A chat interface based on llama.cpp for running Alpaca models. Entirely self-hosted, no API keys needed. Fits on 4GB of RAM and runs on the CPU.
- SvelteKit frontend
- MongoDB for storing chat history & parameters
- FastAPI + beanie for the API, wrapping calls to
llama.cpp
demo.webm
Setting up Serge is very easy. Starting it up can be done in a single command:
docker run -d -v weights:/usr/src/app/weights -v datadb:/data/db/ -p 8008:8008 ghcr.io/nsarrazin/serge:latest
Then just go to http://localhost:8008/ !
Make sure you have docker desktop installed, WSL2 configured and enough free RAM to run models. (see below)
Setting up Serge on Kubernetes or docker compose can be found in the wiki: https://github.com/nsarrazin/serge/wiki/Integrating-Serge-in-your-orchestration#kubernetes-example
Currently only the 7B, 713B and 30B alpaca models are supported. If you have existing weights from another project you can add them to the serge_weights volume using docker cp.
llama will just crash if you don't have enough available memory for your model.
- 7B requires about 4.5GB of free RAM
- 13B requires about 12GB free
- 30B requires about 20GB free
Feel free to join the discord if you need help with the setup: https://discord.gg/62Hc6FEYQH
Serge is always open for contributions! If you catch a bug or have a feature idea, feel free to open an issue or a PR.
If you want to run Serge in development mode (with hot-module reloading for svelte & autoreload for FastAPI) you can do so like this:
git clone https://github.com/nsarrazin/serge.git
DOCKER_BUILDKIT=1 docker compose -f docker-compose.dev.yml up -d --build
You can test the production image with
DOCKER_BUILDKIT=1 docker compose up -d --build
- Front-end to interface with the API
- Pass model parameters when creating a chat
- User profiles & authentication
- Different prompt options
- LangChain integration with a custom LLM
- Support for other llama models, quantization, etc.
And a lot more!