ml-lightning-basics

What problem does it solve?

It solves the complexity of structuring PyTorch training code by showing how to implement LightningModule, LightningDataModule, and Trainer in a way that supports scalable, reproducible ML workflows.

Core Features & Use Cases

• LightningModule patterns: Encapsulates model architecture and training/validation logic with required hooks like training_step, validation_step, and configure_optimizers.
• LightningDataModule for reproducible data: Centralizes prepare_data, setup, and dataloader definitions for consistent loading across distributed runs.
• Trainer orchestration: Demonstrates automatic hardware/distributed configuration (DDP/FSDP/DeepSpeed), callbacks, logging, and mixed-precision settings.
• PyTorch 2.0 optimization: Covers practical torch.compile integration (modes, best practices, and handling graph breaks).
• Production best practices: Emphasizes hyperparameter saving, correct logging via self.log, and guidance for common issues like NaNs, OOM, and slow training.

Quick Start

Use the ml-lightning-basics skill to design your model as a LightningModule and your data as a LightningDataModule, then train it with Trainer using the configuration patterns shown in the skill.