Torch.no_grad() affecting outputs/loss

autograd
3

Could you post a small code snippet showing this problem?
I would like to reproduce this issue, as it might be a bug, and couldn’t reproduce it so far.