llm.training — Training Loop, Configs, Callbacks

The training entry point is uv run llm-train. Most users will only interact with the configs and the public callback hooks.

Note: The training subpackage uses namespace packages (__init__.py-less directories). Some submodules require optional dependencies (tensorboard via the logging group; onnx via the test group) which may not be installed in every docs build environment. The CLI entry point llm.training.train additionally requires tensorboard.

For an overview of the training data flow, see Training Flow. For deep dives on the callback bridge and extending the trainer, see the Development guides.

The source for the RLHF PPO trainer and the core training engine lives in src/llm/training/ — browse it directly for the auto-generated API reference until namespace-package support stabilises upstream.