diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
index 58fa7f1bebc082df11e07f5f97927b417af3e4e8..e02245d05dd9fd5d82a26dcbc00a8b832f93842e 100644
--- a/python/paddle/fluid/backward.py
+++ b/python/paddle/fluid/backward.py
@@ -220,7 +220,10 @@ def _callback_lookup_(op):
     :return: callback function
     """
     if op.type == 'parallel_do' and op.attr('use_nccl'):
+        all_vars = op.block.vars
         param_names = set(op.input('parameters'))
+        param_names = filter(lambda name: all_vars[name].stop_gradient is False,
+                             param_names)
         param_grad_names = [n + "@GRAD" for n in param_names]
 
         class ParallelDoCallBack(object):
diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py
index ff8861856372882f4da9c3c775b3856b21d062db..1bb1aa30ee1019c6f80eb64b6dc20459e7a3073b 100644
--- a/python/paddle/fluid/layers/control_flow.py
+++ b/python/paddle/fluid/layers/control_flow.py
@@ -294,8 +294,7 @@ class ParallelDo(object):
 
         params = list(set(params))
 
-        param_list = [parent_block.var(name) for name in params]
-        return filter(lambda param: param.stop_gradient is False, param_list)
+        return [parent_block.var(name) for name in params]
 
     def complete_op(self):
         main_program = self.helper.main_program