Thanks to visit codestin.com
Credit goes to github.com

Skip to content

add inplace conversion support for f8_e4m3 #359

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 24, 2024
Merged

add inplace conversion support for f8_e4m3 #359

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
66 changes: 65 additions & 1 deletion model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -554,13 +554,62 @@ float bf16_to_f32(uint16_t bfloat16) {
return *reinterpret_cast<float*>(&val_bits);
}

uint16_t f8_e4m3_to_f16(uint8_t f8) {
// do we need to support uz?

const uint32_t exponent_bias = 7;
if (f8 == 0xff) {
return ggml_fp32_to_fp16(-NAN);
} else if (f8 == 0x7f) {
return ggml_fp32_to_fp16(NAN);
}

uint32_t sign = f8 & 0x80;
uint32_t exponent = (f8 & 0x78) >> 3;
uint32_t mantissa = f8 & 0x07;
uint32_t result = sign << 24;
if (exponent == 0) {
if (mantissa > 0) {
exponent = 0x7f - exponent_bias;

// yes, 2 times
if ((mantissa & 0x04) == 0) {
mantissa &= 0x03;
mantissa <<= 1;
exponent -= 1;
}
if ((mantissa & 0x04) == 0) {
mantissa &= 0x03;
mantissa <<= 1;
exponent -= 1;
}

result |= (mantissa & 0x03) << 21;
result |= exponent << 23;
}
} else {
result |= mantissa << 20;
exponent += 0x7f - exponent_bias;
result |= exponent << 23;
}

return ggml_fp32_to_fp16(*reinterpret_cast<const float*>(&result));
}

void bf16_to_f32_vec(uint16_t* src, float* dst, int64_t n) {
// support inplace op
for (int64_t i = n - 1; i >= 0; i--) {
dst[i] = bf16_to_f32(src[i]);
}
}

void f8_e4m3_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
// support inplace op
for (int64_t i = n - 1; i >= 0; i--) {
dst[i] = f8_e4m3_to_f16(src[i]);
}
}

void convert_tensor(void* src,
ggml_type src_type,
void* dst,
Expand Down Expand Up @@ -794,6 +843,8 @@ ggml_type str_to_ggml_type(const std::string& dtype) {
ttype = GGML_TYPE_F32;
} else if (dtype == "F32") {
ttype = GGML_TYPE_F32;
} else if (dtype == "F8_E4M3") {
ttype = GGML_TYPE_F16;
}
return ttype;
}
Expand Down Expand Up @@ -866,7 +917,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const

ggml_type type = str_to_ggml_type(dtype);
if (type == GGML_TYPE_COUNT) {
LOG_ERROR("unsupported dtype '%s'", dtype.c_str());
LOG_ERROR("unsupported dtype '%s' (tensor '%s')", dtype.c_str(), name.c_str());
return false;
}

Expand Down Expand Up @@ -903,6 +954,10 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
if (dtype == "BF16") {
tensor_storage.is_bf16 = true;
GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
} else if (dtype == "F8_E4M3") {
tensor_storage.is_f8_e4m3 = true;
// f8 -> f16
GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
} else {
GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);
}
Expand Down Expand Up @@ -1537,6 +1592,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
if (tensor_storage.is_bf16) {
// inplace op
bf16_to_f32_vec((uint16_t*)dst_tensor->data, (float*)dst_tensor->data, tensor_storage.nelements());
} else if (tensor_storage.is_f8_e4m3) {
// inplace op
f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
}
} else {
read_buffer.resize(tensor_storage.nbytes());
Expand All @@ -1545,6 +1603,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
if (tensor_storage.is_bf16) {
// inplace op
bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
} else if (tensor_storage.is_f8_e4m3) {
// inplace op
f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
}

convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
Expand All @@ -1557,6 +1618,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
if (tensor_storage.is_bf16) {
// inplace op
bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
} else if (tensor_storage.is_f8_e4m3) {
// inplace op
f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
}

if (tensor_storage.type == dst_tensor->type) {
Expand Down
9 changes: 7 additions & 2 deletions model.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@ struct TensorStorage {
std::string name;
ggml_type type = GGML_TYPE_F32;
bool is_bf16 = false;
bool is_f8_e4m3 = false;
int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
int n_dims = 0;

Expand Down Expand Up @@ -61,7 +62,7 @@ struct TensorStorage {
}

int64_t nbytes_to_read() const {
if (is_bf16) {
if (is_bf16 || is_f8_e4m3) {
return nbytes() / 2;
} else {
return nbytes();
Expand Down Expand Up @@ -109,6 +110,8 @@ struct TensorStorage {
const char* type_name = ggml_type_name(type);
if (is_bf16) {
type_name = "bf16";
} else if (is_f8_e4m3) {
type_name = "f8_e4m3";
}
ss << name << " | " << type_name << " | ";
ss << n_dims << " [";
Expand Down Expand Up @@ -160,4 +163,6 @@ class ModelLoader {
static std::string load_merges();
static std::string load_t5_tokenizer_json();
};
#endif // __MODEL_H__

#endif // __MODEL_H__

Loading