diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 254c4a5942db0d6b165237bda3574471a498d5e1..5121987f922a33f5fd93076b395c43770db50c2d 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -505,13 +505,15 @@ All parameter, weight, gradient are variables in Paddle.
                                   scope, local_scopes, allow_op_delay);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
+      // NOTE: even we return a vec<Scope*>* to Python use reference policy.
+      // We still cannot get local_scope from this vector, since the element
+      // of vec<Scope*> will be freed by Python GC. We can only return Scope*
+      // one by one and mark them as reference.
       .def("local_scopes",
            [](ParallelExecutor &self) -> std::vector<Scope *> * {
              return &self.GetLocalScopes();
            },
            py::return_value_policy::reference)
-      .def("local_scopes_len",
-           [](ParallelExecutor &self) { return self.GetLocalScopes().size(); })
       .def("local_scope", [](ParallelExecutor &self,
                              size_t i) { return self.GetLocalScopes()[i]; },
            py::return_value_policy::reference)
diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h
index c9cad15a74bdf1077d4103c5874a2c4556b73e65..159d1d5f4e70033fabf93514bd63b38f83675bff 100644
--- a/paddle/fluid/pybind/tensor_py.h
+++ b/paddle/fluid/pybind/tensor_py.h
@@ -190,6 +190,10 @@ void PyCUDATensorSetFromArray(
       static_cast<const platform::CUDADeviceContext *>(pool.Get(place));
   paddle::platform::GpuMemcpyAsync(dst, array.data(), sizeof(T) * array.size(),
                                    cudaMemcpyHostToDevice, dev_ctx->stream());
+  // NOTE: For safety, here wait the copy complete.
+  // It because the CPU array.data() could be destroyed after this method.
+  // If we make this method async, it could be copied data from a memory buffer
+  // that has been freed.
   dev_ctx->Wait();
 }
 
@@ -217,6 +221,11 @@ void PyCUDATensorSetFromArray(
   paddle::platform::GpuMemcpyAsync(dst, array.data(),
                                    sizeof(uint16_t) * array.size(),
                                    cudaMemcpyHostToDevice, dev_ctx->stream());
+  // NOTE: For safety, here wait the copy complete.
+  // It because the CPU array.data() could be destroyed after this method.
+  // If we make this method async, it could be copied data from a memory buffer
+  // that has been freed.
+  dev_ctx->Wait();
 }
 
 template <typename T>