Skip to content

gmarko/Auto-Llama-cpp

 
 

Repository files navigation

Auto-Llama-cpp: An Autonomous Llama Experiment

This is a fork of Auto-GPT with added support for locally running llama models through llama.cpp. This is more of a proof of concept. It's sloooow and most of the time you're fighting with the too small context window size or the models answer is not valid JSON. But sometimes it works and then it's really quite magical what even such a small model comes up with. But obviously don't expect GPT-4 brilliance here.

Supported Models


Since this uses llama.cpp under the hood it should work with all models they support. As of writing this is

  • LLaMA
  • Alpaca
  • GPT4All
  • Chinese LLaMA / Alpaca
  • Vigogne (French)
  • Vicuna
  • Koala

Model Performance (the experience so far)


Response Quality

So far I have tried

  • Vicuna-13b-4BIT
  • LLama-13B-4BIT

Overall the Vicuna model performed much better than the original LLama model in terms of answering in the required JSON format and how much sense the answers make. I just couldn't get it to stop starting every answer with ### ASSISTANT. I am very curious to hear how well others models perform. The 7B models seemed have problems with grasping what's asked of them in the prompt, but I tried very little in this direction since the inference speed didn't seem to be much faster for me.

Inference Speed

The biggest problem at the moment is indeed inference speed. As the agent is self prompting a lot, a few seconds of infernce that are acceptable in a chatbot scenario become minutes and more. Testing things like different prompts etc is a pain under these conditions.

Discussion

Fell free to add your thoughts and experiences in the discussion area. What models did you try? How well did they work ou for you?

Future Plans


  1. Add GPU Support via GPTQ
  2. Improve Prompts
  3. Remove external API support (This is supposed to be completely self-contained agent)
  4. Add support for Open Assistent models

About

Uses Auto-GPT with Llama.cpp

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.8%
  • Dockerfile 0.2%