Replace the use of the deprecated torch.distributed.reduce_op with to… (#804)

Summary:
…rch.distributed.ReduceOp
Pull Request resolved: https://github.com/pytorch/fairseq/pull/804

Differential Revision: D15877033

Pulled By: myleott

fbshipit-source-id: 58e7c39a88b67345a55b761fee4d9f211a5ee82c

fairseq/optim/bmuf.py

@@ -62,7 +62,7 @@ class FairseqBMUF(FairseqOptimizer):
         if self.get_num_updates() % self.sync_iter == 0:
             size = float(dist.get_world_size())
             for p in self.params:
-                dist.all_reduce(p.data, op=dist.reduce_op.SUM)
+                dist.all_reduce(p.data, op=dist.ReduceOp.SUM)
                 p.data /= size
 
     def step(self, closure=None):