> from llama_models.llama3.reference_impl.model import Transformer

This just imports the Llama reference implementation and patches the device FYI.

There are more robust implementations out there.