What problem does it solve? Raw PyTorch training code mixes research logic with engineering boilerplate like device management, distributed synchronization, and checkpointing, making experiments hard to reproduce and scale across GPUs. ## Core Features & Use Cases - LightningModule Templates: Structure models into training, validation, test, and optimizer configuration steps with automatic metric logging and hyperparameter saving. - Trainer & Distributed Strategies: Configure multi-GPU and multi-node training with DDP, FSDP, or DeepSpeed, plus mixed precision, gradient accumulation, and checkpointing. - DataModules, Callbacks & Logging: Build reusable data pipelines, extend training with callbacks like EarlyStopping and ModelCheckpoint, and log to TensorBoard, W&B, MLflow, or CSV. - Use Case: You need to train a transformer across 8 GPUs with limited memory. Use this Skill to wrap your model in a LightningModule, select the FSDP strategy with bf16 mixed precision, and add checkpointing and early stopping callbacks. ## Quick Start Help me convert my PyTorch training script into a PyTorch Lightning LightningModule with a Trainer configured for multi-GPU DDP training.
