fix new_group for hang problem

PaddlePaddle · Jun 3, 2021 · e4fa0a6 · e4fa0a6 · paddle-bot-old · Jun 4, 2021
1 parent 901796f
commit e4fa0a6
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/python/paddle/distributed/collective.py b/python/paddle/distributed/collective.py
@@ -262,10 +262,12 @@ def new_group(ranks=None, backend=None):
                                          place).init_with_ring_id(ring_id)
             else:
                 assert False, ("no cuda device found")
+        else:
+            return gp
 
     # TODO(shenliang03): This is a temporary solution to solve the problem of 
     # hang caused by cross-creation of new_group
-    tmp = paddle.to_tensor([0])
+    tmp = fill_constant([0], dtype="int32", value="1")
     paddle.distributed.all_reduce(tmp, use_calc_stream=True)
     paddle.distributed.wait(tmp)
     return gp