🐛 Bug

I met the following error when I was using pytorch 1.6 with LSTM
RuntimeError: cuDNN error: CUDNN_STATUS_BAD_PARAM

To Reproduce

I think the root cause was that I was using the mixed precision training mode in pytorch 1.6 following https://pytorch.org/docs/stable/notes/amp_examples.html
and the datatype might be converted to fp16 which causes this error.
When I train my model with full precision, it worked normally.

cc @csarofeen @ptrblck @zou3519