Using learn.to_fp16() Correctly?

When trying to turn on .to_fp16() getting strange results. Upon turning it off things look normal again.

torch.backends.cudnn.enabled =True
learn.to_fp16()

My GPU Info for Reference: