Skip to content

Latest commit

 

History

History
174 lines (129 loc) · 5.87 KB

README.md

File metadata and controls

174 lines (129 loc) · 5.87 KB
Raw

LangPort

GitHub Repo stars License

architecture

LangPort is a open-source large language model serving platform. Our goal is to build a super fast LLM inference service.

This project is inspired by lmsys/fastchat, we hope that the serving platform is lightweight and fast, but fastchat includes other features such as training and evaluation make it complicated.

The core features include:

  • Huggingface transformers support.
  • ggml (llama.cpp) support.
  • A distributed serving system for state-of-the-art models.
  • Streaming generation support with various decoding strategies.
  • Batch inference for higher throughput.
  • Support for encoder-only, decoder-only and encoder-decoder models.
  • OpenAI-compatible RESTful APIs.
  • FauxPilot-compatible RESTful APIs.
  • HuggingFace-compatible RESTful APIs.
  • Tabby-compatible RESTful APIs.

Support Model Architectures

  • LLaMa, LLaMa2, GLM, Bloom, OPT, GPT2, GPT Neo, GPT Big Code and so on.

Tested Models

  • NingYu, LLaMa, LLaMa2, Vicuna, ChatGLM, ChatGLM2, Falcon, Starcoder, WizardLM, InternLM, OpenBuddy, FireFly, CodeGen, Phoenix, RWKV, StableLM and so on.

News

  • [2024/01/13] Introduce the ChatProto.
  • [2023/08/04] Dynamic batch inference.
  • [2023/07/16] Support int4 quantization.
  • [2023/07/13] Support generation logprobs parameter.
  • [2023/06/18] Add ggml (llama.cpp gpt.cpp starcoder.cpp etc.) worker support.
  • [2023/06/09] Add LLama.cpp worker support.
  • [2023/06/01] Add HuggingFace Bert embedding worker support.
  • [2023/06/01] Add HuggingFace text generation API support.
  • [2023/06/01] Add tabby API support.
  • [2023/05/23] Add chat throughput test script.
  • [2023/05/22] New distributed architecture.
  • [2023/05/14] Batch inference supported.
  • [2023/05/10] Langport project started.

Install

Method 1: With pip

pip install langport

or:

pip install git+https://github.com/vtuber-plan/langport.git

If you need ggml generation worker, use this command:

pip install langport[ggml]

If you want to use GPU:

CT_CUBLAS=1 pip install langport[ggml]

Method 2: From source

  1. Clone this repository
git clone https://github.com/vtuber-plan/langport.git
cd langport
  1. Install the Package
pip install --upgrade pip
pip install -e .

Quick start

It is simple to start a local chat API service:

First, start a worker process in the terminal:

python -m langport.service.server.generation_worker --port 21001 --model-path <your model path>

Then, start a API service in another terminal:

python -m langport.service.gateway.openai_api

Now, you can use the inference API by openai protocol.

Start the server

It is simple to start a single node chat API service:

python -m langport.service.server.generation_worker --port 21001 --model-path <your model path>
python -m langport.service.gateway.openai_api

If you need a single node embeddings API server:

python -m langport.service.server.embedding_worker --port 21002 --model-path bert-base-chinese --gpus 0 --num-gpus 1
python -m langport.service.gateway.openai_api --port 8000 --controller-address http://localhost:21002

If you need the embeddings API or other features, you can deploy a distributed inference cluster:

python -m langport.service.server.dummy_worker --port 21001
python -m langport.service.server.generation_worker --model-path <your model path> --neighbors http://localhost:21001
python -m langport.service.server.embedding_worker --model-path <your model path> --neighbors http://localhost:21001
python -m langport.service.gateway.openai_api --controller-address http://localhost:21001

In practice, the gateway can connect to any node to distribute inference tasks:

python -m langport.service.server.dummy_worker --port 21001
python -m langport.service.server.generation_worker --port 21002 --model-path <your model path> --neighbors http://localhost:21001
python -m langport.service.server.generation_worker --port 21003 --model-path <your model path> --neighbors http://localhost:21001 http://localhost:21002
python -m langport.service.server.generation_worker --port 21004 --model-path <your model path> --neighbors http://localhost:21001 http://localhost:21003
python -m langport.service.server.generation_worker --port 21005 --model-path <your model path> --neighbors http://localhost:21001 http://localhost:21004
python -m langport.service.gateway.openai_api --controller-address http://localhost:21003 # 21003 is OK!
python -m langport.service.gateway.openai_api --controller-address http://localhost:21002 # Any worker is also OK!

Run text generation with multi GPUs:

python -m langport.service.server.generation_worker --port 21001 --model-path <your model path> --gpus 0,1 --num-gpus 2
python -m langport.service.gateway.openai_api

Run text generation with ggml worker:

python -m langport.service.server.ggml_generation_worker --port 21001 --model-path <your model path> --gpu-layers <num layer to gpu (resize this for your VRAM)>

Run OpenAI forward server:

python -m langport.service.server.chatgpt_generation_worker --port 21001 --api-url <url> --api-key <key>

License

langport is released under the Apache Software License.

See also

Star History

Star History Chart