ml-lightning-basics

Structure PyTorch Lightning training code by encapsulating model architecture and validation logic in a LightningModule, centralizing data loading in a LightningDataModule, and orchestrating distributed training via the Trainer with strategies like DDP or FSDP.

Configure a LightningDataModule for reproducible data loading by centralizing prepare_data, setup, and dataloader definitions, which ensures consistent data loading across distributed runs and separates data logic from model logic.

Integrate torch.compile with PyTorch Lightning by applying it within your LightningModule configuration to achieve PyTorch 2.0 speedups, handling graph breaks and selecting appropriate compile modes for optimized performance.

Yes, you can use mixed precision and callbacks with PyTorch Lightning Trainer by configuring the Trainer with appropriate precision settings and passing callback lists to monitor metrics, log data, and control training behavior automatically.

PyTorch Lightning training loops can produce NaNs or encounter OOM errors due to incorrect mixed-precision settings or improper batch sizing, requiring adjustments to Trainer configuration and hyperparameter logging to debug effectively.

Yes, you need to define the configure_optimizers hook when building a LightningModule, along with training_step and validation_step, to properly encapsulate your model's optimization logic and training loop responsibilities.