FSDP always sets requires_grad to False? Breaking checkpoint_wrapper.

## ❓ Questions and Help

Hi, I get the following error when trying to use FSDP with checkpoint_wrapper:
RuntimeError: None of the outputs have requires_grad=True, this checkpoint() is not necessary

I found this weird so I started digging around. I found that in the checkpoint wrapper function, all the input tensors had requires_grad=False stashed in the forward pass which led to this error. I checked that the inputs themselves do have requires_grad=True (at least one of them) and tried to find where this got clobbered. It is happening in the call to cast_floats_to_right_precision(True, True, *args, **kwargs) in the forward method of FullyShardedDataParallel.

I am really confused by this method since it seems that it methodically sets requires_grad to False for all inputs. If relevant, the layer being wrapped with a checkpoint is a Resnet-like block that is an intermediate layer in a network, the layer before it is an un-checkpointed Conv2D. So the input to the block is N x C x H x W and it has requires_grad=True, but when the forward call is dispatched (outputs = self.module(*args, **kwargs)), this tensor now has requires_grad=False. This breaks checkpoint_wrapper :(

Help would be much appreciated!