- rtp-llm is a Large Language Model (LLM) inference acceleration engine developed by Alibaba's Foundation Model Inference Team. It is widely used within Alibaba Group, supporting LLM service across multiple business units including Taobao, Tmall, Idlefish, Cainiao, Amap, Ele.me, AE, and Lazada.
- The rtp-llm project is a sub-project of the havenask .
Applied in numerous LLM scenarios, such as:
- Taobao Wenwen
- Alibaba's international AI platform, Aidge
- OpenSearch LLM Smart Q&A Edition
- Large Language Model based Long-tail Query Rewriting in Taobao Search
- Utilizes high-performance CUDA kernels, including PagedAttention, FlashAttention, FlashDecoding, etc.
- Implements WeightOnly INT8 Quantization with automatic quantization at load time
- Adaptive KVCache Quantization
- Detailed optimization of dynamic batching overhead at the framework level
- Specially optimized for the V100 GPU
- Seamless integration with the HuggingFace models, supporting multiple weight formats such as SafeTensors, Pytorch, and Megatron
- Deploys multiple LoRA services with a single model instance
- Handles multimodal inputs (combining images and text)
- Enables multi-machine/multi-GPU tensor parallelism
- Supports P-tuning models
- Loads pruned irregular models
- Contextual Prefix Cache for multi-turn dialogues
- System Prompt Cache
- Speculative Decoding
- Medusa for advanced parallelization strategies
- Operating System: Linux
- Python: 3.10
- NVIDIA GPU: Compute Capability 7.0 or higher (e.g., RTX20xx, RTX30xx, RTX40xx, V100, T4, A10/A30/A100, L4, H100, etc.)
# Install rtp-llm
cd rtp-llm
# For cuda12 environment, please use requirements_torch_gpu_cuda12.txt
pip3 install -r ./open_source/deps/requirements_torch_gpu.txt
# Use the corresponding whl from the release version, here's an example for the cuda11 version 0.1.0, for the cuda12 whl package please check the release page.
pip3 install maga_transformer-0.0.1+cuda118-cp310-cp310-manylinux1_x86_64.whl
# start http service
TOKENIZER_PATH=/path/to/tokenizer CHECKPOINT_PATH=/path/to/model MODEL_TYPE=your_model_type FT_SERVER_TEST=1 python3 -m maga_transformer.start_server
# request to server
curl -XPOST http://localhost:8088 -d '{"prompt": "hello, what is your name", "generate_config": {"max_new_tokens": 1000}}'- Test in Deploy Docker
- Serving Example
- RWKV-Runner Example
- Python Library Example
- Using RTP-LLm in Aliyun Ecs
- Configuration Parameters
- Source Code Build
- Request Format
- Multi GPU Inference
- LoRA
- PTuning
- SystemPrompt
- ReuseKVCache
- Multimodal
- Structured Pruning
- Speculative Sampling
- Roadmap
- Contributing
- Benchmark&Performance
Our project is mainly based on FasterTransformer, and on this basis, we have integrated some kernel implementations from TensorRT-LLM. FasterTransformer and TensorRT-LLM have provided us with reliable performance guarantees. Flash-Attention2 and cutlass have also provided a lot of help in our continuous performance optimization process. Our continuous batching and increment decoding draw on the implementation of vllm; sampling draws on transformers, with speculative sampling integrating Medusa's implementation, and the multimodal part integrating implementations from llava and qwen-vl. We thank these projects for their inspiration and help.
- Taobao Wenda
- Alibaba's International AI Platform Aidge
- OpenSearch LLM Smart Q&A Edition
- Large Language Model based Long-tail Query Rewriting in Taobao Search
- Aquila and Aquila2 (BAAI/AquilaChat2-7B, BAAI/AquilaChat2-34B, BAAI/Aquila-7B, BAAI/AquilaChat-7B, etc.)
- Baichuan and Baichuan2 (baichuan-inc/Baichuan2-13B-Chat, baichuan-inc/Baichuan-7B)
- Bloom (bigscience/bloom, bigscience/bloomz)
- ChatGlm (THUDM/chatglm2-6b, THUDM/chatglm3-6b)
- Falcon (tiiuae/falcon-7b, tiiuae/falcon-40b, tiiuae/falcon-rw-7b, etc.)
- GptNeox (EleutherAI/gpt-neox-20b)
- GPT BigCode (bigcode/starcoder, bigcode/starcoder2)
- LLaMA and LLaMA-2 (meta-llama/Llama-2-7b, meta-llama/Llama-2-13b-hf, meta-llama/Llama-2-70b-hf, lmsys/vicuna-33b-v1.3, 01-ai/Yi-34B, xverse/XVERSE-13B, etc.)
- MPT (mosaicml/mpt-30b-chat, etc.)
- Phi (microsoft/phi-1_5, etc.)
- Qwen (Qwen/Qwen-7B, Qwen/Qwen-7B-Chat, Qwen/Qwen-14B, Qwen/Qwen-14B-Chat, etc.)
- InternLM (internlm/internlm-7b, internlm/internlm-chat-7b, etc.)
- Gemma (google/gemma-it, etc)
- Mixtral (mistralai/Mixtral-8x7B-v0.1, etc)
- LLAVA (liuhaotian/llava-v1.5-13b, liuhaotian/llava-v1.5-7b)
- Qwen-VL (Qwen/Qwen-VL)