What it does PyTorch Lightning is an organizational wrapper around PyTorch. You subclass LightningModule to define your model, training step, and optimizer; the Trainer then handles backpropagation, device placement, mixed precision, multi-GPU, and distributed training. The pitch is that scaling from a single CPU to 10,000 GPUs requires zero changes to your core model code.

The interesting bit The project explicitly offers a dial, not a wall. If you want full abstraction, use PyTorch Lightning. If you want to hand-roll distributed logic but still skip some boilerplate, Lightning Fabric gives you “expert control” at a lower level. The README’s own analogy: PyTorch is JavaScript, Lightning is React.

Key highlights

• LightningModule bundles model, loss, and optimizer configuration into one subclass, keeping research code separate from engineering glue.
• Trainer.fit() automatically handles mixed precision, gradient clipping, checkpointing, and distributed backends.
• Lightning Fabric provides a lower-level API for PyTorch experts who want granular control without starting from scratch.
• Pre-built examples cover LLM fine-tuning (Llama 3.1 8B), diffusion models, audio generation, and standard vision tasks.
• Optional hosted platform (Lightning Cloud) offers one-command GPU access with autoscaling.

Caveats

• The “zero code changes” scaling claim assumes you wrote your model in Lightning’s structure to begin with; migrating existing PyTorch loops still takes work.
• The README is heavy on marketing for Lightning Cloud and LitServe; the core framework documentation is linked but not shown inline.

Verdict Worth a look if you’re tired of rewriting distributed training scaffolding for every new project. Skip it if you already have a battle-tested training loop you don’t want to refactor, or if you need something more framework-agnostic than PyTorch-specific.