Deploy MusicGen as an API with Hugging Face Inference Endpoints

MusicGen, a powerful music generation model from Facebook, can be deployed as an API using Hugging Face's Inference Endpoints. This guide walks through creating a custom handler to serve the model, enabling text-to-music generation with just a few steps.

MusicGen takes a text prompt and optionally a melody to generate audio. While transformers pipelines handle many models out of the box, MusicGen requires a custom inference function. Hugging Face Inference Endpoints support such custom handlers, allowing deployment of any model.

Step-by-Step Deployment

1. Duplicate the MusicGen repository – Use the repository duplicator to copy facebook/musicgen-large to your account.

2. Add custom files – Upload handler.py and requirements.txt to the duplicated repo.

   • handler.py: Defines an EndpointHandler class that loads the model and processor, then processes requests. It uses half-precision (FP16) for efficiency.
   • requirements.txt: Lists dependencies like transformers==4.31.0 and accelerate>=0.20.3.

3. Create an Inference Endpoint – On the Inference Endpoints page, select your repo, choose a GPU instance with at least 16 GB RAM, and deploy. The endpoint will be ready to accept requests.

Querying the Endpoint

Use curl or the huggingface-hub library to send prompts:

curl URL_OF_ENDPOINT \
  -X POST \
  -d '{"inputs":"happy folk song, cheerful and lively"}' \
  -H "Authorization: Bearer {YOUR_TOKEN}" \
  -H "Content-Type: application/json"

The response contains generated audio as a list of floats.

Conclusion

With Inference Endpoints, deploying custom models like MusicGen is straightforward. This approach can be extended to any model without a built-in pipeline. MusicGen's capabilities open doors for creative applications in music generation.

For more details, explore the full guide on the Hugging Face blog.