switch to pytorch DistributedDataParallel

examples/run_squad.py

@@ -902,12 +902,12 @@ def main():
         model.half()
     model.to(device)
     if args.local_rank != -1:
-        try:
-            from apex.parallel import DistributedDataParallel as DDP
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+    #     try:
+    #         from apex.parallel import DistributedDataParallel as DDP
+    #     except ImportError:
+    #         raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
 
-        model = DDP(model)
+        model = torch.nn.parallel.DistributedDataParallel(model)
     elif n_gpu > 1:
         model = torch.nn.DataParallel(model)