What happens inside PyTorch when loss.backward() is called?

loss.backward() triggers reverse traversal of the autograd graph and computes gradients using chain rule.

When backward() is called, PyTorch starts from the scalar loss node and traverses the dynamic computation graph in reverse topological order. Each node has a backward function (grad_fn) that computes local gradients and propagates them to parent tensors. Gradients accumulate in leaf tensors' .grad fields. Intermediate buffers are freed unless retain_graph=True is set.