A huge difference of memory usage on different GPU

8

It seems to me that torch.backends.cudnn.benchmark = False solve my problem) Thank you!

1 Like