diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt
index e5b438cece3cd517c237d2d580d319aa808b7776..55e2ea760158cda631ec07e2c7d318ec1cf79b77 100644
--- a/paddle/fluid/operators/math/CMakeLists.txt
+++ b/paddle/fluid/operators/math/CMakeLists.txt
@@ -68,6 +68,7 @@ cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selec
 cc_test(im2col_test SRCS im2col_test.cc DEPS im2col)
 cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col)
 cc_test(sequence_padding_test SRCS sequence_padding_test.cc DEPS sequence_padding)
+cc_test(sequence_pooling_test SRCS sequence_pooling_test.cc DEPS sequence_pooling)
 if(WITH_GPU)
     nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function)
     nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc
index 235b5405fb7d016f4bd8c738f75b303522183116..7be8539a7b0f1890898fd386a3056601fda8a7c3 100644
--- a/paddle/fluid/operators/math/sequence_pooling.cc
+++ b/paddle/fluid/operators/math/sequence_pooling.cc
@@ -157,6 +157,31 @@ class FirstSeqPoolFunctor {
   }
 };
 
+template <typename T>
+class SumSeqPoolGradFunctor {
+ public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::Tensor& out_grad,
+                  framework::LoDTensor* in_grad) {
+    auto lod = in_grad->lod()[0];
+    int64_t out_w = out_grad.numel() / out_grad.dims()[0];
+    int64_t in_w = in_grad->numel() / in_grad->dims()[0];
+    PADDLE_ENFORCE(in_w == out_w);
+    const T* out_g_data = out_grad.data<T>();
+    T* in_g_data = in_grad->mutable_data<T>(context.GetPlace());
+    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
+    for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
+      int64_t h = static_cast<int64_t>(lod[i + 1] - lod[i]);
+      int64_t in_offset = lod[i] * in_w;
+      const T* out_pos = out_g_data + i * out_w;
+      T* in_pos = in_g_data + in_offset;
+      for (int r = 0; r != h; ++r) {
+        blas.VCOPY(in_w, out_pos, in_pos + r * in_w);
+      }
+    }
+  }
+};
+
 template <typename T>
 class SequencePoolFunctor<platform::CPUDeviceContext, T> {
  public:
@@ -231,9 +256,15 @@ class SequencePoolGradFunctor<platform::CPUDeviceContext, T> {
       math::SetConstant<platform::CPUDeviceContext, T> functor;
       functor(context, in_grad, 0);
     }
+
+    if (pooltype == "SUM") {
+      math::SumSeqPoolGradFunctor<T> sum_pool_grad;
+      sum_pool_grad(context, out_grad, in_grad);
+      return;
+    }
+
     auto lod = in_grad->lod()[0];
     auto& place = *context.eigen_device();
-    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
     for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
       auto in_g_t = in_grad->Slice(static_cast<int>(lod[i]),
                                    static_cast<int>(lod[i + 1]));
@@ -247,12 +278,6 @@ class SequencePoolGradFunctor<platform::CPUDeviceContext, T> {
 
       if (pooltype == "AVERAGE") {
         in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
-      } else if (pooltype == "SUM") {
-        const T* out_g_data = out_g_t.data<T>();
-        T* in_g_data = in_g_t.mutable_data<T>(context.GetPlace());
-        for (int r = 0; r != h; ++r) {
-          blas.VCOPY(w, out_g_data, in_g_data + r * w);
-        }
       } else if (pooltype == "SQRT") {
         in_g_e.device(place) =
             (out_g_e / std::sqrt(static_cast<T>(h))).broadcast(bcast);
diff --git a/paddle/fluid/operators/math/sequence_pooling_test.cc b/paddle/fluid/operators/math/sequence_pooling_test.cc
new file mode 100644
index 0000000000000000000000000000000000000000..2bc008dd34ffcfe93a00bd4a8cde61626d91e235
--- /dev/null
+++ b/paddle/fluid/operators/math/sequence_pooling_test.cc
@@ -0,0 +1,126 @@
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/math/sequence_pooling.h"
+#include <gtest/gtest.h>
+#include <vector>
+
+template <typename DeviceContext, typename Place, typename T>
+void TestSequencePoolingSum(const paddle::framework::LoD& lod) {
+  paddle::framework::LoDTensor cpu_out_grad;
+  paddle::framework::LoDTensor cpu_in_grad;
+  paddle::framework::LoDTensor out_grad;
+  paddle::framework::LoDTensor in_grad;
+  const size_t second_dim = 128u;
+
+  // construct out_grad's tensor in cpu
+  const size_t out_first_dim = lod[0].size() - 1;
+  auto out_dims = paddle::framework::make_ddim(
+      {static_cast<int64_t>(out_first_dim), static_cast<int64_t>(second_dim)});
+
+  cpu_out_grad.mutable_data<T>(out_dims, paddle::platform::CPUPlace());
+  for (int64_t i = 0; i < cpu_out_grad.numel(); ++i) {
+    cpu_out_grad.data<T>()[i] = static_cast<T>(i);
+  }
+
+  // copy to dst out_grad
+  auto* place = new Place();
+  DeviceContext* context = new DeviceContext(*place);
+  if (paddle::platform::is_cpu_place(*place)) {
+    out_grad = cpu_out_grad;
+  } else {
+    TensorCopySync(cpu_out_grad, *place, &out_grad);
+  }
+
+  // construct in_grad
+  in_grad.set_lod(lod);
+  auto in_dims = paddle::framework::make_ddim(
+      {static_cast<int64_t>(lod[0].back()), static_cast<int64_t>(second_dim)});
+  in_grad.mutable_data<T>(in_dims, context->GetPlace());
+
+  // check tensor contruction result
+  PADDLE_ENFORCE_EQ(in_grad.dims().size(), out_grad.dims().size());
+  for (int64_t i = 1; i < out_grad.dims().size(); ++i) {
+    PADDLE_ENFORCE_EQ(in_grad.dims()[i], out_grad.dims()[i]);
+  }
+
+  // call functor
+  paddle::operators::math::SequencePoolGradFunctor<DeviceContext, T>()(
+      *context, "SUM", out_grad, &in_grad);
+
+  if (paddle::platform::is_cpu_place(*place)) {
+    cpu_in_grad = in_grad;
+  } else {
+    TensorCopySync(in_grad, paddle::platform::CPUPlace(), &cpu_in_grad);
+    cpu_in_grad.set_lod(in_grad.lod());
+  }
+
+  EXPECT_EQ(in_grad.numel(), lod[0].back() * second_dim);
+  EXPECT_EQ(in_grad.lod(), lod);
+
+  if (paddle::platform::is_cpu_place(*place)) {
+    for (int64_t i = 0; i < in_grad.lod()[0].size() - 1; ++i) {
+      int64_t begin = in_grad.lod()[0][i];
+      int64_t end = in_grad.lod()[0][i + 1];
+      paddle::framework::Tensor tmp = in_grad.Slice(begin, end);
+      for (int64_t j = 0; j != tmp.numel() / second_dim; ++j) {
+        for (int64_t m = 0; m != second_dim; ++m) {
+          EXPECT_EQ(tmp.data<T>()[m + j * second_dim],
+                    out_grad.data<T>()[m + i * second_dim]);
+        }
+      }
+    }
+  } else {
+    for (int64_t i = 0; i < cpu_in_grad.lod()[0].size() - 1; ++i) {
+      int64_t begin = cpu_in_grad.lod()[0][i];
+      int64_t end = cpu_in_grad.lod()[0][i + 1];
+      paddle::framework::Tensor tmp = cpu_in_grad.Slice(begin, end);
+      for (int64_t j = 0; j != tmp.numel() / second_dim; ++j) {
+        for (int64_t m = 0; m != second_dim; ++m) {
+          EXPECT_EQ(tmp.data<T>()[m + j * second_dim],
+                    cpu_out_grad.data<T>()[m + i * second_dim]);
+        }
+      }
+    }
+  }
+
+  delete place;
+  delete context;
+}
+
+TEST(SequencePoolingGrad, CPU_SUM) {
+  paddle::framework::LoD lod1;
+  lod1.push_back(std::vector<size_t>{0, 10});
+  TestSequencePoolingSum<paddle::platform::CPUDeviceContext,
+                         paddle::platform::CPUPlace, float>(lod1);
+
+  paddle::framework::LoD lod2;
+  lod2.push_back(std::vector<size_t>{0, 2, 7, 10});
+  TestSequencePoolingSum<paddle::platform::CPUDeviceContext,
+                         paddle::platform::CPUPlace, float>(lod2);
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(SequencePoolingGrad, CUDA_SUM) {
+  paddle::framework::LoD lod1;
+  lod1.push_back(std::vector<size_t>{0, 10});
+  TestSequencePoolingSum<paddle::platform::CUDADeviceContext,
+                         paddle::platform::CUDAPlace, float>(lod1);
+
+  paddle::framework::LoD lod2;
+  lod2.push_back(std::vector<size_t>{0, 2, 7, 10});
+  TestSequencePoolingSum<paddle::platform::CUDADeviceContext,
+                         paddle::platform::CUDAPlace, float>(lod2);
+}
+#endif