iblai / ibl-ai-neural-startup

ibl-ai-neural-startup

This project presents different ways to run llms in production

Note

Openllm with gpu support is the preferred means of running llms for high throughput.

OpenLLM

OpenLLM exposes an intuitive api on top of backends like vllm, ctranslate, etc. It scales very well and has very good integration with langchain through langchain's OpenLLM llm class.

Three approaches to deploying with openllm are presented in the openllm directory.

1. deploy.sh: Run openllm directly on the host device. This checks to ensure that python3 and openllm packages are installed. You must always pass a model's repo_id from huggingface as input parameter.
2. docker-deploy-no-gpu.py: This runs opennl model in a non-gpu container.
3. docker-deploy-gpu.py: This runs openllm in docker with gpu support. nvidia-smi driver must be configured from the aws instance.

api

The api module presents a mini fastapi app that expoises endpoints for inference. This has only minimal features.

usage

You can either run the project directly

cd api
pip install -r requirements.txt
uvicorn app:app

Or run with docker using

cd api
docker build . -t api
docker run --rm -d --gpus all -p 3000:3000  api

You may run without gpus but at a performance penalty.e