Training¶
Training in UniLab is config-first. Use the package CLI for day-to-day runs and the script entrypoints when debugging the underlying Hydra composition.
Routes for uv run train, uv run eval, uv run demo, and low-level scripts.
Owner YAML layout, backend selection, and safe override examples.
TensorBoard, W&B, run metadata, and trace options.
How algo.load_run, checkpoint files, and replay commands fit together.
Run UniLab inside the checked-in Linux NVIDIA image workflow.
Current off-policy multi-GPU knobs and their config boundary.
When to Drop to scripts/train_*.py¶
Day-to-day runs should use the unified CLI. Reach for the low-level
scripts/train_*.py entrypoints only when you are:
debugging a specific training stack,
observing Hydra compose behavior directly, or
comparing script-level log directories or adapter behavior.