remove move_state_dict_to_gpu, which is causing cuda oom (#1367)

Summary:
Pull Request resolved: #1367

I keep getting cuda oom in this load_best_model stage, move_state_dict_to_gpu and model.cuda() are not both needed. Looks like it will double gpu memory this way.

Reviewed By: anchit

Differential Revision: D21725316

fbshipit-source-id: 70b5761a25afb19da7f44a3fead37b36d0e122da

pytext/trainers/trainer.py

@@ -333,9 +333,7 @@ def load_best_model(self, state: TrainingState):
         if cuda.CUDA_ENABLED:
             # Move current model to CPU to avoid multiple models in GPU memory
             state.model.cpu()
-            state.model.load_state_dict(
-                self.move_state_dict_to_gpu(state.best_model_state)
-            )
+            state.model.load_state_dict(state.best_model_state)
             # Move model back to GPU
             state.model.cuda()
         else: