cleaning up - speeding up a bit multi-gpu

modeling.py

@@ -467,6 +467,6 @@ class BertForQuestionAnswering(nn.Module):
             start_loss = loss_fct(start_logits, start_positions)
             end_loss = loss_fct(end_logits, end_positions)
             total_loss = (start_loss + end_loss) / 2
-            return total_loss, (start_logits, end_logits)
+            return total_loss
         else:
             return start_logits, end_logits