PyTorch Lightning

Lightning reorganizes raw PyTorch so research logic stays separate from engineering boilerplate. Model code and training steps live in a LightningModule, while the Trainer takes over backpropagation, mixed precision, checkpointing, logging, and distributed coordination, letting the same script scale from a laptop to multi-node GPU clusters without changes. The framework supports CUDA GPUs, Apple Silicon, and TPUs, distributed strategies including DDP, FSDP, and DeepSpeed, 16-bit precision, and integrations with more than 40 experiment trackers such as TensorBoard, Weights and Biases, MLflow, and Comet. Lightning Fabric exposes a lower-level API for teams that want finer control, and the companion LitServe library handles inference serving. Researchers and ML engineers use it to train everything from CNNs and transformers to LLMs and diffusion models with less boilerplate and no loss of flexibility. Open source under the Apache 2.0 license.