Meta Chameleon

Paper | Blog | Model Checkpoint Download | HuggingFace

This repository contains artifacts for the Meta Chameleon model from FAIR, Meta AI. In this repository is:

• Standalone Inference Code — a fast GPU-based inference implementation for running model checkpoints
• Input-Output Viewing — a harness for richly viewing multimodal model inputs and outputs with a browser-based tool
• Evaluation Prompts — mixed-modal and text-only prompts for human evaluation

System Requirements

Running constituent components for inference and the input-output viewer currently requires a CUDA-capable GPU. If you'd like to run inference on other hardware, other inference implementations, including HuggingFace, are platform agnostic.

Getting Started

First, pip install this repository:

pip install -U git+https://github.com/facebookresearch/chameleon.git

Alternatively, if you want access to the full visualizer, you'll need to clone this repository (instead of installing), then pip install from the repository root:

git clone https://github.com/facebookresearch/chameleon.git
cd chameleon
pip install -e .

Model checkpoints and configs must be downloaded before running inference or the viewer. After requesting model access, run the following script, adding pre-signed download URL you were emailed when prompted:

python -m chameleon.download_data [pre-signed URL]

(you can also paste the command given in the email containing the download link)

Running the Viewer

The viewer visualizes multi-modal model input and output. It is most easily run with docker-compose. You'll need to clone the repository, not just a pip install.

The following runs both the service and viewer interface.

By default, this runs the 7B parameter model. You can change the model_path variable in ./config/model_viewer.yaml to select another model and alter other configuration:

docker-compose up --build

You can open the viewer at http://localhost:7654/

Running the MiniViewer

The miniviewer is a light weight debug visualizer, that can be run with:

python -m chameleon.miniviewer

This runs the 7B parameter model. To run the 30B model, use the following command:

python -m chameleon.miniviewer --model-size 30b

You can open the miniviewer at http://localhost:5000/.

License

Use of this repository and related resources are governed by the Chameleon Research License and the LICENSE file.

Citation

To cite the paper, model, or software, please use the below:

@article{Chameleon_Team_Chameleon_Mixed-Modal_Early-Fusion_2024,
  author = {Chameleon Team},
  doi = {10.48550/arXiv.2405.09818},
  journal = {arXiv preprint arXiv:2405.09818},
  title = {Chameleon: Mixed-Modal Early-Fusion Foundation Models},
  url = {https://github.com/facebookresearch/chameleon},
  year = {2024}
}