diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h
index 25e33861c06c9fcd2625e3a4036a04508acbd2ca..0d27be5fc007746d6ca41ff0dbcea5c5f45599ef 100644
--- a/paddle/fluid/framework/ir/graph.h
+++ b/paddle/fluid/framework/ir/graph.h
@@ -142,8 +142,6 @@ class Graph {
     nodes_.erase(node);
   }
 
-  const ProgramDesc &program() const { return program_; }
-
  private:
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {
@@ -154,7 +152,7 @@ class Graph {
   }
 
   // NOTE: program_ shouldn't be exposed to user.
-  const ProgramDesc &program_;
+  const ProgramDesc program_;
   std::map<std::string, boost::any> attrs_;
   std::map<std::string, std::function<void(void)>> attr_dels_;
   std::map<ir::Node *, std::unique_ptr<ir::Node>> nodes_;
diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h
index 063c70fb7b9c0f9b90d872a70f362459ef149391..63277d2d012873847d58ebeb0b3a6b175822809f 100644
--- a/paddle/fluid/framework/ir/node.h
+++ b/paddle/fluid/framework/ir/node.h
@@ -41,8 +41,7 @@ class Node {
   explicit Node(OpDesc* op_desc)
       : name_(op_desc->Type()),
         var_desc_(nullptr),
-        op_desc_(new OpDesc(*op_desc)),  // TODO(panyx0718) the pointer in the
-                                         // original OpDesc might go out.
+        op_desc_(new OpDesc(*op_desc, op_desc->Block())),
         type_(Type::kOperation) {}
 
   Type NodeType() const { return type_; }
diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc
index 2288c7fe6609a765612b468d69ad35101b92b384..9c289243c5a27839f628f3e143ce0363bf75a0b1 100644
--- a/paddle/fluid/framework/op_proto_maker.cc
+++ b/paddle/fluid/framework/op_proto_maker.cc
@@ -129,6 +129,10 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
                                     "Optimized for variable")
       .SetDefault({});
 
+  AddAttr<std::vector<std::string>>(OpCreationCallstackAttrName(),
+                                    "Callstack for Op Creatation.")
+      .SetDefault({});
+
   Validate();
 }
 
diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h
index 80970291c9c234f1306162f4ffa3c2528f88c35f..cb9c8ab1704ab867182079db31a34125669c645b 100644
--- a/paddle/fluid/framework/op_proto_maker.h
+++ b/paddle/fluid/framework/op_proto_maker.h
@@ -39,6 +39,7 @@ class OpProtoAndCheckerMaker {
  public:
   static const char *OpRoleAttrName() { return "op_role"; }
   static const char *OpRoleVarAttrName() { return "op_role_var"; }
+  static const char *OpCreationCallstackAttrName() { return "op_callstack"; }
 
   void operator()(proto::OpProto *proto, OpAttrChecker *attr_checker);
 
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index d04f7744961b2561977f4d36d0073a97557043da..9f8cdf1aeba43d30676cb2adf80a77cab86547a8 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -11,15 +11,17 @@ 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 <gflags/gflags.h>
-#include <glog/logging.h>
-
+#include "paddle/fluid/framework/operator.h"
 #include <algorithm>
-
+#include <sstream>
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
 #include "paddle/fluid/framework/data_transform.h"
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/shape_inference.h"
 #include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -127,19 +129,48 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
 }
 
 void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
-  VLOG(4) << place << " " << DebugStringEx(&scope);
-  if (platform::is_gpu_place(place)) {
+  try {
+    if (VLOG_IS_ON(4)) {
+      VLOG(4) << place << " " << DebugStringEx(&scope);
+    }
+    if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
-    PADDLE_THROW("Cannot run operator on place %s", place);
+      PADDLE_THROW("Cannot run operator on place %s", place);
 #else
-    auto dev_id = boost::get<platform::CUDAPlace>(place).device;
-    platform::SetDeviceId(dev_id);
+      auto dev_id = boost::get<platform::CUDAPlace>(place).device;
+      platform::SetDeviceId(dev_id);
 #endif
+    }
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::RecordEvent record_event(Type(), pool.Get(place));
+    RunImpl(scope, place);
+    if (VLOG_IS_ON(3)) {
+      VLOG(3) << place << " " << DebugStringEx(&scope);
+    }
+  } catch (platform::EnforceNotMet exception) {
+    if (Attrs().count("sub_block") != 0) {
+      throw exception;
+    }
+
+    auto& callstack = Attr<std::vector<std::string>>(
+        OpProtoAndCheckerMaker::OpCreationCallstackAttrName());
+
+    if (callstack.empty()) {
+      throw exception;
+    }
+    std::ostringstream sout;
+    sout << "Invoke operator " << Type() << " error.\n";
+    sout << "Python Callstacks: \n";
+    for (auto& line : callstack) {
+      sout << line;
+    }
+    sout << "C++ Callstacks: \n";
+    sout << exception.err_str_;
+    exception.err_str_ = sout.str();
+    throw exception;
+  } catch (...) {
+    std::rethrow_exception(std::current_exception());
   }
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-  platform::RecordEvent record_event(Type(), pool.Get(place));
-  RunImpl(scope, place);
-  VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 
 bool OperatorBase::HasInputs(const std::string& name) const {
@@ -167,7 +198,7 @@ const std::vector<std::string>& OperatorBase::Inputs(
 }
 
 bool OperatorBase::HasOutputs(const std::string& name) const {
-  if (outputs_.find(name) != outputs_.end()) {
+  if (outputs_.end() != outputs_.find(name)) {
     return true;
   } else {
     return false;
diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h
index a496301526f58875ff51aeaa5b2094c3c656531c..78be2e1e1f06c7a518e35a770c1dc9581b2d10fe 100644
--- a/paddle/fluid/operators/concat_op.h
+++ b/paddle/fluid/operators/concat_op.h
@@ -62,9 +62,21 @@ class ConcatGradKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const {
     auto* out_grad =
         ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
-    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto ins = ctx.MultiInput<framework::LoDTensor>("X");
     auto out_var_names = ctx.Outputs(framework::GradVarName("X"));
-    auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
+    auto outs =
+        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
+
+    {
+      auto dx = outs;
+      auto x = ins;
+      for (size_t i = 0; i < dx.size(); ++i) {
+        if (dx[i] != nullptr) {
+          dx[i]->set_lod(x[i]->lod());
+        }
+      }
+    }
+
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
 
     // get output tensor that the name is not kEmptyVarName
diff --git a/paddle/fluid/operators/elementwise_add_mkldnn_op.cc b/paddle/fluid/operators/elementwise_add_mkldnn_op.cc
index c86cd57316078778e5930c9b524b931d523028d7..9ad82aec8182d6ba06b67391d71317a3d0df1833 100644
--- a/paddle/fluid/operators/elementwise_add_mkldnn_op.cc
+++ b/paddle/fluid/operators/elementwise_add_mkldnn_op.cc
@@ -137,9 +137,10 @@ class EltwiseAddMKLDNNKernel : public framework::OpKernel<T> {
 };
 
 template <typename T>
-class EltwiseAddMKLDNNGradKernel : public framework::OpKernel<T> {
+class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
diff --git a/paddle/fluid/operators/elementwise_add_op.h b/paddle/fluid/operators/elementwise_add_op.h
index 5356105e2e551c0528694091608fc7585dce66d2..c60cb1f92e99329d52f6ed39dccde406a5f83563 100644
--- a/paddle/fluid/operators/elementwise_add_op.h
+++ b/paddle/fluid/operators/elementwise_add_op.h
@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
@@ -136,9 +137,11 @@ elementwise_add_grad(const framework::ExecutionContext& ctx,
 }
 
 template <typename DeviceContext, typename T>
-class ElementwiseAddGradKernel : public framework::OpKernel<T> {
+class ElementwiseAddGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
+
     using Tensor = framework::Tensor;
 
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
diff --git a/paddle/fluid/operators/elementwise_div_op.h b/paddle/fluid/operators/elementwise_div_op.h
index 95649ac46e6bd41b9e1a865794cdec3ae1e6e247..41a7950bf0c598507c0fda48c6a43f2fd38c41d2 100644
--- a/paddle/fluid/operators/elementwise_div_op.h
+++ b/paddle/fluid/operators/elementwise_div_op.h
@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
-
 namespace paddle {
 namespace operators {
 
@@ -53,9 +53,10 @@ struct DivGradDY {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseDivGradKernel : public framework::OpKernel<T> {
+class ElementwiseDivGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");
diff --git a/paddle/fluid/operators/elementwise_max_op.h b/paddle/fluid/operators/elementwise_max_op.h
index 527a18ee3ba88a158a13266a7fbcdafe59ec69d9..bfb5c931958b4ca890ea720af42dad91d5625abb 100644
--- a/paddle/fluid/operators/elementwise_max_op.h
+++ b/paddle/fluid/operators/elementwise_max_op.h
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 
 namespace paddle {
@@ -55,9 +56,10 @@ struct MaxGradDy {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseMaxGradKernel : public framework::OpKernel<T> {
+class ElementwiseMaxGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");
diff --git a/paddle/fluid/operators/elementwise_min_op.h b/paddle/fluid/operators/elementwise_min_op.h
index d4e5831463f3e54c72789b6876ea696cf1b4ef4b..db035ffb52e619b337c8190af4ed0e155aaac48d 100644
--- a/paddle/fluid/operators/elementwise_min_op.h
+++ b/paddle/fluid/operators/elementwise_min_op.h
@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
-
 namespace paddle {
 namespace operators {
 
@@ -55,9 +55,10 @@ struct MinGradDy {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseMinGradKernel : public framework::OpKernel<T> {
+class ElementwiseMinGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");
diff --git a/paddle/fluid/operators/elementwise_mul_op.h b/paddle/fluid/operators/elementwise_mul_op.h
index 329d2d129a9ea450cd211f0c6d2ea5e37ff8491d..4437da4d95f97b5cbbca1650badf9710c26b4380 100644
--- a/paddle/fluid/operators/elementwise_mul_op.h
+++ b/paddle/fluid/operators/elementwise_mul_op.h
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
@@ -84,9 +85,10 @@ struct MulGradDY {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseMulGradKernel : public framework::OpKernel<T> {
+class ElementwiseMulGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");
diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h
index d8a12e800ad733800c1ec333f15d31d4dcd1a3a5..a79b900b9801e6b80e4433a9acdd4dab6c34859d 100644
--- a/paddle/fluid/operators/elementwise_op.h
+++ b/paddle/fluid/operators/elementwise_op.h
@@ -205,6 +205,20 @@ class ElementwiseOpExplicitGrad : public ElementwiseOpGrad {
   }
 };
 
+template <typename T>
+class ElemwiseGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* dx =
+        context.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    if (dx != nullptr) {
+      auto& dout =
+          *context.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+      dx->set_lod(dout.lod());
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
diff --git a/paddle/fluid/operators/elementwise_sub_op.h b/paddle/fluid/operators/elementwise_sub_op.h
index 11c7e3fe628001f095836a788f2bcc7c4ee7ad4b..3385df0897700d37d60d8804a01db777ebc02a7e 100644
--- a/paddle/fluid/operators/elementwise_sub_op.h
+++ b/paddle/fluid/operators/elementwise_sub_op.h
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 
 namespace paddle {
@@ -50,9 +51,10 @@ struct SubGradDY {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseSubGradKernel : public framework::OpKernel<T> {
+class ElementwiseSubGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat.cc
index 55c8a472aca7fe700ef6a3f96bed1496d7b12b80..fbe7c2978385401b35765101c87387ff727be4e0 100644
--- a/paddle/fluid/operators/math/concat.cc
+++ b/paddle/fluid/operators/math/concat.cc
@@ -71,7 +71,7 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
-                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const std::vector<const framework::LoDTensor*>& ref_inputs,
                   const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
     size_t num = outputs->size();
diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat.cu
index 5863d74fca21de8b77bc208fb95d8fd52562f7a7..820e73e779720e4f76168e0a84a254ef645784ee 100644
--- a/paddle/fluid/operators/math/concat.cu
+++ b/paddle/fluid/operators/math/concat.cu
@@ -189,7 +189,7 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
-                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const std::vector<const framework::LoDTensor*>& ref_inputs,
                   const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
     int o_num = outputs->size();
diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat.h
index 9e080f2e8be23768dcea47b577043beef37b2eaf..e5d7d860b371677b3cfc67a57390bdee0d0ecc37 100644
--- a/paddle/fluid/operators/math/concat.h
+++ b/paddle/fluid/operators/math/concat.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 #include <vector>
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/lod_tensor.h"
 
 namespace paddle {
 namespace operators {
@@ -57,7 +57,7 @@ template <typename DeviceContext, typename T>
 class ConcatGradFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor& input,
-                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const std::vector<const framework::LoDTensor*>& ref_inputs,
                   const int axis, std::vector<framework::Tensor*>* outputs);
 };
 
diff --git a/paddle/fluid/operators/mul_op.h b/paddle/fluid/operators/mul_op.h
index 15dd975e3bbf80b2e616e6628555e812d025f70a..f72824806ed6ee3a4490938403d441326f8a3d4a 100644
--- a/paddle/fluid/operators/mul_op.h
+++ b/paddle/fluid/operators/mul_op.h
@@ -62,23 +62,31 @@ class MulGradKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     int x_num_col_dims = ctx.template Attr<int>("x_num_col_dims");
     int y_num_col_dims = ctx.template Attr<int>("y_num_col_dims");
-    const Tensor* x = ctx.Input<Tensor>("X");
-    const Tensor* y = ctx.Input<Tensor>("Y");
-    const Tensor x_matrix = x->dims().size() > 2
-                                ? framework::ReshapeToMatrix(*x, x_num_col_dims)
-                                : *x;
-    const Tensor y_matrix = y->dims().size() > 2
-                                ? framework::ReshapeToMatrix(*y, y_num_col_dims)
-                                : *y;
-    const Tensor* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto* y = ctx.Input<framework::LoDTensor>("Y");
+    auto x_matrix = x->dims().size() > 2
+                        ? framework::ReshapeToMatrix(*x, x_num_col_dims)
+                        : static_cast<const Tensor&>(*x);
+    auto y_matrix = y->dims().size() > 2
+                        ? framework::ReshapeToMatrix(*y, y_num_col_dims)
+                        : static_cast<const Tensor&>(*y);
+    auto* dout = ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
 
     Tensor dout_mat;
     dout_mat.ShareDataWith(*dout);
     dout_mat.Resize({framework::flatten_to_2d(x->dims(), x_num_col_dims)[0],
                      framework::flatten_to_2d(y->dims(), y_num_col_dims)[1]});
 
-    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto* dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+
+    if (dx != nullptr) {
+      dx->set_lod(x->lod());
+    }
+    if (dy != nullptr) {
+      dy->set_lod(y->lod());
+    }
+
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
     if (dx) {
diff --git a/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc b/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc
index 0ddacb57106c090e8f4f9350a65a30ca102f8e0a..7aca9f7111956dba63e2ceee10077879fe092bdf 100644
--- a/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc
+++ b/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc
@@ -68,7 +68,9 @@ class SequenceSoftmaxGradCUDNNKernel : public framework::OpKernel<T> {
     auto* out_grad = ctx.Input<LoDTensor>(framework::GradVarName("Out"));
     auto* x = ctx.Input<LoDTensor>("X");
     auto* x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
-
+    if (x_grad) {
+      x_grad->set_lod(x->lod());
+    }
     auto lod = x->lod();
     const size_t level = lod.size() - 1;
 
diff --git a/paddle/fluid/operators/sequence_softmax_op.h b/paddle/fluid/operators/sequence_softmax_op.h
index cb93a02b8386ed50ff176fc25b88449b7eb16902..bca564e16f9951519eefe25126aadebb4c1326b6 100644
--- a/paddle/fluid/operators/sequence_softmax_op.h
+++ b/paddle/fluid/operators/sequence_softmax_op.h
@@ -66,6 +66,9 @@ class SequenceSoftmaxGradKernel : public framework::OpKernel<T> {
     auto* out_grad = ctx.Input<LoDTensor>(framework::GradVarName("Out"));
     auto* x = ctx.Input<LoDTensor>("X");
     auto* x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    if (x_grad) {
+      x_grad->set_lod(x->lod());
+    }
 
     auto lod = x->lod();
     const size_t level = lod.size() - 1;
diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc
index 4a8ac441cfaf642fde58ee30865a22e83c065498..92a0697e27ba0da66fa3b0f5380e7bd52575640d 100644
--- a/paddle/fluid/operators/top_k_op.cc
+++ b/paddle/fluid/operators/top_k_op.cc
@@ -30,6 +30,8 @@ class TopkOp : public framework::OperatorWithKernel {
                    "Output(Indices) of TopkOp should not be null.");
 
     auto input_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_EQ(input_dims.size(), 2,
+                      "Rank of TopK op's input must be 2.");
     const int k = static_cast<int>(ctx->Attrs().Get<int>("k"));
 
     PADDLE_ENFORCE_GE(k, 1, "k must >= 1");
diff --git a/paddle/fluid/pybind/const_value.cc b/paddle/fluid/pybind/const_value.cc
index 76aa7d2010682416f68e982e9b89da9813abb078..9094f6051c2c695e41ed5882a688bb29e8a9c4fd 100644
--- a/paddle/fluid/pybind/const_value.cc
+++ b/paddle/fluid/pybind/const_value.cc
@@ -40,6 +40,9 @@ void BindConstValue(pybind11::module* m) {
   op_proto_and_checker_maker.def(
       "kOpRoleVarAttrName",
       framework::OpProtoAndCheckerMaker::OpRoleVarAttrName);
+  op_proto_and_checker_maker.def(
+      "kOpCreationCallstackAttrName",
+      framework::OpProtoAndCheckerMaker::OpCreationCallstackAttrName);
 }
 
 }  // namespace pybind
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index 62682d10324c7cfe656c9ddb09f1b61ac1772e69..389fce18741547ab1357018b8fb2a593f6bd35b0 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -18,6 +18,7 @@ import collections
 import contextlib
 import re
 import six
+import traceback
 
 import numpy as np
 
@@ -499,6 +500,10 @@ class Operator(object):
         if role_var_name in op_attrs and len(op_attrs[role_var_name]) == 0:
             del op_attrs[role_var_name]
 
+        callstack_var_name = op_maker.kOpCreationCallstackAttrName()
+        op_attrs[callstack_var_name] = list(
+            reversed(traceback.format_stack()))[1:]
+
         if len(self.desc.type()) != 0:
             return
         if type is None:
diff --git a/python/paddle/fluid/tests/unittests/test_operator_desc.py b/python/paddle/fluid/tests/unittests/test_operator_desc.py
index 6d01955993324498de42462b7f85ef6f8e444505..3ac82680733feb4b82ab98669269160e4aad948f 100644
--- a/python/paddle/fluid/tests/unittests/test_operator_desc.py
+++ b/python/paddle/fluid/tests/unittests/test_operator_desc.py
@@ -67,7 +67,10 @@ class TestOperator(unittest.TestCase):
         self.assertEqual(mul_op.output("Out"), ["mul.out"])
         self.assertEqual(
             set(mul_op.attr_names),
-            set(["x_num_col_dims", "y_num_col_dims", "op_role", "op_role_var"]))
+            set([
+                "x_num_col_dims", "y_num_col_dims", "op_role", "op_role_var",
+                "op_callstack"
+            ]))
         self.assertEqual(mul_op.has_attr("x_num_col_dims"), True)
         self.assertEqual(mul_op.attr_type("x_num_col_dims"), core.AttrType.INT)
         self.assertEqual(mul_op.attr("x_num_col_dims"), 1)