Using vLLM to deploy LLM as an API to accelerate inference

[fx-hit]

Based on practical tests, deploying omost-llama-3-8b on an A100 using torch==2.3.0+cu118, vllm==0.5.0.post1+cu118, and xformers==0.0.26.post1+cu118 works well. if want to speed up the process, can refer to this setup.

vllm: https://docs.vllm.ai/en/stable/getting_started/quickstart.html

[badcookie78]

Hi, Can I know if is possible to run in with ollama and then host the LLM locally?

[zk19971101]

I find comfyui_omost show a way to accelerate inference by TGI(text generation inference).
https://github.com/huchenlei/ComfyUI_omost?tab=readme-ov-file#accelerating-llm

[sudanl]

Based on practical tests, deploying omost-llama-3-8b on an A100 using torch==2.3.0+cu118, vllm==0.5.0.post1+cu118, and xformers==0.0.26.post1+cu118 works well. if want to speed up the process, can refer to this setup.

vllm: https://docs.vllm.ai/en/stable/getting_started/quickstart.html

Good idea! Could you kindly share the code?