diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index 905f499dfce96644765003abf449639f8ee9f445..05731baad24716325fa7b458a9c1959f4b292be0 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -7170,6 +7170,8 @@ class Parameter(Variable, metaclass=ParameterMetaClass):
         )
         self.trainable = kwargs.get('trainable', True)
 
+        self.stop_gradient = not self.trainable
+
         self.optimize_attr = kwargs.get('optimize_attr', {'learning_rate': 1.0})
 
         self.regularizer = kwargs.get('regularizer', None)
diff --git a/test/legacy_test/test_trainable.py b/test/legacy_test/test_trainable.py
index 5f6968af51096fa5f8780d35ea1b89052d622339..07501923885a99c92589db73345fc9453e1202fe 100644
--- a/test/legacy_test/test_trainable.py
+++ b/test/legacy_test/test_trainable.py
@@ -20,6 +20,8 @@ from simple_nets import init_data
 import paddle
 from paddle import fluid
 
+paddle.enable_static()
+
 
 def test_trainable():
     x = paddle.static.data(name='image', shape=[-1, 784], dtype='float32')
@@ -68,12 +70,12 @@ class TestTrainable(unittest.TestCase):
         self.check_trainable(
             test_trainable,
             feed_dict,
-            op_count={'adam': 1, 'scale': 0, 'mul_grad': 1},
+            op_count={'adam': 1, 'scale': 0, 'mul_grad': 0},
         )
         self.check_trainable(
             test_trainable,
             feed_dict,
-            op_count={'adamax': 1, 'scale': 1, 'mul_grad': 1},
+            op_count={'adamax': 1, 'scale': 1, 'mul_grad': 0},
             optimizer=paddle.optimizer.Adamax(learning_rate=0.2),
         )