Contributing

Thanks for taking the time to contribute to this project!

Releases

This section documents the process used internally at Replicate to deploy the many variant Llama models.

Model variants live in the models directory, and deployment is managed by a Makefile.

To release a new model:

Run make select <model-name>, where model name corresponds to the name of a folder in the models directory, like model-llama-2-7b. This will copy stuff around and jigger the local state of the repo to say "use this model".
Run make test-local to test locally (assuming you're on a machine with a GPU).
Run make stage test-stage <model-name> to push to staging. If this passes, the model is ready to be promoted to production.
Run REPLICATE_USER=replicate && make push test-prod <model-name>. This runs the same tests as staging.

After releasing to production:

Search for old instances of the previous version's Docker image id in documentation and replace them with the new version.