Answer

Dynamic computation graph that records operations on tensors. On backward(), traverses graph in reverse computing gradients via chain rule. requires_grad=True enables gradient tracking. Detach tensors to stop gradient flow. torch.no_grad() for inference (no graph overhead).