Add privateuse1 in FSDP's sharded grad scaler (pytorch#126971)

1. add privateuse1 in FSDP's sharded grad scaler
2. support found_inf copy for more devices

Pull Request resolved: pytorch#126971
Approved by: https://github.com/awgu, https://github.com/weifengpy

torch/distributed/fsdp/sharded_grad_scaler.py

@@ -15,7 +15,12 @@ def _refresh_per_optimizer_state() -> Dict[str, Any]:
 
 
 def _is_supported_device(tensor: torch.Tensor) -> bool:
-    return tensor.is_cuda or tensor.device.type in ("xla", "cpu", "hpu")
+    return tensor.is_cuda or tensor.device.type in (
+        "xla",
+        "cpu",
+        "hpu",
+        torch._C._get_privateuse1_backend_name(),
+    )
 
 
 class _GeneralMultiDeviceReplicator(_MultiDeviceReplicator):
@@ -284,16 +289,16 @@ def unscale_(self, optimizer: torch.optim.Optimizer) -> None:
         optimizer_state = self._per_optimizer_states[id(optimizer)]
         works = []
         found_inf_on_cpus = []
-        found_inf_on_cudas = []
+        found_inf_on_devices = []
 
         for found_inf in optimizer_state["found_inf_per_device"].values():
-            if self._device == "cuda" and found_inf.device.type == "cpu":
+            if self._device != "cpu" and found_inf.device.type == "cpu":
                 found_inf_on_cpus.append(found_inf)
-                found_inf_on_cuda = found_inf.cuda()
-                found_inf_on_cudas.append(found_inf_on_cuda)
+                found_inf_on_device = found_inf.to(self._device)
+                found_inf_on_devices.append(found_inf_on_device)
                 works.append(
                     dist.all_reduce(
-                        found_inf_on_cuda, async_op=True, group=self.process_group
+                        found_inf_on_device, async_op=True, group=self.process_group
                     )
                 )
             else:
@@ -303,7 +308,7 @@ def unscale_(self, optimizer: torch.optim.Optimizer) -> None:
         for work in works:
             work.wait()
         if found_inf_on_cpus:
-            torch._foreach_copy_(found_inf_on_cpus, found_inf_on_cudas)
+            torch._foreach_copy_(found_inf_on_cpus, found_inf_on_devices)
 
     def _amp_update_scale_cpu_(self, found_inf: torch.Tensor) -> None:
         """