pytorch-lightning

PyTorch Lightning is a deep learning framework designed to pretrain and finetune AI models of any size, scaling from a single CPU to thousands of GPUs with minimal code changes. It organizes PyTorch code to separate scientific model logic from engineering infrastructure, automating backpropagation, mixed precision, multi-GPU, and distributed training. The project offers two core packages: PyTorch Lightning, which provides a high-level Trainer and LightningModule abstraction, and Lightning Fabric, which gives expert-level control over the training loop and scaling strategy for complex models like LLMs, diffusion models, and transformers. Users can choose the level of abstraction they need. Key features include hardware-agnostic training across CPU, GPU, TPU, and multi-node setups; support for distributed strategies such as DDP, FSDP, and DeepSpeed; mixed precision; experiment logging integrations; early stopping; checkpointing; and export to TorchScript or ONNX. The README notes minimal running speed overhead compared to pure PyTorch. Installation is via pip or conda. The repository includes numerous examples covering image classification, segmentation, object detection, text classification, summarization, audio generation, LLM finetuning, image generation, recommendation systems, and time-series forecasting. It is licensed under Apache 2.0 and is rigorously tested across Python and PyTorch versions, operating systems, and hardware accelerators.