1. 1. 1. 🚀 The feature, motivation and pitch

1. 1. 🐛 `grad_dtype` property lost when using `fully_shard` (FSDP2)

1. 1. 1. Problem
         The `grad_dtype` property of PyTorch tensors, which allows specifying gradient dtype for leaf tensors, is not preserved when applying `torch.distributed.fsdp.fully_shard`. This breaks precision control in distributed training.

1. 1. 1. Alternatives

No response

1. 1. 1. Additional context

No response

cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci @zhaojuanmao @mrshenli @rohan-varma @chauhang @mori360 @ppwwyyxx @penguinwu