PyTorch Lightning — обучение и масштабирование нейросетей

★ 7.7 · engineering

pytorch-lightning is a Claude Code skill that assists with PyTorch Lightning, a high-level training framework built on PyTorch: it covers structuring model code into LightningModule, running distributed training with DDP, FSDP, and DeepSpeed, and scaling experiments from a single laptop to multi-node clusters without changing the codebase. The skill addresses common workflows such as converting raw PyTorch loops into Lightning (reducing 40+ lines to ~15), setting up validation and test steps with automatic TensorBoard logging, and configuring callbacks like ModelCheckpoint, EarlyStopping, and LearningRateMonitor. The Trainer handles device management across GPU, TPU, and CPU, mixed-precision training in FP16 and BF16, gradient accumulation, and checkpointing automatically. It is aimed at ML engineers and researchers who want clean, reproducible training pipelines with distributed training support out of the box.