[fix] OSS restore state to proper device (#46)

* move the restored param groups to the original device

* adding a corresponding test

fairscale/optim/oss.py

@@ -148,7 +148,7 @@ class OSS(Optimizer):
         self.load_local_state_dict(state_dict["state"][self.rank])
 
         # Restore the global param_groups
-        self.param_groups = state_dict["param_groups"]
+        self.param_groups = recursive_copy_to_device(state_dict["param_groups"], non_blocking=True, device=self._device)
 
     def add_param_group(self, param_group: dict) -> None:
         super().add_param_group(param_group)

tests/optim/test_oss.py

@@ -62,6 +62,9 @@ def test_state_dict():
     o.step()
     assert x == torch.tensor([0.9], device=DEVICE)
 
+    # Check that the exposed param_groups are on the proper device
+    assert o.param_groups[0]["params"][0].device == x.device
+
 
 def test_local_state_dict():
     x = torch.tensor([1.0], device=DEVICE, requires_grad=True)