pytorch lightning

Originally created by William Falcon, it provides a structured Trainer API that handles the training loop, validation, testing, and logging across configurations from a single GPU to thousands of TPUs, with zero code changes between scales.

The LightningModule class organizes model code into clear lifecycle hooks (training_step, validation_step, configure_optimizers), making codebases easier to read, reproduce, and collaborate on.

Lightning integrates with the full PyTorch ecosystem Hugging Face Transformers, timm, MONAI for medical imaging and supports logging backends including TensorBoard, Weights & Biases, MLflow, and Comet.

For large-scale training, Lightning provides native FSDP (Fully Sharded Data Parallel), DeepSpeed ZeRO, and SLURM cluster support out of the box.

PyTorch Lightning is open-source under the Apache 2.0 license and maintained by Lightning AI, which also offers a managed cloud platform for training, fine-tuning, and deploying models.

The framework is widely adopted in academic research and production ML engineering, with adoption at major AI labs, technology companies, and universities. It is installable via pip and requires no changes to underlying PyTorch code, making migration from raw PyTorch training loops straightforward.