diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index a06e041c1e8aaa8897ac77f2ec1275824849e7ef..fd901f91a1ddf87b031a55cf5d807990fecf3961 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -57,6 +57,8 @@ import recordio_writer
 import parallel_executor
 from parallel_executor import *
 
+from paddle.fluid.layers.math_op_patch import monkey_patch_variable
+
 Tensor = LoDTensor
 
 __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \
@@ -138,5 +140,5 @@ def __bootstrap__():
 
 # TODO(panyx0718): Avoid doing complex initialization logic in __init__.py.
 # Consider paddle.init(args) or paddle.main(args)
-layers.monkey_patch_variable()
+monkey_patch_variable()
 __bootstrap__()
diff --git a/python/paddle/fluid/layers/__init__.py b/python/paddle/fluid/layers/__init__.py
index cd1492da24d5e9d09a9eaac0b1b9c7aaffac6250..4917e67de0d20ff9e8f9a27f38e1bd2abef5c503 100644
--- a/python/paddle/fluid/layers/__init__.py
+++ b/python/paddle/fluid/layers/__init__.py
@@ -33,7 +33,6 @@ from metric_op import *
 from learning_rate_scheduler import *
 
 __all__ = []
-__all__ += math_op_patch.__all__
 __all__ += nn.__all__
 __all__ += io.__all__
 __all__ += tensor.__all__
diff --git a/python/paddle/fluid/layers/math_op_patch.py b/python/paddle/fluid/layers/math_op_patch.py
index 1754061c4ba6f5b97bced3548bc412dfb1b7932c..f814c41633fbac76eb9411e2f418f521e8e9679d 100644
--- a/python/paddle/fluid/layers/math_op_patch.py
+++ b/python/paddle/fluid/layers/math_op_patch.py
@@ -16,8 +16,6 @@ from ..framework import Variable, unique_name
 from layer_function_generator import OpProtoHolder
 from ..initializer import force_init_on_cpu
 
-__all__ = ['monkey_patch_variable']
-
 
 def monkey_patch_variable():
     def unique_tmp_name():