Skip to content

A custom framework for easy use of LLMs, VLMs, etc. supporting various modes and settings via web-ui

License

Notifications You must be signed in to change notification settings

KT313/assistant_base

Repository files navigation

Free to use under the GNU Affero General Public License v3.0.

This project may use optional libraries like LlaVA-NeXT that are subject to their respective original licenses, which may be required to support certain models or functions. This project can be modified to run without these optional libraries if required.

Asba Web Assistant

working:

  • support for probably all transformer, llama-cpp and exllamav2 models
  • image input support for Llama-3-Llava & phi-3-vision (transformers)
  • custom adaptive beam search
  • image-generation with diffusers backend

WIP:

  • Agent Task Mode (Structuring large / complicated tasks and progressively working on it)
  • txt2img, txt+img2img
  • training support

running it

  1. ./start_infer.sh # this is enough if you only want the api without web ui

  2. ./start_web.sh # for the web-ui

more info

For now you will have to download the model(s) yourself and specify their path in the config file!

note: custom adaptive beam search is slow with llama-cpp models (quantized LLama-3-70B) and Llama-3-Llava (only when using an image) right now because they have to do the beam searches in sequence, couldn't get them to work in batch. For llama-cpp it's because it (for whatever reason???) does not support batch inference. Llama-3-Llava does support it but when you supply an image additionally it gets confused (and so do I).

Web UI