Nan Loss with torch.cuda.amp and CrossEntropyLoss

[bruceyo (Bruce YU)]

In my case, I think nan is caused by the loss is too large to be held by float16.