[bug fix] avoiding multiple initialization of gloo for fleet in dygra…

…ph mode (PaddlePaddle#27706)

* add double grad for expand, test=develop

python/paddle/distributed/fleet/base/role_maker.py

@@ -18,6 +18,7 @@
 import warnings
 from multiprocessing import Process, Manager
 
+import paddle
 import paddle.fluid as fluid
 from paddle.distributed.fleet.base.private_helper_function import wait_server_ready
 
@@ -869,7 +870,8 @@ def _generate_role(self):
             else:
                 self._collective_env()
             self._role_is_generated = True
-            self._gloo_init()
+            if not paddle.fluid.framework.in_dygraph_mode():
+                self._gloo_init()
 
 
 class UserDefinedRoleMaker(PaddleCloudRoleMaker):