The docstring for the Optimizer class notes that when using the .step function with a closure, this closure should not change the parameter gradients:

However, every example I've found seems do be doing exactly that, e.g. also on the torch.optim documentation on the same page:

for input, target in dataset:
    def closure():
        optimizer.zero_grad()
        output = model(input)
        loss = loss_fn(output, target)
        loss.backward()
        return loss
    optimizer.step(closure)

Is the note about not changing the gradients in the closure outdated?

cc @vincentqb