[BUG] Does deepspeed work with torch amp autocast?

[liqiangniu]

During training, fp16 is set to true.
But for some submodules, i want to using fp32.
with torch.cuda.amp.autocast(enabled=False):
x = self.model(x.float())

An error occur, show half vs float type not matching

[tjruwase]

@lqniunjunlper, are you able to share repro steps for this issue? Thanks