/serge

A web interface for chatting with Alpaca through llama.cpp. Fully dockerized, with an easy to use API.

Primary LanguagePythonMIT LicenseMIT

Serge - LLaMA made easy 🦙

License Discord

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
  • Redis for storing chat history & parameters
  • FastAPI + langchain for the API, wrapping calls to llama.cpp using the python bindings
demo.webm

Getting started

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/ and you're good to go!

The API documentation can be found at http://localhost:8008/api/docs

Windows

Make sure you have docker desktop installed, WSL2 configured and enough free RAM to run models. (see below)

Kubernetes & docker compose

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

Models

Currently the following models are supported:

  • GPT4-Alpaca-LoRA-30B
  • Alpaca-LoRA-65B
  • OpenAssistant-30B
  • GPT4All-13B
  • Stable-Vicuna-13B
  • Guanaco-7B
  • Guanaco-13B
  • Guanaco-33B
  • Guanaco-65B

If you have existing weights from another project you can add them to the serge_weights volume using docker cp.

⚠️ A note on memory usage

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

Support

Feel free to join the discord if you need help with the setup: https://discord.gg/62Hc6FEYQH

Contributing

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

What's next

  • Front-end to interface with the API
  • Pass model parameters when creating a chat
  • Manager for model files
  • Support for other models
  • LangChain integration
  • User profiles & authentication

And a lot more!