ANY = 7 - - -# typedef struct llama_grammar_element { -# enum llama_gretype type; -# uint32_t value; // Unicode code point or rule ID -# } llama_grammar_element; -class llama_grammar_element(ctypes.Structure): if TYPE_CHECKING: - type: int - value: int + token: llama_token + bias: float _fields_ = [ - ("type", ctypes.c_int), - ("value", ctypes.c_uint32), + ("token", llama_token), + ("bias", ctypes.c_float), ] -llama_grammar_element_p = ctypes.POINTER(llama_grammar_element) +llama_logit_bias_p = ctypes.POINTER(llama_logit_bias) -# // performance timing information -# struct llama_timings { -# double t_start_ms; -# double t_end_ms; -# double t_load_ms; -# double t_sample_ms; -# double t_p_eval_ms; -# double t_eval_ms; +# typedef struct llama_sampler_chain_params { +# bool no_perf; // whether to measure performance timings +# } llama_sampler_chain_params; +class llama_sampler_chain_params(ctypes.Structure): + """Parameters for llama_sampler_chain + + Attributes: + no_perf (bool): whether to measure performance timings""" -# int32_t n_sample; -# int32_t n_p_eval; -# int32_t n_eval; -# }; -class llama_timings(ctypes.Structure): if TYPE_CHECKING: - t_start_ms: float - t_end_ms: float - t_load_ms: float - t_sample_ms: float - t_p_eval_ms: float - t_eval_ms: float - n_sample: int - n_p_eval: int - n_eval: int + no_perf: bool _fields_ = [ - ("t_start_ms", ctypes.c_double), - ("t_end_ms", ctypes.c_double), - ("t_load_ms", ctypes.c_double), - ("t_sample_ms", ctypes.c_double), - ("t_p_eval_ms", ctypes.c_double), - ("t_eval_ms", ctypes.c_double), - ("n_sample", ctypes.c_int32), - ("n_p_eval", ctypes.c_int32), - ("n_eval", ctypes.c_int32), + ("no_perf", ctypes.c_bool), ] @@ -1126,7 +1076,7 @@ class llama_chat_message(ctypes.Structure): # // Helpers for getting default parameters -# LLAMA_API struct llama_model_params llama_model_default_params(void); +# LLAMA_API struct llama_model_params llama_model_default_params(void); @ctypes_function( "llama_model_default_params", [], @@ -1137,7 +1087,7 @@ def llama_model_default_params() -> llama_model_params: ... -# LLAMA_API struct llama_context_params llama_context_default_params(void); +# LLAMA_API struct llama_context_params llama_context_default_params(void); @ctypes_function( "llama_context_default_params", [], @@ -1148,6 +1098,17 @@ def llama_context_default_params() -> llama_context_params: ... +# LLAMA_API struct llama_sampler_chain_params llama_sampler_chain_default_params(void); +@ctypes_function( + "llama_sampler_chain_default_params", + [], + llama_sampler_chain_params, +) +def llama_sampler_chain_default_params() -> llama_sampler_chain_params: + """Get default parameters for llama_sampler_chain""" + ... + + # LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void); @ctypes_function( "llama_model_quantize_default_params", @@ -1200,8 +1161,7 @@ def llama_backend_init(): [ctypes.c_int], None, ) -def llama_numa_init(numa: int, /): - ... +def llama_numa_init(numa: int, /): ... # // Optional: an auto threadpool gets created in ggml if not passed explicitly @@ -1228,7 +1188,7 @@ def llama_backend_free(): # LLAMA_API struct llama_model * llama_load_model_from_file( # const char * path_model, -# struct llama_model_params params); +# struct llama_model_params params); @ctypes_function( "llama_load_model_from_file", [ctypes.c_char_p, llama_model_params], @@ -1236,8 +1196,7 @@ def llama_backend_free(): ) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / -) -> Optional[llama_model_p]: - ... +) -> Optional[llama_model_p]: ... # LLAMA_API void llama_free_model(struct llama_model * model); @@ -1246,8 +1205,7 @@ def llama_load_model_from_file( [llama_model_p_ctypes], None, ) -def llama_free_model(model: llama_model_p, /): - ... +def llama_free_model(model: llama_model_p, /): ... # LLAMA_API struct llama_context * llama_new_context_with_model( @@ -1260,8 +1218,7 @@ def llama_free_model(model: llama_model_p, /): ) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / -) -> Optional[llama_context_p]: - ... +) -> Optional[llama_context_p]: ... # // Frees all allocated memory @@ -1282,104 +1239,87 @@ def llama_free(ctx: llama_context_p, /): [], ctypes.c_int64, ) -def llama_time_us() -> int: - ... +def llama_time_us() -> int: ... # LLAMA_API size_t llama_max_devices(void); @ctypes_function("llama_max_devices", [], ctypes.c_size_t) -def llama_max_devices() -> int: - ... +def llama_max_devices() -> int: ... # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) -def llama_supports_mmap() -> bool: - ... +def llama_supports_mmap() -> bool: ... # LLAMA_API bool llama_supports_mlock (void); @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) -def llama_supports_mlock() -> bool: - ... +def llama_supports_mlock() -> bool: ... # LLAMA_API bool llama_supports_gpu_offload(void); @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) -def llama_supports_gpu_offload() -> bool: - ... - - -# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); -@ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) -def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: - ... +def llama_supports_gpu_offload() -> bool: ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ctx(ctx: llama_context_p, /) -> int: - ... +def llama_n_ctx(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_batch(ctx: llama_context_p, /) -> int: - ... +def llama_n_batch(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); @ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ubatch(ctx: llama_context_p, /) -> int: - ... +def llama_n_ubatch(ctx: llama_context_p, /) -> int: ... # LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); @ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_seq_max(ctx: llama_context_p, /) -> int: - ... - - -# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); -@ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) -def llama_pooling_type(ctx: llama_context_p, /) -> int: - ... - - -# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); -@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: - ... - - -# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); -@ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: - ... +def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: - ... +def llama_n_vocab(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_ctx_train(model: llama_model_p, /) -> int: - ... +def llama_n_ctx_train(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_embd(model: llama_model_p, /) -> int: - ... +def llama_n_embd(model: llama_model_p, /) -> int: ... # LLAMA_API int32_t llama_n_layer (const struct llama_model * model); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_layer(model: llama_model_p, /) -> int: - ... +def llama_n_layer(model: llama_model_p, /) -> int: ... + + +# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); +@ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... + + +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +@ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) +def llama_pooling_type(ctx: llama_context_p, /) -> int: ... + + +# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); +@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_vocab_type(model: llama_model_p, /) -> int: ... + + +# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); +@ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_rope_type(model: llama_model_p, /) -> int: ... # // Get the model's RoPE frequency scaling factor @@ -2040,7 +1980,7 @@ def llama_kv_cache_update(ctx: llama_context_p, /): # // Returns the *actual* size in bytes of the state -# // (rng, logits, embedding and kv_cache) +# // (logits, embedding and kv_cache) # // Only use when saving the state, not when restoring it, otherwise the size may be too small. # LLAMA_API size_t llama_state_get_size(struct llama_context * ctx); @ctypes_function("llama_state_get_size", [llama_context_p_ctypes], ctypes.c_size_t) @@ -2170,8 +2110,7 @@ def llama_state_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> bool: - ... +) -> bool: ... # LLAMA_API DEPRECATED(bool llama_load_session_file( @@ -2199,8 +2138,7 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: - ... +) -> int: ... # LLAMA_API bool llama_state_save_file( @@ -2224,8 +2162,7 @@ def llama_state_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> bool: - ... +) -> bool: ... # LLAMA_API DEPRECATED(bool llama_save_session_file( @@ -2250,8 +2187,7 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: - ... +) -> int: ... # // Get the exact size needed to copy the KV cache of a single sequence @@ -2349,8 +2285,7 @@ def llama_state_seq_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: - ... +) -> int: ... # LLAMA_API size_t llama_state_seq_load_file( @@ -2380,8 +2315,7 @@ def llama_state_seq_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: - ... +) -> int: ... # // @@ -2686,8 +2620,7 @@ def llama_get_embeddings_seq( ) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / -) -> bytes: - ... +) -> bytes: ... # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); @@ -2696,8 +2629,7 @@ def llama_token_get_text( ) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / -) -> float: - ... +) -> float: ... # LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token); @@ -2706,8 +2638,7 @@ def llama_token_get_score( ) def llama_token_get_attr( model: llama_model_p, token: Union[llama_token, int], / -) -> int: - ... +) -> int: ... # // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) @@ -2772,14 +2703,12 @@ def llama_token_nl(model: llama_model_p, /) -> int: # LLAMA_API bool llama_add_bos_token(const struct llama_model * model); @ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_bos_token(model: llama_model_p, /) -> bool: - ... +def llama_add_bos_token(model: llama_model_p, /) -> bool: ... # LLAMA_API bool llama_add_eos_token(const struct llama_model * model); @ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_eos_token(model: llama_model_p, /) -> bool: - ... +def llama_add_eos_token(model: llama_model_p, /) -> bool: ... # // Codellama infill tokens @@ -2792,20 +2721,17 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: - ... +def llama_token_middle(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: - ... +def llama_token_suffix(model: llama_model_p, /) -> int: ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: - ... +def llama_token_eot(model: llama_model_p, /) -> int: ... # // @@ -3006,419 +2932,296 @@ def llama_chat_apply_template( chat: CtypesArray[llama_chat_message], n_msg: int, /, -) -> int: - ... +) -> int: ... # // -# // Grammar +# // Sampling API +# // +# // Sample usage: +# // +# // // prepare the sampling chain at the start +# // auto sparams = llama_sampler_chain_default_params(); +# // +# // llama_sampler * smpl = llama_sampler_chain_init(sparams); +# // +# // llama_sampler_chain_add(smpl, llama_sampler_init_top_k(50)); +# // llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1)); +# // llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.8)); +# // +# // // typically, the chain should end with a sampler such as "greedy", "dist" or "mirostat" +# // // this sampler will be responsible to select the actual token +# // llama_sampler_chain_add(smpl, llama_sampler_init_dist(seed)); +# // +# // ... +# // +# // // decoding loop: +# // while (...) { +# // ... +# // +# // llama_decode(ctx, batch); +# // +# // // sample from the logits of the last token in the batch +# // const llama_token id = llama_sampler_sample(smpl, ctx, -1); +# // +# // // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.) +# // llama_sampler_accept(smpl, id); +# // ... +# // } +# // +# // llama_sampler_free(smpl); +# // +# // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU). +# // TODO: in the future, the entire sampling API that uses llama_model should start using llama_vocab # // +# typedef void * llama_sampler_context_t; +llama_sampler_context_t = ctypes.c_void_p -# LLAMA_API struct llama_grammar * llama_grammar_init( -# const llama_grammar_element ** rules, -# size_t n_rules, -# size_t start_rule_index); -@ctypes_function( - "llama_grammar_init", - [ - ctypes.POINTER(llama_grammar_element_p), - ctypes.c_size_t, - ctypes.c_size_t, - ], - llama_grammar_p, -) -def llama_grammar_init( - rules: CtypesArray[ - CtypesPointer[llama_grammar_element] - ], # NOTE: This might be wrong type sig - n_rules: Union[ctypes.c_size_t, int], - start_rule_index: Union[ctypes.c_size_t, int], - /, -) -> Optional[llama_grammar_p]: - """Initialize a grammar from a set of rules.""" - ... + +# // user code can implement the interface below in order to create custom llama_sampler +# struct llama_sampler_i { +# const char * (*name) (const struct llama_sampler * smpl); // can be NULL +# void (*accept)( struct llama_sampler * smpl, llama_token token); // can be NULL +# void (*apply) ( struct llama_sampler * smpl, llama_token_data_array * cur_p); // required +# void (*reset) ( struct llama_sampler * smpl); // can be NULL +# struct llama_sampler * (*clone) (const struct llama_sampler * smpl); // can be NULL if ctx is NULL +# void (*free) ( struct llama_sampler * smpl); // can be NULL if ctx is NULL +# +# // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph +# //void (*apply_ggml) (struct llama_sampler * smpl, ...); +# }; +class llama_sampler_i(ctypes.Structure): ... -# LLAMA_API void llama_grammar_free(struct llama_grammar * grammar); -@ctypes_function( - "llama_grammar_free", - [llama_grammar_p], - None, +# struct llama_sampler { +# struct llama_sampler_i * iface; +# llama_sampler_context_t ctx; +# }; +class llama_sampler(ctypes.Structure): + _fields_ = [ + ("iface", ctypes.POINTER(llama_sampler_i)), + ("ctx", llama_sampler_context_t), + ] + + +if TYPE_CHECKING: + llama_sampler_p = CtypesPointer[llama_sampler] + +llama_sampler_p_ctypes = ctypes.POINTER(llama_sampler) + +llama_sampler_i_name = ctypes.CFUNCTYPE(ctypes.c_char_p, llama_sampler_p_ctypes) +llama_sampler_i_accept = ctypes.CFUNCTYPE(None, llama_sampler_p_ctypes, llama_token) +llama_sampler_i_apply = ctypes.CFUNCTYPE( + None, llama_sampler_p_ctypes, llama_token_data_array_p ) -def llama_grammar_free(grammar: llama_grammar_p, /): - """Free a grammar.""" - ... +llama_sampler_i_reset = ctypes.CFUNCTYPE(None, llama_sampler_p_ctypes) +llama_sampler_i_clone = ctypes.CFUNCTYPE(llama_sampler_p_ctypes, llama_sampler_p_ctypes) +llama_sampler_i_free = ctypes.CFUNCTYPE(None, llama_sampler_p_ctypes) +llama_sampler_i._fields_ = [ + ("name", llama_sampler_i_name), + ("accept", llama_sampler_i_accept), + ("apply", llama_sampler_i_apply), + ("reset", llama_sampler_i_reset), + ("clone", llama_sampler_i_clone), + ("free", llama_sampler_i_free), +] -# LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar); + +# // mirror of llama_sampler_i: +# LLAMA_API const char * llama_sampler_name (const struct llama_sampler * smpl); @ctypes_function( - "llama_grammar_copy", - [llama_grammar_p], - llama_grammar_p, + "llama_sampler_name", + [llama_sampler_p_ctypes], + ctypes.c_char_p, ) -def llama_grammar_copy(grammar: llama_grammar_p, /) -> llama_grammar_p: - """Copy a grammar.""" - ... +def llama_sampler_name(smpl: llama_sampler_p, /) -> bytes: ... -# /// @details Apply constraints from grammar -# LLAMA_API void llama_grammar_sample( -# const struct llama_grammar * grammar, -# const struct llama_context * ctx, -# llama_token_data_array * candidates); +# LLAMA_API void llama_sampler_accept( struct llama_sampler * smpl, llama_token token); @ctypes_function( - "llama_grammar_sample", - [ - llama_grammar_p, - llama_context_p_ctypes, - llama_token_data_array_p, - ], + "llama_sampler_accept", + [llama_sampler_p_ctypes, llama_token], None, ) -def llama_grammar_sample( - grammar: llama_grammar_p, - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - /, -): - """Apply constraints from grammar""" - ... +def llama_sampler_accept(smpl: llama_sampler_p, token: Union[llama_token, int], /): ... -# LLAMA_API DEPRECATED(void llama_sample_grammar( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# const struct llama_grammar * grammar), -# "use llama_grammar_sample instead"); +# LLAMA_API void llama_sampler_apply ( struct llama_sampler * smpl, llama_token_data_array * cur_p); @ctypes_function( - "llama_sample_grammar", - [llama_context_p_ctypes, llama_token_data_array_p, llama_grammar_p], + "llama_sampler_apply", + [llama_sampler_p_ctypes, llama_token_data_array_p], None, ) -def llama_sample_grammar( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - grammar, # type: llama_grammar_p - /, -): - """Apply constraints from grammar - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - grammar: A grammar object containing the rules and constraints to apply to the generated text. - """ - ... +def llama_sampler_apply( + smpl: llama_sampler_p, cur_p: CtypesArray[llama_token_data_array], / +): ... -# /// @details Accepts the sampled token into the grammar -# LLAMA_API void llama_grammar_accept_token( -# struct llama_grammar * grammar, -# struct llama_context * ctx, -# llama_token token); +# LLAMA_API void llama_sampler_reset ( struct llama_sampler * smpl); @ctypes_function( - "llama_grammar_accept_token", - [llama_grammar_p, llama_context_p_ctypes, llama_token], + "llama_sampler_reset", + [llama_sampler_p_ctypes], None, ) -def llama_grammar_accept_token( - grammar: llama_grammar_p, - ctx: llama_context_p, - token: Union[llama_token, int], - /, -): - """Accepts the sampled token into the grammar""" - ... +def llama_sampler_reset(smpl: llama_sampler_p, /): ... -# // -# // Sampling functions -# // +# LLAMA_API struct llama_sampler * llama_sampler_clone (const struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_clone", + [llama_sampler_p_ctypes], + llama_sampler_p_ctypes, +) +def llama_sampler_clone(smpl: llama_sampler_p, /) -> llama_sampler_p: ... -# // Sets the current rng seed. -# LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); +# // important: do not free if the sampler has been added to a llama_sampler_chain (via llama_sampler_chain_add) +# LLAMA_API void llama_sampler_free ( struct llama_sampler * smpl); @ctypes_function( - "llama_set_rng_seed", - [llama_context_p_ctypes, ctypes.c_uint32], + "llama_sampler_free", + [llama_sampler_p_ctypes], None, ) -def llama_set_rng_seed(ctx: llama_context_p, seed: Union[ctypes.c_uint32, int], /): - """Sets the current rng seed.""" - ... +def llama_sampler_free(smpl: llama_sampler_p, /): ... -# /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. -# /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. -# LLAMA_API void llama_sample_repetition_penalties( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# const llama_token * last_tokens, -# size_t penalty_last_n, -# float penalty_repeat, -# float penalty_freq, -# float penalty_present); -@ctypes_function( - "llama_sample_repetition_penalties", - [ - llama_context_p_ctypes, - llama_token_data_array_p, - llama_token_p, - ctypes.c_size_t, - ctypes.c_float, - ctypes.c_float, - ctypes.c_float, - ], - None, +# // llama_sampler_chain +# // a type of llama_sampler that can chain multiple samplers one after another +# +# LLAMA_API struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params); +@ctypes_function( + "llama_sampler_chain_init", + [llama_sampler_chain_params], + llama_sampler_p_ctypes, ) -def llama_sample_repetition_penalties( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - last_tokens_data: CtypesArray[llama_token], - penalty_last_n: Union[ctypes.c_size_t, int], - penalty_repeat: Union[ctypes.c_float, float], - penalty_freq: Union[ctypes.c_float, float], - penalty_present: Union[ctypes.c_float, float], - /, -): - """Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. - Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. - """ - ... +def llama_sampler_chain_init( + params: llama_sampler_chain_params, / +) -> llama_sampler_p: ... -# /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 -# /// @param logits Logits extracted from the original generation context. -# /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. -# /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance. -# LLAMA_API void llama_sample_apply_guidance( -# struct llama_context * ctx, -# float * logits, -# float * logits_guidance, -# float scale); +# // important: takes ownership of the sampler object and will free it when llama_sampler_free is called +# LLAMA_API void llama_sampler_chain_add( struct llama_sampler * chain, struct llama_sampler * smpl); @ctypes_function( - "llama_sample_apply_guidance", - [ - llama_context_p_ctypes, - ctypes.POINTER(ctypes.c_float), - ctypes.POINTER(ctypes.c_float), - ctypes.c_float, - ], + "llama_sampler_chain_add", + [llama_sampler_p_ctypes, llama_sampler_p_ctypes], None, ) -def llama_sample_apply_guidance( - ctx: llama_context_p, - logits: CtypesArray[ctypes.c_float], - logits_guidance: CtypesArray[ctypes.c_float], - scale: Union[ctypes.c_float, float], - /, -): - """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" - ... +def llama_sampler_chain_add(chain: llama_sampler_p, smpl: llama_sampler_p, /): ... -# /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. -# LLAMA_API void llama_sample_softmax( -# struct llama_context * ctx, -# llama_token_data_array * candidates); +# LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i); @ctypes_function( - "llama_sample_softmax", - [llama_context_p_ctypes, llama_token_data_array_p], - None, + "llama_sampler_chain_get", + [llama_sampler_p_ctypes, ctypes.c_int32], + llama_sampler_p_ctypes, ) -def llama_sample_softmax( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - /, -): - """Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.""" - ... +def llama_sampler_chain_get( + chain: llama_sampler_p, i: Union[ctypes.c_int32, int], / +) -> llama_sampler_p: ... -# /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 -# LLAMA_API void llama_sample_top_k( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# int32_t k, -# size_t min_keep); +# LLAMA_API int llama_sampler_chain_n (const struct llama_sampler * chain); @ctypes_function( - "llama_sample_top_k", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_int32, ctypes.c_size_t], - None, + "llama_sampler_chain_n", + [llama_sampler_p_ctypes], + ctypes.c_int, ) -def llama_sample_top_k( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - k: Union[ctypes.c_int, int], - min_keep: Union[ctypes.c_size_t, int], - /, -): - """Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" - ... +def llama_sampler_chain_n(chain: llama_sampler_p, /) -> int: ... + + +# // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed +# LLAMA_API struct llama_sampler * llama_sampler_chain_remove( struct llama_sampler * chain, int32_t i); +@ctypes_function( + "llama_sampler_chain_remove", + [llama_sampler_p_ctypes, ctypes.c_int32], + llama_sampler_p_ctypes, +) +def llama_sampler_chain_remove( + chain: llama_sampler_p, i: Union[ctypes.c_int32, int], / +) -> llama_sampler_p: ... + + +# // available samplers: +# +# LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void); +@ctypes_function("llama_sampler_init_greedy", [], llama_sampler_p_ctypes) +def llama_sampler_init_greedy() -> llama_sampler_p: ... + + +# LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed); +@ctypes_function("llama_sampler_init_dist", [ctypes.c_uint32], llama_sampler_p_ctypes) +def llama_sampler_init_dist(seed: int) -> llama_sampler_p: ... + + +# /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. +# LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void); +@ctypes_function("llama_sampler_init_softmax", [], llama_sampler_p_ctypes) +def llama_sampler_init_softmax() -> llama_sampler_p: ... + + +# /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +# LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); +@ctypes_function("llama_sampler_init_top_k", [ctypes.c_int32], llama_sampler_p_ctypes) +def llama_sampler_init_top_k(k: int) -> llama_sampler_p: ... # /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 -# LLAMA_API void llama_sample_top_p( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float p, -# size_t min_keep); +# LLAMA_API struct llama_sampler * llama_sampler_init_top_p (float p, size_t min_keep); @ctypes_function( - "llama_sample_top_p", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], - None, + "llama_sampler_init_top_p", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, ) -def llama_sample_top_p( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - p: Union[ctypes.c_float, float], - min_keep: Union[ctypes.c_size_t, int], - /, -): - """Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" - ... +def llama_sampler_init_top_p(p: float, min_keep: int) -> llama_sampler_p: ... # /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 -# LLAMA_API void llama_sample_min_p( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float p, -# size_t min_keep); +# LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep); @ctypes_function( - "llama_sample_min_p", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], - None, + "llama_sampler_init_min_p", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, ) -def llama_sample_min_p( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - p: Union[ctypes.c_float, float], - min_keep: Union[ctypes.c_size_t, int], - /, -): - """Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841""" - ... +def llama_sampler_init_min_p(p: float, min_keep: int) -> llama_sampler_p: ... # /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. -# LLAMA_API void llama_sample_tail_free( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float z, -# size_t min_keep); +# LLAMA_API struct llama_sampler * llama_sampler_init_tail_free (float z, size_t min_keep); @ctypes_function( - "llama_sample_tail_free", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], - None, + "llama_sampler_init_tail_free", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, ) -def llama_sample_tail_free( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - z: Union[ctypes.c_float, float], - min_keep: Union[ctypes.c_size_t, int], - /, -): - """Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.""" - ... +def llama_sampler_init_tail_free(z: float, min_keep: int) -> llama_sampler_p: ... # /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. -# LLAMA_API void llama_sample_typical( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float p, -# size_t min_keep); +# LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep); @ctypes_function( - "llama_sample_typical", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float, ctypes.c_size_t], - None, + "llama_sampler_init_typical", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, ) -def llama_sample_typical( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - p: Union[ctypes.c_float, float], - min_keep: Union[ctypes.c_size_t, int], - /, -): - """Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.""" - ... +def llama_sampler_init_typical(p: float, min_keep: int) -> llama_sampler_p: ... -# /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772. -# LLAMA_API void llama_sample_entropy( -# struct llama_context * ctx, -# llama_token_data_array * candidates_p, -# float min_temp, -# float max_temp, -# float exponent_val); -@ctypes_function( - "llama_sample_entropy", - [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_float, - ctypes.c_float, - ], - None, -) -def llama_sample_entropy( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - min_temp: Union[ctypes.c_float, float], - max_temp: Union[ctypes.c_float, float], - exponent_val: Union[ctypes.c_float, float], - /, -): - """Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.""" - ... +# LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t); +@ctypes_function("llama_sampler_init_temp", [ctypes.c_float], llama_sampler_p_ctypes) +def llama_sampler_init_temp(t: float) -> llama_sampler_p: ... -# LLAMA_API void llama_sample_temp( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float temp); +# /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772. +# LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext (float t, float delta, float exponent); @ctypes_function( - "llama_sample_temp", - [llama_context_p_ctypes, llama_token_data_array_p, ctypes.c_float], - None, + "llama_sampler_init_temp_ext", + [ctypes.c_float, ctypes.c_float, ctypes.c_float], + llama_sampler_p_ctypes, ) -def llama_sample_temp( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - temp: Union[ctypes.c_float, float], - /, -): - """Temperature sampling described in academic paper "Generating Long Sequences with Sparse Transformers" https://arxiv.org/abs/1904.10509 - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - temp: The temperature value to use for the sampling. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - """ - ... +def llama_sampler_init_temp_ext( + t: float, delta: float, exponent: float +) -> llama_sampler_p: ... # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -3427,46 +3230,20 @@ def llama_sample_temp( # /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. # /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. # /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. -# LLAMA_API llama_token llama_sample_token_mirostat( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float tau, -# float eta, -# int32_t m, -# float * mu); +# LLAMA_API struct llama_sampler * llama_sampler_init_mirostat( +# int32_t n_vocab, +# uint32_t seed, +# float tau, +# float eta, +# int32_t m); @ctypes_function( - "llama_sample_token_mirostat", - [ - llama_context_p_ctypes, - llama_token_data_array_p, - ctypes.c_float, - ctypes.c_float, - ctypes.c_int32, - ctypes.POINTER(ctypes.c_float), - ], - llama_token, + "llama_sampler_init_mirostat", + [ctypes.c_int32, ctypes.c_uint32, ctypes.c_float, ctypes.c_float, ctypes.c_int32], + llama_sampler_p_ctypes, ) -def llama_sample_token_mirostat( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - tau: Union[ctypes.c_float, float], - eta: Union[ctypes.c_float, float], - m: Union[ctypes.c_int, int], - mu: CtypesPointerOrRef[ctypes.c_float], - /, -) -> int: - """Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - m: The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. - mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. - """ - ... +def llama_sampler_init_mirostat( + n_vocab: int, seed: int, tau: float, eta: float, m: int, / +) -> llama_sampler_p: ... # /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -3474,83 +3251,116 @@ def llama_sample_token_mirostat( # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. # /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. # /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. -# LLAMA_API llama_token llama_sample_token_mirostat_v2( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float tau, -# float eta, -# float * mu); -@ctypes_function( - "llama_sample_token_mirostat_v2", +# LLAMA_API struct llama_sampler * llama_sampler_init_mirostat_v2( +# uint32_t seed, +# float tau, +# float eta); +@ctypes_function( + "llama_sampler_init_mirostat_v2", + [ctypes.c_uint32, ctypes.c_float, ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_mirostat_v2( + seed: int, tau: float, eta: float, / +) -> llama_sampler_p: ... + + +# LLAMA_API struct llama_sampler * llama_sampler_init_grammar( +# const struct llama_model * model, +# const char * grammar_str, +# const char * grammar_root); +@ctypes_function( + "llama_sampler_init_grammar", + [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_char_p], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar( + model: llama_model_p, grammar_str: bytes, grammar_root: bytes, / +) -> llama_sampler_p: ... + + +# LLAMA_API struct llama_sampler * llama_sampler_init_penalties( +# int32_t n_vocab, // llama_n_vocab() +# llama_token special_eos_id, // llama_token_eos() +# llama_token linefeed_id, // llama_token_nl() +# int32_t penalty_last_n, // last n tokens to penalize (0 = disable penalty, -1 = context size) +# float penalty_repeat, // 1.0 = disabled +# float penalty_freq, // 0.0 = disabled +# float penalty_present, // 0.0 = disabled +# bool penalize_nl, // consider newlines as a repeatable token +# bool ignore_eos); // ignore the end-of-sequence token +@ctypes_function( + "llama_sampler_init_penalties", [ - llama_context_p_ctypes, - llama_token_data_array_p, + ctypes.c_int32, + llama_token, + llama_token, + ctypes.c_int32, ctypes.c_float, ctypes.c_float, - ctypes.POINTER(ctypes.c_float), - ], - llama_token, -) -def llama_sample_token_mirostat_v2( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] + ctypes.c_float, + ctypes.c_bool, + ctypes.c_bool, ], - tau: Union[ctypes.c_float, float], - eta: Union[ctypes.c_float, float], - mu: CtypesPointerOrRef[ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_penalties( + n_vocab: int, + special_eos_id: int, + linefeed_id: int, + penalty_last_n: int, + penalty_repeat: float, + penalty_freq: float, + penalty_present: float, + penalize_nl: bool, + ignore_eos: bool, /, -) -> int: - """Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. - """ - ... +) -> llama_sampler_p: ... -# /// @details Selects the token with the highest probability. -# /// Does not compute the token probabilities. Use llama_sample_softmax() instead. -# LLAMA_API llama_token llama_sample_token_greedy( -# struct llama_context * ctx, -# llama_token_data_array * candidates); +# LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias( +# int32_t n_vocab, +# int32_t n_logit_bias, +# const llama_logit_bias * logit_bias); @ctypes_function( - "llama_sample_token_greedy", - [llama_context_p_ctypes, llama_token_data_array_p], - llama_token, + "llama_sampler_init_logit_bias", + [ctypes.c_int32, ctypes.c_int32, llama_logit_bias_p], + llama_sampler_p_ctypes, ) -def llama_sample_token_greedy( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - /, -) -> int: - """Selects the token with the highest probability.""" - ... +def llama_sampler_init_logit_bias( + n_vocab: int, n_logit_bias: int, logit_bias: CtypesArray[llama_logit_bias], / +) -> llama_sampler_p: ... -# /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx. -# LLAMA_API llama_token llama_sample_token( -# struct llama_context * ctx, -# llama_token_data_array * candidates); +# // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise +# LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl); @ctypes_function( - "llama_sample_token", - [llama_context_p_ctypes, llama_token_data_array_p], + "llama_sampler_get_seed", + [llama_sampler_p_ctypes], + ctypes.c_uint32, +) +def llama_sampler_get_seed(smpl: llama_sampler_p, /) -> int: ... + + +# /// @details Sample and accept a token from the idx-th output of the last evaluation +# // +# // Shorthand for: +# // const auto * logits = llama_get_logits_ith(ctx, idx); +# // llama_token_data_array cur_p = { ... init from logits ... }; +# // llama_sampler_apply(smpl, &cur_p); +# // auto token = cur_p.data[cur_p.selected].id; +# // llama_sampler_accept(smpl, token); +# // return token; +# // Returns the sampled token +# LLAMA_API llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx); +@ctypes_function( + "llama_sampler_sample", + [llama_sampler_p_ctypes, llama_context_p_ctypes, ctypes.c_int32], llama_token, ) -def llama_sample_token( - ctx: llama_context_p, - candidates: Union[ - CtypesArray[llama_token_data_array], CtypesPointerOrRef[llama_token_data_array] - ], - /, -) -> int: - """Randomly selects a token from the candidates based on their probabilities.""" - ... +def llama_sampler_sample( + smpl: llama_sampler_p, ctx: llama_context_p, idx: int, / +) -> int: ... # // @@ -3600,79 +3410,131 @@ def llama_split_prefix( ... -# Performance information +# // Print system information +# LLAMA_API const char * llama_print_system_info(void); +@ctypes_function("llama_print_system_info", [], ctypes.c_char_p) +def llama_print_system_info() -> bytes: ... -# LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); +# // Set callback for all future logging events. +# // If this is not called, or NULL is supplied, everything is output on stderr. +# LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); @ctypes_function( - "llama_get_timings", - [llama_context_p_ctypes], - llama_timings, + "llama_log_set", + [ctypes.c_void_p, ctypes.c_void_p], + None, ) -def llama_get_timings(ctx: llama_context_p, /) -> llama_timings: - """Get performance information""" +def llama_log_set( + log_callback: Optional[CtypesFuncPointer], + user_data: ctypes.c_void_p, + /, +): + """Set callback for all future logging events. + + If this is not called, or NULL is supplied, everything is output on stderr.""" ... -# LLAMA_API void llama_print_timings(struct llama_context * ctx); +# // +# // Performance utils +# // +# // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements. +# // + + +# struct llama_perf_context_data { +# double t_start_ms; +# double t_load_ms; +# double t_p_eval_ms; +# double t_eval_ms; +# +# int32_t n_p_eval; +# int32_t n_eval; +# }; +class llama_perf_context_data(ctypes.Structure): + _fields_ = [ + ("t_start_ms", ctypes.c_double), + ("t_load_ms", ctypes.c_double), + ("t_p_eval_ms", ctypes.c_double), + ("t_eval_ms", ctypes.c_double), + ("n_p_eval", ctypes.c_int32), + ("n_eval", ctypes.c_int32), + ] + + +# struct llama_perf_sampler_data { +# double t_sample_ms; +# +# int32_t n_sample; +# }; +class llama_perf_sampler_data(ctypes.Structure): + _fields_ = [ + ("t_sample_ms", ctypes.c_double), + ("n_sample", ctypes.c_int32), + ] + + +# LLAMA_API struct llama_perf_context_data llama_perf_context (const struct llama_context * ctx); @ctypes_function( - "llama_print_timings", + "llama_perf_context", + [llama_context_p_ctypes], + llama_perf_context_data, +) +def llama_perf_context(ctx: llama_context_p, /) -> llama_perf_context_data: ... + + +# LLAMA_API void llama_perf_context_print(const struct llama_context * ctx); +@ctypes_function( + "llama_perf_context_print", [llama_context_p_ctypes], None, ) -def llama_print_timings(ctx: llama_context_p, /): - """Print performance information""" - ... +def llama_perf_context_print(ctx: llama_context_p, /): ... -# LLAMA_API void llama_reset_timings(struct llama_context * ctx); +# LLAMA_API void llama_perf_context_reset( struct llama_context * ctx); @ctypes_function( - "llama_reset_timings", + "llama_perf_context_reset", [llama_context_p_ctypes], None, ) -def llama_reset_timings(ctx: llama_context_p, /): - """Reset performance information""" - ... +def llama_perf_context_reset(ctx: llama_context_p, /): ... -# Print system information -# LLAMA_API const char * llama_print_system_info(void); +# // NOTE: the following work only with samplers constructed via llama_sampler_chain_init +# LLAMA_API struct llama_perf_sampler_data llama_perf_sampler (const struct llama_sampler * chain); @ctypes_function( - "llama_print_system_info", - [], - ctypes.c_char_p, + "llama_perf_sampler", + [llama_sampler_p_ctypes], + llama_perf_sampler_data, ) -def llama_print_system_info() -> bytes: - """Print system information""" - ... +def llama_perf_sampler(chain: llama_sampler_p, /) -> llama_perf_sampler_data: ... -# NOTE: THIS IS CURRENTLY BROKEN AS ggml_log_callback IS NOT EXPOSED IN LLAMA.H -# // Set callback for all future logging events. -# // If this is not called, or NULL is supplied, everything is output on stderr. -# LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); +# LLAMA_API void llama_perf_sampler_print(const struct llama_sampler * chain); @ctypes_function( - "llama_log_set", - [ctypes.c_void_p, ctypes.c_void_p], + "llama_perf_sampler_print", + [llama_sampler_p_ctypes], None, ) -def llama_log_set( - log_callback: Optional[CtypesFuncPointer], - user_data: ctypes.c_void_p, - /, -): - """Set callback for all future logging events. +def llama_perf_sampler_print(chain: llama_sampler_p, /): ... - If this is not called, or NULL is supplied, everything is output on stderr.""" - ... + +# LLAMA_API void llama_perf_sampler_reset( struct llama_sampler * chain); +@ctypes_function( + "llama_perf_sampler_reset", + [llama_sampler_p_ctypes], + None, +) +def llama_perf_sampler_reset(chain: llama_sampler_p, /): ... -# LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); +# LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx); @ctypes_function( - "llama_dump_timing_info_yaml", - [ctypes.c_void_p, llama_context_p_ctypes], + "llama_perf_dump_yaml", + [ctypes.POINTER(ctypes.c_void_p), llama_context_p_ctypes], None, ) -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p, /): - ... +def llama_perf_dump_yaml( + stream: ctypes.POINTER(ctypes.c_void_p), ctx: llama_context_p, / +): ... diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 4cd52c2d5..785aa88d7 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -2,11 +2,6 @@ # flake8: noqa from pathlib import Path -import sys -import ctypes -import enum -import typing -import dataclasses from itertools import groupby from typing import ( @@ -18,882 +13,16 @@ Union, ) -import llama_cpp.llama_cpp as llama_cpp - -class GrammarElementType(enum.IntEnum): - END = llama_cpp.LLAMA_GRETYPE_END - ALT = llama_cpp.LLAMA_GRETYPE_ALT - RULE_REF = llama_cpp.LLAMA_GRETYPE_RULE_REF - CHAR = llama_cpp.LLAMA_GRETYPE_CHAR - CHAR_NOT = llama_cpp.LLAMA_GRETYPE_CHAR_NOT - CHAR_RNG_UPPER = llama_cpp.LLAMA_GRETYPE_CHAR_RNG_UPPER - CHAR_ALT = llama_cpp.LLAMA_GRETYPE_CHAR_ALT - CHAR_ANY = llama_cpp.LLAMA_GRETYPE_CHAR_ANY - - -@dataclasses.dataclass -class GrammarElement: - type: GrammarElementType - value: int - - -@dataclasses.dataclass -class ParseState: - symbol_ids: typing.Dict[str, int] = dataclasses.field(default_factory=dict) - rules: typing.List[typing.List[GrammarElement]] = dataclasses.field(default_factory=list) - - -# static std::pair decode_utf8(const char * src) { -# static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; -# uint8_t first_byte = static_cast(*src); -# uint8_t highbits = first_byte >> 4; -# int len = lookup[highbits]; -# uint8_t mask = (1 << (8 - len)) - 1; -# uint32_t value = first_byte & mask; -# const char * end = src + len; // may overrun! -# const char * pos = src + 1; -# for ( ; pos < end && *pos; pos++) { -# value = (value << 6) + (static_cast(*pos) & 0x3F); -# } -# return std::make_pair(value, pos); -# } -def decode_utf8(src: str) -> typing.Tuple[int, str]: - lookup: list[int] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4] - first_byte: int = ord(src[0]) - highbits: int = first_byte >> 4 - length: int = lookup[highbits] - mask: int = (1 << (8 - length)) - 1 - value: int = first_byte & mask - end: int = min(len(src), length) # Prevent overrun - - pos: int = 1 - for pos in range(1, end): - if not src[pos]: - break - value = (value << 6) + (ord(src[pos]) & 0x3F) - - return value, src[pos:] if pos < len(src) else "" - - -# static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) { -# uint32_t next_id = static_cast(state.symbol_ids.size()); -# auto result = state.symbol_ids.emplace(std::string(src, len), next_id); -# return result.first->second; -# } -def get_symbol_id(state: ParseState, name: str) -> int: - next_id = len(state.symbol_ids) - return state.symbol_ids.setdefault(name, next_id) - - -# static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) { -# uint32_t next_id = static_cast(state.symbol_ids.size()); -# state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id; -# return next_id; -# } -def generate_symbol_id(state: ParseState, base_name: str) -> int: - next_id = len(state.symbol_ids) - state.symbol_ids[f"{base_name}_{next_id}"] = next_id - return next_id - - -# static void add_rule( -# parse_state & state, -# uint32_t rule_id, -# const std::vector & rule) { -# if (state.rules.size() <= rule_id) { -# state.rules.resize(rule_id + 1); -# } -# state.rules[rule_id] = rule; -# } -def add_rule(state: ParseState, rule_id: int, rule: typing.List[GrammarElement]) -> None: - if len(state.rules) <= rule_id: - state.rules.extend([[]] * (rule_id + 1 - len(state.rules))) - state.rules[rule_id] = rule - - -# static bool is_digit_char(char c) { -# return '0' <= c && c <= '9'; -# } -def is_digit_char(c: str) -> bool: - return "0" <= c <= "9" - - -# static bool is_word_char(char c) { -# return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c); -# } -def is_word_char(c: str) -> bool: - return ("a" <= c <= "z") or ("A" <= c <= "Z") or c == "-" or is_digit_char(c) - - -# static std::pair parse_hex(const char * src, int size) { -# const char * pos = src; -# const char * end = src + size; -# uint32_t value = 0; -# for ( ; pos < end && *pos; pos++) { -# value <<= 4; -# char c = *pos; -# if ('a' <= c && c <= 'f') { -# value += c - 'a' + 10; -# } else if ('A' <= c && c <= 'F') { -# value += c - 'A' + 10; -# } else if ('0' <= c && c <= '9') { -# value += c - '0'; -# } else { -# break; -# } -# } -# if (pos != end) { -# throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src); -# } -# return std::make_pair(value, pos); -# } -def parse_hex(src: str, size: int) -> typing.Tuple[int, str]: - pos = 0 - value = 0 - for _ in range(size): - value <<= 4 - c = src[pos] - if "a" <= c <= "f": - value += ord(c) - ord("a") + 10 - elif "A" <= c <= "F": - value += ord(c) - ord("A") + 10 - elif "0" <= c <= "9": - value += ord(c) - ord("0") - else: - break - pos += 1 - if pos != size: - raise ValueError(f"expecting {size} hex chars at {src}") - return value, src[pos:] - - -# static const char * parse_space(const char * src, bool newline_ok) { -# const char * pos = src; -# while (*pos == ' ' || *pos == '\t' || *pos == '#' || -# (newline_ok && (*pos == '\r' || *pos == '\n'))) { -# if (*pos == '#') { -# while (*pos && *pos != '\r' && *pos != '\n') { -# pos++; -# } -# } else { -# pos++; -# } -# } -# return pos; -# } -def parse_space(src: str, newline_ok: bool) -> str: - pos = src - while pos and (pos[0] in (' ', '\t', '#') or (newline_ok and pos[0] in ('\r', '\n'))): - if pos[0] == "#": - while pos and pos[0] not in ("\r", "\n"): - pos = pos[1:] - else: - pos = pos[1:] - return pos - - -# static const char * parse_name(const char * src) { -# const char * pos = src; -# while (is_word_char(*pos)) { -# pos++; -# } -# if (pos == src) { -# throw std::runtime_error(std::string("expecting name at ") + src); -# } -# return pos; -# } -def parse_name(src: str) -> typing.Tuple[str, str]: - pos = src - while pos and is_word_char(pos[0]): - pos = pos[1:] - if pos == src: - raise ValueError(f"expecting name at {src}") - return src[:len(src) - len(pos)], pos - -# static const char * parse_int(const char * src) { -# const char * pos = src; -# while (is_digit_char(*pos)) { -# pos++; -# } -# if (pos == src) { -# throw std::runtime_error(std::string("expecting integer at ") + src); -# } -# return pos; -# } -def parse_int(src: str) -> typing.Tuple[int, str]: - pos = src - while pos and is_digit_char(pos[0]): - pos = pos[1:] - if pos == src: - raise ValueError(f"expecting integer at {src}") - return int(src[:len(src) - len(pos)]), pos - - -# static std::pair parse_char(const char * src) { -# if (*src == '\\') { -# switch (src[1]) { -# case 'x': return parse_hex(src + 2, 2); -# case 'u': return parse_hex(src + 2, 4); -# case 'U': return parse_hex(src + 2, 8); -# case 't': return std::make_pair('\t', src + 2); -# case 'r': return std::make_pair('\r', src + 2); -# case 'n': return std::make_pair('\n', src + 2); -# case '\\': -# case '"': -# case '[': -# case ']': -# return std::make_pair(src[1], src + 2); -# default: -# throw std::runtime_error(std::string("unknown escape at ") + src); -# } -# } else if (*src) { -# return decode_utf8(src); -# } -# throw std::runtime_error("unexpected end of input"); -# } -def parse_char(src: str) -> typing.Tuple[int, str]: - if not src: - raise ValueError("unexpected end of input") - if src[0] == "\\": - if src[1] == "x": - return parse_hex(src[2:], 2) - elif src[1] == "u": - return parse_hex(src[2:], 4) - elif src[1] == "U": - return parse_hex(src[2:], 8) - elif src[1] == "t": - return ord("\t"), src[2:] - elif src[1] == "r": - return ord("\r"), src[2:] - elif src[1] == "n": - return ord("\n"), src[2:] - elif src[1] in ('\\', '"', '[', ']'): - return ord(src[1]), src[2:] - else: - raise ValueError(f"unknown escape at {src}") - return decode_utf8(src) - -# static const char * parse_sequence( -# parse_state & state, -# const char * src, -# const std::string & rule_name, -# std::vector & out_elements, -# bool is_nested) { -# size_t last_sym_start = out_elements.size(); -# const char * pos = src; -# -# auto handle_repetitions = [&](int min_times, int max_times) { -# -# if (last_sym_start == out_elements.size()) { -# throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos); -# } -# -# // apply transformation to previous symbol (last_sym_start to end) according to -# // the following rewrite rules: -# // S{m,n} --> S S S (m times) S'(n-m) -# // S'(x) ::= S S'(x-1) | -# // (... n-m definitions of these S' rules ...) -# // S'(1) ::= S | -# // S{m,} --> S S S (m times) S' -# // S' ::= S S' | -# // S* --> S{0,} -# // --> S' ::= S S' | -# // S+ --> S{1,} -# // --> S S' -# // S' ::= S S' | -# // S? --> S{0,1} -# // --> S' -# // S' ::= S | -# -# std::vector previous_elements(out_elements.begin() + last_sym_start, out_elements.end()); -# if (min_times == 0) { -# out_elements.resize(last_sym_start); -# } else { -# // Repeat the previous elements (min_times - 1) times -# for (int i = 1; i < min_times; i++) { -# out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end()); -# } -# } -# -# uint32_t last_rec_rule_id = 0; -# auto n_opt = max_times < 0 ? 1 : max_times - min_times; -# -# std::vector rec_rule(previous_elements); -# for (int i = 0; i < n_opt; i++) { -# rec_rule.resize(previous_elements.size()); -# uint32_t rec_rule_id = generate_symbol_id(state, rule_name); -# if (i > 0 || max_times < 0) { -# rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id}); -# } -# rec_rule.push_back({LLAMA_GRETYPE_ALT, 0}); -# rec_rule.push_back({LLAMA_GRETYPE_END, 0}); -# add_rule(state, rec_rule_id, rec_rule); -# last_rec_rule_id = rec_rule_id; -# } -# if (n_opt > 0) { -# out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id}); -# } -# }; -# -# while (*pos) { -# if (*pos == '"') { // literal string -# pos++; -# last_sym_start = out_elements.size(); -# while (*pos != '"') { -# if (!*pos) { -# throw std::runtime_error("unexpected end of input"); -# } -# auto char_pair = parse_char(pos); -# pos = char_pair.second; -# out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first}); -# } -# pos = parse_space(pos + 1, is_nested); -# } else if (*pos == '[') { // char range(s) -# pos++; -# enum llama_gretype start_type = LLAMA_GRETYPE_CHAR; -# if (*pos == '^') { -# pos++; -# start_type = LLAMA_GRETYPE_CHAR_NOT; -# } -# last_sym_start = out_elements.size(); -# while (*pos != ']') { -# if (!*pos) { -# throw std::runtime_error("unexpected end of input"); -# } -# auto char_pair = parse_char(pos); -# pos = char_pair.second; -# enum llama_gretype type = last_sym_start < out_elements.size() -# ? LLAMA_GRETYPE_CHAR_ALT -# : start_type; -# -# out_elements.push_back({type, char_pair.first}); -# if (pos[0] == '-' && pos[1] != ']') { -# if (!pos[1]) { -# throw std::runtime_error("unexpected end of input"); -# } -# auto endchar_pair = parse_char(pos + 1); -# pos = endchar_pair.second; -# out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first}); -# } -# } -# pos = parse_space(pos + 1, is_nested); -# } else if (is_word_char(*pos)) { // rule reference -# const char * name_end = parse_name(pos); -# uint32_t ref_rule_id = get_symbol_id(state, pos, name_end - pos); -# pos = parse_space(name_end, is_nested); -# last_sym_start = out_elements.size(); -# out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id}); -# } else if (*pos == '(') { // grouping -# // parse nested alternates into synthesized rule -# pos = parse_space(pos + 1, true); -# uint32_t sub_rule_id = generate_symbol_id(state, rule_name); -# pos = parse_alternates(state, pos, rule_name, sub_rule_id, true); -# last_sym_start = out_elements.size(); -# // output reference to synthesized rule -# out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); -# if (*pos != ')') { -# throw std::runtime_error(std::string("expecting ')' at ") + pos); -# } -# pos = parse_space(pos + 1, is_nested); -# } else if (*pos == '.') { // any char -# last_sym_start = out_elements.size(); -# out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0}); -# pos = parse_space(pos + 1, is_nested); -# } else if (*pos == '*') { -# pos = parse_space(pos + 1, is_nested); -# handle_repetitions(0, -1); -# } else if (*pos == '+') { -# pos = parse_space(pos + 1, is_nested); -# handle_repetitions(1, -1); -# } else if (*pos == '?') { -# pos = parse_space(pos + 1, is_nested); -# handle_repetitions(0, 1); -# } else if (*pos == '{') { -# pos = parse_space(pos + 1, is_nested); -# -# if (!is_digit_char(*pos)) { -# throw std::runtime_error(std::string("expecting an int at ") + pos); -# } -# const char * int_end = parse_int(pos); -# int min_times = std::stoul(std::string(pos, int_end - pos)); -# pos = parse_space(int_end, is_nested); -# -# int max_times = -1; -# -# if (*pos == '}') { -# max_times = min_times; -# pos = parse_space(pos + 1, is_nested); -# } else if (*pos == ',') { -# pos = parse_space(pos + 1, is_nested); -# -# if (is_digit_char(*pos)) { -# const char * int_end = parse_int(pos); -# max_times = std::stoul(std::string(pos, int_end - pos)); -# pos = parse_space(int_end, is_nested); -# } -# -# if (*pos != '}') { -# throw std::runtime_error(std::string("expecting '}' at ") + pos); -# } -# pos = parse_space(pos + 1, is_nested); -# } else { -# throw std::runtime_error(std::string("expecting ',' at ") + pos); -# } -# handle_repetitions(min_times, max_times); -# } else { -# break; -# } -# } -# return pos; -# } -def parse_sequence(state: ParseState, src: str, rule_name: str, out_elements: typing.List[GrammarElement], is_nested: bool) -> str: - last_sym_start = len(out_elements) - pos = src - - def handle_repetitions(min_times: int, max_times: int) -> None: - nonlocal state, src, rule_name, out_elements, is_nested, last_sym_start, pos - - if last_sym_start == len(out_elements): - raise ValueError(f"expecting preceding item to */+/?/{{ at {pos}") - - previous_elements = out_elements[last_sym_start:] - if min_times == 0: - del out_elements[last_sym_start:] - else: - for i in range(1, min_times): - out_elements.extend(previous_elements) - - last_rec_rule_id = 0 - n_opt = 1 if max_times < 0 else max_times - min_times - - rec_rule = previous_elements[:] - for i in range(n_opt): - rec_rule = rec_rule[:len(previous_elements)] - rec_rule_id = generate_symbol_id(state, rule_name) - if i > 0 or max_times < 0: - rec_rule.append(GrammarElement(GrammarElementType.RULE_REF, rec_rule_id if max_times < 0 else last_rec_rule_id)) - rec_rule.append(GrammarElement(GrammarElementType.ALT, 0)) - rec_rule.append(GrammarElement(GrammarElementType.END, 0)) - add_rule(state, rec_rule_id, rec_rule) - last_rec_rule_id = rec_rule_id - if n_opt > 0: - out_elements.append(GrammarElement(GrammarElementType.RULE_REF, last_rec_rule_id)) - - while pos: - if pos[0] == '"': - pos = pos[1:] - last_sym_start = len(out_elements) - while not pos.startswith('"'): - if not pos: - raise ValueError("unexpected end of input") - char, pos = parse_char(pos) - out_elements.append(GrammarElement(GrammarElementType.CHAR, char)) - pos = parse_space(pos[1:], is_nested) - elif pos[0] == "[": - pos = pos[1:] - start_type = GrammarElementType.CHAR - if pos[0] == "^": - pos = pos[1:] - start_type = GrammarElementType.CHAR_NOT - last_sym_start = len(out_elements) - while pos[0] != "]": - if not pos: - raise ValueError("unexpected end of input") - char, pos = parse_char(pos) - type = GrammarElementType.CHAR_ALT if last_sym_start < len(out_elements) else start_type - out_elements.append(GrammarElement(type, char)) - if pos[0] == "-" and pos[1] != "]": - if not pos[1]: - raise ValueError("unexpected end of input") - endchar, pos = parse_char(pos[1:]) - out_elements.append(GrammarElement(GrammarElementType.CHAR_RNG_UPPER, endchar)) - pos = parse_space(pos[1:], is_nested) - elif pos and is_word_char(pos[0]): - name, rest = parse_name(pos) - ref_rule_id = get_symbol_id(state, name) - pos = parse_space(rest, is_nested) - last_sym_start = len(out_elements) - out_elements.append(GrammarElement(GrammarElementType.RULE_REF, ref_rule_id)) - elif pos.startswith("("): - pos = parse_space(pos[1:], newline_ok=True) - sub_rule_id = generate_symbol_id(state, rule_name) - pos = parse_alternates(state, pos, rule_name, sub_rule_id, is_nested=True) - last_sym_start = len(out_elements) - out_elements.append(GrammarElement(GrammarElementType.RULE_REF, sub_rule_id)) - if pos[0] != ")": - raise ValueError(f"expecting ')' at {pos}") - pos = parse_space(pos[1:], is_nested) - elif pos.startswith("."): - last_sym_start = len(out_elements) - out_elements.append(GrammarElement(GrammarElementType.CHAR_ANY, 0)) - pos = parse_space(pos[1:], is_nested) - elif pos.startswith("*"): - pos = parse_space(pos[1:], is_nested) - handle_repetitions(0, -1) - elif pos.startswith("+"): - pos = parse_space(pos[1:], is_nested) - handle_repetitions(1, -1) - elif pos.startswith("?"): - pos = parse_space(pos[1:], is_nested) - handle_repetitions(0, 1) - elif pos.startswith("{"): - pos = parse_space(pos[1:], is_nested) - - if not is_digit_char(pos): - raise ValueError(f"expecting an int at {pos}") - min_times, pos = parse_int(pos) - pos = parse_space(pos, is_nested) - - max_times = -1 - - if pos[0] == "}": - max_times = min_times - pos = parse_space(pos[1:], is_nested) - elif pos[0] == ",": - pos = parse_space(pos[1:], is_nested) - - if is_digit_char(pos): - max_times, pos = parse_int(pos) - pos = parse_space(pos, is_nested) - - if pos[0] != "}": - raise ValueError("expecting '}' at {}".format(pos)) - - pos = parse_space(pos[1:], is_nested) - else: - raise ValueError(f"expecting ',' at {pos}") - handle_repetitions(min_times, max_times) - else: - break - return pos - - -# const char * parse_alternates( -# parse_state & state, -# const char * src, -# const std::string & rule_name, -# uint32_t rule_id, -# bool is_nested) { -# std::vector rule; -# const char * pos = parse_sequence(state, src, rule_name, rule, is_nested); -# while (*pos == '|') { -# rule.push_back({LLAMA_GRETYPE_ALT, 0}); -# pos = parse_space(pos + 1, true); -# pos = parse_sequence(state, pos, rule_name, rule, is_nested); -# } -# rule.push_back({LLAMA_GRETYPE_END, 0}); -# add_rule(state, rule_id, rule); -# return pos; -# } -def parse_alternates(state: ParseState, src: str, rule_name: str, rule_id: int, is_nested: bool) -> str: - rule = [] - pos = parse_sequence(state, src, rule_name, rule, is_nested) - while pos.startswith("|"): - rule.append(GrammarElement(GrammarElementType.ALT, 0)) - pos = parse_space(pos[1:], newline_ok=True) - pos = parse_sequence(state, pos, rule_name, rule, is_nested) - rule.append(GrammarElement(GrammarElementType.END, 0)) - add_rule(state, rule_id, rule) - return pos - - -# static const char * parse_rule(parse_state & state, const char * src) { -# const char * name_end = parse_name(src); -# const char * pos = parse_space(name_end, false); -# size_t name_len = name_end - src; -# uint32_t rule_id = get_symbol_id(state, src, name_len); -# const std::string name(src, name_len); -# -# if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) { -# throw std::runtime_error(std::string("expecting ::= at ") + pos); -# } -# pos = parse_space(pos + 3, true); -# -# pos = parse_alternates(state, pos, name, rule_id, false); -# -# if (*pos == '\r') { -# pos += pos[1] == '\n' ? 2 : 1; -# } else if (*pos == '\n') { -# pos++; -# } else if (*pos) { -# throw std::runtime_error(std::string("expecting newline or end at ") + pos); -# } -# return parse_space(pos, true); -# } -def parse_rule(state: ParseState, src: str) -> str: - pos = src - name, pos = parse_name(pos) - pos = parse_space(pos, newline_ok=False) - rule_id = get_symbol_id(state, name) - - if not pos.startswith("::="): - raise ValueError(f"expecting ::= at {pos}") - - pos = parse_space(pos[3:], newline_ok=True) - - pos = parse_alternates(state, pos, name, rule_id, is_nested=False) - - if pos.startswith("\r"): - pos = pos[2:] if pos[1] == "\n" else pos[1:] - elif pos.startswith("\n"): - pos = pos[1:] - elif pos: - raise ValueError(f"expecting newline or end at {pos}") - return parse_space(pos, newline_ok=True) - - -# parse_state parse(const char * src) { -# try { -# parse_state state; -# const char * pos = parse_space(src, true); -# while (*pos) { -# pos = parse_rule(state, pos); -# } -# // Validate the state to ensure that all rules are defined -# for (const auto & rule : state.rules) { -# for (const auto & elem : rule) { -# if (elem.type == LLAMA_GRETYPE_RULE_REF) { -# // Ensure that the rule at that location exists -# if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) { -# // Get the name of the rule that is missing -# for (const auto & kv : state.symbol_ids) { -# if (kv.second == elem.value) { -# throw std::runtime_error("Undefined rule identifier '" + kv.first + "'"); -# } -# } -# } -# } -# } -# } -# return state; -# } catch (const std::exception & err) { -# fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what()); -# return parse_state(); -# } -# } -def parse(src: str) -> ParseState: - state = ParseState() - pos = src - pos = parse_space(pos, newline_ok=True) - while pos: - pos = parse_rule(state, pos) - # validate - for rule in state.rules: - for elem in rule: - if elem.type == GrammarElementType.RULE_REF: - if elem.value >= len(state.rules) or not state.rules[elem.value]: - for k, v in state.symbol_ids.items(): - if v == elem.value: - raise ValueError(f"Undefined rule identifier '{k}'") - return state - - -# static bool is_char_element(llama_grammar_element elem) { -# switch (elem.type) { -# case LLAMA_GRETYPE_CHAR: return true; -# case LLAMA_GRETYPE_CHAR_NOT: return true; -# case LLAMA_GRETYPE_CHAR_ALT: return true; -# case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true; -# case LLAMA_GRETYPE_CHAR_ANY: return true; -# default: return false; -# } -# } -def is_char_element(elem: GrammarElement) -> bool: - return elem.type in ( - GrammarElementType.CHAR, - GrammarElementType.CHAR_NOT, - GrammarElementType.CHAR_ALT, - GrammarElementType.CHAR_RNG_UPPER, - GrammarElementType.CHAR_ANY - ) - - -def print_grammar_char(file: typing.TextIO, c: int) -> None: - if 0x20 <= c <= 0x7f: - print(chr(c), end="", file=file) - else: - print(f"", end="", file=file) - - -# static void print_rule( -# FILE * file, -# uint32_t rule_id, -# const std::vector & rule, -# const std::map & symbol_id_names) { -# if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) { -# throw std::runtime_error( -# "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id)); -# } -# fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str()); -# for (size_t i = 0, end = rule.size() - 1; i < end; i++) { -# llama_grammar_element elem = rule[i]; -# switch (elem.type) { -# case LLAMA_GRETYPE_END: -# throw std::runtime_error( -# "unexpected end of rule: " + std::to_string(rule_id) + "," + -# std::to_string(i)); -# case LLAMA_GRETYPE_ALT: -# fprintf(file, "| "); -# break; -# case LLAMA_GRETYPE_RULE_REF: -# fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str()); -# break; -# case LLAMA_GRETYPE_CHAR: -# fprintf(file, "["); -# print_grammar_char(file, elem.value); -# break; -# case LLAMA_GRETYPE_CHAR_NOT: -# fprintf(file, "[^"); -# print_grammar_char(file, elem.value); -# break; -# case LLAMA_GRETYPE_CHAR_RNG_UPPER: -# if (i == 0 || !is_char_element(rule[i - 1])) { -# throw std::runtime_error( -# "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " + -# std::to_string(rule_id) + "," + std::to_string(i)); -# } -# fprintf(file, "-"); -# print_grammar_char(file, elem.value); -# break; -# case LLAMA_GRETYPE_CHAR_ALT: -# if (i == 0 || !is_char_element(rule[i - 1])) { -# throw std::runtime_error( -# "LLAMA_GRETYPE_CHAR_ALT without preceding char: " + -# std::to_string(rule_id) + "," + std::to_string(i)); -# } -# print_grammar_char(file, elem.value); -# break; -# case LLAMA_GRETYPE_CHAR_ANY: -# fprintf(file, "."); -# break; -# } -# if (is_char_element(elem)) { -# switch (rule[i + 1].type) { -# case LLAMA_GRETYPE_CHAR_ALT: -# case LLAMA_GRETYPE_CHAR_RNG_UPPER: -# case LLAMA_GRETYPE_CHAR_ANY: -# break; -# default: -# fprintf(file, "] "); -# } -# } -# } -# fprintf(file, "\n"); -# } -def print_rule( - file: typing.TextIO, - rule_id: int, - rule: typing.List[GrammarElement], - symbol_id_names: typing.Dict[int, str], -) -> None: - if not rule or rule[-1].type != GrammarElementType.END: - raise ValueError(f"malformed rule, does not end with LLAMA_GRETYPE_END: {rule_id}") - - print(f"{symbol_id_names[rule_id]} ::=", end=" ", file=file) - - for i, elem in enumerate(rule[:-1]): - if elem.type == GrammarElementType.END: - raise ValueError(f"unexpected end of rule: {rule_id}, {i}") - if elem.type == GrammarElementType.ALT: - print("| ", end="", file=file) - elif elem.type == GrammarElementType.RULE_REF: - print(f"{symbol_id_names[elem.value]} ", end="", file=file) - elif elem.type == GrammarElementType.CHAR: - print("[", end="", file=file) - print_grammar_char(file, elem.value) - elif elem.type == GrammarElementType.CHAR_NOT: - print("[^", end="", file=file) - print_grammar_char(file, elem.value) - elif elem.type == GrammarElementType.CHAR_RNG_UPPER: - if i == 0 or not is_char_element(rule[i - 1]): - raise ValueError(f"LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: {rule_id}, {i}") - print(f"-", end="", file=file) - print_grammar_char(file, elem.value) - elif elem.type == GrammarElementType.CHAR_ALT: - if i == 0 or not is_char_element(rule[i - 1]): - raise ValueError(f"LLAMA_GRETYPE_CHAR_ALT without preceding char: {rule_id}, {i}") - print_grammar_char(file, elem.value) - elif elem.type == GrammarElementType.CHAR_ANY: - print(".", end="", file=file) - if is_char_element(elem): - if rule[i + 1].type in (GrammarElementType.CHAR_ALT, GrammarElementType.CHAR_RNG_UPPER, GrammarElementType.CHAR_ANY): - continue - print("] ", end="", file=file) - print(file=file) - - -def print_grammar(file: typing.TextIO, state: ParseState) -> None: - try: - symbol_id_names = {v: k for k, v in state.symbol_ids.items()} - for i, rule in enumerate(state.rules): - print_rule(file, i, rule, symbol_id_names) - except Exception as err: - print(f"\nerror printing grammar: {err}", file=file) - raise err - +LLAMA_GRAMMAR_DEFAULT_ROOT = "root" class LlamaGrammar: - def __init__(self, parse_state: ParseState): - self.parse_state = parse_state - - self._grammar_rules = parse_state.rules - self._n_rules = len(self._grammar_rules) - self._start_rule_index = parse_state.symbol_ids["root"] - - self._element_lists = [ - [ - llama_cpp.llama_grammar_element(ctypes.c_int(elem.type), ctypes.c_uint32(elem.value)) - for elem in subvector - ] - for subvector in self._grammar_rules - ] - - # Step 2: Convert each list to llama_grammar_element array and get pointer - self._element_arrays = [ - (llama_cpp.llama_grammar_element * len(sublist))(*sublist) - for sublist in self._element_lists - ] - - # Step 3: Get pointer of each array - self._element_array_pointers = [ - ctypes.cast(subarray, llama_cpp.llama_grammar_element_p) for subarray in self._element_arrays - ] - - # Step 4: Make array of these pointers and get its pointer - self._rules = (llama_cpp.llama_grammar_element_p * len(self._element_array_pointers))( - *self._element_array_pointers - ) - - self.grammar = None - self._init_grammar() - - - def _init_grammar(self): - grammar = llama_cpp.llama_grammar_init( - self._rules, ctypes.c_size_t(self._n_rules), ctypes.c_size_t(self._start_rule_index) - ) - - if grammar is None: - raise ValueError("Failed to create grammar") - - self.grammar = grammar - - def __del__(self): - if self.grammar is not None: - llama_cpp.llama_grammar_free(self.grammar) - self.grammar = None - - def reset(self): - if self.grammar is not None: - llama_cpp.llama_grammar_free(self.grammar) - self._init_grammar() + def __init__(self, *args, _grammar: str, **kwargs): + self._grammar = _grammar + self._root = LLAMA_GRAMMAR_DEFAULT_ROOT @classmethod def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": - parsed_grammar = parse(grammar) - if verbose: - print_grammar(file=sys.stdout, state=parsed_grammar) - return cls(parsed_grammar) + return cls(_grammar=grammar) @classmethod def from_file(cls, file: Union[str, Path], verbose: bool = True) -> "LlamaGrammar": diff --git a/pyproject.toml b/pyproject.toml index ce50c673f..9983ef777 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -45,6 +45,7 @@ test = [ "sse-starlette>=1.6.1", "starlette-context>=0.3.6,<0.4", "pydantic-settings>=2.0.1", + "huggingface-hub>=0.23.0" ] dev = [ "black>=23.3.0", diff --git a/tests/test_llama.py b/tests/test_llama.py index 469ef91ca..cf134c2e7 100644 --- a/tests/test_llama.py +++ b/tests/test_llama.py @@ -1,14 +1,24 @@ import ctypes +import multiprocessing import numpy as np -import pytest from scipy.special import log_softmax +from huggingface_hub import hf_hub_download + +import pytest + import llama_cpp +import llama_cpp._internals as internals + MODEL = "./vendor/llama.cpp/models/ggml-vocab-llama-spm.gguf" +def test_llama_cpp_version(): + assert llama_cpp.__version__ + + def test_llama_cpp_tokenization(): llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, verbose=False) @@ -47,247 +57,117 @@ def test_llama_cpp_tokenization(): @pytest.fixture -def mock_llama(monkeypatch): - def setup_mock(llama: llama_cpp.Llama, output_text: str): - n_ctx = llama.n_ctx() - n_vocab = llama.n_vocab() - output_tokens = llama.tokenize( - output_text.encode("utf-8"), add_bos=True, special=True - ) - logits = (ctypes.c_float * (n_vocab * n_ctx))(-100.0) - for i in range(n_ctx): - output_idx = i + 1 # logits for first tokens predict second token - if output_idx < len(output_tokens): - logits[i * n_vocab + output_tokens[output_idx]] = 100.0 - else: - logits[i * n_vocab + llama.token_eos()] = 100.0 - n = 0 - last_n_tokens = 0 - - def mock_decode(ctx: llama_cpp.llama_context_p, batch: llama_cpp.llama_batch): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - assert batch.n_tokens > 0, "no tokens in batch" - assert all( - batch.n_seq_id[i] == 1 for i in range(batch.n_tokens) - ), "n_seq >1 not supported by mock_llama" - assert all( - batch.seq_id[i][0] == 0 for i in range(batch.n_tokens) - ), "n_seq >1 not supported by mock_llama" - assert batch.logits[ - batch.n_tokens - 1 - ], "logits not allocated for last token" - # Update the mock context state - nonlocal n - nonlocal last_n_tokens - n = max(batch.pos[i] for i in range(batch.n_tokens)) + 1 - last_n_tokens = batch.n_tokens - return 0 - - def mock_get_logits(ctx: llama_cpp.llama_context_p): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - assert n > 0, "mock_llama_decode not called" - assert last_n_tokens > 0, "mock_llama_decode not called" - # Return view of logits for last_n_tokens - return (ctypes.c_float * (last_n_tokens * n_vocab)).from_address( - ctypes.addressof(logits) - + (n - last_n_tokens) * n_vocab * ctypes.sizeof(ctypes.c_float) - ) - - monkeypatch.setattr("llama_cpp.llama_cpp.llama_decode", mock_decode) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_get_logits", mock_get_logits) - - def mock_kv_cache_clear(ctx: llama_cpp.llama_context_p): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_rm( - ctx: llama_cpp.llama_context_p, - seq_id: llama_cpp.llama_seq_id, - pos0: llama_cpp.llama_pos, - pos1: llama_cpp.llama_pos, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_cp( - ctx: llama_cpp.llama_context_p, - seq_id_src: llama_cpp.llama_seq_id, - seq_id_dst: llama_cpp.llama_seq_id, - pos0: llama_cpp.llama_pos, - pos1: llama_cpp.llama_pos, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_keep( - ctx: llama_cpp.llama_context_p, - seq_id: llama_cpp.llama_seq_id, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_add( - ctx: llama_cpp.llama_context_p, - seq_id: llama_cpp.llama_seq_id, - pos0: llama_cpp.llama_pos, - pos1: llama_cpp.llama_pos, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_clear", mock_kv_cache_clear) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_rm", mock_kv_cache_seq_rm) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_cp", mock_kv_cache_seq_cp) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_keep", mock_kv_cache_seq_keep) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_add", mock_kv_cache_seq_add) - - return setup_mock - - -def test_llama_patch(mock_llama): - n_ctx = 128 - llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, n_ctx=n_ctx) - n_vocab = llama_cpp.llama_n_vocab(llama._model.model) - assert n_vocab == 32000 - - text = "The quick brown fox" - output_text = " jumps over the lazy dog." - all_text = text + output_text - - ## Test basic completion from bos until eos - mock_llama(llama, all_text) - completion = llama.create_completion("", max_tokens=36) - assert completion["choices"][0]["text"] == all_text - assert completion["choices"][0]["finish_reason"] == "stop" - - ## Test basic completion until eos - mock_llama(llama, all_text) - completion = llama.create_completion(text, max_tokens=20) - assert completion["choices"][0]["text"] == output_text - assert completion["choices"][0]["finish_reason"] == "stop" - - ## Test streaming completion until eos - mock_llama(llama, all_text) - chunks = list(llama.create_completion(text, max_tokens=20, stream=True)) - assert "".join(chunk["choices"][0]["text"] for chunk in chunks) == output_text - assert chunks[-1]["choices"][0]["finish_reason"] == "stop" - - ## Test basic completion until stop sequence - mock_llama(llama, all_text) - completion = llama.create_completion(text, max_tokens=20, stop=["lazy"]) - assert completion["choices"][0]["text"] == " jumps over the " - assert completion["choices"][0]["finish_reason"] == "stop" - - ## Test streaming completion until stop sequence - mock_llama(llama, all_text) - chunks = list( - llama.create_completion(text, max_tokens=20, stream=True, stop=["lazy"]) - ) - assert ( - "".join(chunk["choices"][0]["text"] for chunk in chunks) == " jumps over the " +def llama_cpp_model_path(): + repo_id = "Qwen/Qwen2-0.5B-Instruct-GGUF" + filename = "qwen2-0_5b-instruct-q8_0.gguf" + model_path = hf_hub_download(repo_id, filename) + return model_path + + +def test_real_model(llama_cpp_model_path): + import os + assert os.path.exists(llama_cpp_model_path) + + params = llama_cpp.llama_model_default_params() + params.use_mmap = llama_cpp.llama_supports_mmap() + params.use_mlock = llama_cpp.llama_supports_mlock() + params.check_tensors = False + + model = internals.LlamaModel(path_model=llama_cpp_model_path, params=params) + + cparams = llama_cpp.llama_context_default_params() + cparams.n_ctx = 16 + cparams.n_batch = 16 + cparams.n_ubatch = 16 + cparams.n_threads = multiprocessing.cpu_count() + cparams.n_threads_batch = multiprocessing.cpu_count() + cparams.logits_all = False + cparams.flash_attn = True + + context = internals.LlamaContext(model=model, params=cparams) + tokens = model.tokenize(b"Hello, world!", add_bos=True, special=True) + + assert tokens == [9707, 11, 1879, 0] + + tokens = model.tokenize(b"The quick brown fox jumps", add_bos=True, special=True) + + batch = internals.LlamaBatch(n_tokens=len(tokens), embd=0, n_seq_max=1) + + seed = 1337 + sampler = internals.LlamaSampler() + sampler.add_top_k(50) + sampler.add_top_p(0.9, 1) + sampler.add_temp(0.8) + sampler.add_dist(seed) + + result = tokens + n_eval = 0 + for _ in range(4): + batch.set_batch(tokens, n_past=n_eval, logits_all=False) + context.decode(batch) + n_eval += len(tokens) + token_id = sampler.sample(context, -1) + tokens = [token_id] + result += tokens + + output = result[5:] + output_text = model.detokenize(output, special=True) + assert output_text == b" over the lazy dog" + +def test_real_llama(llama_cpp_model_path): + model = llama_cpp.Llama( + llama_cpp_model_path, + n_ctx=32, + n_batch=32, + n_ubatch=32, + n_threads=multiprocessing.cpu_count(), + n_threads_batch=multiprocessing.cpu_count(), + logits_all=False, + flash_attn=True, ) - assert chunks[-1]["choices"][0]["finish_reason"] == "stop" - - ## Test basic completion until length - mock_llama(llama, all_text) - completion = llama.create_completion(text, max_tokens=2) - assert completion["choices"][0]["text"] == " jumps" - assert completion["choices"][0]["finish_reason"] == "length" - - ## Test streaming completion until length - mock_llama(llama, all_text) - chunks = list(llama.create_completion(text, max_tokens=2, stream=True)) - assert "".join(chunk["choices"][0]["text"] for chunk in chunks) == " jumps" - assert chunks[-1]["choices"][0]["finish_reason"] == "length" - - -def test_llama_pickle(): - import pickle - import tempfile - - fp = tempfile.TemporaryFile() - llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True) - pickle.dump(llama, fp) - fp.seek(0) - llama = pickle.load(fp) - - assert llama - assert llama.ctx is not None - - text = b"Hello World" - - assert llama.detokenize(llama.tokenize(text)) == text - - -def test_utf8(mock_llama): - llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, logits_all=True) - - output_text = "😀" - - ## Test basic completion with utf8 multibyte - mock_llama(llama, output_text) - completion = llama.create_completion("", max_tokens=4) - assert completion["choices"][0]["text"] == output_text - - ## Test basic completion with incomplete utf8 multibyte - mock_llama(llama, output_text) - completion = llama.create_completion("", max_tokens=1) - assert completion["choices"][0]["text"] == "" + output = model.create_completion( + "The quick brown fox jumps", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + seed=1337 + ) + assert output["choices"][0]["text"] == " over the lazy dog" + + + output = model.create_completion( + "The capital of france is paris, 'true' or 'false'?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + seed=1337, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "true" | "false" +""") + ) + assert output["choices"][0]["text"] == "true" -def test_llama_server(): - from fastapi.testclient import TestClient - from llama_cpp.server.app import create_app, Settings + suffix = b"rot" + tokens = model.tokenize(suffix, add_bos=True, special=True) + def logit_processor_func(input_ids, logits): + for token in tokens: + logits[token] *= 1000 + return logits - settings = Settings( - model=MODEL, - vocab_only=True, + logit_processors = llama_cpp.LogitsProcessorList( + [logit_processor_func] ) - app = create_app(settings) - client = TestClient(app) - response = client.get("/v1/models") - assert response.json() == { - "object": "list", - "data": [ - { - "id": MODEL, - "object": "model", - "owned_by": "me", - "permissions": [], - } - ], - } - - -@pytest.mark.parametrize( - "size_and_axis", - [ - ((32_000,), -1), # last token's next-token logits - ((10, 32_000), -1), # many tokens' next-token logits, or batch of last tokens - ((4, 10, 32_000), -1), # batch of texts - ], -) -@pytest.mark.parametrize("convert_to_list", [True, False]) -def test_logits_to_logprobs(size_and_axis, convert_to_list: bool, atol: float = 1e-7): - size, axis = size_and_axis - logits: np.ndarray = -np.random.uniform(low=0, high=60, size=size) - logits = logits.astype(np.single) - if convert_to_list: - # Currently, logits are converted from arrays to lists. This may change soon - logits = logits.tolist() - log_probs = llama_cpp.Llama.logits_to_logprobs(logits, axis=axis) - log_probs_correct = log_softmax(logits, axis=axis) - assert log_probs.dtype == np.single - assert log_probs.shape == size - assert np.allclose(log_probs, log_probs_correct, atol=atol) - -def test_llama_cpp_version(): - assert llama_cpp.__version__ + output = model.create_completion( + "The capital of france is par", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + seed=1337, + logits_processor=logit_processors + ) + assert output["choices"][0]["text"].lower().startswith("rot") diff --git a/tests/test_llama_grammar.py b/tests/test_llama_grammar.py index cb221880a..34ef2874d 100644 --- a/tests/test_llama_grammar.py +++ b/tests/test_llama_grammar.py @@ -10,9 +10,9 @@ def test_grammar_from_string(): grammar = llama_cpp.LlamaGrammar.from_string(tree) - assert grammar._n_rules == 3 - assert grammar._start_rule_index == 2 - assert grammar.grammar is not None + # assert grammar._n_rules == 3 + # assert grammar._start_rule_index == 2 + # assert grammar.grammar is not None def test_composed_pydantic_grammar(): @@ -49,7 +49,7 @@ class B(BaseModel): grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(schema)) - assert grammar.grammar is not None + # assert grammar.grammar is not None def test_grammar_anyof(): @@ -75,4 +75,4 @@ def test_grammar_anyof(): grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(sch)) - assert grammar.grammar is not None \ No newline at end of file + # assert grammar.grammar is not None diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 8ebe8ddeb..6262d13e0 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 8ebe8ddebd68526757c631cd019de009697c63c2 +Subproject commit 6262d13e0b2da91f230129a93a996609a2f5a2f2 From 1e64664e0facb7e3595efdf4716f01527f4047ba Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 18 Sep 2024 20:07:48 -0400 Subject: [PATCH 542/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 227 +++++++++++++++++++++++++++-------------- vendor/llama.cpp | 2 +- 2 files changed, 153 insertions(+), 76 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 261d7c6b4..6b82753e1 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1161,7 +1161,8 @@ def llama_backend_init(): [ctypes.c_int], None, ) -def llama_numa_init(numa: int, /): ... +def llama_numa_init(numa: int, /): + ... # // Optional: an auto threadpool gets created in ggml if not passed explicitly @@ -1196,7 +1197,8 @@ def llama_backend_free(): ) def llama_load_model_from_file( path_model: bytes, params: llama_model_params, / -) -> Optional[llama_model_p]: ... +) -> Optional[llama_model_p]: + ... # LLAMA_API void llama_free_model(struct llama_model * model); @@ -1205,7 +1207,8 @@ def llama_load_model_from_file( [llama_model_p_ctypes], None, ) -def llama_free_model(model: llama_model_p, /): ... +def llama_free_model(model: llama_model_p, /): + ... # LLAMA_API struct llama_context * llama_new_context_with_model( @@ -1218,7 +1221,8 @@ def llama_free_model(model: llama_model_p, /): ... ) def llama_new_context_with_model( model: llama_model_p, params: llama_context_params, / -) -> Optional[llama_context_p]: ... +) -> Optional[llama_context_p]: + ... # // Frees all allocated memory @@ -1239,87 +1243,110 @@ def llama_free(ctx: llama_context_p, /): [], ctypes.c_int64, ) -def llama_time_us() -> int: ... +def llama_time_us() -> int: + ... # LLAMA_API size_t llama_max_devices(void); @ctypes_function("llama_max_devices", [], ctypes.c_size_t) -def llama_max_devices() -> int: ... +def llama_max_devices() -> int: + ... # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) -def llama_supports_mmap() -> bool: ... +def llama_supports_mmap() -> bool: + ... # LLAMA_API bool llama_supports_mlock (void); @ctypes_function("llama_supports_mlock", [], ctypes.c_bool) -def llama_supports_mlock() -> bool: ... +def llama_supports_mlock() -> bool: + ... # LLAMA_API bool llama_supports_gpu_offload(void); @ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) -def llama_supports_gpu_offload() -> bool: ... +def llama_supports_gpu_offload() -> bool: + ... # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ctx(ctx: llama_context_p, /) -> int: ... +def llama_n_ctx(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); @ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_batch(ctx: llama_context_p, /) -> int: ... +def llama_n_batch(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); @ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_ubatch(ctx: llama_context_p, /) -> int: ... +def llama_n_ubatch(ctx: llama_context_p, /) -> int: + ... # LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); @ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) -def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... +def llama_n_seq_max(ctx: llama_context_p, /) -> int: + ... # LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); @ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: ... +def llama_n_vocab(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_ctx_train(model: llama_model_p, /) -> int: ... +def llama_n_ctx_train(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_embd (const struct llama_model * model); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_embd(model: llama_model_p, /) -> int: ... +def llama_n_embd(model: llama_model_p, /) -> int: + ... # LLAMA_API int32_t llama_n_layer (const struct llama_model * model); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_layer(model: llama_model_p, /) -> int: ... +def llama_n_layer(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_n_head (const struct llama_model * model); +@ctypes_function("llama_n_head", [llama_model_p_ctypes], ctypes.c_int32) +def llama_n_head(model: llama_model_p, /) -> int: + ... # LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) -def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: + ... # LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) -def llama_pooling_type(ctx: llama_context_p, /) -> int: ... +def llama_pooling_type(ctx: llama_context_p, /) -> int: + ... # LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); @ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: ... +def llama_vocab_type(model: llama_model_p, /) -> int: + ... # LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: ... +def llama_rope_type(model: llama_model_p, /) -> int: + ... # // Get the model's RoPE frequency scaling factor @@ -2110,7 +2137,8 @@ def llama_state_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> bool: ... +) -> bool: + ... # LLAMA_API DEPRECATED(bool llama_load_session_file( @@ -2138,7 +2166,8 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: ... +) -> int: + ... # LLAMA_API bool llama_state_save_file( @@ -2162,7 +2191,8 @@ def llama_state_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> bool: ... +) -> bool: + ... # LLAMA_API DEPRECATED(bool llama_save_session_file( @@ -2187,7 +2217,8 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: ... +) -> int: + ... # // Get the exact size needed to copy the KV cache of a single sequence @@ -2285,7 +2316,8 @@ def llama_state_seq_save_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: ... +) -> int: + ... # LLAMA_API size_t llama_state_seq_load_file( @@ -2315,7 +2347,8 @@ def llama_state_seq_load_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: ... +) -> int: + ... # // @@ -2620,7 +2653,8 @@ def llama_get_embeddings_seq( ) def llama_token_get_text( model: llama_model_p, token: Union[llama_token, int], / -) -> bytes: ... +) -> bytes: + ... # LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); @@ -2629,7 +2663,8 @@ def llama_token_get_text( ) def llama_token_get_score( model: llama_model_p, token: Union[llama_token, int], / -) -> float: ... +) -> float: + ... # LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token); @@ -2638,7 +2673,8 @@ def llama_token_get_score( ) def llama_token_get_attr( model: llama_model_p, token: Union[llama_token, int], / -) -> int: ... +) -> int: + ... # // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) @@ -2703,12 +2739,14 @@ def llama_token_nl(model: llama_model_p, /) -> int: # LLAMA_API bool llama_add_bos_token(const struct llama_model * model); @ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_bos_token(model: llama_model_p, /) -> bool: ... +def llama_add_bos_token(model: llama_model_p, /) -> bool: + ... # LLAMA_API bool llama_add_eos_token(const struct llama_model * model); @ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_eos_token(model: llama_model_p, /) -> bool: ... +def llama_add_eos_token(model: llama_model_p, /) -> bool: + ... # // Codellama infill tokens @@ -2721,17 +2759,20 @@ def llama_token_prefix(model: llama_model_p) -> int: # LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: ... +def llama_token_middle(model: llama_model_p, /) -> int: + ... # LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: ... +def llama_token_suffix(model: llama_model_p, /) -> int: + ... # LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle @ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: ... +def llama_token_eot(model: llama_model_p, /) -> int: + ... # // @@ -2932,7 +2973,8 @@ def llama_chat_apply_template( chat: CtypesArray[llama_chat_message], n_msg: int, /, -) -> int: ... +) -> int: + ... # // @@ -2991,7 +3033,8 @@ def llama_chat_apply_template( # // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph # //void (*apply_ggml) (struct llama_sampler * smpl, ...); # }; -class llama_sampler_i(ctypes.Structure): ... +class llama_sampler_i(ctypes.Structure): + ... # struct llama_sampler { @@ -3036,7 +3079,8 @@ class llama_sampler(ctypes.Structure): [llama_sampler_p_ctypes], ctypes.c_char_p, ) -def llama_sampler_name(smpl: llama_sampler_p, /) -> bytes: ... +def llama_sampler_name(smpl: llama_sampler_p, /) -> bytes: + ... # LLAMA_API void llama_sampler_accept( struct llama_sampler * smpl, llama_token token); @@ -3045,7 +3089,8 @@ def llama_sampler_name(smpl: llama_sampler_p, /) -> bytes: ... [llama_sampler_p_ctypes, llama_token], None, ) -def llama_sampler_accept(smpl: llama_sampler_p, token: Union[llama_token, int], /): ... +def llama_sampler_accept(smpl: llama_sampler_p, token: Union[llama_token, int], /): + ... # LLAMA_API void llama_sampler_apply ( struct llama_sampler * smpl, llama_token_data_array * cur_p); @@ -3056,7 +3101,8 @@ def llama_sampler_accept(smpl: llama_sampler_p, token: Union[llama_token, int], ) def llama_sampler_apply( smpl: llama_sampler_p, cur_p: CtypesArray[llama_token_data_array], / -): ... +): + ... # LLAMA_API void llama_sampler_reset ( struct llama_sampler * smpl); @@ -3065,7 +3111,8 @@ def llama_sampler_apply( [llama_sampler_p_ctypes], None, ) -def llama_sampler_reset(smpl: llama_sampler_p, /): ... +def llama_sampler_reset(smpl: llama_sampler_p, /): + ... # LLAMA_API struct llama_sampler * llama_sampler_clone (const struct llama_sampler * smpl); @@ -3074,7 +3121,8 @@ def llama_sampler_reset(smpl: llama_sampler_p, /): ... [llama_sampler_p_ctypes], llama_sampler_p_ctypes, ) -def llama_sampler_clone(smpl: llama_sampler_p, /) -> llama_sampler_p: ... +def llama_sampler_clone(smpl: llama_sampler_p, /) -> llama_sampler_p: + ... # // important: do not free if the sampler has been added to a llama_sampler_chain (via llama_sampler_chain_add) @@ -3084,7 +3132,8 @@ def llama_sampler_clone(smpl: llama_sampler_p, /) -> llama_sampler_p: ... [llama_sampler_p_ctypes], None, ) -def llama_sampler_free(smpl: llama_sampler_p, /): ... +def llama_sampler_free(smpl: llama_sampler_p, /): + ... # // llama_sampler_chain @@ -3096,9 +3145,8 @@ def llama_sampler_free(smpl: llama_sampler_p, /): ... [llama_sampler_chain_params], llama_sampler_p_ctypes, ) -def llama_sampler_chain_init( - params: llama_sampler_chain_params, / -) -> llama_sampler_p: ... +def llama_sampler_chain_init(params: llama_sampler_chain_params, /) -> llama_sampler_p: + ... # // important: takes ownership of the sampler object and will free it when llama_sampler_free is called @@ -3108,7 +3156,8 @@ def llama_sampler_chain_init( [llama_sampler_p_ctypes, llama_sampler_p_ctypes], None, ) -def llama_sampler_chain_add(chain: llama_sampler_p, smpl: llama_sampler_p, /): ... +def llama_sampler_chain_add(chain: llama_sampler_p, smpl: llama_sampler_p, /): + ... # LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i); @@ -3119,7 +3168,8 @@ def llama_sampler_chain_add(chain: llama_sampler_p, smpl: llama_sampler_p, /): . ) def llama_sampler_chain_get( chain: llama_sampler_p, i: Union[ctypes.c_int32, int], / -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # LLAMA_API int llama_sampler_chain_n (const struct llama_sampler * chain); @@ -3128,7 +3178,8 @@ def llama_sampler_chain_get( [llama_sampler_p_ctypes], ctypes.c_int, ) -def llama_sampler_chain_n(chain: llama_sampler_p, /) -> int: ... +def llama_sampler_chain_n(chain: llama_sampler_p, /) -> int: + ... # // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed @@ -3140,31 +3191,36 @@ def llama_sampler_chain_n(chain: llama_sampler_p, /) -> int: ... ) def llama_sampler_chain_remove( chain: llama_sampler_p, i: Union[ctypes.c_int32, int], / -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # // available samplers: # # LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void); @ctypes_function("llama_sampler_init_greedy", [], llama_sampler_p_ctypes) -def llama_sampler_init_greedy() -> llama_sampler_p: ... +def llama_sampler_init_greedy() -> llama_sampler_p: + ... # LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed); @ctypes_function("llama_sampler_init_dist", [ctypes.c_uint32], llama_sampler_p_ctypes) -def llama_sampler_init_dist(seed: int) -> llama_sampler_p: ... +def llama_sampler_init_dist(seed: int) -> llama_sampler_p: + ... # /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. # LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void); @ctypes_function("llama_sampler_init_softmax", [], llama_sampler_p_ctypes) -def llama_sampler_init_softmax() -> llama_sampler_p: ... +def llama_sampler_init_softmax() -> llama_sampler_p: + ... # /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 # LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); @ctypes_function("llama_sampler_init_top_k", [ctypes.c_int32], llama_sampler_p_ctypes) -def llama_sampler_init_top_k(k: int) -> llama_sampler_p: ... +def llama_sampler_init_top_k(k: int) -> llama_sampler_p: + ... # /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 @@ -3174,7 +3230,8 @@ def llama_sampler_init_top_k(k: int) -> llama_sampler_p: ... [ctypes.c_float, ctypes.c_size_t], llama_sampler_p_ctypes, ) -def llama_sampler_init_top_p(p: float, min_keep: int) -> llama_sampler_p: ... +def llama_sampler_init_top_p(p: float, min_keep: int) -> llama_sampler_p: + ... # /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 @@ -3184,7 +3241,8 @@ def llama_sampler_init_top_p(p: float, min_keep: int) -> llama_sampler_p: ... [ctypes.c_float, ctypes.c_size_t], llama_sampler_p_ctypes, ) -def llama_sampler_init_min_p(p: float, min_keep: int) -> llama_sampler_p: ... +def llama_sampler_init_min_p(p: float, min_keep: int) -> llama_sampler_p: + ... # /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. @@ -3194,7 +3252,8 @@ def llama_sampler_init_min_p(p: float, min_keep: int) -> llama_sampler_p: ... [ctypes.c_float, ctypes.c_size_t], llama_sampler_p_ctypes, ) -def llama_sampler_init_tail_free(z: float, min_keep: int) -> llama_sampler_p: ... +def llama_sampler_init_tail_free(z: float, min_keep: int) -> llama_sampler_p: + ... # /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. @@ -3204,12 +3263,14 @@ def llama_sampler_init_tail_free(z: float, min_keep: int) -> llama_sampler_p: .. [ctypes.c_float, ctypes.c_size_t], llama_sampler_p_ctypes, ) -def llama_sampler_init_typical(p: float, min_keep: int) -> llama_sampler_p: ... +def llama_sampler_init_typical(p: float, min_keep: int) -> llama_sampler_p: + ... # LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t); @ctypes_function("llama_sampler_init_temp", [ctypes.c_float], llama_sampler_p_ctypes) -def llama_sampler_init_temp(t: float) -> llama_sampler_p: ... +def llama_sampler_init_temp(t: float) -> llama_sampler_p: + ... # /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772. @@ -3221,7 +3282,8 @@ def llama_sampler_init_temp(t: float) -> llama_sampler_p: ... ) def llama_sampler_init_temp_ext( t: float, delta: float, exponent: float -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -3243,7 +3305,8 @@ def llama_sampler_init_temp_ext( ) def llama_sampler_init_mirostat( n_vocab: int, seed: int, tau: float, eta: float, m: int, / -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -3262,7 +3325,8 @@ def llama_sampler_init_mirostat( ) def llama_sampler_init_mirostat_v2( seed: int, tau: float, eta: float, / -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # LLAMA_API struct llama_sampler * llama_sampler_init_grammar( @@ -3276,7 +3340,8 @@ def llama_sampler_init_mirostat_v2( ) def llama_sampler_init_grammar( model: llama_model_p, grammar_str: bytes, grammar_root: bytes, / -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # LLAMA_API struct llama_sampler * llama_sampler_init_penalties( @@ -3315,7 +3380,8 @@ def llama_sampler_init_penalties( penalize_nl: bool, ignore_eos: bool, /, -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias( @@ -3329,7 +3395,8 @@ def llama_sampler_init_penalties( ) def llama_sampler_init_logit_bias( n_vocab: int, n_logit_bias: int, logit_bias: CtypesArray[llama_logit_bias], / -) -> llama_sampler_p: ... +) -> llama_sampler_p: + ... # // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise @@ -3339,7 +3406,8 @@ def llama_sampler_init_logit_bias( [llama_sampler_p_ctypes], ctypes.c_uint32, ) -def llama_sampler_get_seed(smpl: llama_sampler_p, /) -> int: ... +def llama_sampler_get_seed(smpl: llama_sampler_p, /) -> int: + ... # /// @details Sample and accept a token from the idx-th output of the last evaluation @@ -3360,7 +3428,8 @@ def llama_sampler_get_seed(smpl: llama_sampler_p, /) -> int: ... ) def llama_sampler_sample( smpl: llama_sampler_p, ctx: llama_context_p, idx: int, / -) -> int: ... +) -> int: + ... # // @@ -3413,7 +3482,8 @@ def llama_split_prefix( # // Print system information # LLAMA_API const char * llama_print_system_info(void); @ctypes_function("llama_print_system_info", [], ctypes.c_char_p) -def llama_print_system_info() -> bytes: ... +def llama_print_system_info() -> bytes: + ... # // Set callback for all future logging events. @@ -3480,7 +3550,8 @@ class llama_perf_sampler_data(ctypes.Structure): [llama_context_p_ctypes], llama_perf_context_data, ) -def llama_perf_context(ctx: llama_context_p, /) -> llama_perf_context_data: ... +def llama_perf_context(ctx: llama_context_p, /) -> llama_perf_context_data: + ... # LLAMA_API void llama_perf_context_print(const struct llama_context * ctx); @@ -3489,7 +3560,8 @@ def llama_perf_context(ctx: llama_context_p, /) -> llama_perf_context_data: ... [llama_context_p_ctypes], None, ) -def llama_perf_context_print(ctx: llama_context_p, /): ... +def llama_perf_context_print(ctx: llama_context_p, /): + ... # LLAMA_API void llama_perf_context_reset( struct llama_context * ctx); @@ -3498,7 +3570,8 @@ def llama_perf_context_print(ctx: llama_context_p, /): ... [llama_context_p_ctypes], None, ) -def llama_perf_context_reset(ctx: llama_context_p, /): ... +def llama_perf_context_reset(ctx: llama_context_p, /): + ... # // NOTE: the following work only with samplers constructed via llama_sampler_chain_init @@ -3508,7 +3581,8 @@ def llama_perf_context_reset(ctx: llama_context_p, /): ... [llama_sampler_p_ctypes], llama_perf_sampler_data, ) -def llama_perf_sampler(chain: llama_sampler_p, /) -> llama_perf_sampler_data: ... +def llama_perf_sampler(chain: llama_sampler_p, /) -> llama_perf_sampler_data: + ... # LLAMA_API void llama_perf_sampler_print(const struct llama_sampler * chain); @@ -3517,7 +3591,8 @@ def llama_perf_sampler(chain: llama_sampler_p, /) -> llama_perf_sampler_data: .. [llama_sampler_p_ctypes], None, ) -def llama_perf_sampler_print(chain: llama_sampler_p, /): ... +def llama_perf_sampler_print(chain: llama_sampler_p, /): + ... # LLAMA_API void llama_perf_sampler_reset( struct llama_sampler * chain); @@ -3526,7 +3601,8 @@ def llama_perf_sampler_print(chain: llama_sampler_p, /): ... [llama_sampler_p_ctypes], None, ) -def llama_perf_sampler_reset(chain: llama_sampler_p, /): ... +def llama_perf_sampler_reset(chain: llama_sampler_p, /): + ... # LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx); @@ -3537,4 +3613,5 @@ def llama_perf_sampler_reset(chain: llama_sampler_p, /): ... ) def llama_perf_dump_yaml( stream: ctypes.POINTER(ctypes.c_void_p), ctx: llama_context_p, / -): ... +): + ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6262d13e0..64c6af319 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6262d13e0b2da91f230129a93a996609a2f5a2f2 +Subproject commit 64c6af3195c3cd4aa3328a1282d29cd2635c34c9 From 9b64bb5b137385cdf535598df5b2c34ed459450f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 18 Sep 2024 20:11:02 -0400 Subject: [PATCH 543/624] misc: Format --- llama_cpp/_internals.py | 66 ++++++++++++++++++------ llama_cpp/llama.py | 98 ++++++++++++++++++++---------------- llama_cpp/llama_cache.py | 6 +-- llama_cpp/llama_grammar.py | 3 +- llama_cpp/llama_tokenizer.py | 33 ++++++++---- llama_cpp/llava_cpp.py | 21 +++++--- 6 files changed, 147 insertions(+), 80 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 6475695c6..0aff34844 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -100,7 +100,6 @@ def n_params(self) -> int: def get_tensor(self, name: str) -> ctypes.c_void_p: return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) - # Vocab def token_get_text(self, token: int) -> str: @@ -460,9 +459,7 @@ def __init__( self.verbose = verbose self._exit_stack = ExitStack() - batch = llama_cpp.llama_batch_init( - self._n_tokens, self.embd, self.n_seq_max - ) + batch = llama_cpp.llama_batch_init(self._n_tokens, self.embd, self.n_seq_max) if batch is None: raise ValueError("Failed to create llama_batch") @@ -541,6 +538,7 @@ def copy_logits(self, logits: npt.NDArray[np.single]): # Embedding functions + def normalize_embedding(embedding): norm = float(np.linalg.norm(embedding)) if norm == 0.0: @@ -713,11 +711,17 @@ def accept(self, ctx_main: LlamaContext, id: int, apply_grammar: bool): import ctypes import llama_cpp + class CustomSampler: - def __init__(self, apply_func: typing.Callable[[llama_cpp.llama_token_data_array], None]): + def __init__( + self, apply_func: typing.Callable[[llama_cpp.llama_token_data_array], None] + ): self.apply_func = apply_func - def apply_wrapper(sampler: llama_cpp.llama_sampler_p, cur_p: llama_cpp.llama_token_data_array_p): + def apply_wrapper( + sampler: llama_cpp.llama_sampler_p, + cur_p: llama_cpp.llama_token_data_array_p, + ): self.apply_func(cur_p) def free_wrapper(sampler: llama_cpp.llama_sampler_p): @@ -740,6 +744,7 @@ def free_wrapper(sampler: llama_cpp.llama_sampler_p): def get_sampler(self) -> llama_cpp.llama_sampler_p: return ctypes.pointer(self.sampler) + class LlamaSampler: def __init__(self): params = llama_cpp.llama_sampler_chain_params() @@ -788,9 +793,7 @@ def add_temp_ext(self, t: float, delta: float, exponent: float): self._add_sampler(sampler) def add_mirostat(self, n_vocab: int, seed: int, tau: float, eta: float, m: int): - sampler = llama_cpp.llama_sampler_init_mirostat( - n_vocab, seed, tau, eta, m - ) + sampler = llama_cpp.llama_sampler_init_mirostat(n_vocab, seed, tau, eta, m) self._add_sampler(sampler) def add_mirostat_v2(self, seed: int, tau: float, eta: float): @@ -798,23 +801,54 @@ def add_mirostat_v2(self, seed: int, tau: float, eta: float): self._add_sampler(sampler) def add_grammar(self, model: LlamaModel, grammar: LlamaGrammar): - sampler = llama_cpp.llama_sampler_init_grammar(model.model, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8")) + sampler = llama_cpp.llama_sampler_init_grammar( + model.model, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8") + ) self._add_sampler(sampler) - def add_penalties(self, n_vocab: int, special_eos_id: int, linefeed_id: int, penalty_last_n: int, penalty_repeat: float, penalty_freq: float, penalty_present: float, penalize_nl: bool, ignore_eos: bool): - sampler = llama_cpp.llama_sampler_init_penalties(n_vocab, special_eos_id, linefeed_id, penalty_last_n, penalty_repeat, penalty_freq, penalty_present, penalize_nl, ignore_eos) + def add_penalties( + self, + n_vocab: int, + special_eos_id: int, + linefeed_id: int, + penalty_last_n: int, + penalty_repeat: float, + penalty_freq: float, + penalty_present: float, + penalize_nl: bool, + ignore_eos: bool, + ): + sampler = llama_cpp.llama_sampler_init_penalties( + n_vocab, + special_eos_id, + linefeed_id, + penalty_last_n, + penalty_repeat, + penalty_freq, + penalty_present, + penalize_nl, + ignore_eos, + ) self._add_sampler(sampler) - def init_logit_bias(self, n_vocab: int, n_logit_bias, logit_bias: llama_cpp.llama_logit_bias_p): - sampler = llama_cpp.llama_sampler_init_logit_bias(n_vocab, n_logit_bias, logit_bias) + def init_logit_bias( + self, n_vocab: int, n_logit_bias, logit_bias: llama_cpp.llama_logit_bias_p + ): + sampler = llama_cpp.llama_sampler_init_logit_bias( + n_vocab, n_logit_bias, logit_bias + ) self._add_sampler(sampler) - def add_custom(self, apply_func: Callable[[llama_cpp.llama_token_data_array], None]): + def add_custom( + self, apply_func: Callable[[llama_cpp.llama_token_data_array], None] + ): custom_sampler = CustomSampler(apply_func) sampler = custom_sampler.get_sampler() self._add_sampler(sampler) # NOTE: Must remove custom samplers before free or llama.cpp will try to free them - self.custom_samplers.append((llama_cpp.llama_sampler_chain_n(self.sampler) - 1, custom_sampler)) + self.custom_samplers.append( + (llama_cpp.llama_sampler_chain_n(self.sampler) - 1, custom_sampler) + ) def _add_sampler(self, sampler: llama_cpp.llama_sampler_p): assert self.sampler is not None diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 557134a6e..7dd862515 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -255,28 +255,28 @@ def __init__( for i, (k, v) in enumerate(kv_overrides.items()): self._kv_overrides_array[i].key = k.encode("utf-8") if isinstance(v, bool): - self._kv_overrides_array[i].tag = ( - llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL - ) + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL self._kv_overrides_array[i].value.val_bool = v elif isinstance(v, int): - self._kv_overrides_array[i].tag = ( - llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT - ) + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT self._kv_overrides_array[i].value.val_i64 = v elif isinstance(v, float): - self._kv_overrides_array[i].tag = ( - llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT - ) + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT self._kv_overrides_array[i].value.val_f64 = v elif isinstance(v, str): # type: ignore v_bytes = v.encode("utf-8") if len(v_bytes) > 128: # TODO: Make this a constant raise ValueError(f"Value for {k} is too long: {v}") v_bytes = v_bytes.ljust(128, b"\0") - self._kv_overrides_array[i].tag = ( - llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR - ) + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR # copy min(v_bytes, 128) to str_value address = typing.cast( int, @@ -292,9 +292,9 @@ def __init__( else: raise ValueError(f"Unknown value type for {k}: {v}") - self._kv_overrides_array[-1].key = ( - b"\0" # ensure sentinel element is zeroed - ) + self._kv_overrides_array[ + -1 + ].key = b"\0" # ensure sentinel element is zeroed self.model_params.kv_overrides = self._kv_overrides_array self.n_batch = min(n_ctx, n_batch) # ??? @@ -431,9 +431,9 @@ def free_lora_adapter(): self.chat_format = chat_format self.chat_handler = chat_handler - self._chat_handlers: Dict[str, llama_chat_format.LlamaChatCompletionHandler] = ( - {} - ) + self._chat_handlers: Dict[ + str, llama_chat_format.LlamaChatCompletionHandler + ] = {} self.draft_model = draft_model @@ -580,7 +580,10 @@ def tokenize( return self.tokenizer_.tokenize(text, add_bos, special) def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, ) -> bytes: """Detokenize a list of tokens. @@ -592,7 +595,9 @@ def detokenize( Returns: The detokenized string. """ - return self.tokenizer_.detokenize(tokens, prev_tokens=prev_tokens, special=special) + return self.tokenizer_.detokenize( + tokens, prev_tokens=prev_tokens, special=special + ) def set_cache(self, cache: Optional[BaseLlamaCache]): """Set the cache. @@ -681,12 +686,16 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): recarray = np.recarray( shape=(size,), dtype=np.dtype( - [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True + [("id", np.intc), ("logit", np.single), ("p", np.single)], + align=True, + ), + buf=(llama_cpp.llama_token_data * size).from_address( + data_soa_address ), - buf=(llama_cpp.llama_token_data * size).from_address(data_soa_address), ) for logit_processor in logits_processor: recarray.logit[:] = logit_processor(self._input_ids, recarray.logit) + sampler.add_custom(apply_func) sampler.add_penalties( @@ -698,7 +707,7 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): penalty_freq=frequency_penalty, penalty_present=presence_penalty, penalize_nl=penalize_nl, - ignore_eos=False + ignore_eos=False, ) if grammar is not None: @@ -841,22 +850,22 @@ def generate( # Reset mirostat sampling self._mirostat_mu = ctypes.c_float(2.0 * mirostat_tau) self._sampler = self._init_sampler( - top_k=top_k, - top_p=top_p, - min_p=min_p, - typical_p=typical_p, - temp=temp, - repeat_penalty=repeat_penalty, - frequency_penalty=frequency_penalty, - presence_penalty=presence_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - penalize_nl=penalize_nl, - logits_processor=logits_processor, - grammar=grammar, - seed=seed, + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + penalize_nl=penalize_nl, + logits_processor=logits_processor, + grammar=grammar, + seed=seed, ) # Check for kv cache prefix match @@ -872,8 +881,11 @@ def generate( tokens = tokens[longest_prefix:] self.n_tokens = longest_prefix if self.verbose: - print(f"Llama.generate: {longest_prefix} prefix-match hit, " - f"remaining {len(tokens)} prompt tokens to eval", file=sys.stderr) + print( + f"Llama.generate: {longest_prefix} prefix-match hit, " + f"remaining {len(tokens)} prompt tokens to eval", + file=sys.stderr, + ) # Reset the model state if reset: @@ -1032,7 +1044,9 @@ def decode_batch(seq_sizes: List[int]): for j in range(size) ] if normalize: - embedding = [internals.normalize_embedding(e) for e in embedding] + embedding = [ + internals.normalize_embedding(e) for e in embedding + ] data.append(embedding) pos += size else: diff --git a/llama_cpp/llama_cache.py b/llama_cpp/llama_cache.py index 5220c7933..e059e98e1 100644 --- a/llama_cpp/llama_cache.py +++ b/llama_cpp/llama_cache.py @@ -52,9 +52,9 @@ class LlamaRAMCache(BaseLlamaCache): def __init__(self, capacity_bytes: int = (2 << 30)): super().__init__(capacity_bytes) self.capacity_bytes = capacity_bytes - self.cache_state: OrderedDict[Tuple[int, ...], "llama_cpp.llama.LlamaState"] = ( - OrderedDict() - ) + self.cache_state: OrderedDict[ + Tuple[int, ...], "llama_cpp.llama.LlamaState" + ] = OrderedDict() @property def cache_size(self): diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index 785aa88d7..b95c77ab5 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -15,6 +15,7 @@ LLAMA_GRAMMAR_DEFAULT_ROOT = "root" + class LlamaGrammar: def __init__(self, *args, _grammar: str, **kwargs): self._grammar = _grammar @@ -23,7 +24,7 @@ def __init__(self, *args, _grammar: str, **kwargs): @classmethod def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": return cls(_grammar=grammar) - + @classmethod def from_file(cls, file: Union[str, Path], verbose: bool = True) -> "LlamaGrammar": try: diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py index 2e7590d14..1375e1392 100644 --- a/llama_cpp/llama_tokenizer.py +++ b/llama_cpp/llama_tokenizer.py @@ -27,7 +27,10 @@ def tokenize( @abc.abstractmethod def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, ) -> bytes: """Detokenize the tokens into text. @@ -49,7 +52,10 @@ def tokenize( return self._model.tokenize(text, add_bos=add_bos, special=special) def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, ) -> bytes: return self._model.detokenize(tokens, special=special) @@ -80,19 +86,24 @@ def tokenize( ) def detokenize( - self, tokens: List[int], prev_tokens: Optional[List[int]] = None, special: bool = False + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, ) -> bytes: - skip_special_tokens = not special + skip_special_tokens = not special if prev_tokens is not None: - text = self.hf_tokenizer.decode(prev_tokens + tokens, skip_special_tokens=skip_special_tokens).encode( - "utf-8", errors="ignore" - ) - prev_text = self.hf_tokenizer.decode(prev_tokens, skip_special_tokens=skip_special_tokens).encode( - "utf-8", errors="ignore" - ) + text = self.hf_tokenizer.decode( + prev_tokens + tokens, skip_special_tokens=skip_special_tokens + ).encode("utf-8", errors="ignore") + prev_text = self.hf_tokenizer.decode( + prev_tokens, skip_special_tokens=skip_special_tokens + ).encode("utf-8", errors="ignore") return text[len(prev_text) :] else: - return self.hf_tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens).encode("utf-8", errors="ignore") + return self.hf_tokenizer.decode( + tokens, skip_special_tokens=skip_special_tokens + ).encode("utf-8", errors="ignore") @classmethod def from_pretrained(cls, pretrained_model_name_or_path: str) -> "LlamaHFTokenizer": diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index b80d85913..7d97dc0fd 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -165,7 +165,8 @@ class llava_image_embed(Structure): ) def llava_validate_embed_size( ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p, / -) -> bool: ... +) -> bool: + ... # /** build an image embed from image file bytes */ @@ -181,7 +182,8 @@ def llava_image_embed_make_with_bytes( image_bytes: CtypesArray[c_uint8], image_bytes_length: Union[c_int, int], /, -) -> "_Pointer[llava_image_embed]": ... +) -> "_Pointer[llava_image_embed]": + ... # /** build an image embed from a path to an image filename */ @@ -193,13 +195,15 @@ def llava_image_embed_make_with_bytes( ) def llava_image_embed_make_with_filename( ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes, / -) -> "_Pointer[llava_image_embed]": ... +) -> "_Pointer[llava_image_embed]": + ... # LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); # /** free an embedding made with llava_image_embed_make_* */ @ctypes_function("llava_image_embed_free", [POINTER(llava_image_embed)], None) -def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): ... +def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): + ... # /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ @@ -220,7 +224,8 @@ def llava_eval_image_embed( n_batch: Union[c_int, int], n_past: "_Pointer[c_int]", /, -) -> bool: ... +) -> bool: + ... ################################################ @@ -233,10 +238,12 @@ def llava_eval_image_embed( @ctypes_function("clip_model_load", [c_char_p, c_int], clip_ctx_p_ctypes) def clip_model_load( fname: bytes, verbosity: Union[c_int, int], / -) -> Optional[clip_ctx_p]: ... +) -> Optional[clip_ctx_p]: + ... # /** free mmproj model */ # CLIP_API void clip_free(struct clip_ctx * ctx); @ctypes_function("clip_free", [clip_ctx_p_ctypes], None) -def clip_free(ctx: clip_ctx_p, /): ... +def clip_free(ctx: clip_ctx_p, /): + ... From 22cedad8a9f010bdea6186ee564da7aaa21b6684 Mon Sep 17 00:00:00 2001 From: Xu Song Date: Thu, 19 Sep 2024 10:11:01 +0800 Subject: [PATCH 544/624] fix: Fix memory allocation of ndarray (#1704) * Fix memory allocation of ndarray * Add basic LlamaState tests * Improve LlamaState test and fix rng / seed --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 34 +++++++++++++++++++--------------- tests/test_llama.py | 45 +++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 64 insertions(+), 15 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 7dd862515..80fadffe2 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -7,6 +7,7 @@ import json import ctypes import typing +import random import fnmatch import warnings import contextlib @@ -301,9 +302,11 @@ def __init__( self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1) self.n_threads_batch = n_threads_batch or multiprocessing.cpu_count() + # Used by the sampler + self._seed = seed or llama_cpp.LLAMA_DEFAULT_SEED + # Context Params self.context_params = llama_cpp.llama_context_default_params() - self.context_params.seed = seed self.context_params.n_ctx = n_ctx self.context_params.n_batch = self.n_batch self.context_params.n_threads = self.n_threads @@ -613,8 +616,7 @@ def set_seed(self, seed: int): Args: seed: The random seed. """ - # TODO: Fix this - # llama_cpp.llama_set_rng_seed(self._ctx.ctx, seed) + self._seed = seed def reset(self): """Reset the model state.""" @@ -672,7 +674,6 @@ def _init_sampler( penalize_nl: bool = True, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - seed: Optional[int] = None, ): sampler = internals.LlamaSampler() @@ -715,7 +716,7 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): if temp < 0.0: sampler.add_softmax() - sampler.add_dist(seed or llama_cpp.LLAMA_DEFAULT_SEED) + sampler.add_dist(self._seed) elif temp == 0.0: sampler.add_greedy() else: @@ -723,14 +724,14 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): mirostat_m = 100 sampler.add_mirostat( self._n_vocab, - seed or llama_cpp.LLAMA_DEFAULT_SEED, + self._seed, mirostat_tau, mirostat_eta, mirostat_m, ) elif mirostat_mode == 2: sampler.add_mirostat_v2( - seed or llama_cpp.LLAMA_DEFAULT_SEED, + self._seed, mirostat_tau, mirostat_eta, ) @@ -743,7 +744,7 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): sampler.add_top_p(top_p, min_keep) sampler.add_min_p(min_p, min_keep) sampler.add_temp(temp) - sampler.add_dist(seed or llama_cpp.LLAMA_DEFAULT_SEED) + sampler.add_dist(self._seed) return sampler def sample( @@ -826,7 +827,6 @@ def generate( logits_processor: Optional[LogitsProcessorList] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, grammar: Optional[LlamaGrammar] = None, - seed: Optional[int] = None, ) -> Generator[int, Optional[Sequence[int]], None]: """Create a generator of tokens from a prompt. @@ -865,7 +865,6 @@ def generate( penalize_nl=penalize_nl, logits_processor=logits_processor, grammar=grammar, - seed=seed, ) # Check for kv cache prefix match @@ -1301,9 +1300,10 @@ def logit_bias_processor( if self.verbose: print("Llama._create_completion: cache miss", file=sys.stderr) - # TODO: Fix this - # if seed is not None: - # self._ctx.set_rng_seed(seed) + if seed is not None: + self.set_seed(seed) + else: + self.set_seed(random.Random(self._seed).randint(0, 2 ** 32)) finish_reason = "length" multibyte_fix = 0 @@ -1324,7 +1324,6 @@ def logit_bias_processor( stopping_criteria=stopping_criteria, logits_processor=logits_processor, grammar=grammar, - seed=seed, ): if llama_cpp.llama_token_is_eog(self._model.model, token): text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) @@ -2136,14 +2135,17 @@ def save_state(self) -> LlamaState: n_tokens=self.n_tokens, llama_state=bytes(llama_state_compact), llama_state_size=n_bytes, + seed=self._seed, ) def load_state(self, state: LlamaState) -> None: # Only filling in up to `n_tokens` and then zero-ing out the rest self.scores[: state.n_tokens, :] = state.scores.copy() - self.scores[state.n_tokens :, :] = 0.0 + rest = self.scores[state.n_tokens :, :] + rest[rest > 0] = 0.0 self.input_ids = state.input_ids.copy() self.n_tokens = state.n_tokens + self._seed = state.seed state_size = state.llama_state_size LLamaStateArrayType = ctypes.c_uint8 * state_size llama_state = LLamaStateArrayType.from_buffer_copy(state.llama_state) @@ -2321,12 +2323,14 @@ def __init__( n_tokens: int, llama_state: bytes, llama_state_size: int, + seed: int, ): self.input_ids = input_ids self.scores = scores self.n_tokens = n_tokens self.llama_state = llama_state self.llama_state_size = llama_state_size + self.seed = seed LogitsProcessor = Callable[ diff --git a/tests/test_llama.py b/tests/test_llama.py index cf134c2e7..fc182ae20 100644 --- a/tests/test_llama.py +++ b/tests/test_llama.py @@ -171,3 +171,48 @@ def logit_processor_func(input_ids, logits): logits_processor=logit_processors ) assert output["choices"][0]["text"].lower().startswith("rot") + + model.set_seed(1337) + + state = model.save_state() + + output = model.create_completion( + "Pick a number from 1 to 10?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" | "10" +""") + ) + number_1 = output["choices"][0]["text"] + + output = model.create_completion( + "Pick a number from 1 to 10?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" | "10" +""") + ) + number_2 = output["choices"][0]["text"] + + model.load_state(state) + + output = model.create_completion( + "Pick a number from 1 to 10?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" | "10" +""") + ) + number_3 = output["choices"][0]["text"] + + assert number_1 != number_2 + assert number_1 == number_3 From 29afcfdff5e75d7df4c13bad0122c98661d251ab Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 18 Sep 2024 23:04:59 -0400 Subject: [PATCH 545/624] fix: Don't store scores internally unless logits_all=True. Reduces memory requirements for large context. Closes #1542 --- llama_cpp/llama.py | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 80fadffe2..70623d813 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -451,7 +451,7 @@ def free_lora_adapter(): self.n_tokens = 0 self.input_ids: npt.NDArray[np.intc] = np.ndarray((n_ctx,), dtype=np.intc) self.scores: npt.NDArray[np.single] = np.ndarray( - (n_ctx, self._n_vocab), dtype=np.single + (n_ctx if logits_all == True else n_batch, self._n_vocab), dtype=np.single ) self._mirostat_mu = ctypes.c_float( @@ -648,12 +648,14 @@ def eval(self, tokens: Sequence[int]): ) self.scores[n_past : n_past + n_tokens, :].reshape(-1)[::] = logits else: - rows = 1 - cols = self._n_vocab - logits = np.ctypeslib.as_array( - self._ctx.get_logits(), shape=(rows * cols,) - ) - self.scores[n_past + n_tokens - 1, :].reshape(-1)[::] = logits + # rows = 1 + # cols = self._n_vocab + # logits = np.ctypeslib.as_array( + # self._ctx.get_logits(), shape=(rows * cols,) + # ) + # self.scores[n_past + n_tokens - 1, :].reshape(-1)[::] = logits + # NOTE: Now that sampling is done inside the sampler, logits are only needed for logprobs which requires logits_all + pass # Update n_tokens self.n_tokens += n_tokens From 84c092063e8f222758dd3d60bdb2d1d342ac292e Mon Sep 17 00:00:00 2001 From: Jonathan Jordan Date: Thu, 19 Sep 2024 05:09:01 +0200 Subject: [PATCH 546/624] feat: Add loading sharded GGUF files from HuggingFace with Llama.from_pretrained(additional_files=[...]) . Closes #1341 Co-authored-by: Andrei --- llama_cpp/llama.py | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 70623d813..3505dcfa1 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2227,6 +2227,7 @@ def from_pretrained( cls, repo_id: str, filename: Optional[str], + additional_files: Optional[List] = None, local_dir: Optional[Union[str, os.PathLike[str]]] = None, local_dir_use_symlinks: Union[bool, Literal["auto"]] = "auto", cache_dir: Optional[Union[str, os.PathLike[str]]] = None, @@ -2239,6 +2240,7 @@ def from_pretrained( Args: repo_id: The model repo id. filename: A filename or glob pattern to match the model file in the repo. + additional_files: A list of filenames or glob patterns to match additional model files in the repo. local_dir: The local directory to save the model to. local_dir_use_symlinks: Whether to use symlinks when downloading the model. **kwargs: Additional keyword arguments to pass to the Llama constructor. @@ -2269,6 +2271,7 @@ def from_pretrained( rel_path = Path(file).relative_to(repo_id) file_list.append(str(rel_path)) + # find the only/first shard file: matching_files = [file for file in file_list if fnmatch.fnmatch(file, filename)] # type: ignore if len(matching_files) == 0: @@ -2298,6 +2301,35 @@ def from_pretrained( cache_dir=cache_dir, ) + if additional_files: + for additonal_file_name in additional_files: + # find the additional shard file: + matching_additional_files = [file for file in file_list if fnmatch.fnmatch(file, additonal_file_name)] + + if len(matching_additional_files) == 0: + raise ValueError( + f"No file found in {repo_id} that match {additonal_file_name}\n\n" + f"Available Files:\n{json.dumps(file_list)}" + ) + + if len(matching_additional_files) > 1: + raise ValueError( + f"Multiple files found in {repo_id} matching {additonal_file_name}\n\n" + f"Available Files:\n{json.dumps(files)}" + ) + + (matching_additional_file,) = matching_additional_files + + # download the additional file + hf_hub_download( + repo_id=repo_id, + filename=matching_additional_file, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cache_dir, + ) + if local_dir is None: model_path = hf_hub_download( repo_id=repo_id, @@ -2311,6 +2343,7 @@ def from_pretrained( else: model_path = os.path.join(local_dir, filename) + # loading the first file of a sharded GGUF loads all remaining shard files in the subfolder return cls( model_path=model_path, **kwargs, From 47d7a621335d4f1bc7b96e4d9a39d663c32db22b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 20 Sep 2024 18:02:11 -0400 Subject: [PATCH 547/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 64c6af319..6026da52d 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 64c6af3195c3cd4aa3328a1282d29cd2635c34c9 +Subproject commit 6026da52d6942b253df835070619775d849d0258 From 6c44a3f36b089239cb6396bb408116aad262c702 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 20 Sep 2024 18:03:37 -0400 Subject: [PATCH 548/624] feat: Add option to configure n_ubatch --- llama_cpp/llama.py | 5 +++++ llama_cpp/server/model.py | 1 + llama_cpp/server/settings.py | 3 +++ 3 files changed, 9 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3505dcfa1..ae712dedb 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -75,6 +75,7 @@ def __init__( seed: int = llama_cpp.LLAMA_DEFAULT_SEED, n_ctx: int = 512, n_batch: int = 512, + n_ubatch: int = 512, n_threads: Optional[int] = None, n_threads_batch: Optional[int] = None, rope_scaling_type: Optional[ @@ -156,6 +157,7 @@ def __init__( seed: RNG seed, -1 for random n_ctx: Text context, 0 = from model n_batch: Prompt processing maximum batch size + n_ubatch: Physical batch size n_threads: Number of threads to use for generation n_threads_batch: Number of threads to use for batch processing rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054 @@ -309,6 +311,7 @@ def __init__( self.context_params = llama_cpp.llama_context_default_params() self.context_params.n_ctx = n_ctx self.context_params.n_batch = self.n_batch + self.context_params.n_ubatch = min(self.n_batch, n_ubatch) self.context_params.n_threads = self.n_threads self.context_params.n_threads_batch = self.n_threads_batch self.context_params.rope_scaling_type = ( @@ -380,6 +383,7 @@ def __init__( self.n_batch = min(n_ctx, n_batch) self.context_params.n_ctx = self._model.n_ctx_train() self.context_params.n_batch = self.n_batch + self.context_params.n_ubatch = min(self.n_batch, n_ubatch) self._ctx = self._stack.enter_context( contextlib.closing( @@ -2071,6 +2075,7 @@ def __getstate__(self): seed=self.context_params.seed, n_ctx=self.context_params.n_ctx, n_batch=self.n_batch, + n_ubatch=self.context_params.n_ubatch, n_threads=self.context_params.n_threads, n_threads_batch=self.context_params.n_threads_batch, rope_scaling_type=self.context_params.rope_scaling_type, diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py index 901998094..c6716f919 100644 --- a/llama_cpp/server/model.py +++ b/llama_cpp/server/model.py @@ -249,6 +249,7 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: seed=settings.seed, n_ctx=settings.n_ctx, n_batch=settings.n_batch, + n_ubatch=settings.n_ubatch, n_threads=settings.n_threads, n_threads_batch=settings.n_threads_batch, rope_scaling_type=settings.rope_scaling_type, diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py index b20655813..13c951241 100644 --- a/llama_cpp/server/settings.py +++ b/llama_cpp/server/settings.py @@ -70,6 +70,9 @@ class ModelSettings(BaseSettings): n_batch: int = Field( default=512, ge=1, description="The batch size to use per eval." ) + n_ubatch: int = Field( + default=512, ge=1, description="The physical batch size used by llama.cpp" + ) n_threads: int = Field( default=max(multiprocessing.cpu_count() // 2, 1), ge=1, From 49b1e7379f5bb308acdba470e561f280008d8f15 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Sat, 21 Sep 2024 00:04:34 +0200 Subject: [PATCH 549/624] docs: Add cuda 12.5 to README.md (#1750) Typo fix Co-authored-by: Andrei --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 12b4b127d..dbaec5077 100644 --- a/README.md +++ b/README.md @@ -138,7 +138,7 @@ Where `` is one of the following: - `cu122`: CUDA 12.2 - `cu123`: CUDA 12.3 - `cu124`: CUDA 12.4 -- `cu124`: CUDA 12.5 +- `cu125`: CUDA 12.5 For example, to install the CUDA 12.1 wheel: From 1324c0c50c53e7987ff2ad1ba6d2bbe3fb13af69 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri, 20 Sep 2024 18:04:59 -0400 Subject: [PATCH 550/624] chore(deps): bump actions/cache from 3 to 4 (#1751) Bumps [actions/cache](https://github.com/actions/cache) from 3 to 4. - [Release notes](https://github.com/actions/cache/releases) - [Changelog](https://github.com/actions/cache/blob/main/RELEASES.md) - [Commits](https://github.com/actions/cache/compare/v3...v4) --- updated-dependencies: - dependency-name: actions/cache dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Andrei --- .github/workflows/build-wheels-cuda.yaml | 2 +- .github/workflows/test.yaml | 8 ++++---- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index b1e5fc1a9..647b171e8 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -72,7 +72,7 @@ jobs: - name: VS Integration Cache id: vs-integration-cache if: runner.os == 'Windows' - uses: actions/cache@v4.0.2 + uses: actions/cache@v4 with: path: ./MSBuildExtensions key: cuda-${{ matrix.cuda }}-vs-integration diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index f24f5b7e1..8f619cf1f 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -46,7 +46,7 @@ jobs: python-version: ${{ matrix.python-version }} cache: 'pip' - name: Restore model cache - uses: actions/cache@v3 + uses: actions/cache@v4 with: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} @@ -86,7 +86,7 @@ jobs: cache: 'pip' - name: Restore model cache - uses: actions/cache@v3 + uses: actions/cache@v4 with: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} @@ -129,7 +129,7 @@ jobs: cache: 'pip' - name: Restore model cache - uses: actions/cache@v3 + uses: actions/cache@v4 with: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} @@ -168,7 +168,7 @@ jobs: python-version: "3.9" - name: Restore model cache - uses: actions/cache@v3 + uses: actions/cache@v4 with: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} From 4744551b3b777d9889ee0fdd25409101d57d59f5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sat, 21 Sep 2024 20:53:16 -0400 Subject: [PATCH 551/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6026da52d..d09770cae 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6026da52d6942b253df835070619775d849d0258 +Subproject commit d09770cae71b416c032ec143dda530f7413c4038 From 926b4143e865bc6ffd405a952639099214385139 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 25 Sep 2024 13:23:18 -0400 Subject: [PATCH 552/624] feat: Update llama.cpp --- llama_cpp/_logger.py | 7 +++++++ vendor/llama.cpp | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/llama_cpp/_logger.py b/llama_cpp/_logger.py index 157af692f..787b3f108 100644 --- a/llama_cpp/_logger.py +++ b/llama_cpp/_logger.py @@ -10,6 +10,7 @@ # GGML_LOG_LEVEL_WARN = 2, # GGML_LOG_LEVEL_ERROR = 3, # GGML_LOG_LEVEL_DEBUG = 4, +# GGML_LOG_LEVEL_CONT = 5, // continue previous log # }; GGML_LOG_LEVEL_TO_LOGGING_LEVEL = { 0: logging.CRITICAL, @@ -17,10 +18,12 @@ 2: logging.WARNING, 3: logging.ERROR, 4: logging.DEBUG, + 5: logging.DEBUG, } logger = logging.getLogger("llama-cpp-python") +_last_log_level = GGML_LOG_LEVEL_TO_LOGGING_LEVEL[0] # typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data); @llama_cpp.llama_log_callback @@ -29,8 +32,12 @@ def llama_log_callback( text: bytes, user_data: ctypes.c_void_p, ): + # TODO: Correctly implement continue previous log + global _last_log_level + log_level = GGML_LOG_LEVEL_TO_LOGGING_LEVEL[level] if level != 5 else _last_log_level if logger.level <= GGML_LOG_LEVEL_TO_LOGGING_LEVEL[level]: print(text.decode("utf-8"), end="", flush=True, file=sys.stderr) + _last_log_level = log_level llama_cpp.llama_log_set(llama_log_callback, ctypes.c_void_p(0)) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index d09770cae..ea9c32be7 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit d09770cae71b416c032ec143dda530f7413c4038 +Subproject commit ea9c32be71b91b42ecc538bd902e93cbb5fb36cb From b3dfb420e79fcdb326bc3bac861f73a152380946 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 25 Sep 2024 14:05:08 -0400 Subject: [PATCH 553/624] chore: Bump version --- CHANGELOG.md | 13 +++++++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 14 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index f01b1b9bf..14f68a456 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,19 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.0] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ea9c32be71b91b42ecc538bd902e93cbb5fb36cb +- feat: Enable detokenizing special tokens with special=True by @benniekiss in #1596 +- feat(ci): Speed up CI workflows using uv, add support for CUDA 12.5 wheels by @Smartappli in e529940f45d42ed8aa31334123b8d66bc67b0e78 +- feat: Add loading sharded GGUF files from HuggingFace with Llama.from_pretrained(additional_files=[...]) by @Gnurro in 84c092063e8f222758dd3d60bdb2d1d342ac292e +- feat: Add option to configure n_ubatch by @abetlen in 6c44a3f36b089239cb6396bb408116aad262c702 +- feat: Update sampling API for llama.cpp. Sampling now uses sampler chain by @abetlen in f8fcb3ea3424bcfba3a5437626a994771a02324b +- fix: Don't store scores internally unless logits_all=True. Reduces memory requirements for large context by @abetlen in 29afcfdff5e75d7df4c13bad0122c98661d251ab +- fix: Fix memory allocation of ndarray in by @xu-song in #1704 +- fix: Use system message in og qwen format by @abetlen in 98eb092d3c6e7c142c4ba2faaca6c091718abbb3 + + ## [0.2.90] - feat: Update llama.cpp to ggerganov/llama.cpp@1d1ccce67613674c75c9c7e3fa4c1e24e428ba48 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index db1f1d0fa..70fde2041 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.90" +__version__ = "0.3.0" From 8e07db0f87f2e0cd543f1449a5139b7026a19e9f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 25 Sep 2024 14:17:04 -0400 Subject: [PATCH 554/624] fix: install build dependency --- .github/workflows/publish.yaml | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index 67816ea03..bb76f5394 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -13,34 +13,36 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" - + - name: Set up Python uses: actions/setup-python@v5 with: python-version: "3.9" - + - name: Install dependencies (Linux/MacOS) if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install build shell: bash - name: Install dependencies (Windows) if: runner.os == 'Windows' env: - RUST_LOG: trace + RUST_LOG: trace run: | python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose + python -m uv pip install build shell: cmd - + - name: Build source distribution run: | python -m build --sdist - + - name: Publish distribution to PyPI # TODO: move to tag based releases # if: startsWith(github.ref, 'refs/tags') From 65222bc4931e8d5cf637a8d5f966da2f46897837 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 25 Sep 2024 14:27:19 -0400 Subject: [PATCH 555/624] fix: install build dependency --- .github/workflows/build-and-release.yaml | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 3f85a30fe..bfa97decd 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -92,7 +92,7 @@ jobs: - uses: actions/checkout@v4 with: submodules: "recursive" - + - uses: actions/setup-python@v5 with: python-version: "3.9" @@ -103,6 +103,7 @@ jobs: python -m pip install --upgrade pip python -m pip install uv RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install build shell: bash - name: Install dependencies (Windows) @@ -113,12 +114,13 @@ jobs: python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose + python -m uv pip install build shell: cmd - + - name: Build source distribution run: | python -m build --sdist - + - uses: actions/upload-artifact@v4 with: name: sdist From 9992c5084a3df2f533e265d10f81d4269b97a1e6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 26 Sep 2024 00:34:21 -0400 Subject: [PATCH 556/624] fix: Fix speculative decoding --- llama_cpp/llama.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index ae712dedb..2ae65a83d 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -807,8 +807,10 @@ def sample( grammar=grammar, ) + ridx = idx - self.n_tokens if idx is not None else -1 + assert self.ctx is not None - token = self._sampler.sample(self._ctx, -1) + token = self._sampler.sample(self._ctx, ridx) if tmp_sampler: self._sampler = None return token From 11d9562fd63473dbf4ec91d942ca5f197deec462 Mon Sep 17 00:00:00 2001 From: Xu Song Date: Thu, 26 Sep 2024 12:37:17 +0800 Subject: [PATCH 557/624] misc: Rename all_text to remaining_text (#1658) Co-authored-by: Andrei --- llama_cpp/llama.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 2ae65a83d..4f22fae84 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1519,15 +1519,15 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - all_text = self.detokenize( + remaining_text = self.detokenize( remaining_tokens, prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], ) - any_stop = [s for s in stop_sequences if s in all_text] + any_stop = [s for s in stop_sequences if s in remaining_text] if len(any_stop) > 0: - end = min(all_text.index(stop) for stop in any_stop) + end = min(remaining_text.index(stop) for stop in any_stop) else: - end = len(all_text) + end = len(remaining_text) token_end_position = 0 for token in remaining_tokens: From e975dabf74b3ad85689c9a07719cbb181313139b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 26 Sep 2024 01:11:02 -0400 Subject: [PATCH 558/624] fix: Additional fixes for speculative decoding --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 4f22fae84..babb30cf0 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -930,7 +930,7 @@ def generate( sample_idx += 1 if stopping_criteria is not None and stopping_criteria( - self._input_ids, self._scores[-1, :] + self._input_ids[: sample_idx], self._scores[sample_idx - self.n_tokens, :] ): return tokens_or_none = yield token From dca0c9a09cea39a8454dd897253be540e82f369d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 26 Sep 2024 10:09:22 -0400 Subject: [PATCH 559/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ea9c32be7..95bc82fbc 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ea9c32be71b91b42ecc538bd902e93cbb5fb36cb +Subproject commit 95bc82fbc0df6d48cf66c857a4dda3d044f45ca2 From 01c76078eea77f12fb01aff605859afc372bf598 Mon Sep 17 00:00:00 2001 From: Andrei Date: Thu, 26 Sep 2024 11:23:00 -0400 Subject: [PATCH 560/624] feat: Expose libggml in internal APIs (#1761) * Expose libggml and refactor ctypes extension * Only expose libggml * Use ctypes_extensions module for libllama and libllava --- llama_cpp/_ctypes_extensions.py | 131 ++++++++++++++++++++++++++++++ llama_cpp/_ggml.py | 12 +++ llama_cpp/llama_cpp.py | 140 ++++---------------------------- llama_cpp/llava_cpp.py | 113 +++----------------------- 4 files changed, 171 insertions(+), 225 deletions(-) create mode 100644 llama_cpp/_ctypes_extensions.py create mode 100644 llama_cpp/_ggml.py diff --git a/llama_cpp/_ctypes_extensions.py b/llama_cpp/_ctypes_extensions.py new file mode 100644 index 000000000..e88ed387d --- /dev/null +++ b/llama_cpp/_ctypes_extensions.py @@ -0,0 +1,131 @@ +from __future__ import annotations + +import sys +import os +import ctypes +import functools +import pathlib + +from typing import ( + Any, + Callable, + List, + Union, + Optional, + TYPE_CHECKING, + TypeVar, + Generic, +) +from typing_extensions import TypeAlias + + +# Load the library +def load_shared_library(lib_base_name: str, base_path: pathlib.Path): + """Platform independent shared library loader""" + # Searching for the library in the current directory under the name "libllama" (default name + # for llamacpp) and "llama" (default name for this repo) + lib_paths: List[pathlib.Path] = [] + # Determine the file extension based on the platform + if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"): + lib_paths += [ + base_path / f"lib{lib_base_name}.so", + ] + elif sys.platform == "darwin": + lib_paths += [ + base_path / f"lib{lib_base_name}.so", + base_path / f"lib{lib_base_name}.dylib", + ] + elif sys.platform == "win32": + lib_paths += [ + base_path / f"{lib_base_name}.dll", + base_path / f"lib{lib_base_name}.dll", + ] + else: + raise RuntimeError("Unsupported platform") + + cdll_args = dict() # type: ignore + + # Add the library directory to the DLL search path on Windows (if needed) + if sys.platform == "win32": + os.add_dll_directory(str(base_path)) + os.environ["PATH"] = str(base_path) + os.pathsep + os.environ["PATH"] + + if sys.platform == "win32" and sys.version_info >= (3, 8): + os.add_dll_directory(str(base_path)) + if "CUDA_PATH" in os.environ: + os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin")) + os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib")) + if "HIP_PATH" in os.environ: + os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "bin")) + os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "lib")) + cdll_args["winmode"] = ctypes.RTLD_GLOBAL + + # Try to load the shared library, handling potential errors + for lib_path in lib_paths: + if lib_path.exists(): + try: + return ctypes.CDLL(str(lib_path), **cdll_args) # type: ignore + except Exception as e: + raise RuntimeError(f"Failed to load shared library '{lib_path}': {e}") + + raise FileNotFoundError( + f"Shared library with base name '{lib_base_name}' not found" + ) + + +# ctypes sane type hint helpers +# +# - Generic Pointer and Array types +# - PointerOrRef type with a type hinted byref function +# +# NOTE: Only use these for static type checking not for runtime checks +# no good will come of that + +if TYPE_CHECKING: + CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore + + CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore + + CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore + + CtypesVoidPointer: TypeAlias = ctypes.c_void_p + + class CtypesRef(Generic[CtypesCData]): + pass + + CtypesPointerOrRef: TypeAlias = Union[ + CtypesPointer[CtypesCData], CtypesRef[CtypesCData] + ] + + CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore + +F = TypeVar("F", bound=Callable[..., Any]) + + +def ctypes_function_for_shared_library(lib: ctypes.CDLL): + """Decorator for defining ctypes functions with type hints""" + + def ctypes_function( + name: str, argtypes: List[Any], restype: Any, enabled: bool = True + ): + def decorator(f: F) -> F: + if enabled: + func = getattr(lib, name) + func.argtypes = argtypes + func.restype = restype + functools.wraps(f)(func) + return func + else: + return f + + return decorator + + return ctypes_function + + +def _byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCData]: + """Type-annotated version of ctypes.byref""" + ... + + +byref = _byref if TYPE_CHECKING else ctypes.byref diff --git a/llama_cpp/_ggml.py b/llama_cpp/_ggml.py new file mode 100644 index 000000000..5bee8a93b --- /dev/null +++ b/llama_cpp/_ggml.py @@ -0,0 +1,12 @@ +"""Internal module use at your own risk + +This module provides a minimal interface for working with ggml tensors from llama-cpp-python +""" +import os +import pathlib + +import llama_cpp._ctypes_extensions as ctypes_ext + +libggml_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" +libggml = ctypes_ext.load_shared_library("ggml", libggml_base_path) + diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 6b82753e1..de427970c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1,151 +1,45 @@ from __future__ import annotations -import sys import os import ctypes -import functools import pathlib from typing import ( - Any, Callable, - List, Union, NewType, Optional, TYPE_CHECKING, - TypeVar, - Generic, ) -from typing_extensions import TypeAlias +from llama_cpp._ctypes_extensions import ( + load_shared_library, + byref, + ctypes_function_for_shared_library, +) -# Load the library -def _load_shared_library(lib_base_name: str): - # Construct the paths to the possible shared library names - _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" - # Searching for the library in the current directory under the name "libllama" (default name - # for llamacpp) and "llama" (default name for this repo) - _lib_paths: List[pathlib.Path] = [] - # Determine the file extension based on the platform - if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"): - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - ] - elif sys.platform == "darwin": - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - _base_path / f"lib{lib_base_name}.dylib", - ] - elif sys.platform == "win32": - _lib_paths += [ - _base_path / f"{lib_base_name}.dll", - _base_path / f"lib{lib_base_name}.dll", - ] - else: - raise RuntimeError("Unsupported platform") - - if "LLAMA_CPP_LIB" in os.environ: - lib_base_name = os.environ["LLAMA_CPP_LIB"] - _lib = pathlib.Path(lib_base_name) - _base_path = _lib.parent.resolve() - _lib_paths = [_lib.resolve()] - - cdll_args = dict() # type: ignore - - # Add the library directory to the DLL search path on Windows (if needed) - if sys.platform == "win32": - os.add_dll_directory(str(_base_path)) - os.environ["PATH"] = str(_base_path) + os.pathsep + os.environ["PATH"] - - if sys.platform == "win32" and sys.version_info >= (3, 8): - os.add_dll_directory(str(_base_path)) - if "CUDA_PATH" in os.environ: - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin")) - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib")) - if "HIP_PATH" in os.environ: - os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "bin")) - os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "lib")) - cdll_args["winmode"] = ctypes.RTLD_GLOBAL - - # Try to load the shared library, handling potential errors - for _lib_path in _lib_paths: - if _lib_path.exists(): - try: - return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore - except Exception as e: - raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") - - raise FileNotFoundError( - f"Shared library with base name '{lib_base_name}' not found" +if TYPE_CHECKING: + from llama_cpp._ctypes_extensions import ( + CtypesCData, + CtypesArray, + CtypesPointer, + CtypesVoidPointer, + CtypesRef, + CtypesPointerOrRef, + CtypesFuncPointer, ) # Specify the base name of the shared library to load _lib_base_name = "llama" - +_override_base_path = os.environ.get("LLAMA_CPP_LIB_PATH") +_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" if _override_base_path is None else pathlib.Path(_override_base_path) # Load the library -_lib = _load_shared_library(_lib_base_name) - - -# ctypes sane type hint helpers -# -# - Generic Pointer and Array types -# - PointerOrRef type with a type hinted byref function -# -# NOTE: Only use these for static type checking not for runtime checks -# no good will come of that - -if TYPE_CHECKING: - CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore - - CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore - - CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore - - CtypesVoidPointer: TypeAlias = ctypes.c_void_p - - class CtypesRef(Generic[CtypesCData]): - pass - - CtypesPointerOrRef: TypeAlias = Union[ - CtypesPointer[CtypesCData], CtypesRef[CtypesCData] - ] - - CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore - -F = TypeVar("F", bound=Callable[..., Any]) - - -def ctypes_function_for_shared_library(lib: ctypes.CDLL): - def ctypes_function( - name: str, argtypes: List[Any], restype: Any, enabled: bool = True - ): - def decorator(f: F) -> F: - if enabled: - func = getattr(lib, name) - func.argtypes = argtypes - func.restype = restype - functools.wraps(f)(func) - return func - else: - return f - - return decorator - - return ctypes_function - +_lib = load_shared_library(_lib_base_name, _base_path) ctypes_function = ctypes_function_for_shared_library(_lib) -def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCData]: - """Type-annotated version of ctypes.byref""" - ... - - -byref = ctypes.byref # type: ignore - # from ggml.h # // NOTE: always add types at the end of the enum to keep backward compatibility # enum ggml_type { diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index 7d97dc0fd..d9dfaf5fd 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -1,9 +1,6 @@ from __future__ import annotations -import sys import os -import ctypes -import functools from ctypes import ( c_bool, c_char_p, @@ -17,121 +14,32 @@ ) import pathlib from typing import ( - List, Union, NewType, Optional, - TypeVar, - Callable, - Any, TYPE_CHECKING, - Generic, ) -from typing_extensions import TypeAlias import llama_cpp.llama_cpp as llama_cpp +from llama_cpp._ctypes_extensions import ( + load_shared_library, + ctypes_function_for_shared_library, +) -# Load the library -def _load_shared_library(lib_base_name: str): - # Construct the paths to the possible shared library names - _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" - # Searching for the library in the current directory under the name "libllama" (default name - # for llamacpp) and "llama" (default name for this repo) - _lib_paths: List[pathlib.Path] = [] - # Determine the file extension based on the platform - if sys.platform.startswith("linux"): - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - ] - elif sys.platform == "darwin": - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - _base_path / f"lib{lib_base_name}.dylib", - ] - elif sys.platform == "win32": - _lib_paths += [ - _base_path / f"{lib_base_name}.dll", - _base_path / f"lib{lib_base_name}.dll", - ] - else: - raise RuntimeError("Unsupported platform") - - if "LLAVA_CPP_LIB" in os.environ: - lib_base_name = os.environ["LLAVA_CPP_LIB"] - _lib = pathlib.Path(lib_base_name) - _base_path = _lib.parent.resolve() - _lib_paths = [_lib.resolve()] - - cdll_args = dict() # type: ignore - # Add the library directory to the DLL search path on Windows (if needed) - if sys.platform == "win32" and sys.version_info >= (3, 8): - os.add_dll_directory(str(_base_path)) - if "CUDA_PATH" in os.environ: - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin")) - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib")) - cdll_args["winmode"] = ctypes.RTLD_GLOBAL - - # Try to load the shared library, handling potential errors - for _lib_path in _lib_paths: - if _lib_path.exists(): - try: - return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore - except Exception as e: - raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") - - raise FileNotFoundError( - f"Shared library with base name '{lib_base_name}' not found" +if TYPE_CHECKING: + from llama_cpp._ctypes_extensions import ( + CtypesArray, ) # Specify the base name of the shared library to load _libllava_base_name = "llava" +_libllava_override_path = os.environ.get("LLAVA_CPP_LIB") +_libllava_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" if _libllava_override_path is None else pathlib.Path() # Load the library -_libllava = _load_shared_library(_libllava_base_name) - -# ctypes helper - -if TYPE_CHECKING: - CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore - - CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore - - CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore - - CtypesVoidPointer: TypeAlias = ctypes.c_void_p - - class CtypesRef(Generic[CtypesCData]): - pass - - CtypesPointerOrRef: TypeAlias = Union[ - CtypesPointer[CtypesCData], CtypesRef[CtypesCData] - ] - - CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore - -F = TypeVar("F", bound=Callable[..., Any]) - - -def ctypes_function_for_shared_library(lib: ctypes.CDLL): - def ctypes_function( - name: str, argtypes: List[Any], restype: Any, enabled: bool = True - ): - def decorator(f: F) -> F: - if enabled: - func = getattr(lib, name) - func.argtypes = argtypes - func.restype = restype - functools.wraps(f)(func) - return func - else: - return f - - return decorator - - return ctypes_function - +_libllava = load_shared_library(_libllava_base_name, _libllava_base_path) ctypes_function = ctypes_function_for_shared_library(_libllava) @@ -247,3 +155,4 @@ def clip_model_load( @ctypes_function("clip_free", [clip_ctx_p_ctypes], None) def clip_free(ctx: clip_ctx_p, /): ... + From 57e70bb5693fb0330b11c50f707adb88a6be7b85 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 29 Sep 2024 15:23:24 -0400 Subject: [PATCH 561/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 16 ++++++++++++---- vendor/llama.cpp | 2 +- 2 files changed, 13 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index de427970c..97c969136 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -220,6 +220,7 @@ # LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, # LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, # LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, +# LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -247,6 +248,7 @@ LLAMA_VOCAB_PRE_TYPE_BLOOM = 23 LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24 LLAMA_VOCAB_PRE_TYPE_EXAONE = 25 +LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 # // note: these values should be synchronized with ggml_rope @@ -404,12 +406,14 @@ # LLAMA_POOLING_TYPE_MEAN = 1, # LLAMA_POOLING_TYPE_CLS = 2, # LLAMA_POOLING_TYPE_LAST = 3, +# LLAMA_POOLING_TYPE_RANK = 4, // used by reranking models to attach the classification head to the graph # }; LLAMA_POOLING_TYPE_UNSPECIFIED = -1 LLAMA_POOLING_TYPE_NONE = 0 LLAMA_POOLING_TYPE_MEAN = 1 LLAMA_POOLING_TYPE_CLS = 2 LLAMA_POOLING_TYPE_LAST = 3 +LLAMA_POOLING_TYPE_RANK = 4 # enum llama_attention_type { # LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1, @@ -420,10 +424,11 @@ LLAMA_ATTENTION_TYPE_CAUSAL = 0 LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1 + # enum llama_split_mode { -# LLAMA_SPLIT_MODE_NONE = 0, // single GPU -# LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs -# LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs +# LLAMA_SPLIT_MODE_NONE = 0, // single GPU +# LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs +# LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs # }; LLAMA_SPLIT_MODE_NONE = 0 LLAMA_SPLIT_MODE_LAYER = 1 @@ -2520,7 +2525,8 @@ def llama_get_embeddings_ith( # // Get the embeddings for a sequence id # // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE -# // shape: [n_embd] (1-dimensional) +# // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence +# // otherwise: float[n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id); @ctypes_function( "llama_get_embeddings_seq", @@ -2672,6 +2678,8 @@ def llama_token_eot(model: llama_model_p, /) -> int: # // # // Tokenization # // +# // The API is thread-safe. +# // # /// @details Convert the provided text into tokens. diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 95bc82fbc..c919d5db3 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 95bc82fbc0df6d48cf66c857a4dda3d044f45ca2 +Subproject commit c919d5db39c8a7fcb64737f008e4b105ee0acd20 From 7c4aead82d349469bbbe7d8c0f4678825873c039 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 29 Sep 2024 15:26:54 -0400 Subject: [PATCH 562/624] chore: Bump version --- CHANGELOG.md | 7 +++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 14f68a456..c68e96b87 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.1] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c919d5db39c8a7fcb64737f008e4b105ee0acd20 +- feat: Expose libggml in internal APIs by @abetlen in #1761 +- fix: Fix speculative decoding by @abetlen in 9992c5084a3df2f533e265d10f81d4269b97a1e6 and e975dabf74b3ad85689c9a07719cbb181313139b +- misc: Rename all_text to remaining_text by @xu-song in #1658 + ## [0.3.0] - feat: Update llama.cpp to ggerganov/llama.cpp@ea9c32be71b91b42ecc538bd902e93cbb5fb36cb diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 70fde2041..c11b879d6 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.0" +__version__ = "0.3.1" From 7403e002b8e033c0a34e93fba2b311e2118487fe Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 22 Oct 2024 09:33:48 -0400 Subject: [PATCH 563/624] feat: Update llama.cpp --- CMakeLists.txt | 18 ++++- llama_cpp/llama_cpp.py | 152 ++++++++++++++++++++++++++++++++--------- vendor/llama.cpp | 2 +- 3 files changed, 139 insertions(+), 33 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index c6b35ed6c..9af5b2180 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -55,6 +55,9 @@ if (LLAMA_BUILD) set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) set(CMAKE_SKIP_RPATH FALSE) + # Enable building of the common library + set(LLAMA_BUILD_COMMON ON CACHE BOOL "Build llama.cpp common library" FORCE) + # Building llama if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") # Need to disable these llama.cpp flags on Apple x86_64, @@ -106,7 +109,7 @@ if (LLAMA_BUILD) # Building llava add_subdirectory(vendor/llama.cpp/examples/llava) set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") - # Set CUDA_ARCHITECTURES to OFF on windows + # Set CUDA_ARCHITECTURES to OFF on Windows if (WIN32) set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) endif() @@ -121,5 +124,18 @@ if (LLAMA_BUILD) DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib ) endif() + + # Fix for llava build: Add include directory for llama.h + # Move these commands after the add_subdirectory call + target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) + + if (BUILD_SHARED_LIBS) + target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) + endif() + + target_include_directories(llama-llava-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + target_include_directories(llama-minicpmv-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) endif() endif() diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 97c969136..66feed896 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -464,6 +464,8 @@ class llama_token_data(ctypes.Structure): # typedef struct llama_token_data_array { +# // TODO: consider SoA +# // NOTE: this pointer can be modified by the samplers # llama_token_data * data; # size_t size; # int64_t selected; // this is the index in the data array (i.e. not the token id) @@ -507,8 +509,11 @@ class llama_token_data_array(ctypes.Structure): # // - token : the token ids of the input (used when embd is NULL) # // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) # // - pos : the positions of the respective token in the sequence +# // (if set to NULL, the token position will be tracked automatically by llama_decode) # // - seq_id : the sequence to which the respective token belongs +# // (if set to NULL, the sequence ID will be assumed to be 0) # // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output +# // (if set to NULL, only the logits for last token will be returned) # // # typedef struct llama_batch { # int32_t n_tokens; @@ -519,16 +524,6 @@ class llama_token_data_array(ctypes.Structure): # int32_t * n_seq_id; # llama_seq_id ** seq_id; # int8_t * logits; // TODO: rename this to "output" - - -# // NOTE: helpers for smooth API transition - can be deprecated in the future -# // for future-proof code, use the above fields instead and ignore everything below -# // -# // pos[i] = all_pos_0 + i*all_pos_1 -# // -# llama_pos all_pos_0; // used if pos == NULL -# llama_pos all_pos_1; // used if pos == NULL -# llama_seq_id all_seq_id; // used if seq_id == NULL # } llama_batch; class llama_batch(ctypes.Structure): """Input data for llama_decode @@ -563,9 +558,6 @@ class llama_batch(ctypes.Structure): ("n_seq_id", ctypes.POINTER(ctypes.c_int32)), ("seq_id", ctypes.POINTER(ctypes.POINTER(llama_seq_id))), ("logits", ctypes.POINTER(ctypes.c_int8)), - ("all_pos_0", llama_pos), - ("all_pos_1", llama_pos), - ("all_seq_id", llama_seq_id), ] @@ -1170,6 +1162,12 @@ def llama_supports_gpu_offload() -> bool: ... +# LLAMA_API bool llama_supports_rpc (void); +@ctypes_function("llama_supports_rpc", [], ctypes.c_bool) +def llama_supports_rpc() -> bool: + ... + + # LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) def llama_n_ctx(ctx: llama_context_p, /) -> int: @@ -2255,30 +2253,26 @@ def llama_state_seq_load_file( # // -# // Return batch for single sequence of tokens starting at pos_0 +# // Return batch for single sequence of tokens +# // The sequence ID will be fixed to 0 +# // The position of the tokens will be tracked automatically by llama_decode # // # // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it # // # LLAMA_API struct llama_batch llama_batch_get_one( # llama_token * tokens, -# int32_t n_tokens, -# llama_pos pos_0, -# llama_seq_id seq_id); +# int32_t n_tokens); @ctypes_function( "llama_batch_get_one", [ llama_token_p, - ctypes.c_int, - llama_pos, - llama_seq_id, + ctypes.c_int32, ], llama_batch, ) def llama_batch_get_one( tokens: CtypesArray[llama_token], n_tokens: Union[ctypes.c_int, int], - pos_0: Union[llama_pos, int], - seq_id: llama_seq_id, /, ) -> llama_batch: """Return batch for single sequence of tokens starting at pos_0 @@ -2616,6 +2610,13 @@ def llama_token_eos(model: llama_model_p, /) -> int: ... +# LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn +@ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) +def llama_token_eot(model: llama_model_p, /) -> int: + """end-of-turn""" + ... + + # LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification @ctypes_function("llama_token_cls", [llama_model_p_ctypes], llama_token) def llama_token_cls(model: llama_model_p, /) -> int: @@ -2650,30 +2651,54 @@ def llama_add_eos_token(model: llama_model_p, /) -> bool: # // Codellama infill tokens -# LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix +# DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead"); @ctypes_function("llama_token_prefix", [llama_model_p_ctypes], llama_token) def llama_token_prefix(model: llama_model_p) -> int: """codellama infill tokens""" ... -# LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle +# DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead"); @ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) def llama_token_middle(model: llama_model_p, /) -> int: ... -# LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix +# DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead"); @ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) def llama_token_suffix(model: llama_model_p, /) -> int: ... -# LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle -@ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model); +@ctypes_function("llama_token_fim_pre", [llama_model_p_ctypes], llama_token) +def llama_token_fim_pre(model: llama_model_p, /) -> int: + ... + +# LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model); +@ctypes_function("llama_token_fim_suf", [llama_model_p_ctypes], llama_token) +def llama_token_fim_suf(model: llama_model_p, /) -> int: ... +# LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model); +@ctypes_function("llama_token_fim_mid", [llama_model_p_ctypes], llama_token) +def llama_token_fim_mid(model: llama_model_p, /) -> int: + ... + +# LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model); +@ctypes_function("llama_token_fim_pad", [llama_model_p_ctypes], llama_token) +def llama_token_fim_pad(model: llama_model_p, /) -> int: + ... + +# LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model); +@ctypes_function("llama_token_fim_rep", [llama_model_p_ctypes], llama_token) +def llama_token_fim_rep(model: llama_model_p, /) -> int: + ... + +# LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model); +@ctypes_function("llama_token_fim_sep", [llama_model_p_ctypes], llama_token) +def llama_token_fim_sep(model: llama_model_p, /) -> int: + ... # // # // Tokenization @@ -2786,6 +2811,23 @@ def llama_token_to_piece( ... +# # // check if token0 is contained as a prefix in token1 +# # LLAMA_API bool llama_token_is_prefix( +# # const struct llama_model * model, +# # llama_token token0, +# # llama_token token1); +# @ctypes_function( +# "llama_token_is_prefix", +# [llama_model_p_ctypes, llama_token, llama_token], +# ctypes.c_bool, +# ) +# def llama_token_is_prefix( +# model: llama_model_p, token0: Union[llama_token, int], token1: Union[llama_token, int], / +# ) -> bool: +# """Check if token0 is contained as a prefix in token1""" +# ... + + # /// @details Convert the provided tokens into text (inverse of llama_tokenize()). # /// @param text The char pointer must be large enough to hold the resulting text. # /// @return Returns the number of chars/bytes on success, no more than text_len_max. @@ -3099,20 +3141,22 @@ def llama_sampler_chain_remove( # // available samplers: # -# LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void); +# LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void); @ctypes_function("llama_sampler_init_greedy", [], llama_sampler_p_ctypes) def llama_sampler_init_greedy() -> llama_sampler_p: ... -# LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed); +# LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed); @ctypes_function("llama_sampler_init_dist", [ctypes.c_uint32], llama_sampler_p_ctypes) def llama_sampler_init_dist(seed: int) -> llama_sampler_p: ... # /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. -# LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void); +# /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first. +# DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void), +# "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)"); @ctypes_function("llama_sampler_init_softmax", [], llama_sampler_p_ctypes) def llama_sampler_init_softmax() -> llama_sampler_p: ... @@ -3188,6 +3232,19 @@ def llama_sampler_init_temp_ext( ... +# /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335 +# LLAMA_API struct llama_sampler * llama_sampler_init_xtc (float p, float t, size_t min_keep, uint32_t seed); +@ctypes_function( + "llama_sampler_init_xtc", + [ctypes.c_float, ctypes.c_float, ctypes.c_size_t, ctypes.c_uint32], + llama_sampler_p_ctypes, +) +def llama_sampler_init_xtc( + p: float, t: float, min_keep: int, seed: int, / +) -> llama_sampler_p: + ... + + # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. # /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -3301,6 +3358,39 @@ def llama_sampler_init_logit_bias( ... +# // this sampler is meant to be used for fill-in-the-middle infilling +# // it's supposed to be used after top_k + top_p sampling +# // +# // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG +# // 2. combine probs of tokens that have the same prefix +# // +# // example: +# // +# // - before: +# // "hel": 0.5 +# // "hell": 0.2 +# // "hello": 0.1 +# // "dummy": 0.1 +# // +# // - after: +# // "hel": 0.8 +# // "dummy": 0.1 +# // +# // 3. discard non-EOG tokens with low prob +# // 4. if no tokens are left -> pick EOT +# // +# LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model); +@ctypes_function( + "llama_sampler_init_infill", + [llama_model_p_ctypes], + llama_sampler_p_ctypes, +) +def llama_sampler_init_infill(model: llama_model_p, /) -> llama_sampler_p: + """This sampler is meant to be used for fill-in-the-middle infilling. + """ + ... + + # // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise # LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl); @ctypes_function( diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c919d5db3..c421ac072 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c919d5db39c8a7fcb64737f008e4b105ee0acd20 +Subproject commit c421ac072d46172ab18924e1e8be53680b54ed3b From e712cffc0c22d5406c6fe02a1769992f2181f229 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 31 Oct 2024 13:50:18 -0400 Subject: [PATCH 564/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 14 ------------- llama_cpp/llama.py | 1 - llama_cpp/llama_cpp.py | 46 +++++++++++++++++++++++++++++++---------- vendor/llama.cpp | 2 +- 4 files changed, 36 insertions(+), 27 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 0aff34844..994d5f149 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -362,13 +362,6 @@ def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: i self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep ) - def sample_tail_free( - self, candidates: "_LlamaTokenDataArray", z: float, min_keep: int - ): - llama_cpp.llama_sample_tail_free( - self.ctx, llama_cpp.byref(candidates.candidates), z, min_keep - ) - def sample_typical( self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int ): @@ -685,9 +678,6 @@ def sample( ctx_main.sample_top_k( token_data_array, self.params.top_k, min_keep=min_keep ) - ctx_main.sample_tail_free( - token_data_array, self.params.tfs_z, min_keep=min_keep - ) ctx_main.sample_typical( token_data_array, self.params.typical_p, min_keep=min_keep ) @@ -776,10 +766,6 @@ def add_min_p(self, p: float, min_keep: int): sampler = llama_cpp.llama_sampler_init_min_p(p, min_keep) self._add_sampler(sampler) - def add_tail_free(self, z: float, min_keep: int): - sampler = llama_cpp.llama_sampler_init_tail_free(z, min_keep) - self._add_sampler(sampler) - def add_typical(self, p: float, min_keep: int): sampler = llama_cpp.llama_sampler_init_typical(p, min_keep) self._add_sampler(sampler) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index babb30cf0..d15a88b00 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -745,7 +745,6 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): n_probs = 0 min_keep = max(1, n_probs) sampler.add_top_k(top_k) - sampler.add_tail_free(tfs_z, min_keep) sampler.add_typical(typical_p, min_keep) sampler.add_top_p(top_p, min_keep) sampler.add_min_p(min_p, min_keep) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 66feed896..457c6dddb 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -3191,17 +3191,6 @@ def llama_sampler_init_min_p(p: float, min_keep: int) -> llama_sampler_p: ... -# /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. -# LLAMA_API struct llama_sampler * llama_sampler_init_tail_free (float z, size_t min_keep); -@ctypes_function( - "llama_sampler_init_tail_free", - [ctypes.c_float, ctypes.c_size_t], - llama_sampler_p_ctypes, -) -def llama_sampler_init_tail_free(z: float, min_keep: int) -> llama_sampler_p: - ... - - # /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. # LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep); @ctypes_function( @@ -3343,6 +3332,41 @@ def llama_sampler_init_penalties( ... +# /// @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982 +# LLAMA_API struct llama_sampler * llama_sampler_init_dry( +# const struct llama_model * model, +# float dry_multiplier, +# float dry_base, +# int32_t dry_allowed_length, +# int32_t dry_penalty_last_n, +# const char ** seq_breakers, +# size_t num_breakers); +@ctypes_function( + "llama_sampler_init_dry", + [ + llama_model_p_ctypes, + ctypes.c_float, + ctypes.c_float, + ctypes.c_int32, + ctypes.c_int32, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_dry( + model: llama_model_p, + dry_multiplier: float, + dry_base: float, + dry_allowed_length: int, + dry_penalty_last_n: int, + seq_breakers: CtypesArray[bytes], + num_breakers: int, + /, +) -> llama_sampler_p: + ... + + # LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias( # int32_t n_vocab, # int32_t n_logit_bias, diff --git a/vendor/llama.cpp b/vendor/llama.cpp index c421ac072..0a683e808 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit c421ac072d46172ab18924e1e8be53680b54ed3b +Subproject commit 0a683e8088d849626e7471f9e2ed381f7dbdf2e9 From cafa33e1a8646258f2e44dd95782f4903d6944c0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 15 Nov 2024 18:00:30 -0500 Subject: [PATCH 565/624] feat: Update llama.cpp --- CMakeLists.txt | 16 +++++++++++++++- vendor/llama.cpp | 2 +- 2 files changed, 16 insertions(+), 2 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 9af5b2180..7d31ffd6c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -75,7 +75,21 @@ if (LLAMA_BUILD) add_subdirectory(vendor/llama.cpp) llama_cpp_python_install_target(llama) llama_cpp_python_install_target(ggml) - + llama_cpp_python_install_target(ggml-cpu) + llama_cpp_python_install_target(ggml-base) + + if (GGML_METAL) + llama_cpp_python_install_target(ggml-metal) + endif() + + if (GGML_CUDA) + llama_cpp_python_install_target(ggml-cuda) + endif() + + if (GGML_VULKAN) + llama_cpp_python_install_target(ggml-vulkan) + endif() + # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 if (WIN32) install( diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 0a683e808..74d73dc85 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 0a683e8088d849626e7471f9e2ed381f7dbdf2e9 +Subproject commit 74d73dc85cc2057446bf63cc37ff649ae7cebd80 From d1cb50b49fafad171bca623f346cbc5438376fb3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 15 Nov 2024 19:26:16 -0500 Subject: [PATCH 566/624] Add missing ggml dependency --- CMakeLists.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 7d31ffd6c..f8c55d339 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -77,6 +77,7 @@ if (LLAMA_BUILD) llama_cpp_python_install_target(ggml) llama_cpp_python_install_target(ggml-cpu) llama_cpp_python_install_target(ggml-base) + llama_cpp_python_install_target(ggml-amx) if (GGML_METAL) llama_cpp_python_install_target(ggml-metal) From 2796f4e62f810780d89a3485a0a122a9e014e8ae Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 15 Nov 2024 19:31:12 -0500 Subject: [PATCH 567/624] Add all missing ggml dependencies --- CMakeLists.txt | 30 +++++++++++++++++------------- 1 file changed, 17 insertions(+), 13 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index f8c55d339..c9f3c2b36 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -6,6 +6,10 @@ option(LLAMA_BUILD "Build llama.cpp shared library and install alongside python option(LLAVA_BUILD "Build llava shared library and install alongside python package" ON) function(llama_cpp_python_install_target target) + if(NOT TARGET ${target}) + return() + endif() + install( TARGETS ${target} LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib @@ -75,21 +79,21 @@ if (LLAMA_BUILD) add_subdirectory(vendor/llama.cpp) llama_cpp_python_install_target(llama) llama_cpp_python_install_target(ggml) - llama_cpp_python_install_target(ggml-cpu) - llama_cpp_python_install_target(ggml-base) - llama_cpp_python_install_target(ggml-amx) - - if (GGML_METAL) - llama_cpp_python_install_target(ggml-metal) - endif() - if (GGML_CUDA) - llama_cpp_python_install_target(ggml-cuda) - endif() + llama_cpp_python_install_target(ggml-base) - if (GGML_VULKAN) - llama_cpp_python_install_target(ggml-vulkan) - endif() + llama_cpp_python_install_target(ggml-amx) + llama_cpp_python_install_target(ggml-blas) + llama_cpp_python_install_target(ggml-can) + llama_cpp_python_install_target(ggml-cpu) + llama_cpp_python_install_target(ggml-cuda) + llama_cpp_python_install_target(ggml-hip) + llama_cpp_python_install_target(ggml-kompute) + llama_cpp_python_install_target(ggml-metal) + llama_cpp_python_install_target(ggml-musa) + llama_cpp_python_install_target(ggml-rpc) + llama_cpp_python_install_target(ggml-sycl) + llama_cpp_python_install_target(ggml-vulkan) # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 if (WIN32) From 7ecdd944624cbd49e4af0a5ce1aa402607d58dcc Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 15 Nov 2024 19:55:09 -0500 Subject: [PATCH 568/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index c68e96b87..8a194fece 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.2] + +- feat: Update llama.cpp to ggerganov/llama.cpp@74d73dc85cc2057446bf63cc37ff649ae7cebd80 + ## [0.3.1] - feat: Update llama.cpp to ggerganov/llama.cpp@c919d5db39c8a7fcb64737f008e4b105ee0acd20 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index c11b879d6..cda38a841 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.1" +__version__ = "0.3.2" From f3fb90b114835cc50c4816787d56bac2fe1180c3 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 28 Nov 2024 18:27:55 -0500 Subject: [PATCH 569/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 22 ++++++++++++---------- vendor/llama.cpp | 2 +- 2 files changed, 13 insertions(+), 11 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 457c6dddb..bb2ba1993 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -615,6 +615,9 @@ class llama_model_kv_override(ctypes.Structure): # struct llama_model_params { +# // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) +# ggml_backend_dev_t * devices; + # int32_t n_gpu_layers; // number of layers to store in VRAM # enum llama_split_mode split_mode; // how to split the model across multiple GPUs @@ -680,6 +683,7 @@ class llama_model_params(ctypes.Structure): check_tensors: bool _fields_ = [ + ("devices", ctypes.c_void_p), # NOTE: unnused ("n_gpu_layers", ctypes.c_int32), ("split_mode", ctypes.c_int), ("main_gpu", ctypes.c_int32), @@ -1898,6 +1902,14 @@ def llama_kv_cache_update(ctx: llama_context_p, /): ... +# // Check if the context supports KV cache shifting +# LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); +@ctypes_function("llama_kv_cache_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +def llama_kv_cache_can_shift(ctx: llama_context_p, /) -> bool: + """Check if the context supports KV cache shifting""" + ... + + # // # // State / sessions # // @@ -3621,13 +3633,3 @@ def llama_perf_sampler_reset(chain: llama_sampler_p, /): ... -# LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx); -@ctypes_function( - "llama_perf_dump_yaml", - [ctypes.POINTER(ctypes.c_void_p), llama_context_p_ctypes], - None, -) -def llama_perf_dump_yaml( - stream: ctypes.POINTER(ctypes.c_void_p), ctx: llama_context_p, / -): - ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 74d73dc85..dc2234408 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 74d73dc85cc2057446bf63cc37ff649ae7cebd80 +Subproject commit dc22344088a7ee81a1e4f096459b03a72f24ccdc From 7ba257e5dac58f8796a12260776f27fd989a1182 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 02:33:55 -0500 Subject: [PATCH 570/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 32 ++++++++++++++++++++++++++++++++ vendor/llama.cpp | 2 +- 2 files changed, 33 insertions(+), 1 deletion(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index bb2ba1993..24663146c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -221,6 +221,7 @@ # LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, # LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, # LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, +# LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -249,6 +250,7 @@ LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24 LLAMA_VOCAB_PRE_TYPE_EXAONE = 25 LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 +LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 # // note: these values should be synchronized with ggml_rope @@ -392,12 +394,14 @@ # LLAMA_ROPE_SCALING_TYPE_NONE = 0, # LLAMA_ROPE_SCALING_TYPE_LINEAR = 1, # LLAMA_ROPE_SCALING_TYPE_YARN = 2, +# LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3, # LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN, # }; LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1 LLAMA_ROPE_SCALING_TYPE_NONE = 0 LLAMA_ROPE_SCALING_TYPE_LINEAR = 1 LLAMA_ROPE_SCALING_TYPE_YARN = 2 +LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3 LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN # enum llama_pooling_type { @@ -2933,6 +2937,34 @@ def llama_chat_apply_template( ... +# // Get list of built-in chat templates +# LLAMA_API int32_t llama_chat_builtin_templates(const char ** output, size_t len); +@ctypes_function( + "llama_chat_builtin_templates", + [ + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ], + ctypes.c_int32, +) +def llama_chat_builtin_templates( + output: CtypesArray[bytes], + len: Union[ctypes.c_size_t, int], + /, +) -> int: + """Get list of built-in chat templates. + + Args: + output: Output buffer to store template names. + len: Length of the output buffer. + + Returns: + Number of templates available. + Returns a negative number on error. + """ + ... + + # // # // Sampling API # // diff --git a/vendor/llama.cpp b/vendor/llama.cpp index dc2234408..7736837d6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit dc22344088a7ee81a1e4f096459b03a72f24ccdc +Subproject commit 7736837d62efed1dbebfe579472fca041eda12d6 From 9d06e36c00aea381dd995abc800ce9727da13a82 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 02:49:20 -0500 Subject: [PATCH 571/624] fix(ci): Explicitly install arm64 python version --- .github/workflows/test.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 8f619cf1f..01b2f4c40 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -126,6 +126,7 @@ jobs: uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} + architecture: "arm64" cache: 'pip' - name: Restore model cache @@ -166,6 +167,7 @@ jobs: uses: actions/setup-python@v5 with: python-version: "3.9" + architecture: "arm64" - name: Restore model cache uses: actions/cache@v4 From fb0b8fefd41dc8c467b0e353638f7345254274b5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 02:53:06 -0500 Subject: [PATCH 572/624] fix(ci): Explicitly set cmake osx architecture --- CMakeLists.txt | 2 ++ 1 file changed, 2 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index c9f3c2b36..584e54d39 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -70,6 +70,8 @@ if (LLAMA_BUILD) set(GGML_AVX2 "Off" CACHE BOOL "ggml: enable AVX2" FORCE) set(GGML_FMA "Off" CACHE BOOL "gml: enable FMA" FORCE) set(GGML_F16C "Off" CACHE BOOL "gml: enable F16C" FORCE) + + set(CMAKE_OSX_ARCHITECTURES "arm64" CACHE STRING "Build architecture for OS X" FORCE) endif() if (APPLE) From 72ed7b8855a168a454026b98c7f86a452524a7d1 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 02:55:57 -0500 Subject: [PATCH 573/624] fix(ci): Explicitly test on arm64 macos runner --- .github/workflows/test.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 01b2f4c40..445554e6b 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -113,7 +113,7 @@ jobs: build-macos: needs: download-model - runs-on: macos-latest + runs-on: macos-14-arm64 strategy: matrix: python-version: ["3.9", "3.10", "3.11", "3.12"] @@ -157,7 +157,7 @@ jobs: build-macos-metal: needs: download-model - runs-on: macos-latest + runs-on: macos-14-arm64 steps: - uses: actions/checkout@v4 with: From 8988aaf7f64a4ffbe30fb1b94d786848ea1efd3c Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:02:04 -0500 Subject: [PATCH 574/624] fix(ci): Use macos-14 runner --- .github/workflows/test.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 445554e6b..f5a08f6cb 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -113,7 +113,7 @@ jobs: build-macos: needs: download-model - runs-on: macos-14-arm64 + runs-on: macos-14 strategy: matrix: python-version: ["3.9", "3.10", "3.11", "3.12"] @@ -157,7 +157,7 @@ jobs: build-macos-metal: needs: download-model - runs-on: macos-14-arm64 + runs-on: macos-14 steps: - uses: actions/checkout@v4 with: From f11a781d86c914ab12dc620edf64c049c6fb0ee7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:04:48 -0500 Subject: [PATCH 575/624] fix(ci): Use macos-13 runner --- .github/workflows/test.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index f5a08f6cb..1631d63de 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -113,7 +113,7 @@ jobs: build-macos: needs: download-model - runs-on: macos-14 + runs-on: macos-13 strategy: matrix: python-version: ["3.9", "3.10", "3.11", "3.12"] @@ -157,7 +157,7 @@ jobs: build-macos-metal: needs: download-model - runs-on: macos-14 + runs-on: macos-13 steps: - uses: actions/checkout@v4 with: From 9a09fc78487b08423f4e2c9f02b7bc308dc0e2e7 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:19:36 -0500 Subject: [PATCH 576/624] fix(ci): Debug print python system architecture --- .github/workflows/test.yaml | 49 ++++++++++++------------------------- 1 file changed, 15 insertions(+), 34 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 1631d63de..ee8680a9b 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -51,19 +51,11 @@ jobs: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - name: Install dependencies (Linux/MacOS) - if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose shell: bash - - name: Install dependencies (Windows) - if: runner.os == 'Windows' - run: | - python -m pip install --upgrade pip - python -m pip install uv - python -m uv pip install -e .[all] --verbose - shell: cmd - name: Test with pytest run: | python -m pytest @@ -90,22 +82,13 @@ jobs: with: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - - - name: Install dependencies (Linux/MacOS) - if: runner.os != 'Windows' - run: | - python -m pip install --upgrade pip - python -m pip install uv - python -m uv pip install -e .[all] --verbose - shell: bash - name: Install dependencies (Windows) - if: runner.os == 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose - shell: cmd + shell: cmd - name: Test with pytest run: | @@ -129,6 +112,12 @@ jobs: architecture: "arm64" cache: 'pip' + - name: System Info + run: | + uname -a + sysctl -n machdep.cpu.brand_string + python3 -c "import platform; print(platform.machine(), platform.architecture())" + - name: Restore model cache uses: actions/cache@v4 with: @@ -136,21 +125,13 @@ jobs: key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - name: Install dependencies (Linux/MacOS) - if: runner.os != 'Windows' run: | python -m pip install --upgrade pip python -m pip install uv python -m uv pip install -e .[all] --verbose + CMAKE_ARGS="-DLLAMA_METAL=off" python -m uv pip install .[all] --verbose shell: bash - - name: Install dependencies (Windows) - if: runner.os == 'Windows' - run: | - python -m pip install --upgrade pip - python -m pip install uv - python -m uv pip install -e .[all] --verbose - shell: cmd - - name: Test with pytest run: | python -m pytest @@ -169,25 +150,25 @@ jobs: python-version: "3.9" architecture: "arm64" + - name: System Info + run: | + uname -a + sysctl -n machdep.cpu.brand_string + python3 -c "import platform; print(platform.machine(), platform.architecture())" + - name: Restore model cache uses: actions/cache@v4 with: path: ~/.cache/huggingface/hub key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} - - name: Install dependencies (Linux/MacOS) - if: runner.os != 'Windows' + - name: Install dependencies run: | python -m pip install --upgrade pip python -m pip install uv CMAKE_ARGS="-DLLAMA_METAL=on" python -m uv pip install .[all] --verbose shell: bash - - name: Install dependencies (Windows) - if: runner.os == 'Windows' - run: | - python -m pip install --upgrade pip - CMAKE_ARGS="-DGGML_METAL=on" python -m pip install .[all] --verbose - name: Test with pytest run: | python -m pytest From a412ba5539be6a5499acfcb0a2c234665fcbda11 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:38:12 -0500 Subject: [PATCH 577/624] fix(ci): Update config --- .github/workflows/test.yaml | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index ee8680a9b..f8b7893f8 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -109,7 +109,6 @@ jobs: uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - architecture: "arm64" cache: 'pip' - name: System Info @@ -138,7 +137,7 @@ jobs: build-macos-metal: needs: download-model - runs-on: macos-13 + runs-on: macos-14 steps: - uses: actions/checkout@v4 with: From df05096fc56347efe14ee550f065a1b15933c322 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:42:33 -0500 Subject: [PATCH 578/624] fix(ci): Install with regular pip --- .github/workflows/test.yaml | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index f8b7893f8..9332e0830 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -164,8 +164,7 @@ jobs: - name: Install dependencies run: | python -m pip install --upgrade pip - python -m pip install uv - CMAKE_ARGS="-DLLAMA_METAL=on" python -m uv pip install .[all] --verbose + CMAKE_ARGS="-DLLAMA_METAL=on" python -m pip install .[all] --verbose shell: bash - name: Test with pytest From 1cd3f2cc6aceaa3167c943eb33b4076ae7a9ba53 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:52:54 -0500 Subject: [PATCH 579/624] fix(ci): gg --- .github/workflows/test.yaml | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 9332e0830..a4d56110a 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -110,6 +110,7 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' + architecture: "arm64" - name: System Info run: | @@ -137,7 +138,7 @@ jobs: build-macos-metal: needs: download-model - runs-on: macos-14 + runs-on: macos-13 steps: - uses: actions/checkout@v4 with: @@ -163,10 +164,10 @@ jobs: - name: Install dependencies run: | - python -m pip install --upgrade pip - CMAKE_ARGS="-DLLAMA_METAL=on" python -m pip install .[all] --verbose + python3 -m pip install --upgrade pip + CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install .[all] --verbose shell: bash - name: Test with pytest run: | - python -m pytest + python3 -m pytest From b34f20046b9fc058ed8581291a724c6723b06fe4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 03:59:51 -0500 Subject: [PATCH 580/624] fix(ci): Use python3 --- .github/workflows/test.yaml | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index a4d56110a..7f4774959 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -126,15 +126,15 @@ jobs: - name: Install dependencies (Linux/MacOS) run: | - python -m pip install --upgrade pip - python -m pip install uv - python -m uv pip install -e .[all] --verbose - CMAKE_ARGS="-DLLAMA_METAL=off" python -m uv pip install .[all] --verbose + python3 -m pip install --upgrade pip + python3 -m pip install uv + python3 -m uv pip install -e .[all] --verbose + CMAKE_ARGS="-DLLAMA_METAL=off" python3 -m uv pip install .[all] --verbose shell: bash - name: Test with pytest run: | - python -m pytest + python3 -m pytest build-macos-metal: needs: download-model From d8cc231fd00b67528e7e4563c2b4c514acbcd8f9 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 04:06:07 -0500 Subject: [PATCH 581/624] fix(ci): Use default architecture chosen by action --- .github/workflows/test.yaml | 2 -- 1 file changed, 2 deletions(-) diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 7f4774959..95f6e5a27 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -110,7 +110,6 @@ jobs: with: python-version: ${{ matrix.python-version }} cache: 'pip' - architecture: "arm64" - name: System Info run: | @@ -148,7 +147,6 @@ jobs: uses: actions/setup-python@v5 with: python-version: "3.9" - architecture: "arm64" - name: System Info run: | From d5d50990d016333982ff350c9c151efe5732d76b Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 04:14:59 -0500 Subject: [PATCH 582/624] fix(ci): Update CMakeLists.txt for macos --- CMakeLists.txt | 43 +++++++++++++++++++++++++++++-------------- 1 file changed, 29 insertions(+), 14 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 584e54d39..64a0304a1 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -62,20 +62,35 @@ if (LLAMA_BUILD) # Enable building of the common library set(LLAMA_BUILD_COMMON ON CACHE BOOL "Build llama.cpp common library" FORCE) - # Building llama - if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") - # Need to disable these llama.cpp flags on Apple x86_64, - # otherwise users may encounter invalid instruction errors - set(GGML_AVX "Off" CACHE BOOL "ggml: enable AVX" FORCE) - set(GGML_AVX2 "Off" CACHE BOOL "ggml: enable AVX2" FORCE) - set(GGML_FMA "Off" CACHE BOOL "gml: enable FMA" FORCE) - set(GGML_F16C "Off" CACHE BOOL "gml: enable F16C" FORCE) - - set(CMAKE_OSX_ARCHITECTURES "arm64" CACHE STRING "Build architecture for OS X" FORCE) - endif() - + # Architecture detection and settings for Apple platforms if (APPLE) - set(GGML_METAL_EMBED_LIBRARY "On" CACHE BOOL "llama: embed metal library" FORCE) + # Get the target architecture + execute_process( + COMMAND uname -m + OUTPUT_VARIABLE HOST_ARCH + OUTPUT_STRIP_TRAILING_WHITESPACE + ) + + # If CMAKE_OSX_ARCHITECTURES is not set, use the host architecture + if(NOT CMAKE_OSX_ARCHITECTURES) + set(CMAKE_OSX_ARCHITECTURES ${HOST_ARCH} CACHE STRING "Build architecture for macOS" FORCE) + endif() + + message(STATUS "Host architecture: ${HOST_ARCH}") + message(STATUS "Target architecture: ${CMAKE_OSX_ARCHITECTURES}") + + # Configure based on target architecture + if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64") + # Intel Mac settings + set(GGML_AVX "OFF" CACHE BOOL "ggml: enable AVX" FORCE) + set(GGML_AVX2 "OFF" CACHE BOOL "ggml: enable AVX2" FORCE) + set(GGML_FMA "OFF" CACHE BOOL "ggml: enable FMA" FORCE) + set(GGML_F16C "OFF" CACHE BOOL "ggml: enable F16C" FORCE) + endif() + + # Metal settings (enable for both architectures) + set(GGML_METAL "ON" CACHE BOOL "ggml: enable Metal" FORCE) + set(GGML_METAL_EMBED_LIBRARY "ON" CACHE BOOL "ggml: embed metal library" FORCE) endif() add_subdirectory(vendor/llama.cpp) @@ -130,7 +145,7 @@ if (LLAMA_BUILD) # Building llava add_subdirectory(vendor/llama.cpp/examples/llava) set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") - # Set CUDA_ARCHITECTURES to OFF on Windows + if (WIN32) set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) endif() From 4f17ae5af073f3741be7235f86e2b183db62a260 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?P=CE=9BBL=C3=98=20=E1=84=83=CE=9E?= Date: Fri, 6 Dec 2024 04:31:01 -0500 Subject: [PATCH 583/624] fix(ci): Remove cuda version 12.5.0 incompatibility with VS (#1838) Co-authored-by: Andrei --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 647b171e8..bde626828 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -22,7 +22,7 @@ jobs: $matrix = @{ 'os' = @('ubuntu-latest', 'windows-2019') 'pyver' = @("3.9", "3.10", "3.11", "3.12") - 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.0") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") 'releasetag' = @("basic") } From 991d9cde1b81dfdda4d3117b508c16c6fe7dbdca Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 04:32:19 -0500 Subject: [PATCH 584/624] fix(ci): Remove CUDA 12.5 from index --- .github/workflows/generate-index-from-release.yaml | 1 - 1 file changed, 1 deletion(-) diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 8d80e7e47..51961088d 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -44,7 +44,6 @@ jobs: ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' ./scripts/releases-to-pep-503.sh index/whl/cu123 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu123$' ./scripts/releases-to-pep-503.sh index/whl/cu124 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' - ./scripts/releases-to-pep-503.sh index/whl/cu125 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu125$' ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' - name: Upload artifact uses: actions/upload-pages-artifact@v3 From 27953036b1bb1da9e4b71b875f66ba8f1aefbf6f Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri, 6 Dec 2024 04:32:37 -0500 Subject: [PATCH 585/624] chore(deps): bump pypa/cibuildwheel from 2.21.1 to 2.22.0 (#1844) Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.21.1 to 2.22.0. - [Release notes](https://github.com/pypa/cibuildwheel/releases) - [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md) - [Commits](https://github.com/pypa/cibuildwheel/compare/v2.21.1...v2.22.0) --- updated-dependencies: - dependency-name: pypa/cibuildwheel dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-and-release.yaml | 4 ++-- .github/workflows/build-wheels-metal.yaml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index bfa97decd..1d549ed68 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -42,7 +42,7 @@ jobs: shell: cmd - name: Build wheels - uses: pypa/cibuildwheel@v2.21.1 + uses: pypa/cibuildwheel@v2.22.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" @@ -69,7 +69,7 @@ jobs: platforms: linux/arm64 - name: Build wheels - uses: pypa/cibuildwheel@v2.21.1 + uses: pypa/cibuildwheel@v2.22.0 env: CIBW_SKIP: "*musllinux* pp*" CIBW_REPAIR_WHEEL_COMMAND: "" diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index 8e6fb5cb7..01da3cef5 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -43,7 +43,7 @@ jobs: shell: cmd - name: Build wheels - uses: pypa/cibuildwheel@v2.21.1 + uses: pypa/cibuildwheel@v2.22.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" From 2523472c3eccb9ab9277117cc4ff705212b6888a Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 04:36:22 -0500 Subject: [PATCH 586/624] fix: Fix pickling of Llama class by setting seed from _seed member. Closes #1769 --- llama_cpp/llama.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index d15a88b00..ce1a26649 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -2073,7 +2073,7 @@ def __getstate__(self): use_mlock=self.model_params.use_mlock, kv_overrides=self.kv_overrides, # Context Params - seed=self.context_params.seed, + seed=self._seed, n_ctx=self.context_params.n_ctx, n_batch=self.n_batch, n_ubatch=self.context_params.n_ubatch, From ddac04c78e504196207292aeca4c3adccc655a88 Mon Sep 17 00:00:00 2001 From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri, 6 Dec 2024 04:39:27 -0500 Subject: [PATCH 587/624] chore(deps): bump conda-incubator/setup-miniconda from 3.0.4 to 3.1.0 (#1821) Bumps [conda-incubator/setup-miniconda](https://github.com/conda-incubator/setup-miniconda) from 3.0.4 to 3.1.0. - [Release notes](https://github.com/conda-incubator/setup-miniconda/releases) - [Changelog](https://github.com/conda-incubator/setup-miniconda/blob/main/CHANGELOG.md) - [Commits](https://github.com/conda-incubator/setup-miniconda/compare/v3.0.4...v3.1.0) --- updated-dependencies: - dependency-name: conda-incubator/setup-miniconda dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index bde626828..3d9c5d9d4 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -59,7 +59,7 @@ jobs: cache: 'pip' - name: Setup Mamba - uses: conda-incubator/setup-miniconda@v3.0.4 + uses: conda-incubator/setup-miniconda@v3.1.0 with: activate-environment: "build" python-version: ${{ matrix.pyver }} From fa04cdc497e2c72888cccf6268836fe3291c3d42 Mon Sep 17 00:00:00 2001 From: ddh0 Date: Fri, 6 Dec 2024 03:40:47 -0600 Subject: [PATCH 588/624] fix logit-bias type hint (#1802) `Optional[Dict[str, float]]` --> `Optional[Dict[int, float]]` --- llama_cpp/llama.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index ce1a26649..2fd7ff193 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1141,7 +1141,7 @@ def _create_completion( stopping_criteria: Optional[StoppingCriteriaList] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, ) -> Union[ Iterator[CreateCompletionResponse], Iterator[CreateCompletionStreamResponse] ]: @@ -1761,7 +1761,7 @@ def create_completion( stopping_criteria: Optional[StoppingCriteriaList] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, ) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]: """Generate text from a prompt. @@ -1858,7 +1858,7 @@ def __call__( stopping_criteria: Optional[StoppingCriteriaList] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, ) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]: """Generate text from a prompt. @@ -1951,7 +1951,7 @@ def create_chat_completion( model: Optional[str] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, logprobs: Optional[bool] = None, top_logprobs: Optional[int] = None, ) -> Union[ From 38fbd2968d3e937c9f4157a54372b02fa527233b Mon Sep 17 00:00:00 2001 From: Rich Dougherty Date: Fri, 6 Dec 2024 22:41:52 +1300 Subject: [PATCH 589/624] docs: Remove ref to llama_eval in llama_cpp.py docs (#1819) The llama_eval function was removed in 0.2.14 (0d37ce5) but the docs still reference it. Updated to match the llama.h docs from upstream. --- llama_cpp/llama_cpp.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 24663146c..cd59bb430 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -772,7 +772,7 @@ class llama_context_params(ctypes.Structure): cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache type_v (int): data type for V cache - logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) + logits_all (bool): the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) embeddings (bool): if true, extract embeddings (together with logits) offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU flash_attn (bool): whether to use flash attention @@ -2469,10 +2469,10 @@ def llama_synchronize(ctx: llama_context_p, /): "llama_get_logits", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) ) def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: - """Token logits obtained from the last call to llama_eval() - The logits for the last token are stored in the last row - Logits for which llama_batch.logits[i] == 0 are undefined - Rows: n_tokens provided with llama_batch + """Token logits obtained from the last call to llama_decode() + The logits for which llama_batch.logits[i] != 0 are stored contiguously + in the order they have appeared in the batch. + Rows: number of tokens for which llama_batch.logits[i] != 0 Cols: n_vocab Returns: From 4192210232ee20ceb028ebd0ee68a6b964f5ffc1 Mon Sep 17 00:00:00 2001 From: Philippe Martin Date: Fri, 6 Dec 2024 10:47:38 +0100 Subject: [PATCH 590/624] fix: make content not required in ChatCompletionRequestAssistantMessage (#1807) --- llama_cpp/llama_types.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index bbb58afc3..57836ee76 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -214,7 +214,7 @@ class ChatCompletionRequestAssistantMessageFunctionCall(TypedDict): class ChatCompletionRequestAssistantMessage(TypedDict): role: Literal["assistant"] - content: Optional[str] + content: NotRequired[str] tool_calls: NotRequired[ChatCompletionMessageToolCalls] function_call: NotRequired[ ChatCompletionRequestAssistantMessageFunctionCall From 77a12a39663bbe5a5d22a7cc3d5d4f1c591c2851 Mon Sep 17 00:00:00 2001 From: Olivier DEBAUCHE Date: Fri, 6 Dec 2024 11:01:17 +0100 Subject: [PATCH 591/624] fix: Re-add suport for CUDA 12.5, add CUDA 12.6 (#1775) Add support for Cuda 12.6.1 Update version of Cuda 12.5.0 to 12.5.1 Co-authored-by: Andrei --- .github/workflows/build-wheels-cuda.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 3d9c5d9d4..8624ec6d9 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -22,7 +22,7 @@ jobs: $matrix = @{ 'os' = @('ubuntu-latest', 'windows-2019') 'pyver' = @("3.9", "3.10", "3.11", "3.12") - 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "12.6.1") 'releasetag' = @("basic") } From 073b7e421e419ec389d8c064bbdae26c482ec17c Mon Sep 17 00:00:00 2001 From: "Ignaz \"Ian\" Kraft" Date: Fri, 6 Dec 2024 11:02:34 +0100 Subject: [PATCH 592/624] fix: added missing exit_stack.close() to /v1/chat/completions (#1796) * fix: added missing exit_stack.close() to /v1/chat/completions * fix: added missing exit_stack.close() to /v1/completions --- llama_cpp/server/app.py | 23 ++++++++++++++++------- llama_cpp/server/errors.py | 6 ++++-- 2 files changed, 20 insertions(+), 9 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index cd3255176..9b8dce418 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -314,10 +314,14 @@ async def create_completion( else: kwargs["logits_processor"].extend(_min_tokens_logits_processor) - iterator_or_completion: Union[ - llama_cpp.CreateCompletionResponse, - Iterator[llama_cpp.CreateCompletionStreamResponse], - ] = await run_in_threadpool(llama, **kwargs) + try: + iterator_or_completion: Union[ + llama_cpp.CreateCompletionResponse, + Iterator[llama_cpp.CreateCompletionStreamResponse], + ] = await run_in_threadpool(llama, **kwargs) + except Exception as err: + exit_stack.close() + raise err if isinstance(iterator_or_completion, Iterator): # EAFP: It's easier to ask for forgiveness than permission @@ -344,6 +348,7 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: ping_message_factory=_ping_message_factory, ) else: + exit_stack.close() return iterator_or_completion @@ -508,9 +513,13 @@ async def create_chat_completion( else: kwargs["logits_processor"].extend(_min_tokens_logits_processor) - iterator_or_completion: Union[ - llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] - ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) + try: + iterator_or_completion: Union[ + llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] + ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) + except Exception as err: + exit_stack.close() + raise err if isinstance(iterator_or_completion, Iterator): # EAFP: It's easier to ask for forgiveness than permission diff --git a/llama_cpp/server/errors.py b/llama_cpp/server/errors.py index fbf9fd80d..d0eda5664 100644 --- a/llama_cpp/server/errors.py +++ b/llama_cpp/server/errors.py @@ -134,8 +134,6 @@ def error_message_wrapper( ] = None, ) -> Tuple[int, ErrorResponse]: """Wraps error message in OpenAI style error response""" - print(f"Exception: {str(error)}", file=sys.stderr) - traceback.print_exc(file=sys.stderr) if body is not None and isinstance( body, ( @@ -149,6 +147,10 @@ def error_message_wrapper( if match is not None: return callback(body, match) + # Only print the trace on unexpected exceptions + print(f"Exception: {str(error)}", file=sys.stderr) + traceback.print_exc(file=sys.stderr) + # Wrap other errors as internal server error return 500, ErrorResponse( message=str(error), From 9bd0c95424aad6a80a7305761b465006af3dde0d Mon Sep 17 00:00:00 2001 From: Graeme Power Date: Fri, 6 Dec 2024 10:03:41 +0000 Subject: [PATCH 593/624] fix: Avoid thread starvation on many concurrent requests by making use of asyncio to lock llama_proxy context (#1798) * fix: make use of asyncio to lock llama_proxy context * fix: use aclose instead of close for AsyncExitStack * fix: don't call exit stack close in stream iterator as it will be called by finally from on_complete anyway * fix: use anyio.Lock instead of asyncio.Lock --------- Co-authored-by: Andrei --- llama_cpp/server/app.py | 34 ++++++++++++++-------------------- 1 file changed, 14 insertions(+), 20 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index 9b8dce418..f7c028475 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -5,7 +5,7 @@ import typing import contextlib -from threading import Lock +from anyio import Lock from functools import partial from typing import Iterator, List, Optional, Union, Dict @@ -70,14 +70,14 @@ def set_llama_proxy(model_settings: List[ModelSettings]): _llama_proxy = LlamaProxy(models=model_settings) -def get_llama_proxy(): +async def get_llama_proxy(): # NOTE: This double lock allows the currently streaming llama model to # check if any other requests are pending in the same thread and cancel # the stream if so. - llama_outer_lock.acquire() + await llama_outer_lock.acquire() release_outer_lock = True try: - llama_inner_lock.acquire() + await llama_inner_lock.acquire() try: llama_outer_lock.release() release_outer_lock = False @@ -159,7 +159,7 @@ async def get_event_publisher( request: Request, inner_send_chan: MemoryObjectSendStream[typing.Any], iterator: Iterator[typing.Any], - on_complete: typing.Optional[typing.Callable[[], None]] = None, + on_complete: typing.Optional[typing.Callable[[], typing.Awaitable[None]]] = None, ): server_settings = next(get_server_settings()) interrupt_requests = ( @@ -182,7 +182,7 @@ async def get_event_publisher( raise e finally: if on_complete: - on_complete() + await on_complete() def _logit_bias_tokens_to_input_ids( @@ -267,10 +267,8 @@ async def create_completion( request: Request, body: CreateCompletionRequest, ) -> llama_cpp.Completion: - exit_stack = contextlib.ExitStack() - llama_proxy = await run_in_threadpool( - lambda: exit_stack.enter_context(contextlib.contextmanager(get_llama_proxy)()) - ) + exit_stack = contextlib.AsyncExitStack() + llama_proxy = await exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) if llama_proxy is None: raise HTTPException( status_code=status.HTTP_503_SERVICE_UNAVAILABLE, @@ -332,7 +330,6 @@ async def create_completion( def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: yield first_response yield from iterator_or_completion - exit_stack.close() send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( @@ -342,13 +339,13 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: request=request, inner_send_chan=send_chan, iterator=iterator(), - on_complete=exit_stack.close, + on_complete=exit_stack.aclose, ), sep="\n", ping_message_factory=_ping_message_factory, ) else: - exit_stack.close() + await exit_stack.aclose() return iterator_or_completion @@ -477,10 +474,8 @@ async def create_chat_completion( # where the dependency is cleaned up before a StreamingResponse # is complete. # https://github.com/tiangolo/fastapi/issues/11143 - exit_stack = contextlib.ExitStack() - llama_proxy = await run_in_threadpool( - lambda: exit_stack.enter_context(contextlib.contextmanager(get_llama_proxy)()) - ) + exit_stack = contextlib.AsyncExitStack() + llama_proxy = exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) if llama_proxy is None: raise HTTPException( status_code=status.HTTP_503_SERVICE_UNAVAILABLE, @@ -530,7 +525,6 @@ async def create_chat_completion( def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: yield first_response yield from iterator_or_completion - exit_stack.close() send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( @@ -540,13 +534,13 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: request=request, inner_send_chan=send_chan, iterator=iterator(), - on_complete=exit_stack.close, + on_complete=exit_stack.aclose, ), sep="\n", ping_message_factory=_ping_message_factory, ) else: - exit_stack.close() + await exit_stack.aclose() return iterator_or_completion From 1ea615424b7647a56946ce4fc3ea72b6ebaa867c Mon Sep 17 00:00:00 2001 From: Florents Tselai Date: Fri, 6 Dec 2024 12:11:50 +0200 Subject: [PATCH 594/624] fix(docs): Update development instructions (#1833) --- README.md | 18 ++++++++++++++++-- 1 file changed, 16 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index dbaec5077..e00456580 100644 --- a/README.md +++ b/README.md @@ -752,15 +752,29 @@ pip install --upgrade pip pip install -e . # if you want to use the fastapi / openapi server -pip install -e .[server] +pip install -e '.[server]' # to install all optional dependencies -pip install -e .[all] +pip install -e '.[all]' # to clear the local build cache make clean ``` +Now try running the tests + +```bash +pytest +``` + +There's a `Makefile` available with useful targets. +A typical workflow would look like this: + +```bash +make build +make test +``` + You can also test out specific commits of `llama.cpp` by checking out the desired commit in the `vendor/llama.cpp` submodule and then running `make clean` and `pip install -e .` again. Any changes in the `llama.h` API will require changes to the `llama_cpp/llama_cpp.py` file to match the new API (additional changes may be required elsewhere). From d610477a0659fd7f56c5712adccebde4135a250c Mon Sep 17 00:00:00 2001 From: Luke Stanley <306671+lukestanley@users.noreply.github.com> Date: Fri, 6 Dec 2024 10:22:26 +0000 Subject: [PATCH 595/624] fix(examples): Refactor Batching notebook to use new sampler chain API (#1793) - Replaced deprecated llama_sample_token with llama_sampler_sample - Updated llama_token_to_piece signature to include lstrip and special arguments - Changed llama_load_model_from_file to use positional argument for params (currently does not accept a keyword argument for it) --- examples/notebooks/Batching.ipynb | 557 ++++++++---------------------- 1 file changed, 146 insertions(+), 411 deletions(-) diff --git a/examples/notebooks/Batching.ipynb b/examples/notebooks/Batching.ipynb index 73b28c744..be7fe9b52 100644 --- a/examples/notebooks/Batching.ipynb +++ b/examples/notebooks/Batching.ipynb @@ -6,6 +6,7 @@ "metadata": {}, "outputs": [], "source": [ + "import ctypes\n", "import llama_cpp" ] }, @@ -13,18 +14,7 @@ "cell_type": "code", "execution_count": 2, "metadata": {}, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no\n", - "ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes\n", - "ggml_init_cublas: found 1 CUDA devices:\n", - " Device 0: NVIDIA GeForce RTX 2060, compute capability 7.5\n" - ] - } - ], + "outputs": [], "source": [ "llama_cpp.llama_backend_init(numa=False)" ] @@ -38,354 +28,94 @@ "name": "stderr", "output_type": "stream", "text": [ - "llama_model_loader: loaded meta data with 16 key-value pairs and 291 tensors from ../../models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf (version GGUF V2)\n", - "llama_model_loader: - tensor 0: token_embd.weight q4_K [ 4096, 32000, 1, 1 ]\n", - "llama_model_loader: - tensor 1: output_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 2: output.weight q6_K [ 4096, 32000, 1, 1 ]\n", - "llama_model_loader: - tensor 3: blk.0.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 4: blk.0.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 5: blk.0.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 6: blk.0.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 7: blk.0.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 8: blk.0.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 9: blk.0.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 10: blk.0.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 11: blk.0.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 12: blk.1.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 13: blk.1.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 14: blk.1.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 15: blk.1.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 16: blk.1.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 17: blk.1.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 18: blk.1.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 19: blk.1.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 20: blk.1.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 21: blk.2.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 22: blk.2.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 23: blk.2.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 24: blk.2.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 25: blk.2.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 26: blk.2.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 27: blk.2.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 28: blk.2.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 29: blk.2.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 30: blk.3.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 31: blk.3.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 32: blk.3.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 33: blk.3.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 34: blk.3.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 35: blk.3.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 36: blk.3.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 37: blk.3.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 38: blk.3.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 39: blk.4.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 40: blk.4.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 41: blk.4.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 42: blk.4.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 43: blk.4.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 44: blk.4.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 45: blk.4.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 46: blk.4.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 47: blk.4.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 48: blk.5.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 49: blk.5.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 50: blk.5.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 51: blk.5.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 52: blk.5.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 53: blk.5.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 54: blk.5.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 55: blk.5.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 56: blk.5.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 57: blk.6.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 58: blk.6.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 59: blk.6.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 60: blk.6.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 61: blk.6.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 62: blk.6.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 63: blk.6.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 64: blk.6.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 65: blk.6.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 66: blk.7.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 67: blk.7.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 68: blk.7.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 69: blk.7.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 70: blk.7.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 71: blk.7.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 72: blk.7.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 73: blk.7.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 74: blk.7.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 75: blk.8.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 76: blk.8.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 77: blk.8.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 78: blk.8.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 79: blk.8.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 80: blk.8.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 81: blk.8.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 82: blk.8.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 83: blk.8.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 84: blk.9.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 85: blk.9.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 86: blk.9.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 87: blk.9.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 88: blk.9.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 89: blk.9.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 90: blk.9.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 91: blk.9.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 92: blk.9.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 93: blk.10.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 94: blk.10.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 95: blk.10.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 96: blk.10.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 97: blk.10.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 98: blk.10.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 99: blk.10.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 100: blk.10.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 101: blk.10.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 102: blk.11.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 103: blk.11.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 104: blk.11.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 105: blk.11.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 106: blk.11.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 107: blk.11.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 108: blk.11.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 109: blk.11.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 110: blk.11.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 111: blk.12.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 112: blk.12.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 113: blk.12.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 114: blk.12.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 115: blk.12.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 116: blk.12.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 117: blk.12.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 118: blk.12.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 119: blk.12.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 120: blk.13.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 121: blk.13.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 122: blk.13.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 123: blk.13.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 124: blk.13.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 125: blk.13.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 126: blk.13.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 127: blk.13.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 128: blk.13.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 129: blk.14.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 130: blk.14.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 131: blk.14.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 132: blk.14.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 133: blk.14.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 134: blk.14.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 135: blk.14.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 136: blk.14.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 137: blk.14.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 138: blk.15.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 139: blk.15.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 140: blk.15.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 141: blk.15.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 142: blk.15.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 143: blk.15.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 144: blk.15.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 145: blk.15.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 146: blk.15.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 147: blk.16.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 148: blk.16.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 149: blk.16.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 150: blk.16.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 151: blk.16.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 152: blk.16.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 153: blk.16.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 154: blk.16.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 155: blk.16.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 156: blk.17.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 157: blk.17.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 158: blk.17.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 159: blk.17.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 160: blk.17.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 161: blk.17.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 162: blk.17.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 163: blk.17.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 164: blk.17.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 165: blk.18.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 166: blk.18.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 167: blk.18.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 168: blk.18.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 169: blk.18.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 170: blk.18.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 171: blk.18.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 172: blk.18.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 173: blk.18.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 174: blk.19.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 175: blk.19.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 176: blk.19.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 177: blk.19.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 178: blk.19.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 179: blk.19.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 180: blk.19.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 181: blk.19.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 182: blk.19.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 183: blk.20.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 184: blk.20.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 185: blk.20.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 186: blk.20.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 187: blk.20.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 188: blk.20.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 189: blk.20.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 190: blk.20.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 191: blk.20.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 192: blk.21.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 193: blk.21.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 194: blk.21.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 195: blk.21.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 196: blk.21.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 197: blk.21.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 198: blk.21.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 199: blk.21.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 200: blk.21.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 201: blk.22.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 202: blk.22.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 203: blk.22.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 204: blk.22.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 205: blk.22.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 206: blk.22.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 207: blk.22.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 208: blk.22.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 209: blk.22.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 210: blk.23.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 211: blk.23.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 212: blk.23.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 213: blk.23.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 214: blk.23.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 215: blk.23.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 216: blk.23.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 217: blk.23.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 218: blk.23.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 219: blk.24.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 220: blk.24.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 221: blk.24.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 222: blk.24.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 223: blk.24.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 224: blk.24.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 225: blk.24.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 226: blk.24.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 227: blk.24.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 228: blk.25.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 229: blk.25.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 230: blk.25.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 231: blk.25.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 232: blk.25.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 233: blk.25.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 234: blk.25.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 235: blk.25.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 236: blk.25.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 237: blk.26.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 238: blk.26.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 239: blk.26.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 240: blk.26.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 241: blk.26.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 242: blk.26.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 243: blk.26.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 244: blk.26.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 245: blk.26.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 246: blk.27.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 247: blk.27.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 248: blk.27.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 249: blk.27.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 250: blk.27.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 251: blk.27.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 252: blk.27.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 253: blk.27.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 254: blk.27.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 255: blk.28.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 256: blk.28.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 257: blk.28.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 258: blk.28.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 259: blk.28.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 260: blk.28.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 261: blk.28.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 262: blk.28.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 263: blk.28.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 264: blk.29.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 265: blk.29.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 266: blk.29.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 267: blk.29.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 268: blk.29.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 269: blk.29.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 270: blk.29.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 271: blk.29.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 272: blk.29.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 273: blk.30.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 274: blk.30.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 275: blk.30.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 276: blk.30.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 277: blk.30.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 278: blk.30.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 279: blk.30.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 280: blk.30.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 281: blk.30.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 282: blk.31.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 283: blk.31.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 284: blk.31.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 285: blk.31.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 286: blk.31.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 287: blk.31.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 288: blk.31.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 289: blk.31.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 290: blk.31.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - kv 0: general.architecture str \n", - "llama_model_loader: - kv 1: general.name str \n", - "llama_model_loader: - kv 2: llama.context_length u32 \n", - "llama_model_loader: - kv 3: llama.embedding_length u32 \n", - "llama_model_loader: - kv 4: llama.block_count u32 \n", - "llama_model_loader: - kv 5: llama.feed_forward_length u32 \n", - "llama_model_loader: - kv 6: llama.rope.dimension_count u32 \n", - "llama_model_loader: - kv 7: llama.attention.head_count u32 \n", - "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 \n", - "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 \n", - "llama_model_loader: - kv 10: general.file_type u32 \n", - "llama_model_loader: - kv 11: tokenizer.ggml.model str \n", - "llama_model_loader: - kv 12: tokenizer.ggml.tokens arr \n", - "llama_model_loader: - kv 13: tokenizer.ggml.scores arr \n", - "llama_model_loader: - kv 14: tokenizer.ggml.token_type arr \n", - "llama_model_loader: - kv 15: general.quantization_version u32 \n", + "llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from /workspaces/llama-cpp-python/mistral-7b-v0.1.Q2_K.gguf (version GGUF V2)\n", + "llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", + "llama_model_loader: - kv 0: general.architecture str = llama\n", + "llama_model_loader: - kv 1: general.name str = mistralai_mistral-7b-v0.1\n", + "llama_model_loader: - kv 2: llama.context_length u32 = 32768\n", + "llama_model_loader: - kv 3: llama.embedding_length u32 = 4096\n", + "llama_model_loader: - kv 4: llama.block_count u32 = 32\n", + "llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336\n", + "llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128\n", + "llama_model_loader: - kv 7: llama.attention.head_count u32 = 32\n", + "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 8\n", + "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010\n", + "llama_model_loader: - kv 10: llama.rope.freq_base f32 = 10000.000000\n", + "llama_model_loader: - kv 11: general.file_type u32 = 10\n", + "llama_model_loader: - kv 12: tokenizer.ggml.model str = llama\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32000] = [\"\", \"\", \"\", \"<0x00>\", \"<...\n", + "llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32000] = [0.000000, 0.000000, 0.000000, 0.0000...\n", + "llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32000] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...\n", + "llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1\n", + "llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 2\n", + "llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 0\n", + "llama_model_loader: - kv 19: general.quantization_version u32 = 2\n", "llama_model_loader: - type f32: 65 tensors\n", - "llama_model_loader: - type q4_K: 193 tensors\n", - "llama_model_loader: - type q6_K: 33 tensors\n", - "llm_load_vocab: special tokens definition check successful ( 259/32000 ).\n", + "llama_model_loader: - type q2_K: 65 tensors\n", + "llama_model_loader: - type q3_K: 160 tensors\n", + "llama_model_loader: - type q6_K: 1 tensors\n", + "llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect\n", + "llm_load_vocab: special tokens cache size = 3\n", + "llm_load_vocab: token to piece cache size = 0.1637 MB\n", "llm_load_print_meta: format = GGUF V2\n", "llm_load_print_meta: arch = llama\n", "llm_load_print_meta: vocab type = SPM\n", "llm_load_print_meta: n_vocab = 32000\n", "llm_load_print_meta: n_merges = 0\n", - "llm_load_print_meta: n_ctx_train = 4096\n", + "llm_load_print_meta: vocab_only = 0\n", + "llm_load_print_meta: n_ctx_train = 32768\n", "llm_load_print_meta: n_embd = 4096\n", + "llm_load_print_meta: n_layer = 32\n", "llm_load_print_meta: n_head = 32\n", "llm_load_print_meta: n_head_kv = 8\n", - "llm_load_print_meta: n_layer = 32\n", "llm_load_print_meta: n_rot = 128\n", + "llm_load_print_meta: n_swa = 0\n", + "llm_load_print_meta: n_embd_head_k = 128\n", + "llm_load_print_meta: n_embd_head_v = 128\n", "llm_load_print_meta: n_gqa = 4\n", + "llm_load_print_meta: n_embd_k_gqa = 1024\n", + "llm_load_print_meta: n_embd_v_gqa = 1024\n", "llm_load_print_meta: f_norm_eps = 0.0e+00\n", "llm_load_print_meta: f_norm_rms_eps = 1.0e-05\n", "llm_load_print_meta: f_clamp_kqv = 0.0e+00\n", "llm_load_print_meta: f_max_alibi_bias = 0.0e+00\n", + "llm_load_print_meta: f_logit_scale = 0.0e+00\n", "llm_load_print_meta: n_ff = 14336\n", + "llm_load_print_meta: n_expert = 0\n", + "llm_load_print_meta: n_expert_used = 0\n", + "llm_load_print_meta: causal attn = 1\n", + "llm_load_print_meta: pooling type = 0\n", + "llm_load_print_meta: rope type = 0\n", + "llm_load_print_meta: rope scaling = linear\n", "llm_load_print_meta: freq_base_train = 10000.0\n", "llm_load_print_meta: freq_scale_train = 1\n", + "llm_load_print_meta: n_ctx_orig_yarn = 32768\n", + "llm_load_print_meta: rope_finetuned = unknown\n", + "llm_load_print_meta: ssm_d_conv = 0\n", + "llm_load_print_meta: ssm_d_inner = 0\n", + "llm_load_print_meta: ssm_d_state = 0\n", + "llm_load_print_meta: ssm_dt_rank = 0\n", + "llm_load_print_meta: ssm_dt_b_c_rms = 0\n", "llm_load_print_meta: model type = 7B\n", - "llm_load_print_meta: model ftype = mostly Q4_K - Medium\n", + "llm_load_print_meta: model ftype = Q2_K - Medium\n", "llm_load_print_meta: model params = 7.24 B\n", - "llm_load_print_meta: model size = 4.07 GiB (4.83 BPW) \n", - "llm_load_print_meta: general.name = LLaMA v2\n", - "llm_load_print_meta: BOS token = 1 ''\n", - "llm_load_print_meta: EOS token = 2 ''\n", - "llm_load_print_meta: UNK token = 0 ''\n", - "llm_load_print_meta: LF token = 13 '<0x0A>'\n", - "llm_load_tensors: ggml ctx size = 0.10 MB\n", - "llm_load_tensors: using CUDA for GPU acceleration\n", - "llm_load_tensors: mem required = 70.41 MB\n", - "llm_load_tensors: offloading 32 repeating layers to GPU\n", - "llm_load_tensors: offloading non-repeating layers to GPU\n", - "llm_load_tensors: offloaded 35/35 layers to GPU\n", - "llm_load_tensors: VRAM used: 4095.05 MB\n", - ".................................................................................................\n" + "llm_load_print_meta: model size = 2.87 GiB (3.41 BPW) \n", + "llm_load_print_meta: general.name = mistralai_mistral-7b-v0.1\n", + "llm_load_print_meta: BOS token = 1 ''\n", + "llm_load_print_meta: EOS token = 2 ''\n", + "llm_load_print_meta: UNK token = 0 ''\n", + "llm_load_print_meta: LF token = 13 '<0x0A>'\n", + "llm_load_print_meta: EOG token = 2 ''\n", + "llm_load_print_meta: max token length = 48\n", + "llm_load_tensors: ggml ctx size = 0.14 MiB\n", + "llm_load_tensors: CPU buffer size = 2939.57 MiB\n", + "..................................................................................................\n" ] } ], @@ -393,7 +123,7 @@ "params = llama_cpp.llama_model_default_params()\n", "params.n_gpu_layers = 35\n", "model = llama_cpp.llama_load_model_from_file(\n", - " b\"../../models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf\", params=params\n", + " b\"/workspaces/llama-cpp-python/mistral-7b-v0.1.Q2_K.gguf\", params\n", ") # Update this to whatever" ] }, @@ -406,7 +136,7 @@ "name": "stdout", "output_type": "stream", "text": [ - "[1, 1014, 2936, 9060, 285, 1142]\n", + "[1, 415, 2936, 9060, 285, 1142]\n", "58\n" ] } @@ -436,17 +166,18 @@ "name": "stderr", "output_type": "stream", "text": [ - "llama_new_context_with_model: n_ctx = 58\n", + "llama_new_context_with_model: n_ctx = 64\n", + "llama_new_context_with_model: n_batch = 32\n", + "llama_new_context_with_model: n_ubatch = 32\n", + "llama_new_context_with_model: flash_attn = 0\n", "llama_new_context_with_model: freq_base = 10000.0\n", "llama_new_context_with_model: freq_scale = 1\n", - "llama_kv_cache_init: offloading v cache to GPU\n", - "llama_kv_cache_init: offloading k cache to GPU\n", - "llama_kv_cache_init: VRAM kv self = 7.25 MB\n", - "llama_new_context_with_model: kv self size = 7.25 MB\n", - "llama_build_graph: non-view tensors processed: 740/740\n", - "llama_new_context_with_model: compute buffer total size = 10.63 MB\n", - "llama_new_context_with_model: VRAM scratch buffer: 4.51 MB\n", - "llama_new_context_with_model: total VRAM used: 4106.81 MB (model: 4095.05 MB, context: 11.76 MB)\n" + "llama_kv_cache_init: CPU KV buffer size = 8.00 MiB\n", + "llama_new_context_with_model: KV self size = 8.00 MiB, K (f16): 4.00 MiB, V (f16): 4.00 MiB\n", + "llama_new_context_with_model: CPU output buffer size = 0.12 MiB\n", + "llama_new_context_with_model: CPU compute buffer size = 5.01 MiB\n", + "llama_new_context_with_model: graph nodes = 1030\n", + "llama_new_context_with_model: graph splits = 1\n" ] } ], @@ -476,7 +207,7 @@ "metadata": {}, "outputs": [], "source": [ - "import ctypes\n", + "\n", "\n", "batch.n_tokens = tokens_len\n", "for i in range(tokens_len):\n", @@ -502,10 +233,35 @@ " llama_cpp.llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens)" ] }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + }, { "cell_type": "code", "execution_count": 9, "metadata": {}, + "outputs": [], + "source": [ + "\n", + "# Initialize sampler chain with default parameters\n", + "sparams = llama_cpp.llama_sampler_chain_default_params()\n", + "sampler_chain = llama_cpp.llama_sampler_chain_init(sparams)\n", + "\n", + "# Add top_k, top_p, temperature, and final distribution-based sampler\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_top_k(40))\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_top_p(0.9, 1))\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_temp(0.4))\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_dist(1234)) # Final \"dist\" sampler" + ] + }, + { + "cell_type": "code", + "execution_count": 10, + "metadata": {}, "outputs": [ { "name": "stdout", @@ -514,61 +270,59 @@ "7\n", "[' j', ' jumped']\n", "8\n", - "[' jumps', ' jumped over']\n", + "[' j over', ' jumped over']\n", "9\n", - "[' jumps over', ' jumped over the']\n", + "[' j over the', ' jumped over the']\n", "10\n", - "[' jumps over the', ' jumped over the lazy']\n", + "[' j over the lazy', ' jumped over the lazy']\n", "11\n", - "[' jumps over the lazy', ' jumped over the lazy dog']\n", + "[' j over the lazy dog', ' jumped over the lazy dog']\n", "12\n", - "[' jumps over the lazy dog', ' jumped over the lazy dog.']\n", + "[' j over the lazy dog.', ' jumped over the lazy dog\\n']\n", "13\n", - "[' jumps over the lazy dog.', ' jumped over the lazy dog.\\n']\n", + "[' j over the lazy dog. También', ' jumped over the lazy dog\\nGroupLayout']\n", "14\n", - "[' jumps over the lazy dog.\\n', ' jumped over the lazy dog.\\n\\n']\n", + "[' j over the lazy dog. También:', ' jumped over the lazy dog\\nGroupLayouting']\n", "15\n", - "[' jumps over the lazy dog.\\n\\n', ' jumped over the lazy dog.\\n\\nThe']\n", + "[' j over the lazy dog. También: is', ' jumped over the lazy dog\\nGroupLayouting is']\n", "16\n", - "[' jumps over the lazy dog.\\n\\nI', ' jumped over the lazy dog.\\n\\nThe quick']\n", + "[' j over the lazy dog. También: is a', ' jumped over the lazy dog\\nGroupLayouting is a']\n", "17\n", - "[' jumps over the lazy dog.\\n\\nI’', ' jumped over the lazy dog.\\n\\nThe quick brown']\n", + "[' j over the lazy dog. También: is a technique', ' jumped over the lazy dog\\nGroupLayouting is a common']\n", "18\n", - "[' jumps over the lazy dog.\\n\\nI’m', ' jumped over the lazy dog.\\n\\nThe quick brown f']\n", + "[' j over the lazy dog. También: is a technique practice', ' jumped over the lazy dog\\nGroupLayouting is a common practice']\n", "19\n", - "[' jumps over the lazy dog.\\n\\nI’m not', ' jumped over the lazy dog.\\n\\nThe quick brown fox']\n", + "[' j over the lazy dog. También: is a technique practice in', ' jumped over the lazy dog\\nGroupLayouting is a common practice in']\n", "20\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped']\n", + "[' j over the lazy dog. También: is a technique practice in the', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the']\n", "21\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over']\n", + "[' j over the lazy dog. También: is a technique practice in the real', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media']\n", "22\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the']\n", + "[' j over the lazy dog. También: is a technique practice in the real-', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry']\n", "23\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy']\n", + "[' j over the lazy dog. También: is a technique practice in the real-.', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry.']\n", "24\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However']\n", "25\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We,', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However,']\n", "26\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there']\n", "27\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\n']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has']\n", "28\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been']\n", "29\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little']\n", "30\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little research', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little emp']\n", "31\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the English', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown f']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little researchirical', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little empirical']\n", "32\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the English language', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown fox']\n" + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little researchirical research', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little empirical research']\n" ] } ], "source": [ - "import ctypes\n", - "\n", "streams = [\"\"] * n_parallel\n", "i_batch = [batch.n_tokens - 1] * n_parallel\n", "\n", @@ -581,36 +335,17 @@ " if i_batch[i] < 0:\n", " continue\n", "\n", - " n_vocab = llama_cpp.llama_n_vocab(model)\n", - " logits = llama_cpp.llama_get_logits_ith(ctx, i_batch[i])\n", - "\n", - " candidates = (llama_cpp.llama_token_data * n_vocab)()\n", - "\n", - " for token_id in range(n_vocab):\n", - " candidates[token_id].id = token_id\n", - " candidates[token_id].logit = logits[token_id]\n", - " candidates[token_id].p = 0.0\n", - "\n", - " candidates_p = llama_cpp.llama_token_data_array(\n", - " candidates, len(candidates), False\n", - " )\n", - "\n", - " top_k = 40\n", - " top_p = 0.9\n", - " temp = 0.4\n", - "\n", - " llama_cpp.llama_sample_top_k(ctx, ctypes.byref(candidates_p), top_k, 1)\n", - " llama_cpp.llama_sample_top_p(ctx, ctypes.byref(candidates_p), top_p, 1)\n", - " llama_cpp.llama_sample_temp(ctx, ctypes.byref(candidates_p), temp)\n", - "\n", - " new_token_id = llama_cpp.llama_sample_token(ctx, ctypes.byref(candidates_p))\n", + " # Sample the next token using the sampler chain\n", + " new_token_id = llama_cpp.llama_sampler_sample(sampler_chain, ctx, -1)\n", "\n", " if new_token_id == llama_cpp.llama_token_eos(ctx) or n_cur == n_len:\n", " i_batch[i] = -1\n", " continue\n", "\n", " buf = (ctypes.c_char * 32)()\n", - " outlen = llama_cpp.llama_token_to_piece(model, new_token_id, buf, len(buf))\n", + " \n", + " # Convert token ID to text\n", + " outlen = llama_cpp.llama_token_to_piece(model, new_token_id, buf, len(buf), 0, False)\n", " streams[i] += bytes(buf[:outlen]).decode(\"utf-8\")\n", "\n", " batch.token[batch.n_tokens] = new_token_id\n", @@ -637,14 +372,14 @@ }, { "cell_type": "code", - "execution_count": 10, + "execution_count": 11, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the English language', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown fox']\n" + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little researchirical research', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little empirical research']\n" ] } ], @@ -654,7 +389,7 @@ }, { "cell_type": "code", - "execution_count": 11, + "execution_count": 12, "metadata": {}, "outputs": [], "source": [ @@ -663,7 +398,7 @@ }, { "cell_type": "code", - "execution_count": 12, + "execution_count": 13, "metadata": {}, "outputs": [], "source": [ @@ -672,7 +407,7 @@ }, { "cell_type": "code", - "execution_count": 13, + "execution_count": 14, "metadata": {}, "outputs": [], "source": [ @@ -681,7 +416,7 @@ }, { "cell_type": "code", - "execution_count": 14, + "execution_count": 15, "metadata": {}, "outputs": [], "source": [ @@ -705,7 +440,7 @@ ], "metadata": { "kernelspec": { - "display_name": ".venv", + "display_name": "Python 3", "language": "python", "name": "python3" }, @@ -719,7 +454,7 @@ "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", - "version": "3.11.5+" + "version": "3.12.1" }, "orig_nbformat": 4 }, From 4f0ec659f9519d3009873f6adf6aa8af5ab32d8b Mon Sep 17 00:00:00 2001 From: Adam Jones Date: Fri, 6 Dec 2024 12:35:46 +0000 Subject: [PATCH 596/624] fix: chat API logprobs format (#1788) * fix: chat API logprobs format * Fix optional properties --- Makefile | 2 +- llama_cpp/llama_chat_format.py | 53 ++++++++++++++++++++++++---------- llama_cpp/llama_types.py | 22 ++++++++++++-- 3 files changed, 59 insertions(+), 18 deletions(-) diff --git a/Makefile b/Makefile index be9b55a3e..52b17b61a 100644 --- a/Makefile +++ b/Makefile @@ -62,7 +62,7 @@ docker: docker build -t llama-cpp-python:latest -f docker/simple/Dockerfile . run-server: - uvicorn --factory llama.server:app --host ${HOST} --port ${PORT} + python llama_cpp/server --model ${MODEL} clean: - cd vendor/llama.cpp && make clean diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index dfb0af65e..17575c700 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -259,6 +259,31 @@ def to_chat_handler(self) -> LlamaChatCompletionHandler: return chat_formatter_to_chat_completion_handler(self) +def _convert_text_completion_logprobs_to_chat( + logprobs: Optional[llama_types.CompletionLogprobs], +) -> llama_types.ChatCompletionLogprobs: + if logprobs is None: + return None + + return { + "content": [ + { + "token": token, + "bytes": None, + "logprob": logprob, + "top_logprobs": [ + { + "token": top_token, + "logprob": top_logprob, + "bytes": None, + } + for top_token, top_logprob in top_logprobs.items() + ], + } for (token, logprob, top_logprobs) in zip(logprobs["tokens"], logprobs["token_logprobs"], logprobs["top_logprobs"]) + ], + "refusal": None, + } + def _convert_text_completion_to_chat( completion: llama_types.Completion, ) -> llama_types.ChatCompletion: @@ -275,7 +300,7 @@ def _convert_text_completion_to_chat( "role": "assistant", "content": completion["choices"][0]["text"], }, - "logprobs": completion["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "finish_reason": completion["choices"][0]["finish_reason"], } ], @@ -319,7 +344,7 @@ def _convert_text_completion_chunks_to_chat( if chunk["choices"][0]["finish_reason"] is None else {} ), - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "finish_reason": chunk["choices"][0]["finish_reason"], } ], @@ -382,7 +407,7 @@ def _convert_completion_to_chat_function( } ], }, - "logprobs": completion["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "finish_reason": "tool_calls", } ], @@ -435,7 +460,7 @@ def _stream_response_to_function_stream( { "index": 0, "finish_reason": None, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": None, "content": None, @@ -472,7 +497,7 @@ def _stream_response_to_function_stream( { "index": 0, "finish_reason": None, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": None, "content": None, @@ -1716,7 +1741,7 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): } ], }, - "logprobs": completion["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "finish_reason": "tool_calls", } ], @@ -2128,7 +2153,7 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": None, "content": None, @@ -2230,7 +2255,7 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": "assistant", "content": None, @@ -2268,9 +2293,7 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": chunk["choices"][0][ - "logprobs" - ], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": "assistant", "content": buffer.pop(0), @@ -2293,7 +2316,7 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": "assistant", "content": ( @@ -2379,7 +2402,7 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": chunk["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "delta": { "role": None, "content": None, @@ -2613,7 +2636,7 @@ def generate_streaming(tools, functions, function_call, prompt): choices=[ { "index": 0, - "logprobs": completion["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "message": { "role": "assistant", "content": None if content == "" else content, @@ -3745,7 +3768,7 @@ def chatml_function_calling( { "finish_reason": "tool_calls", "index": 0, - "logprobs": completion["choices"][0]["logprobs"], + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "message": { "role": "assistant", "content": None, diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 57836ee76..f647822ff 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -82,10 +82,28 @@ class ChatCompletionFunction(TypedDict): parameters: Dict[str, JsonType] # TODO: make this more specific +class ChatCompletionTopLogprobToken(TypedDict): + token: str + logprob: float + bytes: Optional[List[int]] + + +class ChatCompletionLogprobToken(ChatCompletionTopLogprobToken): + token: str + logprob: float + bytes: Optional[List[int]] + top_logprobs: List[ChatCompletionTopLogprobToken] + + +class ChatCompletionLogprobs(TypedDict): + content: Optional[List[ChatCompletionLogprobToken]] + refusal: Optional[List[ChatCompletionLogprobToken]] + + class ChatCompletionResponseChoice(TypedDict): index: int message: "ChatCompletionResponseMessage" - logprobs: Optional[CompletionLogprobs] + logprobs: Optional[ChatCompletionLogprobs] finish_reason: Optional[str] @@ -134,7 +152,7 @@ class ChatCompletionStreamResponseChoice(TypedDict): ChatCompletionStreamResponseDelta, ChatCompletionStreamResponseDeltaEmpty ] finish_reason: Optional[Literal["stop", "length", "tool_calls", "function_call"]] - logprobs: NotRequired[Optional[CompletionLogprobs]] + logprobs: NotRequired[Optional[ChatCompletionLogprobs]] class CreateChatCompletionStreamResponse(TypedDict): From df136cb3c7d3a7a9a4e30ef2d16c5e42e0e36e43 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 07:36:47 -0500 Subject: [PATCH 597/624] misc: Update development Makefile --- Makefile | 23 ++++++++++++++++------- 1 file changed, 16 insertions(+), 7 deletions(-) diff --git a/Makefile b/Makefile index be9b55a3e..37ef9b1d6 100644 --- a/Makefile +++ b/Makefile @@ -21,6 +21,15 @@ build.debug: --config-settings=cmake.args="-DCMAKE_BUILD_TYPE=Debug;-DCMAKE_C_FLAGS='-ggdb -O0';-DCMAKE_CXX_FLAGS='-ggdb -O0'" \ --editable . +build.debug.extra: + python3 -m pip install \ + --verbose \ + --config-settings=cmake.verbose=true \ + --config-settings=logging.level=INFO \ + --config-settings=install.strip=false \ + --config-settings=cmake.args="-DCMAKE_BUILD_TYPE=Debug;-DCMAKE_C_FLAGS='-fsanitize=address -ggdb -O0';-DCMAKE_CXX_FLAGS='-fsanitize=address -ggdb -O0'" \ + --editable . + build.cuda: CMAKE_ARGS="-DGGML_CUDA=on" python3 -m pip install --verbose -e . @@ -46,7 +55,7 @@ build.rpc: CMAKE_ARGS="-DGGML_RPC=on" python3 -m pip install --verbose -e . build.sdist: - python3 -m build --sdist + python3 -m build --sdist --verbose deploy.pypi: python3 -m twine upload dist/* @@ -56,7 +65,7 @@ deploy.gh-docs: mkdocs gh-deploy test: - python3 -m pytest + python3 -m pytest --full-trace -v docker: docker build -t llama-cpp-python:latest -f docker/simple/Dockerfile . @@ -68,11 +77,11 @@ clean: - cd vendor/llama.cpp && make clean - cd vendor/llama.cpp && rm libllama.so - rm -rf _skbuild - - rm llama_cpp/*.so - - rm llama_cpp/*.dylib - - rm llama_cpp/*.metal - - rm llama_cpp/*.dll - - rm llama_cpp/*.lib + - rm llama_cpp/lib/*.so + - rm llama_cpp/lib/*.dylib + - rm llama_cpp/lib/*.metal + - rm llama_cpp/lib/*.dll + - rm llama_cpp/lib/*.lib .PHONY: \ update \ From b9b50e50b9d29b2034c50e1b0ee09dc7c47bbc63 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 6 Dec 2024 07:37:53 -0500 Subject: [PATCH 598/624] misc: Update run server command --- Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Makefile b/Makefile index 78b11cf11..26ddf2c7a 100644 --- a/Makefile +++ b/Makefile @@ -71,7 +71,7 @@ docker: docker build -t llama-cpp-python:latest -f docker/simple/Dockerfile . run-server: - python llama_cpp/server --model ${MODEL} + python3 -m llama_cpp.server --model ${MODEL} clean: - cd vendor/llama.cpp && make clean From 5585f8afe1aea74dca7364f18a76f0426381001d Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 8 Dec 2024 22:40:19 -0500 Subject: [PATCH 599/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 12 ++++++------ vendor/llama.cpp | 2 +- 2 files changed, 7 insertions(+), 7 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index cd59bb430..44a1461db 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -344,9 +344,9 @@ # LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors # LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors # LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors -# LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // except 1d tensors -# LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors -# LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors +# //LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // removed from gguf files, use Q4_0 and runtime repack +# //LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // removed from gguf files, use Q4_0 and runtime repack +# //LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // removed from gguf files, use Q4_0 and runtime repack # LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors # LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors # @@ -382,9 +382,9 @@ LLAMA_FTYPE_MOSTLY_IQ4_XS = 30 LLAMA_FTYPE_MOSTLY_IQ1_M = 31 LLAMA_FTYPE_MOSTLY_BF16 = 32 -LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33 -LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34 -LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35 +# LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33 +# LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34 +# LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35 LLAMA_FTYPE_MOSTLY_TQ1_0 = 36 LLAMA_FTYPE_MOSTLY_TQ2_0 = 37 LLAMA_FTYPE_GUESSED = 1024 diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 7736837d6..ce8784bdb 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 7736837d62efed1dbebfe579472fca041eda12d6 +Subproject commit ce8784bdb153ff7794dde5a50b0ebfa51baa6171 From 61508c265a4e295d69aaee7533bde5b51278e574 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 8 Dec 2024 22:55:03 -0500 Subject: [PATCH 600/624] Add CUDA 12.5 and 12.6 to generated output wheels --- .github/workflows/generate-index-from-release.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 51961088d..5c7171bc1 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -44,6 +44,8 @@ jobs: ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' ./scripts/releases-to-pep-503.sh index/whl/cu123 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu123$' ./scripts/releases-to-pep-503.sh index/whl/cu124 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + ./scripts/releases-to-pep-503.sh index/whl/cu125 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + ./scripts/releases-to-pep-503.sh index/whl/cu126 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' - name: Upload artifact uses: actions/upload-pages-artifact@v3 From a9fe0f872527ffc1bf64e125b7996ec4dfa598eb Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 8 Dec 2024 22:56:45 -0500 Subject: [PATCH 601/624] chore: Bump version --- CHANGELOG.md | 14 ++++++++++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 15 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 8a194fece..331161a0c 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,20 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.3] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ce8784bdb153ff7794dde5a50b0ebfa51baa6171 +- fix: chat API logprobs format by @domdomegg in #1788 +- feat: Add support for CUDA 12.6, fix CUDA 12.5 by @Smartappli in #1775 +- fix: Make content not required in ChatCompletionRequestAssistantMessage by @feloy in #1807 +- fix: Fix pickling of Llama class by setting seed from _seed member by @abetlen in 2523472c3eccb9ab9277117cc4ff705212b6888a +- fix: Fix logit-bias type hint by @ddh0 in #1802 +- fix(server): Avoid thread starvation on many concurrent requests by making use of asyncio to lock llama_proxy context by @gjpower in #1798 +- fix(server): Added missing exit_stack.close() to /v1/chat/completions by @Ian321 in #1796 +- fix(examples): Refactor Batching notebook to use new sampler chain API by @lukestanley in #1793 +- fix(docs): Update development instructions by @Florents-Tselai in #1833 +- fix(docs): Remove ref to llama_eval in llama_cpp.py docs by @richdougherty in #1819 + ## [0.3.2] - feat: Update llama.cpp to ggerganov/llama.cpp@74d73dc85cc2057446bf63cc37ff649ae7cebd80 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index cda38a841..36c0eafc0 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.2" +__version__ = "0.3.3" From ca808028bd16b8327bd84128d48015a4b1304690 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 8 Dec 2024 23:51:49 -0500 Subject: [PATCH 602/624] fix(ci): hotfix for wheels --- .github/workflows/build-wheels-cuda.yaml | 2 +- .github/workflows/build-wheels-metal.yaml | 2 +- .github/workflows/generate-index-from-release.yaml | 4 ++-- 3 files changed, 4 insertions(+), 4 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 8624ec6d9..3d410148f 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -22,7 +22,7 @@ jobs: $matrix = @{ 'os' = @('ubuntu-latest', 'windows-2019') 'pyver' = @("3.9", "3.10", "3.11", "3.12") - 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "12.6.1") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") #, "12.5.1", "12.6.1") 'releasetag' = @("basic") } diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml index 01da3cef5..9b97bf2f5 100644 --- a/.github/workflows/build-wheels-metal.yaml +++ b/.github/workflows/build-wheels-metal.yaml @@ -11,7 +11,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [macos-12, macos-13, macos-14] + os: [macos-13, macos-14, macos-15] steps: - uses: actions/checkout@v4 diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml index 5c7171bc1..255ee67d6 100644 --- a/.github/workflows/generate-index-from-release.yaml +++ b/.github/workflows/generate-index-from-release.yaml @@ -44,8 +44,8 @@ jobs: ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' ./scripts/releases-to-pep-503.sh index/whl/cu123 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu123$' ./scripts/releases-to-pep-503.sh index/whl/cu124 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' - ./scripts/releases-to-pep-503.sh index/whl/cu125 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' - ./scripts/releases-to-pep-503.sh index/whl/cu126 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + # ./scripts/releases-to-pep-503.sh index/whl/cu125 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + # ./scripts/releases-to-pep-503.sh index/whl/cu126 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' - name: Upload artifact uses: actions/upload-pages-artifact@v3 From 002f58300fdaf865cf1d305b3721c5d96003c440 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Sun, 8 Dec 2024 23:53:30 -0500 Subject: [PATCH 603/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 331161a0c..2a1d3225f 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.4] + +- fix(ci): Build wheels for macos 13-15, cuda 12.1-12.4 by @abetlen in ca808028bd16b8327bd84128d48015a4b1304690 + ## [0.3.3] - feat: Update llama.cpp to ggerganov/llama.cpp@ce8784bdb153ff7794dde5a50b0ebfa51baa6171 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 36c0eafc0..dd8f1fc58 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.3" +__version__ = "0.3.4" From ea4d86ac25e0f39370ff3f54550387a02467ab77 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 9 Dec 2024 00:42:15 -0500 Subject: [PATCH 604/624] fix(ci): update macos runner image to non-deprecated version --- .github/workflows/build-and-release.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 1d549ed68..7307c85ab 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -11,7 +11,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-20.04, windows-2019, macos-12] + os: [ubuntu-20.04, windows-2019, macos-13] steps: - uses: actions/checkout@v4 From afedfc888462f9a6e809dc9455eb3b663764cc3f Mon Sep 17 00:00:00 2001 From: Graeme Power Date: Mon, 9 Dec 2024 23:47:18 +0000 Subject: [PATCH 605/624] fix: add missing await statements for async exit_stack handling (#1858) --- llama_cpp/server/app.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index f7c028475..b6db453b8 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -318,7 +318,7 @@ async def create_completion( Iterator[llama_cpp.CreateCompletionStreamResponse], ] = await run_in_threadpool(llama, **kwargs) except Exception as err: - exit_stack.close() + await exit_stack.aclose() raise err if isinstance(iterator_or_completion, Iterator): @@ -475,7 +475,7 @@ async def create_chat_completion( # is complete. # https://github.com/tiangolo/fastapi/issues/11143 exit_stack = contextlib.AsyncExitStack() - llama_proxy = exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) + llama_proxy = await exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) if llama_proxy is None: raise HTTPException( status_code=status.HTTP_503_SERVICE_UNAVAILABLE, @@ -513,7 +513,7 @@ async def create_chat_completion( llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) except Exception as err: - exit_stack.close() + await exit_stack.aclose() raise err if isinstance(iterator_or_completion, Iterator): From 801a73a621c6ba8f9a8ffe110766c9e721b09520 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 9 Dec 2024 18:49:27 -0500 Subject: [PATCH 606/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ce8784bdb..26a8406ba 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ce8784bdb153ff7794dde5a50b0ebfa51baa6171 +Subproject commit 26a8406ba9198eb6fdd8329fa717555b4f77f05f From 803924b02cc8ca7531b70e6497d1ff7b83f75737 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 9 Dec 2024 18:59:22 -0500 Subject: [PATCH 607/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 2a1d3225f..6f6355893 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.5] + +- feat: Update llama.cpp to ggerganov/llama.cpp@26a8406ba9198eb6fdd8329fa717555b4f77f05f +- fix(ci): Fix release by updating macos runner image to non-deprecated version by @abetlen in afedfc888462f9a6e809dc9455eb3b663764cc3f +- fix(server): add missing await statements for async exit_stack handling by @gjpower in #1858 + ## [0.3.4] - fix(ci): Build wheels for macos 13-15, cuda 12.1-12.4 by @abetlen in ca808028bd16b8327bd84128d48015a4b1304690 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index dd8f1fc58..d7d9be881 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.4" +__version__ = "0.3.5" From 2bc1d97c9672320828e70dc8293d5f8754682109 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 19 Dec 2024 01:55:12 -0500 Subject: [PATCH 608/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 5 ---- llama_cpp/llama_cpp.py | 54 +++++++++++------------------------------ vendor/llama.cpp | 2 +- 3 files changed, 15 insertions(+), 46 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 994d5f149..8fa2b447f 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -805,15 +805,10 @@ def add_penalties( ignore_eos: bool, ): sampler = llama_cpp.llama_sampler_init_penalties( - n_vocab, - special_eos_id, - linefeed_id, penalty_last_n, penalty_repeat, penalty_freq, penalty_present, - penalize_nl, - ignore_eos, ) self._add_sampler(sampler) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 44a1461db..0481cdbcd 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -256,13 +256,17 @@ # // note: these values should be synchronized with ggml_rope # // TODO: maybe move this enum to ggml.h (ggml_rope_type) # enum llama_rope_type { -# LLAMA_ROPE_TYPE_NONE = -1, -# LLAMA_ROPE_TYPE_NORM = 0, -# LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX, +# LLAMA_ROPE_TYPE_NONE = -1, +# LLAMA_ROPE_TYPE_NORM = 0, +# LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX, +# LLAMA_ROPE_TYPE_MROPE = GGML_ROPE_TYPE_MROPE, +# LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION, # }; LLAMA_ROPE_TYPE_NONE = -1 LLAMA_ROPE_TYPE_NORM = 0 LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX = 2 +LLAMA_ROPE_TYPE_MROPE = GGML_ROPE_TYPE_MROPE = 8 +LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION = 24 # enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file @@ -1265,6 +1269,7 @@ def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: # // Functions to access the model's GGUF metadata scalar values # // - The functions return the length of the string on success, or -1 on failure # // - The output string is always null-terminated and cleared on failure +# // - When retrieving a string, an extra byte must be allocated to account for the null terminator # // - GGUF array values are not supported by these functions @@ -1378,18 +1383,6 @@ def llama_model_n_params(model: llama_model_p, /) -> int: ... -# // Get a llama model tensor -# LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); -@ctypes_function( - "llama_get_model_tensor", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_void_p -) -def llama_get_model_tensor( - model: llama_model_p, name: Union[ctypes.c_char_p, bytes], / -) -> ctypes.c_void_p: - """Get a llama model tensor""" - ... - - # // Returns true if the model contains an encoder that requires llama_encode() call # LLAMA_API bool llama_model_has_encoder(const struct llama_model * model); @ctypes_function("llama_model_has_encoder", [llama_model_p_ctypes], ctypes.c_bool) @@ -3336,41 +3329,22 @@ def llama_sampler_init_grammar( ... +# /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first. # LLAMA_API struct llama_sampler * llama_sampler_init_penalties( -# int32_t n_vocab, // llama_n_vocab() -# llama_token special_eos_id, // llama_token_eos() -# llama_token linefeed_id, // llama_token_nl() -# int32_t penalty_last_n, // last n tokens to penalize (0 = disable penalty, -1 = context size) -# float penalty_repeat, // 1.0 = disabled -# float penalty_freq, // 0.0 = disabled -# float penalty_present, // 0.0 = disabled -# bool penalize_nl, // consider newlines as a repeatable token -# bool ignore_eos); // ignore the end-of-sequence token +# int32_t penalty_last_n, // last n tokens to penalize (0 = disable penalty, -1 = context size) +# float penalty_repeat, // 1.0 = disabled +# float penalty_freq, // 0.0 = disabled +# float penalty_present); // 0.0 = disabled @ctypes_function( "llama_sampler_init_penalties", - [ - ctypes.c_int32, - llama_token, - llama_token, - ctypes.c_int32, - ctypes.c_float, - ctypes.c_float, - ctypes.c_float, - ctypes.c_bool, - ctypes.c_bool, - ], + [ctypes.c_int32, ctypes.c_float, ctypes.c_float, ctypes.c_float], llama_sampler_p_ctypes, ) def llama_sampler_init_penalties( - n_vocab: int, - special_eos_id: int, - linefeed_id: int, penalty_last_n: int, penalty_repeat: float, penalty_freq: float, penalty_present: float, - penalize_nl: bool, - ignore_eos: bool, /, ) -> llama_sampler_p: ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 26a8406ba..7909e8588 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 26a8406ba9198eb6fdd8329fa717555b4f77f05f +Subproject commit 7909e8588ddf70820adf1f325490eb3f67b32875 From c9dfad4b6fdebc85f8f7914e8e46c6d9299ff4b4 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Mon, 30 Dec 2024 13:43:46 -0500 Subject: [PATCH 609/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 7909e8588..716bd6dec 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 7909e8588ddf70820adf1f325490eb3f67b32875 +Subproject commit 716bd6dec3e044e5c325386b5b0483392b24cefe From 1d5f5340e49d90bc097c1f75605a4dd7408c2378 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 Jan 2025 16:33:53 -0500 Subject: [PATCH 610/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 31 +++++++++++++++++++++++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 30 insertions(+), 3 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 0481cdbcd..16c6b9d73 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -222,6 +222,7 @@ # LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, # LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, # LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -251,6 +252,7 @@ LLAMA_VOCAB_PRE_TYPE_EXAONE = 25 LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 # // note: these values should be synchronized with ggml_rope @@ -1090,9 +1092,10 @@ def llama_backend_free(): ... -# LLAMA_API struct llama_model * llama_load_model_from_file( +# DEPRECATED(LLAMA_API struct llama_model * llama_load_model_from_file( # const char * path_model, -# struct llama_model_params params); +# struct llama_model_params params), +# "use llama_model_load_from_file instead"); @ctypes_function( "llama_load_model_from_file", [ctypes.c_char_p, llama_model_params], @@ -1104,6 +1107,20 @@ def llama_load_model_from_file( ... +# LLAMA_API struct llama_model * llama_model_load_from_file( +# const char * path_model, +# struct llama_model_params params); +@ctypes_function( + "llama_model_load_from_file", + [ctypes.c_char_p, llama_model_params], + llama_model_p_ctypes, +) +def llama_model_load_from_file( + path_model: bytes, params: llama_model_params, / +) -> Optional[llama_model_p]: + ... + + # LLAMA_API void llama_free_model(struct llama_model * model); @ctypes_function( "llama_free_model", @@ -1114,6 +1131,16 @@ def llama_free_model(model: llama_model_p, /): ... +# LLAMA_API void llama_model_free(struct llama_model * model); +@ctypes_function( + "llama_model_free", + [llama_model_p_ctypes], + None, +) +def llama_model_free(model: llama_model_p, /): + ... + + # LLAMA_API struct llama_context * llama_new_context_with_model( # struct llama_model * model, # struct llama_context_params params); diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 716bd6dec..f7cd13301 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 716bd6dec3e044e5c325386b5b0483392b24cefe +Subproject commit f7cd13301c2a88f97073fd119072b4cc92c08df1 From e8f14ceff9b3279c67368d85f9457bb27168d126 Mon Sep 17 00:00:00 2001 From: Graeme Power Date: Wed, 8 Jan 2025 21:46:43 +0000 Subject: [PATCH 611/624] fix: streaming resource lock (#1879) * fix: correct issue with handling lock during streaming move locking for streaming into get_event_publisher call so it is locked and unlocked in the correct task for the streaming reponse * fix: simplify exit stack management for create_chat_completion and create_completion * fix: correct missing `async with` and format code * fix: remove unnecessary explicit use of AsyncExitStack fix: correct type hints for body_model --------- Co-authored-by: Andrei --- llama_cpp/server/app.py | 224 ++++++++++++++++++---------------------- 1 file changed, 103 insertions(+), 121 deletions(-) diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index b6db453b8..5120f2416 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -7,7 +7,7 @@ from anyio import Lock from functools import partial -from typing import Iterator, List, Optional, Union, Dict +from typing import List, Optional, Union, Dict import llama_cpp @@ -155,34 +155,71 @@ def create_app( return app +def prepare_request_resources( + body: CreateCompletionRequest | CreateChatCompletionRequest, + llama_proxy: LlamaProxy, + body_model: str | None, + kwargs, +) -> llama_cpp.Llama: + if llama_proxy is None: + raise HTTPException( + status_code=status.HTTP_503_SERVICE_UNAVAILABLE, + detail="Service is not available", + ) + llama = llama_proxy(body_model) + if body.logit_bias is not None: + kwargs["logit_bias"] = ( + _logit_bias_tokens_to_input_ids(llama, body.logit_bias) + if body.logit_bias_type == "tokens" + else body.logit_bias + ) + + if body.grammar is not None: + kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) + + if body.min_tokens > 0: + _min_tokens_logits_processor = llama_cpp.LogitsProcessorList( + [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())] + ) + if "logits_processor" not in kwargs: + kwargs["logits_processor"] = _min_tokens_logits_processor + else: + kwargs["logits_processor"].extend(_min_tokens_logits_processor) + return llama + + async def get_event_publisher( request: Request, inner_send_chan: MemoryObjectSendStream[typing.Any], - iterator: Iterator[typing.Any], - on_complete: typing.Optional[typing.Callable[[], typing.Awaitable[None]]] = None, + body: CreateCompletionRequest | CreateChatCompletionRequest, + body_model: str | None, + llama_call, + kwargs, ): server_settings = next(get_server_settings()) interrupt_requests = ( server_settings.interrupt_requests if server_settings else False ) - async with inner_send_chan: - try: - async for chunk in iterate_in_threadpool(iterator): - await inner_send_chan.send(dict(data=json.dumps(chunk))) - if await request.is_disconnected(): - raise anyio.get_cancelled_exc_class()() - if interrupt_requests and llama_outer_lock.locked(): - await inner_send_chan.send(dict(data="[DONE]")) - raise anyio.get_cancelled_exc_class()() - await inner_send_chan.send(dict(data="[DONE]")) - except anyio.get_cancelled_exc_class() as e: - print("disconnected") - with anyio.move_on_after(1, shield=True): - print(f"Disconnected from client (via refresh/close) {request.client}") - raise e - finally: - if on_complete: - await on_complete() + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + async with inner_send_chan: + try: + iterator = await run_in_threadpool(llama_call, llama, **kwargs) + async for chunk in iterate_in_threadpool(iterator): + await inner_send_chan.send(dict(data=json.dumps(chunk))) + if await request.is_disconnected(): + raise anyio.get_cancelled_exc_class()() + if interrupt_requests and llama_outer_lock.locked(): + await inner_send_chan.send(dict(data="[DONE]")) + raise anyio.get_cancelled_exc_class()() + await inner_send_chan.send(dict(data="[DONE]")) + except anyio.get_cancelled_exc_class() as e: + print("disconnected") + with anyio.move_on_after(1, shield=True): + print( + f"Disconnected from client (via refresh/close) {request.client}" + ) + raise e def _logit_bias_tokens_to_input_ids( @@ -267,18 +304,11 @@ async def create_completion( request: Request, body: CreateCompletionRequest, ) -> llama_cpp.Completion: - exit_stack = contextlib.AsyncExitStack() - llama_proxy = await exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) - if llama_proxy is None: - raise HTTPException( - status_code=status.HTTP_503_SERVICE_UNAVAILABLE, - detail="Service is not available", - ) if isinstance(body.prompt, list): assert len(body.prompt) <= 1 body.prompt = body.prompt[0] if len(body.prompt) > 0 else "" - llama = llama_proxy( + body_model = ( body.model if request.url.path != "/v1/engines/copilot-codex/completions" else "copilot-codex" @@ -293,44 +323,8 @@ async def create_completion( } kwargs = body.model_dump(exclude=exclude) - if body.logit_bias is not None: - kwargs["logit_bias"] = ( - _logit_bias_tokens_to_input_ids(llama, body.logit_bias) - if body.logit_bias_type == "tokens" - else body.logit_bias - ) - - if body.grammar is not None: - kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) - - if body.min_tokens > 0: - _min_tokens_logits_processor = llama_cpp.LogitsProcessorList( - [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())] - ) - if "logits_processor" not in kwargs: - kwargs["logits_processor"] = _min_tokens_logits_processor - else: - kwargs["logits_processor"].extend(_min_tokens_logits_processor) - - try: - iterator_or_completion: Union[ - llama_cpp.CreateCompletionResponse, - Iterator[llama_cpp.CreateCompletionStreamResponse], - ] = await run_in_threadpool(llama, **kwargs) - except Exception as err: - await exit_stack.aclose() - raise err - - if isinstance(iterator_or_completion, Iterator): - # EAFP: It's easier to ask for forgiveness than permission - first_response = await run_in_threadpool(next, iterator_or_completion) - - # If no exception was raised from first_response, we can assume that - # the iterator is valid and we can use it to stream the response. - def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: - yield first_response - yield from iterator_or_completion - + # handle streaming request + if kwargs.get("stream", False): send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( recv_chan, @@ -338,15 +332,29 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: get_event_publisher, request=request, inner_send_chan=send_chan, - iterator=iterator(), - on_complete=exit_stack.aclose, + body=body, + body_model=body_model, + llama_call=llama_cpp.Llama.__call__, + kwargs=kwargs, ), sep="\n", ping_message_factory=_ping_message_factory, ) - else: - await exit_stack.aclose() - return iterator_or_completion + + # handle regular request + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + + if await request.is_disconnected(): + print( + f"Disconnected from client (via refresh/close) before llm invoked {request.client}" + ) + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail="Client closed request", + ) + + return await run_in_threadpool(llama, **kwargs) @router.post( @@ -474,13 +482,8 @@ async def create_chat_completion( # where the dependency is cleaned up before a StreamingResponse # is complete. # https://github.com/tiangolo/fastapi/issues/11143 - exit_stack = contextlib.AsyncExitStack() - llama_proxy = await exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) - if llama_proxy is None: - raise HTTPException( - status_code=status.HTTP_503_SERVICE_UNAVAILABLE, - detail="Service is not available", - ) + + body_model = body.model exclude = { "n", "logit_bias_type", @@ -488,44 +491,9 @@ async def create_chat_completion( "min_tokens", } kwargs = body.model_dump(exclude=exclude) - llama = llama_proxy(body.model) - if body.logit_bias is not None: - kwargs["logit_bias"] = ( - _logit_bias_tokens_to_input_ids(llama, body.logit_bias) - if body.logit_bias_type == "tokens" - else body.logit_bias - ) - - if body.grammar is not None: - kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) - - if body.min_tokens > 0: - _min_tokens_logits_processor = llama_cpp.LogitsProcessorList( - [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())] - ) - if "logits_processor" not in kwargs: - kwargs["logits_processor"] = _min_tokens_logits_processor - else: - kwargs["logits_processor"].extend(_min_tokens_logits_processor) - - try: - iterator_or_completion: Union[ - llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] - ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) - except Exception as err: - await exit_stack.aclose() - raise err - - if isinstance(iterator_or_completion, Iterator): - # EAFP: It's easier to ask for forgiveness than permission - first_response = await run_in_threadpool(next, iterator_or_completion) - - # If no exception was raised from first_response, we can assume that - # the iterator is valid and we can use it to stream the response. - def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: - yield first_response - yield from iterator_or_completion + # handle streaming request + if kwargs.get("stream", False): send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( recv_chan, @@ -533,15 +501,29 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: get_event_publisher, request=request, inner_send_chan=send_chan, - iterator=iterator(), - on_complete=exit_stack.aclose, + body=body, + body_model=body_model, + llama_call=llama_cpp.Llama.create_chat_completion, + kwargs=kwargs, ), sep="\n", ping_message_factory=_ping_message_factory, ) - else: - await exit_stack.aclose() - return iterator_or_completion + + # handle regular request + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + + if await request.is_disconnected(): + print( + f"Disconnected from client (via refresh/close) before llm invoked {request.client}" + ) + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail="Client closed request", + ) + + return await run_in_threadpool(llama.create_chat_completion, **kwargs) @router.get( From 0580cf273debf4a7f2efcdfd5ef092ff5cedf9b0 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 8 Jan 2025 16:53:54 -0500 Subject: [PATCH 612/624] chore: Bump version --- CHANGELOG.md | 5 +++++ llama_cpp/__init__.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 6f6355893..b4f49c904 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.6] + +- feat: Update llama.cpp to ggerganov/llama.cpp@f7cd13301c2a88f97073fd119072b4cc92c08df1 +- fix(server): streaming resource lock by @gjpower in #1879 + ## [0.3.5] - feat: Update llama.cpp to ggerganov/llama.cpp@26a8406ba9198eb6fdd8329fa717555b4f77f05f diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index d7d9be881..8c6118fb4 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.5" +__version__ = "0.3.6" From 80be68ac36b458e02f5e6af8bddd005a2f680e39 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 28 Jan 2025 21:38:05 -0500 Subject: [PATCH 613/624] feat: Update llama.cpp --- llama_cpp/_internals.py | 181 +++++----- llama_cpp/llama.py | 16 +- llama_cpp/llama_cpp.py | 707 +++++++++++++++++++++++++++++----------- vendor/llama.cpp | 2 +- 4 files changed, 626 insertions(+), 280 deletions(-) diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 8fa2b447f..343581dce 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -55,7 +55,13 @@ def __init__( if model is None: raise ValueError(f"Failed to load model from file: {path_model}") + vocab = llama_cpp.llama_model_get_vocab(model) + + if vocab is None: + raise ValueError(f"Failed to get vocab from model: {path_model}") + self.model = model + self.vocab = vocab def free_model(): if self.model is None: @@ -75,7 +81,7 @@ def vocab_type(self) -> int: return llama_cpp.llama_vocab_type(self.model) def n_vocab(self) -> int: - return llama_cpp.llama_n_vocab(self.model) + return llama_cpp.llama_n_vocab(self.vocab) def n_ctx_train(self) -> int: return llama_cpp.llama_n_ctx_train(self.model) @@ -84,7 +90,7 @@ def n_embd(self) -> int: return llama_cpp.llama_n_embd(self.model) def rope_freq_scale_train(self) -> float: - return llama_cpp.llama_rope_freq_scale_train(self.model) + return llama_cpp.llama_model_rope_freq_scale_train(self.model) def desc(self) -> str: buf = ctypes.create_string_buffer(1024) @@ -98,53 +104,53 @@ def n_params(self) -> int: return llama_cpp.llama_model_n_params(self.model) def get_tensor(self, name: str) -> ctypes.c_void_p: - return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) + raise NotImplementedError("get_tensor is not implemented in llama.cpp") # Vocab def token_get_text(self, token: int) -> str: - return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8") + return llama_cpp.llama_token_get_text(self.vocab, token).decode("utf-8") def token_get_score(self, token: int) -> float: - return llama_cpp.llama_token_get_score(self.model, token) + return llama_cpp.llama_token_get_score(self.vocab, token) def token_get_attr(self, token: int) -> int: - return llama_cpp.llama_token_get_attr(self.model, token) + return llama_cpp.llama_token_get_attr(self.vocab, token) # Special tokens def token_bos(self) -> int: - return llama_cpp.llama_token_bos(self.model) + return llama_cpp.llama_token_bos(self.vocab) def token_eos(self) -> int: - return llama_cpp.llama_token_eos(self.model) + return llama_cpp.llama_token_eos(self.vocab) def token_cls(self) -> int: - return llama_cpp.llama_token_cls(self.model) + return llama_cpp.llama_token_cls(self.vocab) def token_sep(self) -> int: - return llama_cpp.llama_token_sep(self.model) + return llama_cpp.llama_token_sep(self.vocab) def token_nl(self) -> int: - return llama_cpp.llama_token_nl(self.model) + return llama_cpp.llama_token_nl(self.vocab) def token_prefix(self) -> int: - return llama_cpp.llama_token_prefix(self.model) + raise NotImplementedError("token_prefix is not implemented in llama.cpp") def token_middle(self) -> int: - return llama_cpp.llama_token_middle(self.model) + raise NotImplementedError("token_middle is not implemented in llama.cpp") def token_suffix(self) -> int: - return llama_cpp.llama_token_suffix(self.model) + raise NotImplementedError("token_suffix is not implemented in llama.cpp") def token_eot(self) -> int: - return llama_cpp.llama_token_eot(self.model) + return llama_cpp.llama_token_eot(self.vocab) def add_bos_token(self) -> bool: - return llama_cpp.llama_add_bos_token(self.model) + return llama_cpp.llama_add_bos_token(self.vocab) def add_eos_token(self) -> bool: - return llama_cpp.llama_add_eos_token(self.model) + return llama_cpp.llama_add_eos_token(self.vocab) # Tokenization @@ -152,13 +158,13 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): n_ctx = self.n_ctx_train() tokens = (llama_cpp.llama_token * n_ctx)() n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_ctx, add_bos, special + self.vocab, text, len(text), tokens, n_ctx, add_bos, special ) if n_tokens < 0: n_tokens = abs(n_tokens) tokens = (llama_cpp.llama_token * n_tokens)() n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_tokens, add_bos, special + self.vocab, text, len(text), tokens, n_tokens, add_bos, special ) if n_tokens < 0: raise RuntimeError( @@ -168,7 +174,7 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): def token_to_piece(self, token: int, special: bool = False) -> bytes: buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32, 0, special) + llama_cpp.llama_token_to_piece(self.vocab, token, buf, 32, 0, special) return bytes(buf) def detokenize(self, tokens: List[int], special: bool = False) -> bytes: @@ -177,7 +183,7 @@ def detokenize(self, tokens: List[int], special: bool = False) -> bytes: buffer = (ctypes.c_char * size)() for token in tokens: n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token), buffer, size, 0, special + self.vocab, llama_cpp.llama_token(token), buffer, size, 0, special ) assert n <= size output += bytes(buffer[:n]) @@ -320,7 +326,8 @@ def get_embeddings(self): def set_rng_seed(self, seed: int): # TODO: Fix - llama_cpp.llama_set_rng_seed(self.ctx, seed) + # llama_cpp.llama_set_rng_seed(self.ctx, seed) + raise NotImplementedError("set_rng_seed is not implemented in llama.cpp") def sample_repetition_penalties( self, @@ -331,55 +338,63 @@ def sample_repetition_penalties( penalty_freq: float, penalty_present: float, ): - llama_cpp.llama_sample_repetition_penalties( - self.ctx, - llama_cpp.byref(candidates.candidates), - last_tokens_data, - penalty_last_n, - penalty_repeat, - penalty_freq, - penalty_present, - ) + # llama_cpp.llama_sample_repetition_penalties( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # last_tokens_data, + # penalty_last_n, + # penalty_repeat, + # penalty_freq, + # penalty_present, + # ) + raise NotImplementedError("sample_repetition_penalties is not implemented in llama.cpp") def sample_softmax(self, candidates: "_LlamaTokenDataArray"): - llama_cpp.llama_sample_softmax( - self.ctx, - llama_cpp.byref(candidates.candidates), - ) + # llama_cpp.llama_sample_softmax( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # ) + raise NotImplementedError("sample_softmax is not implemented in llama.cpp") def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): - llama_cpp.llama_sample_top_k( - self.ctx, llama_cpp.byref(candidates.candidates), k, min_keep - ) + # llama_cpp.llama_sample_top_k( + # self.ctx, llama_cpp.byref(candidates.candidates), k, min_keep + # ) + raise NotImplementedError("sample_top_k is not implemented in llama.cpp") def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - llama_cpp.llama_sample_top_p( - self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep - ) + # llama_cpp.llama_sample_top_p( + # self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep + # ) + raise NotImplementedError("sample_top_p is not implemented in llama.cpp") def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - llama_cpp.llama_sample_min_p( - self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep - ) + # llama_cpp.llama_sample_min_p( + # self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep + # ) + raise NotImplementedError("sample_min_p is not implemented in llama.cpp") def sample_typical( self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int ): - llama_cpp.llama_sample_typical( - self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep - ) + # llama_cpp.llama_sample_typical( + # self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep + # ) + raise NotImplementedError("sample_typical is not implemented in llama.cpp") def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): - llama_cpp.llama_sample_temp( - self.ctx, llama_cpp.byref(candidates.candidates), temp - ) + # llama_cpp.llama_sample_temp( + # self.ctx, llama_cpp.byref(candidates.candidates), temp + # ) + raise NotImplementedError("sample_temp is not implemented in llama.cpp") def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): - llama_cpp.llama_sample_grammar( - self.ctx, - llama_cpp.byref(candidates.candidates), - grammar.grammar, - ) + # llama_cpp.llama_sample_grammar( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # grammar.grammar, + # ) + raise NotImplementedError("sample_grammar is not implemented in llama.cpp") def sample_token_mirostat( self, @@ -389,14 +404,15 @@ def sample_token_mirostat( m: int, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: - return llama_cpp.llama_sample_token_mirostat( - self.ctx, - llama_cpp.byref(candidates.candidates), - tau, - eta, - m, - mu, - ) + raise NotImplementedError("sample_token_mirostat is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token_mirostat( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # tau, + # eta, + # m, + # mu, + # ) def sample_token_mirostat_v2( self, @@ -405,29 +421,33 @@ def sample_token_mirostat_v2( eta: float, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: - return llama_cpp.llama_sample_token_mirostat_v2( - self.ctx, - llama_cpp.byref(candidates.candidates), - tau, - eta, - mu, - ) + raise NotImplementedError("sample_token_mirostat_v2 is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token_mirostat_v2( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # tau, + # eta, + # mu, + # ) def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: - return llama_cpp.llama_sample_token_greedy( - self.ctx, - llama_cpp.byref(candidates.candidates), - ) + raise NotImplementedError("sample_token_greedy is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token_greedy( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # ) def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: - return llama_cpp.llama_sample_token( - self.ctx, - llama_cpp.byref(candidates.candidates), - ) + raise NotImplementedError("sample_token is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # ) # Grammar def grammar_accept_token(self, grammar: LlamaGrammar, token: int): - llama_cpp.llama_grammar_accept_token(grammar.grammar, self.ctx, token) + raise NotImplementedError("grammar_accept_token is not implemented in llama.cpp") + # llama_cpp.llama_grammar_accept_token(grammar.grammar, self.ctx, token) def reset_timings(self): llama_cpp.llama_perf_context_reset(self.ctx) @@ -788,7 +808,7 @@ def add_mirostat_v2(self, seed: int, tau: float, eta: float): def add_grammar(self, model: LlamaModel, grammar: LlamaGrammar): sampler = llama_cpp.llama_sampler_init_grammar( - model.model, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8") + model.vocab, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8") ) self._add_sampler(sampler) @@ -842,6 +862,7 @@ def get_seed(self) -> int: def sample(self, ctx: LlamaContext, idx: int) -> int: assert self.sampler is not None + assert ctx.ctx is not None return llama_cpp.llama_sampler_sample(self.sampler, ctx.ctx, idx) def close(self): diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 2fd7ff193..11fc744a0 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -406,10 +406,10 @@ def __init__( ) ) - self._lora_adapter: Optional[llama_cpp.llama_lora_adapter_p] = None + self._lora_adapter: Optional[llama_cpp.llama_adapter_lora_p] = None if self.lora_path: - self._lora_adapter = llama_cpp.llama_lora_adapter_init( + self._lora_adapter = llama_cpp.llama_adapter_lora_init( self._model.model, self.lora_path.encode("utf-8"), ) @@ -421,12 +421,12 @@ def __init__( def free_lora_adapter(): if self._lora_adapter is None: return - llama_cpp.llama_lora_adapter_free(self._lora_adapter) + llama_cpp.llama_adapter_lora_free(self._lora_adapter) self._lora_adapter = None self._stack.callback(free_lora_adapter) - if llama_cpp.llama_lora_adapter_set( + if llama_cpp.llama_set_adapter_lora( self._ctx.ctx, self._lora_adapter, self.lora_scale ): raise RuntimeError( @@ -1152,9 +1152,9 @@ def _create_completion( bos_token_id: int = self.token_bos() cls_token_id: int = self._model.token_cls() sep_token_id: int = self._model.token_sep() - prefix_token_id: int = self._model.token_prefix() - middle_token_id: int = self._model.token_middle() - suffix_token_id: int = self._model.token_suffix() + prefix_token_id: int = 0 # self._model.token_prefix() # TODO: Fix + middle_token_id: int = 0 # self._model.token_middle() # TODO: Fix + suffix_token_id: int = 0 # self._model.token_suffix() # TODO: Fix add_space_prefix: bool = ( self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true" ) @@ -1332,7 +1332,7 @@ def logit_bias_processor( logits_processor=logits_processor, grammar=grammar, ): - if llama_cpp.llama_token_is_eog(self._model.model, token): + if llama_cpp.llama_token_is_eog(self._model.vocab, token): text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" break diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 16c6b9d73..7845f0d1c 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -149,6 +149,10 @@ # define LLAMA_STATE_SEQ_VERSION 2 LLAMA_STATE_SEQ_VERSION = 2 +# struct llama_vocab; +llama_vocab_p = NewType("llama_vocab_p", int) +llama_vocab_p_ctypes = ctypes.c_void_p + # struct llama_model; llama_model_p = NewType("llama_model_p", int) llama_model_p_ctypes = ctypes.c_void_p @@ -640,9 +644,6 @@ class llama_model_kv_override(ctypes.Structure): # // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() # const float * tensor_split; -# // comma separated list of RPC servers to use for offloading -# const char * rpc_servers; - # // Called with a progress value between 0.0 and 1.0. Pass NULL to disable. # // If the provided progress_callback returns true, model loading continues. # // If it returns false, model loading is immediately aborted. @@ -669,7 +670,6 @@ class llama_model_params(ctypes.Structure): split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored tensor_split (ctypes.Array[ctypes.ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() - rpc_servers (ctypes.c_char_p): comma separated list of RPC servers to use for offloading progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. progress_callback_user_data (ctypes.ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data @@ -683,7 +683,6 @@ class llama_model_params(ctypes.Structure): split_mode: int main_gpu: int tensor_split: CtypesArray[ctypes.c_float] - rpc_servers: ctypes.c_char_p progress_callback: Callable[[float, ctypes.c_void_p], bool] progress_callback_user_data: ctypes.c_void_p kv_overrides: CtypesArray[llama_model_kv_override] @@ -698,7 +697,6 @@ class llama_model_params(ctypes.Structure): ("split_mode", ctypes.c_int), ("main_gpu", ctypes.c_int32), ("tensor_split", ctypes.POINTER(ctypes.c_float)), - ("rpc_servers", ctypes.c_char_p), ("progress_callback", llama_progress_callback), ("progress_callback_user_data", ctypes.c_void_p), ("kv_overrides", ctypes.POINTER(llama_model_kv_override)), @@ -975,9 +973,9 @@ class llama_chat_message(ctypes.Structure): # // lora adapter -# struct llama_lora_adapter; -llama_lora_adapter_p = ctypes.c_void_p -llama_lora_adapter_p_ctypes = ctypes.POINTER(ctypes.c_void_p) +# struct llama_adapter_lora; +llama_adapter_lora_p = ctypes.c_void_p +llama_adapter_lora_p_ctypes = ctypes.POINTER(ctypes.c_void_p) # // Helpers for getting default parameters @@ -1059,6 +1057,18 @@ def llama_backend_init(): GGML_NUMA_STRATEGY_COUNT = 5 +# // Call once at the end of the program - currently only used for MPI +# LLAMA_API void llama_backend_free(void); +@ctypes_function( + "llama_backend_free", + [], + None, +) +def llama_backend_free(): + """Call once at the end of the program - currently only used for MPI""" + ... + + # //optional: # LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); @ctypes_function( @@ -1072,24 +1082,14 @@ def llama_numa_init(numa: int, /): # // Optional: an auto threadpool gets created in ggml if not passed explicitly # LLAMA_API void llama_attach_threadpool( -# struct llama_context * ctx, -# ggml_threadpool_t threadpool, -# ggml_threadpool_t threadpool_batch); +# struct llama_context * ctx, +# ggml_threadpool_t threadpool, +# ggml_threadpool_t threadpool_batch); +# TODO: Add llama_attach_threadpool # LLAMA_API void llama_detach_threadpool(struct llama_context * ctx); - - -# // Call once at the end of the program - currently only used for MPI -# LLAMA_API void llama_backend_free(void); -@ctypes_function( - "llama_backend_free", - [], - None, -) -def llama_backend_free(): - """Call once at the end of the program - currently only used for MPI""" - ... +# TODO: Add llama_detach_threadpool # DEPRECATED(LLAMA_API struct llama_model * llama_load_model_from_file( @@ -1107,6 +1107,9 @@ def llama_load_model_from_file( ... +# // Load the model from a file +# // If the file is split into multiple parts, the file name must follow this pattern: -%05d-of-%05d.gguf +# // If the split file name does not follow this pattern, use llama_model_load_from_splits # LLAMA_API struct llama_model * llama_model_load_from_file( # const char * path_model, # struct llama_model_params params); @@ -1118,6 +1121,31 @@ def llama_load_model_from_file( def llama_model_load_from_file( path_model: bytes, params: llama_model_params, / ) -> Optional[llama_model_p]: + """Load the model from a file + + If the file is split into multiple parts, the file name must follow this pattern: -%05d-of-%05d.gguf + + If the split file name does not follow this pattern, use llama_model_load_from_splits""" + ... + + +# // Load the model from multiple splits (support custom naming scheme) +# // The paths must be in the correct order +# LLAMA_API struct llama_model * llama_model_load_from_splits( +# const char ** paths, +# size_t n_paths, +# struct llama_model_params params); +@ctypes_function( + "llama_model_load_from_splits", + [ctypes.POINTER(ctypes.c_char_p), ctypes.c_size_t, llama_model_params], + llama_model_p_ctypes, +) +def llama_model_load_from_splits( + paths: List[bytes], n_paths: int, params: llama_model_params, / +) -> Optional[llama_model_p]: + """Load the model from multiple splits (support custom naming scheme) + + The paths must be in the correct order""" ... @@ -1141,9 +1169,24 @@ def llama_model_free(model: llama_model_p, /): ... -# LLAMA_API struct llama_context * llama_new_context_with_model( +# LLAMA_API struct llama_context * llama_init_from_model( # struct llama_model * model, # struct llama_context_params params); +@ctypes_function( + "llama_init_from_model", + [llama_model_p_ctypes, llama_context_params], + llama_context_p_ctypes, +) +def llama_init_from_model( + model: llama_model_p, params: llama_context_params, / +) -> Optional[llama_context_p]: + ... + + +# DEPRECATED(LLAMA_API struct llama_context * llama_new_context_with_model( +# struct llama_model * model, +# struct llama_context_params params), +# "use llama_init_from_model instead"); @ctypes_function( "llama_new_context_with_model", [llama_model_p_ctypes, llama_context_params], @@ -1231,65 +1274,102 @@ def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... -# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); -@ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: - ... -# LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model), "use llama_model_n_ctx_train instead"); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_ctx_train(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_embd (const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_embd (const struct llama_model * model), "use llama_model_n_embd instead"); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_embd(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_layer (const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_layer (const struct llama_model * model), "use llama_model_n_layer instead"); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_layer(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_head (const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_head (const struct llama_model * model), "use llama_model_n_head instead"); @ctypes_function("llama_n_head", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_head(model: llama_model_p, /) -> int: ... -# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); +# DEPRECATED(LLAMA_API int32_t llama_n_vocab (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead"); +@ctypes_function("llama_n_vocab", [llama_vocab_p_ctypes], ctypes.c_int32) +def llama_n_vocab(model: llama_vocab_p, /) -> int: + ... + + +# LLAMA_API const struct llama_model * llama_get_model (const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... -# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) def llama_pooling_type(ctx: llama_context_p, /) -> int: ... -# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); -@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: +# LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model); +@ctypes_function("llama_model_get_vocab", [llama_model_p_ctypes], llama_vocab_p_ctypes) +def llama_model_get_vocab(model: llama_model_p, /) -> Optional[llama_vocab_p]: + ... + + +# LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model); +@ctypes_function("llama_model_rope_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_model_rope_type(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model); +@ctypes_function("llama_model_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_ctx_train(model: llama_model_p, /) -> int: ... -# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); -@ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: +# LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model); +@ctypes_function("llama_model_n_embd", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_embd(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model); +@ctypes_function("llama_model_n_layer", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_layer(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_head (const struct llama_model * model); +@ctypes_function("llama_model_n_head", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head(model: llama_model_p, /) -> int: ... # // Get the model's RoPE frequency scaling factor -# LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); -@ctypes_function("llama_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) -def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: - """Get the model's RoPE frequency scaling factor""" +# LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); +@ctypes_function("llama_model_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) +def llama_model_rope_freq_scale_train(model: llama_model_p, /) -> float: + ... + + +# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); +@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_vocab_type(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab); +@ctypes_function("llama_vocab_n_tokens", [llama_vocab_p_ctypes], ctypes.c_int32) +def llama_vocab_n_tokens(vocab: llama_vocab_p, /) -> int: ... @@ -1402,6 +1482,14 @@ def llama_model_size(model: llama_model_p, /) -> int: ... +# // Get the default chat template. Returns nullptr if not available +# LLAMA_API const char * llama_model_chat_template(const struct llama_model * model); +@ctypes_function("llama_model_chat_template", [llama_model_p_ctypes], ctypes.c_char_p) +def llama_model_chat_template(model: llama_model_p, /) -> Optional[bytes]: + """Get the default chat template. Returns None if not available""" + ... + + # // Returns the total number of parameters in the model # LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); @ctypes_function("llama_model_n_params", [llama_model_p_ctypes], ctypes.c_uint64) @@ -1472,37 +1560,48 @@ def llama_model_quantize( # // Load a LoRA adapter from file -# // The loaded adapter will be associated to the given model, and will be free when the model is deleted -# LLAMA_API struct llama_lora_adapter * llama_lora_adapter_init( +# LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init( # struct llama_model * model, # const char * path_lora); @ctypes_function( - "llama_lora_adapter_init", + "llama_adapter_lora_init", [llama_model_p_ctypes, ctypes.c_char_p], - llama_lora_adapter_p_ctypes, + llama_adapter_lora_p_ctypes, ) -def llama_lora_adapter_init( +def llama_adapter_lora_init( model: llama_model_p, path_lora: bytes, / -) -> Optional[llama_lora_adapter_p]: - """Load a LoRA adapter from file - The loaded adapter will be associated to the given model, and will be free when the model is deleted - """ +) -> Optional[llama_adapter_lora_p]: + ... + + +# // Manually free a LoRA adapter +# // Note: loaded adapters will be free when the associated model is deleted +# LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter); +@ctypes_function( + "llama_adapter_lora_free", + [llama_adapter_lora_p_ctypes], + None, +) +def llama_adapter_lora_free(adapter: llama_adapter_lora_p, /): ... +# // The following functions operate on a llama_context, hence the naming: llama_verb_... + + # // Add a loaded LoRA adapter to given context # // This will not modify model's weight -# LLAMA_API int32_t llama_lora_adapter_set( +# LLAMA_API int32_t llama_set_adapter_lora( # struct llama_context * ctx, -# struct llama_lora_adapter * adapter, +# struct llama_adapter_lora * adapter, # float scale); @ctypes_function( - "llama_lora_adapter_set", - [llama_context_p_ctypes, llama_lora_adapter_p_ctypes, ctypes.c_float], + "llama_set_adapter_lora", + [llama_context_p_ctypes, llama_adapter_lora_p_ctypes, ctypes.c_float], ctypes.c_int32, ) -def llama_lora_adapter_set( - ctx: llama_context_p, adapter: llama_lora_adapter_p, scale: float, / +def llama_set_adapter_lora( + ctx: llama_context_p, adapter: llama_adapter_lora_p, scale: float, / ) -> int: """Add a loaded LoRA adapter to given context This will not modify model's weight""" @@ -1511,64 +1610,49 @@ def llama_lora_adapter_set( # // Remove a specific LoRA adapter from given context # // Return -1 if the adapter is not present in the context -# LLAMA_API int32_t llama_lora_adapter_remove( +# LLAMA_API int32_t llama_rm_adapter_lora( # struct llama_context * ctx, -# struct llama_lora_adapter * adapter); +# struct llama_adapter_lora * adapter); @ctypes_function( - "llama_lora_adapter_remove", - [llama_context_p_ctypes, llama_lora_adapter_p_ctypes], + "llama_rm_adapter_lora", + [llama_context_p_ctypes, llama_adapter_lora_p_ctypes], ctypes.c_int32, ) -def llama_lora_adapter_remove( - ctx: llama_context_p, adapter: llama_lora_adapter_p, / +def llama_rm_adapter_lora( + ctx: llama_context_p, adapter: llama_adapter_lora_p, / ) -> int: - """Remove a LoRA adapter from given context + """Remove a specific LoRA adapter from given context Return -1 if the adapter is not present in the context""" ... # // Remove all LoRA adapters from given context -# LLAMA_API void llama_lora_adapter_clear( -# struct llama_context * ctx); +# LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx); @ctypes_function( - "llama_lora_adapter_clear", + "llama_clear_adapter_lora", [llama_context_p_ctypes], None, ) -def llama_lora_adapter_clear(ctx: llama_context_p, /): +def llama_clear_adapter_lora(ctx: llama_context_p, /): """Remove all LoRA adapters from given context""" ... -# // Manually free a LoRA adapter -# // Note: loaded adapters will be free when the associated model is deleted -# LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter); -@ctypes_function( - "llama_lora_adapter_free", - [llama_lora_adapter_p_ctypes], - None, -) -def llama_lora_adapter_free(adapter: llama_lora_adapter_p, /): - """Manually free a LoRA adapter - Note: loaded adapters will be free when the associated model is deleted""" - ... - - # // Apply a loaded control vector to a llama_context, or if data is NULL, clear # // the currently loaded vector. # // n_embd should be the size of a single layer's control, and data should point # // to an n_embd x n_layers buffer starting from layer 1. # // il_start and il_end are the layer range the vector should apply to (both inclusive) # // See llama_control_vector_load in common to load a control vector. -# LLAMA_API int32_t llama_control_vector_apply( -# struct llama_context * lctx, +# LLAMA_API int32_t llama_apply_adapter_cvec( +# struct llama_context * ctx, # const float * data, # size_t len, # int32_t n_embd, # int32_t il_start, # int32_t il_end); @ctypes_function( - "llama_control_vector_apply", + "llama_apply_adapter_cvec", [ llama_context_p_ctypes, ctypes.POINTER(ctypes.c_float), @@ -1579,8 +1663,8 @@ def llama_lora_adapter_free(adapter: llama_lora_adapter_p, /): ], ctypes.c_int32, ) -def llama_control_vector_apply( - lctx: llama_context_p, +def llama_apply_adapter_cvec( + ctx: llama_context_p, data: CtypesPointerOrRef[ctypes.c_float], len: int, n_embd: int, @@ -2577,53 +2661,53 @@ def llama_get_embeddings_seq( # // -# LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); +# LLAMA_API const char * llama_vocab_get_text(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_get_text", [llama_model_p_ctypes, llama_token], ctypes.c_char_p + "llama_vocab_get_text", [llama_vocab_p_ctypes, llama_token], ctypes.c_char_p ) -def llama_token_get_text( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_get_text( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> bytes: ... -# LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); +# LLAMA_API float llama_vocab_get_score(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_get_score", [llama_model_p_ctypes, llama_token], ctypes.c_float + "llama_vocab_get_score", [llama_vocab_p_ctypes, llama_token], ctypes.c_float ) -def llama_token_get_score( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_get_score( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> float: ... -# LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token); +# LLAMA_API enum llama_token_attr llama_vocab_get_attr(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_get_attr", [llama_model_p_ctypes, llama_token], ctypes.c_int + "llama_vocab_get_attr", [llama_vocab_p_ctypes, llama_token], ctypes.c_int ) -def llama_token_get_attr( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_get_attr( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> int: ... # // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) -# LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token); +# LLAMA_API bool llama_vocab_is_eog(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_is_eog", [llama_model_p_ctypes, llama_token], ctypes.c_bool + "llama_vocab_is_eog", [llama_vocab_p_ctypes, llama_token], ctypes.c_bool ) -def llama_token_is_eog(model: llama_model_p, token: Union[llama_token, int], /) -> bool: +def llama_vocab_is_eog(vocab: llama_vocab_p, token: Union[llama_token, int], /) -> bool: """Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)""" ... # // Identify if Token Id is a control token or a render-able token -# LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token); +# LLAMA_API bool llama_vocab_is_control(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_is_control", [llama_model_p_ctypes, llama_token], ctypes.c_bool + "llama_vocab_is_control", [llama_vocab_p_ctypes, llama_token], ctypes.c_bool ) -def llama_token_is_control( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_is_control( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> bool: """Identify if Token Id is a control token or a render-able token""" ... @@ -2632,110 +2716,335 @@ def llama_token_is_control( # // Special tokens -# LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence -@ctypes_function("llama_token_bos", [llama_model_p_ctypes], llama_token) -def llama_token_bos(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_bos(const struct llama_vocab * vocab); // beginning-of-sentence +@ctypes_function("llama_vocab_bos", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_bos(vocab: llama_vocab_p, /) -> llama_token: """beginning-of-sentence""" ... -# LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence -@ctypes_function("llama_token_eos", [llama_model_p_ctypes], llama_token) -def llama_token_eos(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_eos(const struct llama_vocab * vocab); // end-of-sentence +@ctypes_function("llama_vocab_eos", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_eos(vocab: llama_vocab_p, /) -> llama_token: """end-of-sentence""" ... -# LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn -@ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_eot(const struct llama_vocab * vocab); // end-of-turn +@ctypes_function("llama_vocab_eot", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_eot(vocab: llama_vocab_p, /) -> llama_token: """end-of-turn""" ... -# LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification -@ctypes_function("llama_token_cls", [llama_model_p_ctypes], llama_token) -def llama_token_cls(model: llama_model_p, /) -> int: - """classification""" +# LLAMA_API llama_token llama_vocab_sep(const struct llama_vocab * vocab); // sentence separator +@ctypes_function("llama_vocab_sep", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_sep(vocab: llama_vocab_p, /) -> llama_token: + """sentence separator""" ... -# LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator -@ctypes_function("llama_token_sep", [llama_model_p_ctypes], llama_token) -def llama_token_sep(model: llama_model_p, /) -> int: - """sentence separator""" +# LLAMA_API llama_token llama_vocab_nl (const struct llama_vocab * vocab); // next-line +@ctypes_function("llama_vocab_nl", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_nl(vocab: llama_vocab_p, /) -> llama_token: + """next-line""" ... -# LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line -@ctypes_function("llama_token_nl", [llama_model_p_ctypes], llama_token) -def llama_token_nl(model: llama_model_p, /) -> int: - """next-line""" +# LLAMA_API llama_token llama_vocab_pad(const struct llama_vocab * vocab); // padding +@ctypes_function("llama_vocab_pad", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_pad(vocab: llama_vocab_p, /) -> llama_token: + """padding""" + ... + +# LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_bos", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_bos(vocab: llama_vocab_p, /) -> bool: ... -# LLAMA_API bool llama_add_bos_token(const struct llama_model * model); -@ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_bos_token(model: llama_model_p, /) -> bool: +# LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_eos", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_eos(vocab: llama_vocab_p, /) -> bool: ... -# LLAMA_API bool llama_add_eos_token(const struct llama_model * model); -@ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_eos_token(model: llama_model_p, /) -> bool: +# LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_pre", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_pre(vocab: llama_vocab_p, /) -> llama_token: ... -# // Codellama infill tokens -# DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead"); -@ctypes_function("llama_token_prefix", [llama_model_p_ctypes], llama_token) -def llama_token_prefix(model: llama_model_p) -> int: - """codellama infill tokens""" +# LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_suf", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_suf(vocab: llama_vocab_p, /) -> llama_token: ... -# DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead"); -@ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_fim_mid(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_mid", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_mid(vocab: llama_vocab_p, /) -> llama_token: ... -# DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead"); -@ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_fim_pad(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_pad(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model); -@ctypes_function("llama_token_fim_pre", [llama_model_p_ctypes], llama_token) -def llama_token_fim_pre(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_fim_rep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_rep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_rep(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model); -@ctypes_function("llama_token_fim_suf", [llama_model_p_ctypes], llama_token) -def llama_token_fim_suf(model: llama_model_p, /) -> int: + +# LLAMA_API llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_sep(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model); -@ctypes_function("llama_token_fim_mid", [llama_model_p_ctypes], llama_token) -def llama_token_fim_mid(model: llama_model_p, /) -> int: + + +# DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_text instead"); +@ctypes_function( + "llama_token_get_text", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_char_p, +) +def llama_token_get_text( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bytes: + ... + + +# DEPRECATED(LLAMA_API float llama_token_get_score(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_score instead"); +@ctypes_function( + "llama_token_get_score", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_float, +) +def llama_token_get_score( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> float: + ... + +# DEPRECATED(LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_attr instead"); +@ctypes_function( + "llama_token_get_attr", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_int, +) +def llama_token_get_attr( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> int: + ... + +# DEPRECATED(LLAMA_API bool llama_token_is_eog(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_eog instead"); +@ctypes_function( + "llama_token_is_eog", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_bool, +) +def llama_token_is_eog( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: + ... + +# DEPRECATED(LLAMA_API bool llama_token_is_control(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_control instead"); +@ctypes_function( + "llama_token_is_control", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_bool, +) +def llama_token_is_control( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_bos(const struct llama_vocab * vocab), "use llama_vocab_bos instead"); +@ctypes_function( + "llama_token_bos", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_bos(vocab: llama_vocab_p, /) -> int: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_eos(const struct llama_vocab * vocab), "use llama_vocab_eos instead"); +@ctypes_function( + "llama_token_eos", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_eos(vocab: llama_vocab_p, /) -> int: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_eot(const struct llama_vocab * vocab), "use llama_vocab_eot instead"); +@ctypes_function( + "llama_token_eot", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_eot(vocab: llama_vocab_p, /) -> int: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_cls(const struct llama_vocab * vocab), "use llama_vocab_cls instead"); +@ctypes_function( + "llama_token_cls", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_cls(vocab: llama_vocab_p, /) -> int: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_sep(const struct llama_vocab * vocab), "use llama_vocab_sep instead"); +@ctypes_function( + "llama_token_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_sep(vocab: llama_vocab_p, /) -> int: + ... + + +# DEPRECATED(LLAMA_API llama_token llama_token_nl (const struct llama_vocab * vocab), "use llama_vocab_nl instead"); +@ctypes_function( + "llama_token_nl", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_nl(vocab: llama_vocab_p, /) -> int: + ... + + +# DEPRECATED(LLAMA_API llama_token llama_token_pad(const struct llama_vocab * vocab), "use llama_vocab_pad instead"); +@ctypes_function( + "llama_token_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_pad(vocab: llama_vocab_p, /) -> int: + ... + + +# DEPRECATED(LLAMA_API bool llama_add_bos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_bos instead"); +@ctypes_function( + "llama_add_bos_token", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_add_bos_token(vocab: llama_vocab_p, /) -> bool: + ... + +# DEPRECATED(LLAMA_API bool llama_add_eos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_eos instead"); +@ctypes_function( + "llama_add_eos_token", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_add_eos_token(vocab: llama_vocab_p, /) -> bool: + ... + + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_pre(const struct llama_vocab * vocab), "use llama_vocab_fim_pre instead"); +@ctypes_function( + "llama_token_fim_pre", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_pre(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model); -@ctypes_function("llama_token_fim_pad", [llama_model_p_ctypes], llama_token) -def llama_token_fim_pad(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_fim_suf(const struct llama_vocab * vocab), "use llama_vocab_fim_suf instead"); +@ctypes_function( + "llama_token_fim_suf", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_suf(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model); -@ctypes_function("llama_token_fim_rep", [llama_model_p_ctypes], llama_token) -def llama_token_fim_rep(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_fim_mid(const struct llama_vocab * vocab), "use llama_vocab_fim_mid instead"); +@ctypes_function( + "llama_token_fim_mid", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_mid(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model); -@ctypes_function("llama_token_fim_sep", [llama_model_p_ctypes], llama_token) -def llama_token_fim_sep(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_fim_pad(const struct llama_vocab * vocab), "use llama_vocab_fim_pad instead"); +@ctypes_function( + "llama_token_fim_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_pad(vocab: llama_vocab_p, /) -> llama_token: ... +# DEPRECATED(LLAMA_API llama_token llama_token_fim_rep(const struct llama_vocab * vocab), "use llama_vocab_fim_rep instead"); +@ctypes_function( + "llama_token_fim_rep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_rep(vocab: llama_vocab_p, /) -> llama_token: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_sep(const struct llama_vocab * vocab), "use llama_vocab_fim_sep instead"); +@ctypes_function( + "llama_token_fim_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_sep(vocab: llama_vocab_p, /) -> llama_token: + ... + +# // CLS is equivalent to BOS +# DEPRECATED(LLAMA_API llama_token llama_vocab_cls(const struct llama_vocab * vocab), // classification +# "use llama_vocab_bos instead"); +@ctypes_function( + "llama_vocab_cls", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_cls(vocab: llama_vocab_p, /) -> llama_token: + ... + + # // # // Tokenization # // @@ -2751,7 +3060,7 @@ def llama_token_fim_sep(model: llama_model_p, /) -> int: # /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated # /// as plaintext. Does not insert a leading space. # LLAMA_API int32_t llama_tokenize( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # const char * text, # int32_t text_len, # llama_token * tokens, @@ -2761,7 +3070,7 @@ def llama_token_fim_sep(model: llama_model_p, /) -> int: @ctypes_function( "llama_tokenize", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, ctypes.c_char_p, ctypes.c_int32, llama_token_p, @@ -2772,7 +3081,7 @@ def llama_token_fim_sep(model: llama_model_p, /) -> int: ctypes.c_int32, ) def llama_tokenize( - model: llama_model_p, + vocab: llama_vocab_p, text: bytes, text_len: Union[ctypes.c_int, int], tokens: CtypesArray[llama_token], @@ -2784,7 +3093,7 @@ def llama_tokenize( """Convert the provided text into tokens. Args: - model: The model to use for tokenization. + vocab: The vocabulary to use for tokenization. text: The text to tokenize. text_len: The length of the text. tokens: The tokens pointer must be large enough to hold the resulting tokens. @@ -2805,7 +3114,7 @@ def llama_tokenize( # // User can skip up to 'lstrip' leading spaces before copying (useful when encoding/decoding multiple tokens with 'add_space_prefix') # // @param special If true, special tokens are rendered in the output. # LLAMA_API int32_t llama_token_to_piece( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # llama_token token, # char * buf, # int32_t length, @@ -2814,7 +3123,7 @@ def llama_tokenize( @ctypes_function( "llama_token_to_piece", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, llama_token, ctypes.c_char_p, ctypes.c_int32, @@ -2824,7 +3133,7 @@ def llama_tokenize( ctypes.c_int32, ) def llama_token_to_piece( - model: llama_model_p, + vocab: llama_vocab_p, token: Union[llama_token, int], buf: Union[ctypes.c_char_p, bytes, CtypesArray[ctypes.c_char]], length: Union[ctypes.c_int, int], @@ -2838,7 +3147,7 @@ def llama_token_to_piece( User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. Args: - model: The model to use for tokenization. + vocab: The vocabulary to use for tokenization. token: The token to convert. buf: The buffer to write the token to. length: The length of the buffer. @@ -2930,7 +3239,6 @@ def llama_detokenize( # /// @param length The size of the allocated buffer # /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template. # LLAMA_API int32_t llama_chat_apply_template( -# const struct llama_model * model, # const char * tmpl, # const struct llama_chat_message * chat, # size_t n_msg, @@ -2940,20 +3248,37 @@ def llama_detokenize( @ctypes_function( "llama_chat_apply_template", [ - ctypes.c_void_p, - ctypes.c_char_p, - ctypes.POINTER(llama_chat_message), - ctypes.c_size_t, + ctypes.c_char_p, # tmpl + ctypes.POINTER(llama_chat_message), # chat + ctypes.c_size_t, # n_msg + ctypes.c_bool, # add_ass (added) + ctypes.c_char_p, # buf + ctypes.c_int32, # length ], ctypes.c_int32, ) def llama_chat_apply_template( - model: llama_model_p, tmpl: bytes, chat: CtypesArray[llama_chat_message], n_msg: int, + add_ass: bool, # Added parameter + buf: bytes, + length: int, /, ) -> int: + """Apply chat template. + + Args: + tmpl: Template to use. If None, uses model's default + chat: Array of chat messages + n_msg: Number of messages + add_ass: Whether to end prompt with assistant token + buf: Output buffer + length: Buffer length + + Returns: + Number of bytes written, or needed if buffer too small + """ ... @@ -3022,7 +3347,6 @@ def llama_chat_builtin_templates( # // llama_sampler_free(smpl); # // # // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU). -# // TODO: in the future, the entire sampling API that uses llama_model should start using llama_vocab # // # typedef void * llama_sampler_context_t; @@ -3342,16 +3666,16 @@ def llama_sampler_init_mirostat_v2( # LLAMA_API struct llama_sampler * llama_sampler_init_grammar( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # const char * grammar_str, # const char * grammar_root); @ctypes_function( "llama_sampler_init_grammar", - [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_char_p], + [llama_vocab_p_ctypes, ctypes.c_char_p, ctypes.c_char_p], llama_sampler_p_ctypes, ) def llama_sampler_init_grammar( - model: llama_model_p, grammar_str: bytes, grammar_root: bytes, / + vocab: llama_vocab_p, grammar_str: bytes, grammar_root: bytes, / ) -> llama_sampler_p: ... @@ -3379,7 +3703,8 @@ def llama_sampler_init_penalties( # /// @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982 # LLAMA_API struct llama_sampler * llama_sampler_init_dry( -# const struct llama_model * model, +# const struct llama_vocab * vocab, +# int32_t n_ctx_train, # float dry_multiplier, # float dry_base, # int32_t dry_allowed_length, @@ -3389,7 +3714,8 @@ def llama_sampler_init_penalties( @ctypes_function( "llama_sampler_init_dry", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, + ctypes.c_int32, ctypes.c_float, ctypes.c_float, ctypes.c_int32, @@ -3400,7 +3726,8 @@ def llama_sampler_init_penalties( llama_sampler_p_ctypes, ) def llama_sampler_init_dry( - model: llama_model_p, + vocab: llama_vocab_p, + n_ctx_train: int, dry_multiplier: float, dry_base: float, dry_allowed_length: int, @@ -3448,15 +3775,13 @@ def llama_sampler_init_logit_bias( # // 3. discard non-EOG tokens with low prob # // 4. if no tokens are left -> pick EOT # // -# LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model); +# LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab); @ctypes_function( "llama_sampler_init_infill", - [llama_model_p_ctypes], + [llama_vocab_p_ctypes], llama_sampler_p_ctypes, ) -def llama_sampler_init_infill(model: llama_model_p, /) -> llama_sampler_p: - """This sampler is meant to be used for fill-in-the-middle infilling. - """ +def llama_sampler_init_infill(vocab: llama_vocab_p, /) -> llama_sampler_p: ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index f7cd13301..ef6dada60 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit f7cd13301c2a88f97073fd119072b4cc92c08df1 +Subproject commit ef6dada60ca710f4edfbb6fd9e1258685d8ea49d From 0b89fe48ad26ffcff76451bd87642d916a1a3385 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 28 Jan 2025 21:40:53 -0500 Subject: [PATCH 614/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 10 ++++++---- vendor/llama.cpp | 2 +- 2 files changed, 7 insertions(+), 5 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 7845f0d1c..2b9edbb99 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1483,10 +1483,12 @@ def llama_model_size(model: llama_model_p, /) -> int: # // Get the default chat template. Returns nullptr if not available -# LLAMA_API const char * llama_model_chat_template(const struct llama_model * model); -@ctypes_function("llama_model_chat_template", [llama_model_p_ctypes], ctypes.c_char_p) -def llama_model_chat_template(model: llama_model_p, /) -> Optional[bytes]: - """Get the default chat template. Returns None if not available""" +# // If name is NULL, returns the default chat template +# LLAMA_API const char * llama_model_chat_template(const struct llama_model * model, const char * name); +@ctypes_function("llama_model_chat_template", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_char_p) +def llama_model_chat_template(model: llama_model_p, name: Optional[bytes], /) -> Optional[bytes]: + """Get the default chat template. Returns None if not available + If name is None, returns the default chat template""" ... diff --git a/vendor/llama.cpp b/vendor/llama.cpp index ef6dada60..794fe23f2 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit ef6dada60ca710f4edfbb6fd9e1258685d8ea49d +Subproject commit 794fe23f29fb40104975c91fe19f23798f7c726e From 14879c7819eb023a6a272e0f18aca887ae605c0f Mon Sep 17 00:00:00 2001 From: oobabooga Date: Tue, 28 Jan 2025 23:43:29 -0300 Subject: [PATCH 615/624] fix(ci): Fix the CUDA workflow (#1894) --- .github/workflows/build-wheels-cuda.yaml | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 3d410148f..745b2e602 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -61,11 +61,9 @@ jobs: - name: Setup Mamba uses: conda-incubator/setup-miniconda@v3.1.0 with: - activate-environment: "build" + activate-environment: "llamacpp" python-version: ${{ matrix.pyver }} - miniforge-variant: Mambaforge miniforge-version: latest - use-mamba: true add-pip-as-python-dependency: true auto-activate-base: false From 4442ff8e0a09feafcf7c91045ebbe0ba6f026feb Mon Sep 17 00:00:00 2001 From: Shaka Huang Date: Wed, 29 Jan 2025 11:57:37 +0800 Subject: [PATCH 616/624] fix: error showing time spent in llama perf context print (#1898) * feat: Sync with llama.cpp Add `no_perf` field to `llama_context_params` to optionally disable performance timing measurements. * fix: Display performance metrics by default --------- Co-authored-by: Andrei --- llama_cpp/llama.py | 4 ++++ llama_cpp/llama_cpp.py | 3 +++ 2 files changed, 7 insertions(+) diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 11fc744a0..7e9a6af23 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -94,6 +94,7 @@ def __init__( offload_kqv: bool = True, flash_attn: bool = False, # Sampling Params + no_perf: bool = False, last_n_tokens_size: int = 64, # LoRA Params lora_base: Optional[str] = None, @@ -173,6 +174,7 @@ def __init__( embedding: Embedding mode only. offload_kqv: Offload K, Q, V to GPU. flash_attn: Use flash attention. + no_perf: Measure performance timings. last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque. lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model. lora_path: Path to a LoRA file to apply to the model. @@ -351,6 +353,7 @@ def __init__( if type_v is not None: self.context_params.type_v = type_v # Sampling Params + self.context_params.no_perf = no_perf self.last_n_tokens_size = last_n_tokens_size self.cache: Optional[BaseLlamaCache] = None @@ -2093,6 +2096,7 @@ def __getstate__(self): offload_kqv=self.context_params.offload_kqv, flash_attn=self.context_params.flash_attn, # Sampling Params + no_perf=self.context_params.no_perf, last_n_tokens_size=self.last_n_tokens_size, # LoRA Params lora_base=self.lora_base, diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 2b9edbb99..205d89a0b 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -780,6 +780,7 @@ class llama_context_params(ctypes.Structure): embeddings (bool): if true, extract embeddings (together with logits) offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU flash_attn (bool): whether to use flash attention + no_perf (bool): whether to measure performance timings abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback """ @@ -810,6 +811,7 @@ class llama_context_params(ctypes.Structure): embeddings: bool offload_kqv: bool flash_attn: bool + no_perf: bool abort_callback: Callable[[ctypes.c_void_p], bool] abort_callback_data: ctypes.c_void_p @@ -839,6 +841,7 @@ class llama_context_params(ctypes.Structure): ("embeddings", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), ("flash_attn", ctypes.c_bool), + ("no_perf", ctypes.c_bool), ("abort_callback", ggml_abort_callback), ("abort_callback_data", ctypes.c_void_p), ] From 710e19a81284e5af0d5db93cef7a9063b3e8534f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Tue, 28 Jan 2025 23:19:04 -0500 Subject: [PATCH 617/624] chore: Bump version --- CHANGELOG.md | 6 ++++++ llama_cpp/__init__.py | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index b4f49c904..53365e368 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.7] + +- feat: Update llama.cpp to ggerganov/llama.cpp@794fe23f29fb40104975c91fe19f23798f7c726e +- fix(ci): Fix the CUDA workflow by @oobabooga in #1894 +- fix: error showing time spent in llama perf context print, adds `no_perf` flag to `Llama` class by @shakalaca in #1898 + ## [0.3.6] - feat: Update llama.cpp to ggerganov/llama.cpp@f7cd13301c2a88f97073fd119072b4cc92c08df1 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 8c6118fb4..fc1fcbcf6 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.6" +__version__ = "0.3.7" From 344c106a6961b50f734b95f084fbf02057d4b475 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 12 Mar 2025 04:11:42 -0400 Subject: [PATCH 618/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 74 ++++++++++++++++++++++++++++++++++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 72 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 205d89a0b..f3985ad2f 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -227,6 +227,7 @@ # LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, # LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, # LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, +# LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -257,6 +258,7 @@ LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 +LLAMA_VOCAB_PRE_TYPE_GPT4O = 29 # // note: these values should be synchronized with ggml_rope @@ -1357,6 +1359,12 @@ def llama_model_n_head(model: llama_model_p, /) -> int: ... +# LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); +@ctypes_function("llama_model_n_head_kv", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head_kv(model: llama_model_p, /) -> int: + ... + + # // Get the model's RoPE frequency scaling factor # LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); @ctypes_function("llama_model_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) @@ -3375,8 +3383,8 @@ class llama_sampler_i(ctypes.Structure): # struct llama_sampler { -# struct llama_sampler_i * iface; -# llama_sampler_context_t ctx; +# const struct llama_sampler_i * iface; +# llama_sampler_context_t ctx; # }; class llama_sampler(ctypes.Structure): _fields_ = [ @@ -3410,6 +3418,18 @@ class llama_sampler(ctypes.Structure): # // mirror of llama_sampler_i: +# LLAMA_API struct llama_sampler * llama_sampler_init (const struct llama_sampler_i * iface, llama_sampler_context_t ctx); +@ctypes_function( + "llama_sampler_init", + [ctypes.POINTER(llama_sampler_i), llama_sampler_context_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init( + iface: ctypes.POINTER(llama_sampler_i), ctx: llama_sampler_context_t, / +) -> llama_sampler_p: + ... + + # LLAMA_API const char * llama_sampler_name (const struct llama_sampler * smpl); @ctypes_function( "llama_sampler_name", @@ -3627,6 +3647,17 @@ def llama_sampler_init_xtc( ... +# /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641 +# LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float n); +@ctypes_function( + "llama_sampler_init_top_n_sigma", + [ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_top_n_sigma(n: float, /) -> llama_sampler_p: + ... + + # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. # /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -3685,6 +3716,43 @@ def llama_sampler_init_grammar( ... +# /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639 +# /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group. +# /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included. +# LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns( +# const struct llama_vocab * vocab, +# const char * grammar_str, +# const char * grammar_root, +# const char ** trigger_patterns, +# size_t num_trigger_patterns, +# const llama_token * trigger_tokens, +# size_t num_trigger_tokens); +@ctypes_function( + "llama_sampler_init_grammar_lazy_patterns", + [ + llama_vocab_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ctypes.POINTER(llama_token), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar_lazy_patterns( + vocab: llama_vocab_p, + grammar_str: bytes, + grammar_root: bytes, + trigger_patterns: CtypesArray[bytes], + num_trigger_patterns: int, + trigger_tokens: CtypesArray[llama_token], + num_trigger_tokens: int, + /, +) -> llama_sampler_p: + ... + + # /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first. # LLAMA_API struct llama_sampler * llama_sampler_init_penalties( # int32_t penalty_last_n, // last n tokens to penalize (0 = disable penalty, -1 = context size) @@ -3737,7 +3805,7 @@ def llama_sampler_init_dry( dry_base: float, dry_allowed_length: int, dry_penalty_last_n: int, - seq_breakers: CtypesArray[bytes], + seq_breakers, num_breakers: int, /, ) -> llama_sampler_p: diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 794fe23f2..2c9f833d1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 794fe23f29fb40104975c91fe19f23798f7c726e +Subproject commit 2c9f833d17bb5b8ea89dec663b072b5420fc5438 From e232fae1a35c0bdaabf230a5d71780c0178e5d8f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 12 Mar 2025 04:44:18 -0400 Subject: [PATCH 619/624] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 2c9f833d1..7841fc723 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 2c9f833d17bb5b8ea89dec663b072b5420fc5438 +Subproject commit 7841fc723e059d1fd9640e5c0ef19050fcc7c698 From 37eb5f0a4c2a8706b89ead1406b1577c4602cdec Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 12 Mar 2025 05:30:21 -0400 Subject: [PATCH 620/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 53365e368..605370e7d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.8] + +- feat: Update llama.cpp to ggerganov/llama.cpp@7841fc723e059d1fd9640e5c0ef19050fcc7c698 + ## [0.3.7] - feat: Update llama.cpp to ggerganov/llama.cpp@794fe23f29fb40104975c91fe19f23798f7c726e diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index fc1fcbcf6..b1a8b9baa 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.7" +__version__ = "0.3.8" From 99f2ebfde18912adeb7f714b49c1ddb624df3087 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 11 Apr 2025 03:28:20 -0400 Subject: [PATCH 621/624] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 268 ++++++++++++++++++++++++++++++++++++++--- vendor/llama.cpp | 2 +- 2 files changed, 255 insertions(+), 15 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index f3985ad2f..710bd83c8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -165,6 +165,10 @@ # llama_sampler_p = NewType("llama_sampler_p", int) # llama_sampler_p_ctypes = ctypes.c_void_p +# struct llama_kv_cache; +llama_kv_cache_p = NewType("llama_kv_cache_p", int) +llama_kv_cache_p_ctypes = ctypes.c_void_p + # typedef int32_t llama_pos; llama_pos = ctypes.c_int32 # typedef int32_t llama_token; @@ -228,6 +232,9 @@ # LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, # LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, # LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, +# LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, +# LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, +# LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -259,6 +266,9 @@ LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 LLAMA_VOCAB_PRE_TYPE_GPT4O = 29 +LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30 +LLAMA_VOCAB_PRE_TYPE_TRILLION = 31 +LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32 # // note: these values should be synchronized with ggml_rope @@ -630,10 +640,19 @@ class llama_model_kv_override(ctypes.Structure): value: Union[int, float, bool, bytes] +# struct llama_model_tensor_buft_override { +# const char * pattern; +# ggml_backend_buffer_type_t buft; +# }; + + # struct llama_model_params { # // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) # ggml_backend_dev_t * devices; +# // NULL-terminated list of buffer types to use for tensors that match a pattern +# const struct llama_model_tensor_buft_override * tensor_buft_overrides; + # int32_t n_gpu_layers; // number of layers to store in VRAM # enum llama_split_mode split_mode; // how to split the model across multiple GPUs @@ -668,6 +687,8 @@ class llama_model_params(ctypes.Structure): """Parameters for llama_model Attributes: + devices (ctypes.Array[ggml_backend_dev_t]): NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) + tensor_buft_overrides (ctypes.Array[llama_model_tensor_buft_override]): NULL-terminated list of buffer types to use for tensors that match a pattern n_gpu_layers (int): number of layers to store in VRAM split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored @@ -681,6 +702,8 @@ class llama_model_params(ctypes.Structure): check_tensors (bool): validate model tensor data""" if TYPE_CHECKING: + devices: CtypesArray[ctypes.c_void_p] # NOTE: unused + tensor_buft_overrides: CtypesArray[llama_model_tensor_buft_override] # NOTE: unused n_gpu_layers: int split_mode: int main_gpu: int @@ -695,6 +718,7 @@ class llama_model_params(ctypes.Structure): _fields_ = [ ("devices", ctypes.c_void_p), # NOTE: unnused + ("tensor_buft_overrides", ctypes.c_void_p), # NOTE: unused ("n_gpu_layers", ctypes.c_int32), ("split_mode", ctypes.c_int), ("main_gpu", ctypes.c_int32), @@ -1317,7 +1341,18 @@ def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... -# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +# LLAMA_API struct llama_kv_cache * llama_get_kv_self ( struct llama_context * ctx); +@ctypes_function( + "llama_get_kv_self", + [llama_context_p_ctypes], + llama_kv_cache_p_ctypes, +) +def llama_get_kv_self(ctx: llama_context_p, /) -> Optional[llama_kv_cache_p]: + """Get the KV cache for self-attention""" + ... + + +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) def llama_pooling_type(ctx: llama_context_p, /) -> int: ... @@ -1810,7 +1845,19 @@ def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[ll # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times -# LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); +# LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx); +@ctypes_function( + "llama_kv_self_n_tokens", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_n_tokens(ctx: llama_context_p, /) -> int: + """Returns the number of tokens in the KV cache (slow, use only for debug) + If a KV cell has multiple sequences assigned to it, it will be counted multiple times + """ + ... + + +# DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx), +# "use llama_kv_self_n_tokens instead"); @ctypes_function( "llama_get_kv_cache_token_count", [llama_context_p_ctypes], ctypes.c_int32 ) @@ -1822,7 +1869,17 @@ def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) -# LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); +# LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx); +@ctypes_function( + "llama_kv_self_used_cells", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_used_cells(ctx: llama_context_p, /) -> int: + """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" + ... + + +# DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx), +# "use llama_kv_self_used_cells instead"); @ctypes_function( "llama_get_kv_cache_used_cells", [llama_context_p_ctypes], ctypes.c_int32 ) @@ -1832,9 +1889,17 @@ def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: # // Clear the KV cache - both cell info is erased and KV data is zeroed -# LLAMA_API void llama_kv_cache_clear( +# LLAMA_API void llama_kv_self_clear( # struct llama_context * ctx); -@ctypes_function("llama_kv_cache_clear", [llama_context_p_ctypes], None) +@ctypes_function( + "llama_kv_self_clear", [llama_context_p_ctypes], None +) +def llama_kv_self_clear(ctx: llama_context_p, /): + """Clear the KV cache - both cell info is erased and KV data is zeroed""" + ... + +# NOTE: Deprecated +@ctypes_function("llama_kv_self_clear", [llama_context_p_ctypes], None) def llama_kv_cache_clear(ctx: llama_context_p, /): """Clear the KV cache""" ... @@ -1881,14 +1946,41 @@ def llama_kv_cache_seq_rm( # // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_cp( +# LLAMA_API void llama_kv_self_seq_cp( # struct llama_context * ctx, # llama_seq_id seq_id_src, # llama_seq_id seq_id_dst, # llama_pos p0, # llama_pos p1); @ctypes_function( - "llama_kv_cache_seq_cp", + "llama_kv_self_seq_cp", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_cp( + ctx: llama_context_p, + seq_id_src: Union[llama_seq_id, int], + seq_id_dst: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + /, +): + """Copy all tokens that belong to the specified sequence to another sequence + Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# NOTE: Deprecated +@ctypes_function( + "llama_kv_self_seq_cp", [ llama_context_p_ctypes, llama_seq_id, @@ -1914,17 +2006,27 @@ def llama_kv_cache_seq_cp( # // Removes all tokens that do not belong to the specified sequence -# LLAMA_API void llama_kv_cache_seq_keep( +# LLAMA_API void llama_kv_self_seq_keep( # struct llama_context * ctx, # llama_seq_id seq_id); @ctypes_function( - "llama_kv_cache_seq_keep", [llama_context_p_ctypes, llama_seq_id], None + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None +) +def llama_kv_self_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): + """Removes all tokens that do not belong to the specified sequence""" + ... + + +# NOTE: Deprecated +@ctypes_function( + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None ) def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): """Removes all tokens that do not belong to the specified sequence""" ... + # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) # // If the KV cache is RoPEd, the KV data is updated accordingly: # // - lazily on next llama_decode() @@ -1938,7 +2040,48 @@ def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, in # llama_pos p1, # llama_pos delta); @ctypes_function( - "llama_kv_cache_seq_add", + "llama_kv_self_seq_add", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_add( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + delta: Union[llama_pos, int], + /, +): + """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) + If the KV cache is RoPEd, the KV data is updated accordingly: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update() + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // NOTE: Deprecated +# // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_cache_update() +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_add( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# llama_pos delta); +@ctypes_function( + "llama_kv_self_seq_add", [ llama_context_p_ctypes, llama_seq_id, @@ -1976,7 +2119,44 @@ def llama_kv_cache_seq_add( # llama_pos p1, # int d); @ctypes_function( - "llama_kv_cache_seq_div", + "llama_kv_self_seq_div", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ctypes.c_int, + ], + None, +) +def llama_kv_self_seq_div( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + d: Union[ctypes.c_int, int], + /, +): + """Integer division of the positions by factor of `d > 1` + If the KV cache is RoPEd, the KV data is updated accordingly + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // NOTE: Deprecated +# // Integer division of the positions by factor of `d > 1` +# // If the KV cache is RoPEd, the KV data is updated accordingly +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_div( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d); +@ctypes_function( + "llama_kv_self_seq_div", [ llama_context_p_ctypes, llama_seq_id, @@ -2001,10 +2181,39 @@ def llama_kv_cache_seq_div( ... +# // Returns the largest position present in the KV cache for the specified sequence +# LLAMA_API llama_pos llama_kv_self_seq_pos_max( +# struct llama_context * ctx, +# llama_seq_id seq_id); +@ctypes_function( + "llama_kv_self_seq_pos_max", [llama_context_p_ctypes, llama_seq_id], llama_pos +) +def llama_kv_self_seq_pos_max( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the largest position present in the KV cache for the specified sequence""" + ... + + # // Defragment the KV cache # // This will be applied: # // - lazily on next llama_decode() -# // - explicitly with llama_kv_cache_update() +# // - explicitly with llama_kv_self_update() +# LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx); +@ctypes_function("llama_kv_self_defrag", [llama_context_p_ctypes], None) +def llama_kv_self_defrag(ctx: llama_context_p, /): + """Defragment the KV cache + This will be applied: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update()""" + ... + + +# NOTE: Deprecated +# // Defragment the KV cache +# // This will be applied: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_self_update() # LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); @ctypes_function("llama_kv_cache_defrag", [llama_context_p_ctypes], None) def llama_kv_cache_defrag(ctx: llama_context_p, /): @@ -2017,7 +2226,15 @@ def llama_kv_cache_defrag(ctx: llama_context_p, /): # // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) # LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_update", [llama_context_p_ctypes], None) +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) +def llama_kv_self_update(ctx: llama_context_p, /): + """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" + ... + +# // NOTE: Deprecated +# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) +# LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) def llama_kv_cache_update(ctx: llama_context_p, /): """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" ... @@ -2025,7 +2242,16 @@ def llama_kv_cache_update(ctx: llama_context_p, /): # // Check if the context supports KV cache shifting # LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +def llama_kv_self_can_shift(ctx: llama_context_p, /) -> bool: + """Check if the context supports KV cache shifting""" + ... + + +# // NOTE: Deprecated +# // Check if the context supports KV cache shifting +# LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) def llama_kv_cache_can_shift(ctx: llama_context_p, /) -> bool: """Check if the context supports KV cache shifting""" ... @@ -2547,6 +2773,16 @@ def llama_set_causal_attn(ctx: llama_context_p, causal_attn: bool, /): ... +# // Set whether the model is in warmup mode or not +# // If true, all model tensors are activated during llama_decode() to load and cache their weights. +# LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup); +@ctypes_function("llama_set_warmup", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_warmup(ctx: llama_context_p, warmup: bool, /): + """Set whether the model is in warmup mode or not + If true, all model tensors are activated during llama_decode() to load and cache their weights.""" + ... + + # // Set abort callback # LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); @ctypes_function( @@ -3701,6 +3937,10 @@ def llama_sampler_init_mirostat_v2( ... +# /// @details Intializes a GBNF grammar, see grammars/README.md for details. +# /// @param vocab The vocabulary that this grammar will be used with. +# /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails. +# /// @param grammar_root The name of the start symbol for the grammar. # LLAMA_API struct llama_sampler * llama_sampler_init_grammar( # const struct llama_vocab * vocab, # const char * grammar_str, diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 7841fc723..6bf28f011 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 7841fc723e059d1fd9640e5c0ef19050fcc7c698 +Subproject commit 6bf28f0111ff9f21b3c1b1eace20c590281e7ba6 From 4c6514d365cb4f00910424c2feb8e45e68a870a6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 May 2025 06:52:52 -0400 Subject: [PATCH 622/624] feat: Update llama.cpp --- CMakeLists.txt | 2 +- llama_cpp/llama_cpp.py | 33 +++++++++++++++++++++------------ vendor/llama.cpp | 2 +- 3 files changed, 23 insertions(+), 14 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 64a0304a1..443d1ee53 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -143,7 +143,7 @@ if (LLAMA_BUILD) endif() # Building llava - add_subdirectory(vendor/llama.cpp/examples/llava) + add_subdirectory(vendor/llama.cpp/tools/mtmd) set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") if (WIN32) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 710bd83c8..63de3a93a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -235,6 +235,8 @@ # LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, # LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, # LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, +# LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, +# LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -252,7 +254,7 @@ LLAMA_VOCAB_PRE_TYPE_DBRX = 13 LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 LLAMA_VOCAB_PRE_TYPE_PORO = 15 -LLAMA_VOCAV_PRE_TYPE_CHATGLM3 = 16 +LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16 LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17 LLAMA_VOCAB_PRE_TYPE_VIKING = 18 LLAMA_VOCAB_PRE_TYPE_JAIS = 19 @@ -269,6 +271,8 @@ LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30 LLAMA_VOCAB_PRE_TYPE_TRILLION = 31 LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32 +LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33 +LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34 # // note: these values should be synchronized with ggml_rope @@ -891,17 +895,18 @@ class llama_context_params(ctypes.Structure): # // model quantization parameters # typedef struct llama_model_quantize_params { -# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() -# enum llama_ftype ftype; // quantize to this llama_ftype -# enum ggml_type output_tensor_type; // output tensor type -# enum ggml_type token_embedding_type; // token embeddings tensor type -# bool allow_requantize; // allow quantizing non-f32/f16 tensors -# bool quantize_output_tensor; // quantize output.weight -# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored -# bool pure; // quantize all tensors to the default type -# bool keep_split; // quantize to the same number of shards -# void * imatrix; // pointer to importance matrix data -# void * kv_overrides; // pointer to vector containing overrides +# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() +# enum llama_ftype ftype; // quantize to this llama_ftype +# enum ggml_type output_tensor_type; // output tensor type +# enum ggml_type token_embedding_type; // token embeddings tensor type +# bool allow_requantize; // allow quantizing non-f32/f16 tensors +# bool quantize_output_tensor; // quantize output.weight +# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored +# bool pure; // quantize all tensors to the default type +# bool keep_split; // quantize to the same number of shards +# void * imatrix; // pointer to importance matrix data +# void * kv_overrides; // pointer to vector containing overrides +# void * tensor_types; // pointer to vector containing tensor types # } llama_model_quantize_params; class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize @@ -918,6 +923,7 @@ class llama_model_quantize_params(ctypes.Structure): keep_split (bool): quantize to the same number of shards imatrix (ctypes.c_void_p): pointer to importance matrix data kv_overrides (ctypes.c_void_p): pointer to vector containing overrides + tensor_types (ctypes.c_void_p): pointer to vector containing tensor types """ if TYPE_CHECKING: @@ -932,6 +938,7 @@ class llama_model_quantize_params(ctypes.Structure): keep_split: bool imatrix: ctypes.c_void_p kv_overrides: ctypes.c_void_p + tensor_types: ctypes.c_void_p _fields_ = [ ("nthread", ctypes.c_int32), @@ -945,6 +952,7 @@ class llama_model_quantize_params(ctypes.Structure): ("keep_split", ctypes.c_bool), ("imatrix", ctypes.c_void_p), ("kv_overrides", ctypes.c_void_p), + ("tensor_types", ctypes.c_void_p), ] @@ -3812,6 +3820,7 @@ def llama_sampler_init_softmax() -> llama_sampler_p: # /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +# /// Setting k <= 0 makes this a noop # LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); @ctypes_function("llama_sampler_init_top_k", [ctypes.c_int32], llama_sampler_p_ctypes) def llama_sampler_init_top_k(k: int) -> llama_sampler_p: diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6bf28f011..8733e0cf6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6bf28f0111ff9f21b3c1b1eace20c590281e7ba6 +Subproject commit 8733e0cf6eefc7c7752297cc22d0836706f4222c From cb2edb9c9fef90fe680e528fadc86806c6f385e5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 May 2025 06:59:55 -0400 Subject: [PATCH 623/624] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 605370e7d..affbd5db7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.9] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8733e0cf6eefc7c7752297cc22d0836706f4222c + ## [0.3.8] - feat: Update llama.cpp to ggerganov/llama.cpp@7841fc723e059d1fd9640e5c0ef19050fcc7c698 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index b1a8b9baa..2c9c527cd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.8" +__version__ = "0.3.9" From b1d23df0bbd327b774083b5cf88e67ca0dd52b92 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 May 2025 07:02:20 -0400 Subject: [PATCH 624/624] hotfix: Disable curl support --- CMakeLists.txt | 3 +++ 1 file changed, 3 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 443d1ee53..b9178e856 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -62,6 +62,9 @@ if (LLAMA_BUILD) # Enable building of the common library set(LLAMA_BUILD_COMMON ON CACHE BOOL "Build llama.cpp common library" FORCE) + # Disable building curl support + set(LLAMA_CURL OFF CACHE BOOL "llama.cpp: enable curl" FORCE) + # Architecture detection and settings for Apple platforms if (APPLE) # Get the target architecture