leeloolee/lvlm-interpret

LVLM-Interpret: An Interpretability Tool for Large Vision-Language Models

[Project Page] [Paper]

Setup

• Update submodules

  git submodule update --init --recursive

• Install dependencies

  pip install -r requirements.txt

Usage

Start the Gradio server:

python app.py --model_name_or_path Intel/llava-gemma-2b --load_8bit 

or

python app.py --model_name_or_path llava-hf/llava-1.5-7b-hf --load_8bit

Options:

usage: app.py [-h] [--model_name_or_path MODEL_NAME_OR_PATH] [--host HOST] [--port PORT] [--share] [--embed] [--load_4bit] [--load_8bit]

options:
  -h, --help            show this help message and exit
  --model_name_or_path MODEL_NAME_OR_PATH
                        Model name or path to load the model from
  --host HOST           Host to run the server on
  --port PORT           Port to run the server on
  --share               Whether to share the server on Gradio's public server
  --embed               Whether to run the server in an iframe
  --load_4bit           Whether to load the model in 4bit
  --load_8bit           Whether to load the model in 8bit