PyTorch Model Development AI Skill Guide (Claude)

Overview & Engine Architecture

PyTorch provides tensor compute + autograd, nn.Module for models, and DataLoader for batches. A training step is forward → loss → backward → optimizer; inference uses eval() and torch.inference_mode(). Modern stacks use torch.amp for mixed precision and must handle torch.load(..., weights_only=True) defaults (PyTorch 2.6+). Agents keep device placement consistent, checkpoint reproducibly (model + optimizer + scaler), and never confuse train-time Dropout/BN with eval behavior.