How does PyTorch optimizer.step() interact with autograd gradients internally?

optimizer.step() reads gradients from .grad fields and updates parameters without touching the computation graph.

After loss.backward(), gradients are stored in leaf tensors' .grad attributes. optimizer.step() iterates over parameter groups, applies optimization rules (SGD, Adam, etc.), and updates parameters in-place. It does not interact with autograd graph directly but relies on accumulated gradients. The graph is already freed unless retain_graph=True.