README.md

---
title: SCOPE Chat Demo
emoji: 🚀
colorFrom: purple
colorTo: red
sdk: gradio
sdk_version: 4.43.0
app_file: app.py
pinned: false
license: mit
short_description: An interactive chating demo for the SCOPE.
---

# SCOPE Demo-Chat Code
In this Demo, we show the chat results of different visual token compression methods.


## Deploy the Demo Locally

To run the demo, follow the same steps as LLaVA:

1. In **Terminal 1**, start the controller:
   ```bash
   python -m llava.serve.controller --host 0.0.0.0 --port 10000
   ```

2. In **Terminal 2**, launch the Gradio web server:
   ```bash
   python -m llava.serve.gradio_web_server_SCOPE --controller http://localhost:10000 --model-list-mode reload
   ```

3. In **Terminal 3**, start the model worker:
   ```bash
   python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-7b
   ```

## Citation
If you find this project useful in your research, please consider citing:
```
@article{deng2025scope,
  title={SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs},
  author={Deng Jinhong, Li Wen, and Zhou, Joey Tianyi, and He, Yang},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2025}
}
```

## Acknowledgement
This demo-chat is based on [VisionZIP](https://github.com/dvlab-research/VisionZip.git), many thanks.