docker compose up

The demo will run on CPU by default and will be available at http://localhost:7860.

You need to change the docker-compose.yml to do so. See it for details on how to run on GPU.

By default, model weights will be stored in a docker volume mapping to the default location of huggingface cache: ~/.cache/huggingface/ on the local machine to prevent re-downloading. You can change the location of the cache by setting the environment variable DOCKER_HF_CACHE_PATH. Example:

export DOCKER_HF_CACHE_PATH=/path/to/cache
mkdir -p $DOCKER_HF_CACHE_PATH
docker compose up

Important: make sure the DOCKER_HF_CACHE_PATH exists before running the container!

The default limit is 1, set the environment variable GRADIO_DEFAULT_CONCURRENCY_LIMIT to change it or edit the docker-compose.yml file. Possible commands:

GRADIO_DEFAULT_CONCURRENCY_LIMIT=2 docker compose up

GRADIO_DEFAULT_CONCURRENCY_LIMIT=2 python app.py

• Model: MolmoE-1B-0924 from AllenAI
• Technical report: ArXiv

This demo showcases two capabilities of the Molmo model:

• Pointing: Identify and locate objects in images.
• Image Description: Generate detailed descriptions of images.