#optimizer

Question 1: Single choice

A hand-written SGD step updates a weight in place:

w = torch.ones(3, requires_grad=True)
w -= 0.1 * torch.ones(3)

What happens on the last line?

It raises a RuntimeError, because a leaf tensor that requires grad is modified in place1
w becomes [0.9, 0.9, 0.9] and remains a leaf that requires grad2
w becomes [0.9, 0.9, 0.9] and silently loses requires_grad3
The update is recorded in the graph, so the next backward() also differentiates through it4