diff --git a/paddle/framework/variable.h b/paddle/framework/variable.h
index 3720393601a5458fb43217d05257c4acb8be28cd..e5a94759f9230ab4ce9d2cc24849a2debb8a5e2f 100644
--- a/paddle/framework/variable.h
+++ b/paddle/framework/variable.h
@@ -35,7 +35,6 @@ class Variable {
   template <typename T>
   T* GetMutable() {
     if (!IsType<T>()) {
-      VLOG(10) << "Resetting " << *this->name_;
       holder_.reset(new PlaceholderImpl<T>(new T()));
     }
     return static_cast<T*>(holder_->Ptr());
diff --git a/paddle/operators/sum_op.h b/paddle/operators/sum_op.h
index d1277d3edd897b951fe20ee2a0294bd9c615a059..552b48f608b7e0248f03dbea940a83f112a67712 100644
--- a/paddle/operators/sum_op.h
+++ b/paddle/operators/sum_op.h
@@ -37,10 +37,12 @@ class SumKernel : public framework::OpKernel<T> {
     bool in_place = out_var == in_vars[0];
 
     if (out_var->IsType<framework::LoDTensor>()) {
-      auto *out = context.Output<Tensor>("Out");
-      auto result = EigenVector<T>::Flatten(*out);
+      auto *out = context.Output<LoDTensor>("Out");
       if (!in_place) {
         out->mutable_data<T>(context.GetPlace());
+      }
+      auto result = EigenVector<T>::Flatten(*out);
+      if (!in_place) {
         math::SetConstant<DeviceContext, T> constant_functor;
         constant_functor(context.template device_context<DeviceContext>(), out,
                          0.0);
diff --git a/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py b/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py
index 7f61b966fd53dd9abd768fc198255d636781e395..238fd1a8cba81f37e9d63b0d6ea09582d6b0fe83 100644
--- a/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py
+++ b/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py
@@ -242,7 +242,7 @@ class TestSimpleMul(unittest.TestCase):
         out = rnn()
         out = fluid.layers.sequence_pool(out, pool_type='last')
         loss = fluid.layers.mean(x=out)
-        fluid.backward.append_backward_ops(loss)
+        fluid.backward.append_backward(loss)
 
         cpu = fluid.CPUPlace()
         exe = fluid.Executor(cpu)
@@ -317,7 +317,7 @@ class TestSimpleMulWithMemory(unittest.TestCase):
         out = rnn()
         last = fluid.layers.sequence_pool(input=out, pool_type='last')
         loss = fluid.layers.mean(x=last)
-        fluid.backward.append_backward_ops(loss)
+        fluid.backward.append_backward(loss)
 
         cpu = fluid.CPUPlace()
         exe = fluid.Executor(cpu)
@@ -330,6 +330,7 @@ class TestSimpleMulWithMemory(unittest.TestCase):
                                         ],
                                         return_numpy=False))
         last_by_py, = py_rnn.exe().values()
+        print w_g[0]
         self.assertTrue(numpy.allclose(last_np, last_by_py))
         w_g_num = py_rnn.get_numeric_gradient_of_param(self.PARAM_NAME)
         # print w_g_num[0], w_g[0]