diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 04087cb241c9cd4975773e646bc0ef6e1287518f..ae0b50461a113dd90a0b457b5a26030394657f20 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -1663,6 +1663,9 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("is_compiled_with_mkldnn", IsCompiledWithMKLDNN);
   m.def("is_compiled_with_brpc", IsCompiledWithBrpc);
   m.def("is_compiled_with_dist", IsCompiledWithDIST);
+  m.def("_cuda_synchronize", [](const platform::CUDAPlace &place) {
+    platform::DeviceContextPool::Instance().Get(place)->Wait();
+  });
 
   m.def("get_float_stats", []() {
     std::vector<paddle::platform::ExportedStatValue<float>> float_stats;
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index e8cc6ce99016075a950f13d9e23f2957c9686471..881fb86266fced5ea7d9f64a3507e58047c7f477 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -90,6 +90,8 @@ from .io import save, load, load_program_state, set_program_state
 from .dygraph.checkpoint import save_dygraph, load_dygraph
 from .dygraph.varbase_patch_methods import monkey_patch_varbase
 from . import generator
+from .core import _cuda_synchronize
+
 Tensor = LoDTensor
 enable_imperative = enable_dygraph
 disable_imperative = disable_dygraph
@@ -133,7 +135,8 @@ __all__ = framework.__all__ + executor.__all__ + \
         'install_check',
         'save',
         'load',
-        'VarBase'
+        'VarBase',
+        '_cuda_synchronize'
     ]
 
 
diff --git a/python/paddle/fluid/core.py b/python/paddle/fluid/core.py
index 2e3bb6b00218a47467c59a2a88ea45ec80c32419..9a14c4cdf14a40bf5c23476b7bf21d8838b91251 100644
--- a/python/paddle/fluid/core.py
+++ b/python/paddle/fluid/core.py
@@ -263,6 +263,7 @@ if avx_supported():
         from .core_avx import _save_dygraph_dict
         from .core_avx import _load_dygraph_dict
         from .core_avx import _create_loaded_parameter
+        from .core_avx import _cuda_synchronize
         if sys.platform != 'win32':
             from .core_avx import _set_process_pids
             from .core_avx import _erase_process_pids
@@ -307,6 +308,7 @@ if load_noavx:
         from .core_noavx import _save_dygraph_dict
         from .core_noavx import _load_dygraph_dict
         from .core_noavx import _create_loaded_parameter
+        from .core_noavx import _cuda_synchronize
         if sys.platform != 'win32':
             from .core_noavx import _set_process_pids
             from .core_noavx import _erase_process_pids