fix ddp train mode (#794)

Summary:
Pull Request resolved: #794

Since D53659696, train mode was only applied to the inner module of DDP. This diff fixes this

Reviewed By: diego-urgell

Differential Revision: D56424247

fbshipit-source-id: 179e14180cdb8bbc08fde595220bb76f75a37c02

Author: JKSenthil

torchtnt/framework/_loop_utils.py

@@ -64,13 +64,19 @@ def _set_module_training_mode(
     prior_module_train_states = {}
     for name, module in modules.items():
         prior_module_train_states[name] = module.training
-        if isinstance(module, DistributedDataParallel):
-            module = module.module
-        if torch.ao.quantization.pt2e.export_utils.model_is_exported(module):
+        is_ddp = isinstance(module, DistributedDataParallel)
+
+        if torch.ao.quantization.pt2e.export_utils.model_is_exported(
+            module.module if is_ddp else module
+        ):
             if mode:
-                module = torch.ao.quantization.move_exported_model_to_train(module)
+                module = torch.ao.quantization.move_exported_model_to_train(
+                    module.module if is_ddp else module
+                )
             else:
-                module = torch.ao.quantization.move_exported_model_to_eval(module)
+                module = torch.ao.quantization.move_exported_model_to_eval(
+                    module.module if is_ddp else module
+                )
         else:
             module.train(mode)