vLLM Integration

You can use vLLM as an optimized worker implementation in FastChat. It offers advanced continuous batching and a much higher (~10x) throughput. See the supported models here.

Instructions

1. Install vLLM.

   pip install vllm
   

2. When you launch a model worker, replace the normal worker (fastchat.serve.model_worker) with the vLLM worker (fastchat.serve.vllm_worker). All other commands such as controller, gradio web server, and OpenAI API server are kept the same.

   python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-7b-v1.3
   

   If you see tokenizer errors, try

   python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-7b-v1.3 --tokenizer hf-internal-testing/llama-tokenizer