I know it's possible to check for NaN values of torch tensors by using the numpy.isnan() function on CPU tensors, but I think a native torch.isnan() function would be nice to have. I would also propose a constant torch.nan similar to numpy.nan that can be assigned (or compared) to torch tensors for testing purposes.

My main use case for this is that I want to automatically check losses for NaN values during training and warn the user or terminate training when encountering them.