nemo_automodel.recipes.llm.train_eagle2

Minimal Llama-only EAGLE-2 training recipe.

EAGLE-2 reuses the same draft-model training objective as EAGLE-1. The difference is in the speculative decoding tree policy at inference time. This recipe exists as an explicit entrypoint so users can train a draft checkpoint intended for an EAGLE-2 deployment workflow without having to remember that the underlying optimization is identical to EAGLE-1.

Module Contents

Classes

Functions

API

class nemo_automodel.recipes.llm.train_eagle2.TrainEagle2Recipe()

nemo_automodel.recipes.llm.train_eagle2.main(
    config_path: str | None = None
)