diff --git a/paddle/fluid/eager/accumulation/accumulation_node.cc b/paddle/fluid/eager/accumulation/accumulation_node.cc
index 8b6752dfec743d774c8656d36421f31e7dab9799..9b0e784c0efb173f0957516d9abd2a0a2a73b82d 100644
--- a/paddle/fluid/eager/accumulation/accumulation_node.cc
+++ b/paddle/fluid/eager/accumulation/accumulation_node.cc
@@ -47,6 +47,7 @@ void GradNodeAccumulation::RetainGrad(
 std::vector<std::vector<paddle::experimental::Tensor>> GradNodeAccumulation::
 operator()(
     const std::vector<std::vector<paddle::experimental::Tensor>>& grads) {
+  VLOG(3) << "Running Eager Backward Node: GradNodeAccumulation";
   PADDLE_ENFORCE(grads.size() == 1,
                  paddle::platform::errors::Fatal(
                      "GradNodeAccumulation should take exactly 1 grad tensor"
diff --git a/paddle/fluid/eager/accumulation/accumulation_node.h b/paddle/fluid/eager/accumulation/accumulation_node.h
index be2ccc263e806d8874b0e18f93376bd62745940c..3f53517204a5a109226265b2df463d5e0ffa3eaf 100644
--- a/paddle/fluid/eager/accumulation/accumulation_node.h
+++ b/paddle/fluid/eager/accumulation/accumulation_node.h
@@ -35,6 +35,8 @@ class GradNodeAccumulation : public GradNodeBase {
 
   paddle::experimental::Tensor* Grad() { return &accumulated_grad; }
 
+  std::string name() { return "GradNodeAccumulation"; }
+
   /**
    * Register ReduceHook
    * **/
diff --git a/paddle/fluid/eager/api/utils/hook_utils.cc b/paddle/fluid/eager/api/utils/hook_utils.cc
index ee6a3afc6ffd39e264b039adaf8bae716c9e483e..7d2997eb884c86c247ec29aae2af2d2e2f94cc21 100644
--- a/paddle/fluid/eager/api/utils/hook_utils.cc
+++ b/paddle/fluid/eager/api/utils/hook_utils.cc
@@ -70,12 +70,8 @@ void RetainGradForTensor(const paddle::experimental::Tensor& tensor) {
             grad_tensor->set_impl(t.impl());
             return *grad_tensor.get();
           } else {
-            PADDLE_THROW(paddle::platform::errors::Fatal(
-                "Detected uninitialized variable, causing segmentation "
-                "fault "
-                "inside the hook."
-                "Tensor has to be initialized while we need to set it."
-                "please check tensor initialization status."));
+            VLOG(7) << "Retain NULL paddle::experimental::Tensor in Grad Hook";
+            return paddle::experimental::Tensor();
           }
         } else {
           VLOG(7) << "Retain NULL paddle::experimental::Tensor in Grad Hook";
diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
index 63f25f5528100cccdacf7c1c1ca67095b448160c..7cddfd9c1c7dc11c01f820b56673f2a8f995f722 100644
--- a/paddle/fluid/eager/auto_code_generator/eager_generator.cc
+++ b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
@@ -554,6 +554,21 @@ static bool CheckOpProto(proto::OpProto* op_proto) {
   return true;
 }
 
+static bool BeSameAsInput(const std::string& output_name,
+                          const std::set<std::string>& input_names) {
+  if (output_name.size() < 4) {
+    return false;
+  }
+
+  if (output_name.substr(output_name.size() - 3, 3) == "Out") {
+    if (input_names.count(output_name.substr(0, output_name.size() - 3))) {
+      return true;
+    }
+  }
+
+  return false;
+}
+
 /* --------------------------------------- */
 /* --------- Preprocess Ins/Outs --------- */
 /* --------------------------------------- */
@@ -1022,27 +1037,12 @@ static std::string GenerateGradNodeCreationContent(
           "egr::EagerUtils::autograd_meta(&%s);\n";
       get_autograd_meta_str += paddle::string::Sprintf(
           GET_MULTI_AUTOGRAD_META_TEMPLATE, output_autograd_name, output_name);
-      if (op_passing_outs_map[op_type].count(output_name)) {
-        const std::string output_var_args_name = output_name + "Var";
-        const char* FWD_OUT_SYNC_BACK_TEMPLATE =
-            "  egr::EagerUtils::OverwriteOutputs(%s, %s);\n";
-        get_autograd_meta_str += paddle::string::Sprintf(
-            FWD_OUT_SYNC_BACK_TEMPLATE, output_name, output_var_args_name);
-      }
     } else {
       const char* GET_SINGLE_AUTOGRAD_META_TEMPLATE =
           "  egr::AutogradMeta* %s = "
           "egr::EagerUtils::autograd_meta(&%s);\n";
       get_autograd_meta_str += paddle::string::Sprintf(
           GET_SINGLE_AUTOGRAD_META_TEMPLATE, output_autograd_name, output_name);
-
-      if (op_passing_outs_map[op_type].count(output_name)) {
-        const std::string output_var_args_name = output_name + "Var";
-        const char* FWD_OUT_SYNC_BACK_TEMPLATE =
-            "  egr::EagerUtils::OverwriteOutputs(%s, %s);\n";
-        get_autograd_meta_str += paddle::string::Sprintf(
-            FWD_OUT_SYNC_BACK_TEMPLATE, output_name, output_var_args_name);
-      }
     }
   }
   VLOG(6) << "Generated outputs autograd_meta";
@@ -1180,11 +1180,13 @@ static std::string GenerateGradNodeCreationContent(
           SET_GRAD_IN_META_TEMPLATE, output_autograd_name, output_position);
     }
 
-    VLOG(6) << "Generated Call RetainGradForTensor";
-    const char* RETAIN_GRAD_TEMPLATE =
-        "    egr::EagerUtils::CheckAndRetainGrad(%s);\n";
-    grad_node_creation_str +=
-        paddle::string::Sprintf(RETAIN_GRAD_TEMPLATE, output_name);
+    if (!output.intermediate()) {
+      VLOG(6) << "Generated Call RetainGradForTensor";
+      const char* RETAIN_GRAD_TEMPLATE =
+          "    egr::EagerUtils::CheckAndRetainGrad(%s);\n";
+      grad_node_creation_str +=
+          paddle::string::Sprintf(RETAIN_GRAD_TEMPLATE, output_name);
+    }
   }
   VLOG(6) << "Generated SetGradIn/OutMeta";
 
@@ -1324,19 +1326,21 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
   generated_function_body += "\n";
 
   // Handle Dispensable Inputs
+  std::set<std::string> input_names;
   for (const proto::OpProto::Var& input : in_vars) {
     const std::string& input_name = input.name();
+    input_names.insert(input_name);
     if (input.dispensable()) {
       if (input.duplicable()) {
         const char* FWD_INS_CONTENT_TEMPLATE =
             "  if(%s.size() > 0) "
-            "ins[\"%s\"] = egr::EagerUtils::TrySyncToVars(%s)\n;";
+            "ins[\"%s\"] = egr::EagerUtils::TrySyncToVars(%s);\n";
         generated_function_body += paddle::string::Sprintf(
             FWD_INS_CONTENT_TEMPLATE, input_name, input_name, input_name);
       } else {
         const char* FWD_INS_CONTENT_TEMPLATE =
             "  if(%s.initialized()) "
-            "ins[\"%s\"] = egr::EagerUtils::TrySyncToVars(%s)\n;";
+            "ins[\"%s\"] = egr::EagerUtils::TrySyncToVars(%s);\n";
         generated_function_body += paddle::string::Sprintf(
             FWD_INS_CONTENT_TEMPLATE, input_name, input_name, input_name);
       }
@@ -1372,11 +1376,21 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
 
         core_ops_args_type_info[op_type].push_back("tensor");
       }
-      const char* FWD_OUTS_CONTENT_TEMPLATE =
-          "{ \"%s\", egr::EagerUtils::TrySyncToVars(%s) },";
-      outs_contents_str += paddle::string::Sprintf(
-          FWD_OUTS_CONTENT_TEMPLATE, output_name, output_var_name);
 
+      if (BeSameAsInput(output_name, input_names)) {
+        if (!output.dispensable()) {
+          std::string input_name =
+              output_name.substr(0, output_name.size() - 3);
+          const char* FWD_OUTS_CONTENT_TEMPLATE = "{ \"%s\", ins[\"%s\"] },";
+          outs_contents_str += paddle::string::Sprintf(
+              FWD_OUTS_CONTENT_TEMPLATE, output_name, input_name);
+        }
+      } else {
+        const char* FWD_OUTS_CONTENT_TEMPLATE =
+            "{ \"%s\", egr::EagerUtils::TrySyncToVars(%s) },";
+        outs_contents_str += paddle::string::Sprintf(
+            FWD_OUTS_CONTENT_TEMPLATE, output_name, output_var_name);
+      }
       core_ops_args_info[op_type].push_back(output_var_name);
 
     } else {
@@ -1415,6 +1429,23 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
   generated_function_body += outs_map_str;
   generated_function_body += "\n";
 
+  for (const proto::OpProto::Var& output : out_vars) {
+    const std::string& output_name = output.name();
+    if (op_passing_outs_map[op_type].count(output_name)) {
+      if (BeSameAsInput(output_name, input_names)) {
+        if (output.dispensable()) {
+          std::string input_name =
+              output_name.substr(0, output_name.size() - 3);
+          const char* FWD_OUTS_CONTENT_TEMPLATE =
+              "  if (ins.count(\"%s\")) outs[\"%s\"] = ins[\"%s\"];\n";
+          generated_function_body += paddle::string::Sprintf(
+              FWD_OUTS_CONTENT_TEMPLATE, input_name, output_name, input_name);
+        }
+      }
+    }
+  }
+  generated_function_body += "\n";
+
   VLOG(6) << "Generated Outs Map";
 
   // [Generation] Get Attrs
@@ -1448,33 +1479,61 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
     std::string output_varname = LegalizeVariableName(output_name);
 
     if (output.duplicable()) {
-      const char* FWD_OUT_TENSORS_TEMPLATE =
-          "  std::vector<paddle::experimental::Tensor> %s = "
-          "egr::EagerUtils::GetOutputs(outs[\"%s\"]);\n";
-      out_tensor_str = paddle::string::Sprintf(FWD_OUT_TENSORS_TEMPLATE,
-                                               output_varname, output_name);
+      if (op_passing_outs_map[op_type].count(output_name)) {
+        if (output.dispensable()) {
+          const char* FWD_OUT_TENSORS_TEMPLATE =
+              "  std::vector<paddle::experimental::Tensor> %s;\n"
+              "  if (outs.count(\"%s\"))  "
+              "egr::EagerUtils::GetOutputs(outs[\"%s\"], %s);\n"
+              "  egr::EagerUtils::Output2Result(%s, &%s);\n";
+          out_tensor_str = paddle::string::Sprintf(
+              FWD_OUT_TENSORS_TEMPLATE, output_varname, output_name,
+              output_name, output_var_args_name, output_var_args_name,
+              output_varname);
+        } else {
+          const char* FWD_OUT_TENSORS_TEMPLATE =
+              "  std::vector<paddle::experimental::Tensor> %s;\n"
+              "  egr::EagerUtils::GetOutputs(outs[\"%s\"], %s);\n"
+              "  egr::EagerUtils::Output2Result(%s, &%s);\n";
+          out_tensor_str = paddle::string::Sprintf(
+              FWD_OUT_TENSORS_TEMPLATE, output_varname, output_name,
+              output_var_args_name, output_var_args_name, output_varname);
+        }
+      } else {
+        const char* FWD_OUT_TENSORS_TEMPLATE =
+            "  std::vector<paddle::experimental::Tensor> %s;\n"
+            "  egr::EagerUtils::GetOutputs(outs[\"%s\"], &%s);\n";
+        out_tensor_str =
+            paddle::string::Sprintf(FWD_OUT_TENSORS_TEMPLATE, output_varname,
+                                    output_name, output_varname);
+      }
       return_types[return_position] =
           "std::vector<paddle::experimental::Tensor>";
-      if (op_passing_outs_map[op_type].count(output_name) &&
-          bwd_info.GenerateForwardOnly()) {
-        const char* FWD_OUT_SYNC_BACK_TEMPLATE =
-            "  egr::EagerUtils::OverwriteOutputs(outs[\"%s\"], %s);\n";
-        out_tensor_str += paddle::string::Sprintf(
-            FWD_OUT_SYNC_BACK_TEMPLATE, output_name, output_var_args_name);
-      }
     } else {
-      const char* FWD_OUT_TENSOR_TEMPLATE =
-          "  paddle::experimental::Tensor %s = "
-          "egr::EagerUtils::GetOutput(outs[\"%s\"][0]);\n";
-      out_tensor_str = paddle::string::Sprintf(FWD_OUT_TENSOR_TEMPLATE,
-                                               output_varname, output_name);
-
-      if (op_passing_outs_map[op_type].count(output_name) &&
-          bwd_info.GenerateForwardOnly()) {
-        const char* FWD_OUT_SYNC_BACK_TEMPLATE =
-            "  egr::EagerUtils::OverwriteOutputs(outs[\"%s\"][0], %s);\n";
-        out_tensor_str += paddle::string::Sprintf(
-            FWD_OUT_SYNC_BACK_TEMPLATE, output_name, output_var_args_name);
+      if (op_passing_outs_map[op_type].count(output_name)) {
+        if (output.dispensable()) {
+          const char* FWD_OUT_TENSOR_TEMPLATE =
+              "  if (outs.count(\"%s\"))  "
+              "egr::EagerUtils::GetOutput(outs[\"%s\"][0], %s);\n"
+              "  paddle::experimental::Tensor& %s = *%s;\n";
+          out_tensor_str = paddle::string::Sprintf(
+              FWD_OUT_TENSOR_TEMPLATE, output_name, output_name,
+              output_var_args_name, output_varname, output_var_args_name);
+        } else {
+          const char* FWD_OUT_TENSOR_TEMPLATE =
+              "  egr::EagerUtils::GetOutput(outs[\"%s\"][0], %s);\n"
+              "  paddle::experimental::Tensor& %s = *%s;\n";
+          out_tensor_str = paddle::string::Sprintf(
+              FWD_OUT_TENSOR_TEMPLATE, output_name, output_var_args_name,
+              output_varname, output_var_args_name);
+        }
+      } else {
+        const char* FWD_OUT_TENSOR_TEMPLATE =
+            "  paddle::experimental::Tensor %s;\n"
+            "  egr::EagerUtils::GetOutput(outs[\"%s\"][0], &%s);\n";
+        out_tensor_str =
+            paddle::string::Sprintf(FWD_OUT_TENSOR_TEMPLATE, output_varname,
+                                    output_name, output_varname);
       }
       return_types[return_position] = "paddle::experimental::Tensor";
     }
@@ -1494,6 +1553,7 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
         GenerateGradNodeCreationContent(fwd_info, bwd_info);
     generated_function_body += grad_node_creation_body_str;
     generated_function_body += "\n";
+
     // [Generation] Call RetainGradForTensor
     VLOG(6) << "Generated GradNode Creation codes";
   }
@@ -1588,12 +1648,25 @@ static std::string GenerateSingleOpBase(
   const std::string& attrs_name = "attrs_map" + std::to_string(*outs_size);
 
   // [Generation] Get Ins Map
+  std::unordered_set<std::string> dispensable_input_name_set;
+  for (const auto& in : in_vars) {
+    if (in.dispensable()) dispensable_input_name_set.insert(in.name());
+  }
+  std::unordered_set<std::string> duplicable_input_name_set;
+  for (const auto& in : in_vars) {
+    if (in.duplicable()) duplicable_input_name_set.insert(in.name());
+  }
   std::string ins_contents_str = "";
   for (auto iter : grad_ins) {
     const std::string& grad_input_name = iter.first;
 
     if (grad_ins_fwd_slotname_map.count(grad_input_name)) {
       // Fwd Tensor
+      const std::string& fwd_name =
+          grad_ins_fwd_slotname_map.at(grad_input_name);
+      if (dispensable_input_name_set.count(fwd_name)) {
+        continue;
+      }
       std::string struct_fwd_input_name =
           grad_ins_fwd_slotname_map.at(grad_input_name) + "_";
       const char* GRAD_INS_FWD_CONTENT_TEMPLATE =
@@ -1634,14 +1707,41 @@ static std::string GenerateSingleOpBase(
       paddle::string::Sprintf(BWD_INS_MAP_TEMPLATE, ins_name, ins_contents_str);
   generated_grad_function_body += ins_map_str;
 
-  VLOG(6) << "Generated Ins Map";
+  for (auto iter : grad_ins) {
+    const std::string& grad_input_name = iter.first;
 
-  // [Generation] Get Outs Map
-  std::unordered_set<std::string> duplicable_input_name_set;
-  for (const auto& in : in_vars) {
-    if (in.duplicable()) duplicable_input_name_set.insert(in.name());
+    if (grad_ins_fwd_slotname_map.count(grad_input_name)) {
+      // Fwd Tensor
+      const std::string& fwd_name =
+          grad_ins_fwd_slotname_map.at(grad_input_name);
+      if (dispensable_input_name_set.count(fwd_name)) {
+        std::string struct_fwd_input_name =
+            grad_ins_fwd_slotname_map.at(grad_input_name) + "_";
+        if (duplicable_input_name_set.count(fwd_name)) {
+          const char* DISPENSABLE_GRAD_INS_FWD_CONTENT_TEMPLATE =
+              "  if(this->%s.size() > 0) %s[\"%s\"] = "
+              "egr::EagerUtils::TrySyncToVars(egr::EagerUtils::"
+              "RecoverTensorWrapper(&this->%s, nullptr));\n";
+          generated_grad_function_body += paddle::string::Sprintf(
+              DISPENSABLE_GRAD_INS_FWD_CONTENT_TEMPLATE, struct_fwd_input_name,
+              ins_name, grad_input_name, struct_fwd_input_name);
+        } else {
+          const char* DISPENSABLE_GRAD_INS_FWD_CONTENT_TEMPLATE =
+              "  auto %s = egr::EagerUtils::RecoverTensorWrapper(&this->%s, "
+              "nullptr);\n  if(%s.initialized()) %s[\"%s\"] = "
+              "egr::EagerUtils::TrySyncToVars(%s);\n";
+          generated_grad_function_body += paddle::string::Sprintf(
+              DISPENSABLE_GRAD_INS_FWD_CONTENT_TEMPLATE, grad_input_name,
+              struct_fwd_input_name, grad_input_name, ins_name, grad_input_name,
+              grad_input_name);
+        }
+      }
+    }
   }
 
+  VLOG(6) << "Generated Ins Map";
+
+  // [Generation] Get Outs Map
   std::string outs_contents_str = "";
   for (auto iter : grad_outs) {
     const std::string& grad_output_name = iter.first;
@@ -1987,6 +2087,7 @@ static std::string GenerateGradNodeHeaderContents(
       "%s\n"
       "  // SetAttrMap\n"
       "%s\n"
+      "  std::string name() { return \"GradNode%s\"; }\n"
       "\n"
       " private:\n"
       "   // TensorWrappers\n"
@@ -2085,8 +2186,8 @@ static std::string GenerateGradNodeHeaderContents(
 
   std::string grad_node_str = paddle::string::Sprintf(
       GRAD_NODE_TEMPLATE, op_type, op_type, op_type, op_type,
-      set_tensor_wrappers_str, set_attr_map_str, tensor_wrapper_members_str,
-      attr_members_str);
+      set_tensor_wrappers_str, set_attr_map_str, op_type,
+      tensor_wrapper_members_str, attr_members_str);
 
   return grad_node_str;
 }
diff --git a/paddle/fluid/eager/autograd_meta.h b/paddle/fluid/eager/autograd_meta.h
index f4b2b8e08d4fa465c1c3d868659d69f55c4223ea..53f17a4ffe58c586553c9699c610d09df7271808 100644
--- a/paddle/fluid/eager/autograd_meta.h
+++ b/paddle/fluid/eager/autograd_meta.h
@@ -127,6 +127,12 @@ class AutogradMeta : public AbstractAutogradMeta {
     stop_gradient_ = static_cast<int>(stop_gradient);
   }
 
+  void WeakSetStopGradient(bool stop_gradient) {
+    if (stop_gradient_ == -1) {
+      stop_gradient_ = static_cast<int>(stop_gradient);
+    }
+  }
+
   bool Persistable() const { return persistable_; }
 
   void SetPersistable(bool persistable) { persistable_ = persistable; }
diff --git a/paddle/fluid/eager/grad_node_info.cc b/paddle/fluid/eager/grad_node_info.cc
index 598b368c6426a0fde9b286fc92d8e5a01660ef0a..d83fa916db66c3aa5461230e2d5eaaa5557dd451 100644
--- a/paddle/fluid/eager/grad_node_info.cc
+++ b/paddle/fluid/eager/grad_node_info.cc
@@ -69,13 +69,16 @@ void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
           "adj_edges is designed to has the same size of grad "
           "inputs's slot num."));
   if (meta && !meta->StopGradient()) {
-    VLOG(6) << "Add Edges for slot: " << slot_id;
     auto node = meta->GetMutableGradNode();
     if (node) {
+      VLOG(6) << "Add Edges for slot: " << slot_id << ", the Edge is from "
+              << this->name() << " to " << meta->GetMutableGradNode()->name();
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                        meta->OutRankInfo());
     } else {
       meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
+      VLOG(6) << "Add Edges for slot: " << slot_id << ", the Edge is from "
+              << this->name() << " to " << meta->GetMutableGradNode()->name();
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                        meta->OutRankInfo());
     }
diff --git a/paddle/fluid/eager/grad_node_info.h b/paddle/fluid/eager/grad_node_info.h
index 8603d84fe8df597a69f041e2fec41d05dfe16448..f699f9ab28e2d37c893e7a4fdec9acfa6c5a280f 100644
--- a/paddle/fluid/eager/grad_node_info.h
+++ b/paddle/fluid/eager/grad_node_info.h
@@ -147,6 +147,8 @@ class GradNodeBase {
   std::vector<std::vector<paddle::experimental::Tensor>> ApplyGradientHooks(
       const std::vector<std::vector<paddle::experimental::Tensor>>& tensors);
 
+  virtual std::string name() { return "GradNodeBase"; }
+
  private:
   // TODO(jiabin): Use SmallVector instead after merge PR from develop
 
diff --git a/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc b/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc
index 990f700056158fabe2314aa2f5bc9946c0e5076c..217055e4e9e4a19e695f42bf57c2331c9b98e2bd 100644
--- a/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc
+++ b/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc
@@ -159,7 +159,7 @@ TEST(EagerUtils, PassStopGradient) {
   CHECK(auto_grad0->StopGradient() == false);
   egr::EagerUtils::PassStopGradient(true, auto_grad0.get(), auto_grad1.get(),
                                     auto_grad2.get(), auto_grad3.get());
-  CHECK(auto_grad0->StopGradient() == true);
+  CHECK(auto_grad0->StopGradient() == false);
   CHECK(auto_grad1->StopGradient() == true);
   CHECK(auto_grad2->StopGradient() == true);
   CHECK(auto_grad3->StopGradient() == true);
diff --git a/paddle/fluid/eager/utils.cc b/paddle/fluid/eager/utils.cc
index 9c6c7d4d540c6b4a42b5bb9b266d7175c55b15ad..5d8dff5cd5b245720d88104c50f66b7ddb21f32d 100644
--- a/paddle/fluid/eager/utils.cc
+++ b/paddle/fluid/eager/utils.cc
@@ -220,53 +220,62 @@ paddle::experimental::Tensor EagerUtils::GetOutput(
   return paddle::experimental::Tensor(out->GetTensorBase(), out->name());
 }
 
-void EagerUtils::OverwriteOutputs(const std::shared_ptr<EagerVariable>& out,
-                                  paddle::experimental::Tensor* tensor) {
+void EagerUtils::GetOutput(const std::shared_ptr<EagerVariable>& out,
+                           paddle::experimental::Tensor* out_var) {
   PADDLE_ENFORCE_NOT_NULL(
-      tensor, paddle::platform::errors::Fatal(
-                  "Tensor is null and cannot be copied. "
-                  "We are tring to OverwriteOutput from its "
-                  "shared_ptr, this error may indicate some outputs "
-                  "are nullptr"));
-  tensor->set_impl(out->GetTensorBase());
+      out_var, paddle::platform::errors::Fatal(
+                   "Tensor is null and cannot be copied. "
+                   "We are tring to OverwriteOutput from its "
+                   "shared_ptr, this error may indicate some outputs "
+                   "are nullptr"));
+  out_var->set_impl(out->GetTensorBase());
 }
 
-void EagerUtils::OverwriteOutputs(
+void EagerUtils::GetOutputs(
     const std::vector<std::shared_ptr<EagerVariable>>& outs,
-    const std::vector<paddle::experimental::Tensor*>& tensors) {
-  PADDLE_ENFORCE_EQ(
-      outs.size(), tensors.size(),
-      paddle::platform::errors::Fatal(
-          "We are tring to OverwriteOutputs which passed in and it expected "
-          "elements num of outs and origin outputs are equal, but we got outs "
-          "size of: %d, and tensors passed in size is: %d",
-          outs.size(), tensors.size()));
+    std::vector<paddle::experimental::Tensor>* result) {
   for (size_t i = 0; i < outs.size(); i++) {
-    OverwriteOutputs(outs[i], tensors[i]);
+    result->emplace_back(outs[i]->GetTensorBase());
   }
 }
 
-void EagerUtils::OverwriteOutputs(const paddle::experimental::Tensor& out,
-                                  paddle::experimental::Tensor* tensor) {
-  PADDLE_ENFORCE_NOT_NULL(
-      tensor, paddle::platform::errors::Fatal(
-                  "Tensor is null and cannot be copied. "
-                  "We are tring to OverwriteOutput from its "
-                  "shared_ptr, this error may indicate some outputs "
-                  "are nullptr"));
-  *tensor = out;
-}
-void EagerUtils::OverwriteOutputs(
-    const std::vector<paddle::experimental::Tensor>& outs,
-    const std::vector<paddle::experimental::Tensor*>& tensors) {
+void EagerUtils::GetOutputs(
+    const std::vector<std::shared_ptr<EagerVariable>>& outs,
+    const std::vector<paddle::experimental::Tensor*>& out_var) {
   for (size_t i = 0; i < outs.size(); i++) {
     PADDLE_ENFORCE_NOT_NULL(
-        tensors[i], paddle::platform::errors::Fatal(
+        out_var[i], paddle::platform::errors::Fatal(
                         "Tensor is null and cannot be copied. "
                         "We are tring to OverwriteOutput from its "
                         "shared_ptr, this error may indicate some outputs "
                         "are nullptr"));
-    *tensors[i] = outs[i];
+    out_var[i]->set_impl(outs[i]->GetTensorBase());
+  }
+}
+
+void EagerUtils::GetOutputs(const std::shared_ptr<EagerVariable>& out,
+                            std::vector<paddle::experimental::Tensor>* result) {
+  result->emplace_back(out->GetTensorBase());
+}
+
+void EagerUtils::GetOutputs(
+    const std::shared_ptr<EagerVariable>& out,
+    const std::vector<paddle::experimental::Tensor*>& out_var) {
+  PADDLE_ENFORCE_NOT_NULL(
+      out_var[0], paddle::platform::errors::Fatal(
+                      "Tensor is null and cannot be copied. "
+                      "We are tring to OverwriteOutput from its "
+                      "shared_ptr, this error may indicate some outputs "
+                      "are nullptr"));
+  out_var[0]->set_impl(out->GetTensorBase());
+}
+
+void EagerUtils::Output2Result(
+    const std::vector<paddle::experimental::Tensor*>& out_var,
+    std::vector<paddle::experimental::Tensor>* result) {
+  result->reserve(out_var.size());
+  for (size_t i = 0; i < out_var.size(); i++) {
+    result->emplace_back(*out_var[i]);
   }
 }
 
diff --git a/paddle/fluid/eager/utils.h b/paddle/fluid/eager/utils.h
index 00013faa345e213a125a2fe6c70eef1e0b9160ef..b74d68db2a6d5022735e8774e77d783c3044db9e 100644
--- a/paddle/fluid/eager/utils.h
+++ b/paddle/fluid/eager/utils.h
@@ -77,7 +77,7 @@ class PassStopGradientIter : public IterHelper<AutogradMeta*> {
       VLOG(2) << "Tensor is NULL";
       return;
     }
-    element->SetStopGradient(stop_gradient_);
+    element->WeakSetStopGradient(stop_gradient_);
   }
 
   bool stop_gradient_ = true;
@@ -173,17 +173,24 @@ class EagerUtils {
       const std::vector<std::shared_ptr<EagerVariable>>& outs);
   static paddle::experimental::Tensor GetOutput(
       const std::shared_ptr<EagerVariable>& out);
-  // Sync Back to origin output Tensor
-  static void OverwriteOutputs(const std::shared_ptr<EagerVariable>& out,
-                               paddle::experimental::Tensor* tensor);
-  static void OverwriteOutputs(const paddle::experimental::Tensor& out,
-                               paddle::experimental::Tensor* tensor);
-  static void OverwriteOutputs(
+  static void GetOutput(const std::shared_ptr<EagerVariable>& out,
+                        paddle::experimental::Tensor* out_var);
+  static void GetOutputs(
       const std::vector<std::shared_ptr<EagerVariable>>& outs,
-      const std::vector<paddle::experimental::Tensor*>& tensors);
-  static void OverwriteOutputs(
-      const std::vector<paddle::experimental::Tensor>& outs,
-      const std::vector<paddle::experimental::Tensor*>& tensors);
+      std::vector<paddle::experimental::Tensor>* result);
+  static void GetOutputs(
+      const std::vector<std::shared_ptr<EagerVariable>>& outs,
+      const std::vector<paddle::experimental::Tensor*>& out_var);
+  static void GetOutputs(const std::shared_ptr<EagerVariable>& out,
+                         std::vector<paddle::experimental::Tensor>* result);
+  static void GetOutputs(
+      const std::shared_ptr<EagerVariable>& out,
+      const std::vector<paddle::experimental::Tensor*>& out_var);
+
+  static void Output2Result(
+      const std::vector<paddle::experimental::Tensor*>& out_var,
+      std::vector<paddle::experimental::Tensor>* result);
+
   // end Intermidate needed
 
   static void CheckAndRetainGrad(const paddle::experimental::Tensor& tensor);
diff --git a/paddle/fluid/pybind/eager.cc b/paddle/fluid/pybind/eager.cc
index 241e9f9058dfe35fa36df12515e3ffc1a2f38a6b..3867336764834121e64c571d853506f3e57911f3 100644
--- a/paddle/fluid/pybind/eager.cc
+++ b/paddle/fluid/pybind/eager.cc
@@ -50,7 +50,6 @@ PyObject* TensorNew(PyTypeObject* type, PyObject* args, PyObject* kwargs) {
   if (obj) {
     auto v = reinterpret_cast<TensorObject*>(obj);
     new (&(v->tensor)) paddle::experimental::Tensor();
-    Py_INCREF(obj);
   }
   return obj;
 }
@@ -58,7 +57,7 @@ PyObject* TensorNew(PyTypeObject* type, PyObject* args, PyObject* kwargs) {
 // TODO(jiabin): Overload this once we need more constructor in Python
 void EmptyTensorInitializer(TensorObject* self, const std::string& name,
                             const paddle::platform::Place& place,
-                            bool persistable = false, bool stop_gradient = true,
+                            bool persistable = false, int stop_gradient = -1,
                             framework::proto::VarType::Type dtype =
                                 paddle::framework::proto::VarType::FP32,
                             const std::vector<int>& dims = {},
@@ -74,7 +73,9 @@ void EmptyTensorInitializer(TensorObject* self, const std::string& name,
   self->tensor.set_name(name);
   auto autograd_meta = egr::EagerUtils::autograd_meta(&(self->tensor));
   autograd_meta->SetPersistable(persistable);
-  autograd_meta->SetStopGradient(stop_gradient);
+  if (stop_gradient != -1) {
+    autograd_meta->SetStopGradient(static_cast<bool>(stop_gradient));
+  }
   if (var_type == paddle::framework::proto::VarType::LOD_TENSOR) {
     // TODO(jiabin): Maybe support LOD later
     std::shared_ptr<phi::DenseTensor> dense_tensor =
@@ -143,13 +144,12 @@ void InitTensorWithTensor(TensorObject* self,
         src.copy_to(phi::TransToPtenBackend(place), true).impl());
     VLOG(4) << "Different place, do TensorCopy";
   }
-  egr::EagerUtils::autograd_meta(&(self->tensor))->SetStopGradient(true);
   if (src.get_autograd_meta()) {
-    egr::EagerUtils::unsafe_autograd_meta(self->tensor)
+    egr::EagerUtils::autograd_meta(&(self->tensor))
         ->SetPersistable(
             egr::EagerUtils::unsafe_autograd_meta(src)->Persistable());
   } else {
-    egr::EagerUtils::unsafe_autograd_meta(self->tensor)->SetPersistable(false);
+    egr::EagerUtils::autograd_meta(&(self->tensor))->SetPersistable(false);
   }
 }
 
@@ -168,8 +168,7 @@ void InitTensorWithFrameworkTensor(TensorObject* self,
         temp.copy_to(phi::TransToPtenBackend(place), true).impl());
     VLOG(4) << "Different place, do TensorCopy";
   }
-  egr::EagerUtils::autograd_meta(&(self->tensor))->SetStopGradient(true);
-  egr::EagerUtils::unsafe_autograd_meta(self->tensor)->SetPersistable(false);
+  egr::EagerUtils::autograd_meta(&(self->tensor))->SetPersistable(false);
 }
 
 py::object ParsePyArray(
@@ -218,21 +217,18 @@ paddle::platform::Place ParsePlace(
 }
 
 // boolean arguments: zero_copy, stop_gradient, persistable
-bool ParseBooleanArgs(std::string key,
-                      std::unordered_map<std::string, PyObject*> kws_map,
-                      std::unordered_map<std::string, Py_ssize_t> kw_order_map,
-                      PyObject* args, bool flag_kwargs, Py_ssize_t args_num) {
-  bool res = false;
-  if (key == "stop_gradient") res = true;
+int ParseBooleanArgs(std::string key,
+                     std::unordered_map<std::string, PyObject*> kws_map,
+                     std::unordered_map<std::string, Py_ssize_t> kw_order_map,
+                     PyObject* args, bool flag_kwargs, Py_ssize_t args_num) {
+  int res = -1;
 
   if (kw_order_map[key] <= args_num) {
-    res = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, kw_order_map[key] - 1),
-                                kw_order_map[key] - 1);
+    res = static_cast<int>(CastPyArg2AttrBoolean(
+        PyTuple_GET_ITEM(args, kw_order_map[key] - 1), kw_order_map[key] - 1));
   } else {
     if (flag_kwargs && kws_map[key] != NULL) {
-      res = CastPyArg2AttrBoolean(kws_map[key], 0);
-    } else {
-      return res;
+      res = static_cast<int>(CastPyArg2AttrBoolean(kws_map[key], 0));
     }
   }
   return res;
@@ -288,15 +284,15 @@ void AutoInitTensorByPyArray(TensorObject* py_tensor_ptr,
   bool persistable = false;
   bool zero_copy = false;
   std::string act_name = "";
-  bool stop_gradient = true;
+  int stop_gradient = -1;
 
   numpy_value =
       ParsePyArray(kws_map, kw_order_map, args, flag_kwargs, args_num);
   place = ParsePlace(kws_map, kw_order_map, args, flag_kwargs, args_num);
-  persistable = ParseBooleanArgs("persistable", kws_map, kw_order_map, args,
-                                 flag_kwargs, args_num);
-  zero_copy = ParseBooleanArgs("zero_copy", kws_map, kw_order_map, args,
-                               flag_kwargs, args_num);
+  persistable = (1 == ParseBooleanArgs("persistable", kws_map, kw_order_map,
+                                       args, flag_kwargs, args_num));
+  zero_copy = (1 == ParseBooleanArgs("zero_copy", kws_map, kw_order_map, args,
+                                     flag_kwargs, args_num));
   act_name = ParseName(kws_map, kw_order_map, args, flag_kwargs, args_num);
   stop_gradient = ParseBooleanArgs("stop_gradient", kws_map, kw_order_map, args,
                                    flag_kwargs, args_num);
@@ -571,7 +567,7 @@ int TensorInit(PyObject* self, PyObject* args, PyObject* kwargs) {
         EmptyTensorInitializer(py_tensor_ptr, act_name,
                                egr::Controller::Instance().GetExpectedPlace(),
                                persistable,
-                               /* stop_gradient */ true, dtype, dims, var_type);
+                               /* stop_gradient */ -1, dtype, dims, var_type);
 
         return 0;
       } else {
@@ -655,7 +651,7 @@ int TensorInit(PyObject* self, PyObject* args, PyObject* kwargs) {
         bool persistable = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 4), 4);
         EmptyTensorInitializer(py_tensor_ptr, act_name,
                                egr::Controller::Instance().GetExpectedPlace(),
-                               persistable, true, dtype, dims, var_type);
+                               persistable, -1, dtype, dims, var_type);
         return 0;
       } else if (pybind11::detail::npy_api::get().PyArray_Check_(arg0_ptr)) {
         VLOG(6) << "Calling case3's initializer.";
@@ -726,9 +722,8 @@ PyMappingMethods mapping_methods;
 void BindEager(pybind11::module* module) {
   auto m = module->def_submodule("eager");
 
-  auto& internals = pybind11::detail::get_internals();
   auto heap_type = reinterpret_cast<PyHeapTypeObject*>(
-      internals.default_metaclass->tp_alloc(internals.default_metaclass, 0));
+      PyType_Type.tp_alloc(&PyType_Type, 0));
   heap_type->ht_name = ToPyObject("Tensor");
   heap_type->ht_qualname = ToPyObject("Tensor");
   auto type = &heap_type->ht_type;
@@ -742,8 +737,8 @@ void BindEager(pybind11::module* module) {
   type->tp_getset = variable_properties;
   type->tp_init = TensorInit;
   type->tp_new = TensorNew;
-  Py_INCREF(internals.instance_base);
-  type->tp_base = reinterpret_cast<PyTypeObject*>(internals.instance_base);
+  Py_INCREF(&PyBaseObject_Type);
+  type->tp_base = reinterpret_cast<PyTypeObject*>(&PyBaseObject_Type);
   type->tp_flags |=
       Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
 #if PY_VERSION_HEX >= 0x03050000
diff --git a/paddle/fluid/pybind/eager_method.cc b/paddle/fluid/pybind/eager_method.cc
index b1aef8fc08fea818045a97e29e5b4f2d0e30e222..27328bea692af64918323d12ee38829ca5bb518d 100644
--- a/paddle/fluid/pybind/eager_method.cc
+++ b/paddle/fluid/pybind/eager_method.cc
@@ -186,11 +186,17 @@ static PyObject* tensor_retain_grads(TensorObject* self, PyObject* args,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
-static PyObject* tensor__clear_gradient(TensorObject* self, PyObject* args,
-                                        PyObject* kwargs) {
+static PyObject* tensor_clear_gradient(TensorObject* self, PyObject* args,
+                                       PyObject* kwargs) {
   EAGER_TRY
   VLOG(4) << "ClearGradient " << self->tensor.name();
 
+  Py_ssize_t args_num = PyTuple_Size(args);
+  bool set_to_zero = true;
+  if (args_num == (Py_ssize_t)1) {
+    CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 0), 0);
+  }
+
   paddle::experimental::Tensor* grad;
   if (egr::egr_utils_api::IsLeafTensor(self->tensor)) {
     // Add RetainGrad as PostHook to AccumulationNode
@@ -209,13 +215,27 @@ static PyObject* tensor__clear_gradient(TensorObject* self, PyObject* args,
     grad = meta->MutableGrad();
   }
 
-  if (grad->initialized()) {
-    VLOG(4) << "Gradient of " << self->tensor.name()
-            << " is initialized, will be released.";
-    auto dense_tensor =
-        std::dynamic_pointer_cast<phi::DenseTensor>(grad->impl());
-    dense_tensor->MoveMemoryHolder();
+  if (grad->is_selected_rows()) {
+    auto selected_rows =
+        std::dynamic_pointer_cast<phi::SelectedRows>(grad->impl());
+    if (selected_rows->mutable_value()->IsInitialized()) {
+      selected_rows->mutable_rows()->clear();
+      selected_rows->mutable_value()->clear();
+    }
+  } else if (grad->is_dense_tensor()) {
+    if (grad->initialized()) {
+      if (set_to_zero) {
+        grad->set_impl(paddle::experimental::zeros_like(*grad).impl());
+      } else {
+        VLOG(4) << "Gradient of " << self->tensor.name()
+                << " is initialized, will be released.";
+        auto dense_tensor =
+            std::dynamic_pointer_cast<phi::DenseTensor>(grad->impl());
+        dense_tensor->MoveMemoryHolder();
+      }
+    }
   }
+
   Py_INCREF(Py_None);
   return Py_None;
   EAGER_CATCH_AND_THROW_RETURN_NULL
@@ -407,7 +427,7 @@ PyMethodDef variable_methods[] = {
      METH_VARARGS | METH_KEYWORDS, NULL},
     {"retain_grads", (PyCFunction)(void (*)(void))tensor_retain_grads,
      METH_VARARGS | METH_KEYWORDS, NULL},
-    {"_clear_gradient", (PyCFunction)(void (*)(void))tensor__clear_gradient,
+    {"clear_gradient", (PyCFunction)(void (*)(void))tensor_clear_gradient,
      METH_VARARGS | METH_KEYWORDS, NULL},
     {"_zero_grads", (PyCFunction)(void (*)(void))tensor__zero_grads,
      METH_VARARGS | METH_KEYWORDS, NULL},
diff --git a/python/paddle/fluid/dygraph/nn.py b/python/paddle/fluid/dygraph/nn.py
index e1857a34f03f514e04e83e9596c9826569e2a90d..5bb1aef6d6e9b96a8492fe9fc76c7448a053e3bf 100644
--- a/python/paddle/fluid/dygraph/nn.py
+++ b/python/paddle/fluid/dygraph/nn.py
@@ -240,7 +240,8 @@ class Conv2D(layers.Layer):
             is_bias=True)
 
     def forward(self, input):
-        if in_dygraph_mode() and self._l_type == 'conv2d':
+        if in_dygraph_mode() and (self._l_type == 'conv2d' or
+                                  self._l_type == 'depthwise_conv2d'):
             attrs = ('strides', self._stride, 'paddings', self._padding,
                      'dilations', self._dilation, 'groups', self._groups
                      if self._groups else 1, 'use_cudnn', self._use_cudnn,
diff --git a/python/paddle/fluid/dygraph/varbase_patch_methods.py b/python/paddle/fluid/dygraph/varbase_patch_methods.py
index 6f0305f4774d6429951ee69a5b3a9db1bed18131..c4ea751ed92f8d12fd0b141b597e9810dac5200d 100644
--- a/python/paddle/fluid/dygraph/varbase_patch_methods.py
+++ b/python/paddle/fluid/dygraph/varbase_patch_methods.py
@@ -779,13 +779,6 @@ def monkey_patch_varbase():
             raise TypeError(
                 "_set_grad_ivar is only supported for Parameter Tensor")
 
-    @framework.dygraph_only
-    def clear_gradient(self, set_to_zero=True):
-        if set_to_zero:
-            self._zero_grads()
-        else:
-            self._clear_gradient()
-
     @framework.dygraph_only
     def clone(self):
         return _C_ops_.assign(self)
@@ -815,7 +808,6 @@ def monkey_patch_varbase():
     if core._in_eager_mode():
         setattr(core.eager.Tensor, "_grad_ivar", _grad_ivar)
         setattr(core.eager.Tensor, "_set_grad_ivar", _set_grad_ivar)
-        setattr(core.eager.Tensor, "clear_gradient", clear_gradient)
         setattr(core.eager.Tensor, "clone", clone)
         setattr(core.eager.Tensor, "value", value)
     else:
diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py
index ea17d029b6cc2b16d0b426f5202076db6177b7a7..9e3add621795847f84ee9906bd5bb1d2062525bd 100644
--- a/python/paddle/fluid/initializer.py
+++ b/python/paddle/fluid/initializer.py
@@ -138,11 +138,11 @@ class ConstantInitializer(Initializer):
         assert isinstance(block, framework.Block)
 
         if framework.in_dygraph_mode():
-            var = _C_ops.fill_constant(
-                var, 'value',
-                float(self._value), 'force_cpu', self._force_cpu, 'dtype',
-                int(var.dtype), 'str_value',
-                str(float(self._value)), 'shape', var.shape)
+            _C_ops.fill_constant(var, 'value',
+                                 float(self._value), 'force_cpu',
+                                 self._force_cpu, 'dtype',
+                                 int(var.dtype), 'str_value',
+                                 str(float(self._value)), 'shape', var.shape)
             return None
         else:
             # fill constant should set the "str_value" to preserve precision
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index d2653b75eafba2675c9e1b7de86873c72cdd0db2..f022e1791daefb7cbe18434aae8ac1dbc63d39c5 100755
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -26,7 +26,7 @@ import six
 import paddle
 from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant, NumpyArrayInitializer
-from ..framework import Variable, OpProtoHolder, in_dygraph_mode, dygraph_only, _dygraph_tracer, default_main_program, _varbase_creator, static_only, _global_flags
+from ..framework import Variable, OpProtoHolder, in_dygraph_mode, dygraph_only, _dygraph_tracer, default_main_program, _varbase_creator, static_only, _global_flags, _in_eager_mode
 from .. import dygraph_utils
 from ..param_attr import ParamAttr
 from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
@@ -6254,6 +6254,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
             # the shape of reshaped_3 is [6,8].
     """
     if in_dygraph_mode():
+        if _in_eager_mode():
+            tmp_tensor_type = core.eager.Tensor
+        else:
+            tmp_tensor_type = Variable
         #TODO(zhiqiu): enable inplace in dygraph mode.
         if inplace:
             warnings.warn(
@@ -6265,7 +6269,7 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
                 for item in shape
             ]
             out, _ = _C_ops.reshape2(x, None, 'shape', shape)
-        elif isinstance(shape, Variable):
+        elif isinstance(shape, tmp_tensor_type):
             shape.stop_gradient = True
             out, _ = _C_ops.reshape2(x, shape)
         else:
@@ -11132,24 +11136,30 @@ def slice(input, axes, starts, ends):
 
         infer_flags = list(1 for i in range(len(axes)))
 
+        if _in_eager_mode():
+            tmp_tensor_type = core.eager.Tensor
+        else:
+            tmp_tensor_type = Variable
+
         if isinstance(starts, (list, tuple)):
             starts = [
-                item.numpy().item(0) if isinstance(item, Variable) else item
+                item.numpy().item(0)
+                if isinstance(item, tmp_tensor_type) else item
                 for item in starts
             ]
             attrs += ('starts', starts)
-        elif isinstance(starts, Variable):
+        elif isinstance(starts, tmp_tensor_type):
             starts_tensor = starts
             starts.stop_gradient = True
             infer_flags = list(-1 for i in range(len(axes)))
 
         if isinstance(ends, (list, tuple)):
             ends = [
-                item.numpy().item(0) if isinstance(item, Variable) else item
-                for item in ends
+                item.numpy().item(0)
+                if isinstance(item, tmp_tensor_type) else item for item in ends
             ]
             attrs += ('ends', ends)
-        elif isinstance(ends, Variable):
+        elif isinstance(ends, tmp_tensor_type):
             ends_tensor = ends
             ends_tensor.stop_gradient = True
             infer_flags = list(-1 for i in range(len(axes)))
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_basic.py b/python/paddle/fluid/tests/unittests/test_imperative_basic.py
index a36b10f58ffaa503b6ccca580843f07b4bbfc2ac..4734e67140a8d0c05b92556c987208280c4a76de 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_basic.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_basic.py
@@ -953,7 +953,7 @@ class TestMetaclass(unittest.TestCase):
         self.assertNotEqual(type(MyLayer).__name__, 'pybind11_type')
         if core._in_eager_mode():
             self.assertEqual(
-                type(paddle.fluid.core.eager.Tensor).__name__, 'pybind11_type')
+                type(paddle.fluid.core.eager.Tensor).__name__, 'type')
         else:
             self.assertEqual(
                 type(paddle.fluid.core.VarBase).__name__, 'pybind11_type')
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_framework.py b/python/paddle/fluid/tests/unittests/test_imperative_framework.py
index 68628918391cb3a152cf52b0b49dab8235e46756..2d900d65976e70d1c4457bc432f8b7139d414c5c 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_framework.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_framework.py
@@ -18,6 +18,7 @@ import unittest
 import paddle.fluid as fluid
 import numpy as np
 from test_imperative_base import new_program_scope
+from paddle.fluid.framework import _test_eager_guard
 
 
 class MLP(fluid.Layer):
@@ -46,7 +47,7 @@ class MLP(fluid.Layer):
 
 
 class TestDygraphFramework(unittest.TestCase):
-    def test_dygraph_backward(self):
+    def func_test_dygraph_backward(self):
         with new_program_scope():
             mlp = MLP(input_size=2)
             var_inp = fluid.layers.data(
@@ -59,8 +60,18 @@ class TestDygraphFramework(unittest.TestCase):
             except AssertionError as e:
                 self.assertTrue((e is not None))
 
-    def test_dygraph_to_string(self):
+    def test_dygraph_backward(self):
+        with _test_eager_guard():
+            self.func_test_dygraph_backward()
+        self.func_test_dygraph_backward()
+
+    def func_test_dygraph_to_string(self):
         np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
         with fluid.dygraph.guard():
             var_inp = fluid.dygraph.to_variable(np_inp)
             print(str(var_inp))
+
+    def test_dygraph_to_string(self):
+        with _test_eager_guard():
+            self.func_test_dygraph_to_string()
+        self.func_test_dygraph_to_string()
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_gan.py b/python/paddle/fluid/tests/unittests/test_imperative_gan.py
index 189745e7295a8e234fd689c724c0953981c33c16..39b7f941c4bba6115994d466148958045a4b43f6 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_gan.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_gan.py
@@ -25,6 +25,7 @@ from paddle.fluid.optimizer import SGDOptimizer
 from paddle.fluid import Conv2D, Pool2D, Linear
 from test_imperative_base import new_program_scope
 from paddle.fluid.dygraph.base import to_variable
+from paddle.fluid.framework import _test_eager_guard
 
 
 class Discriminator(fluid.Layer):
@@ -54,7 +55,7 @@ class Generator(fluid.Layer):
 
 
 class TestDygraphGAN(unittest.TestCase):
-    def test_gan_float32(self):
+    def func_test_gan_float32(self):
         seed = 90
         paddle.seed(1)
         paddle.framework.random._manual_program_seed(1)
@@ -227,6 +228,11 @@ class TestDygraphGAN(unittest.TestCase):
         for k, v in six.iteritems(dy_params2):
             self.assertTrue(np.allclose(v, static_params[k]))
 
+    def test_gan_float32(self):
+        with _test_eager_guard():
+            self.func_test_gan_float32()
+        self.func_test_gan_float32()
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_mnist.py b/python/paddle/fluid/tests/unittests/test_imperative_mnist.py
index 1e509960c076339d2d56ccfcdd7a795fa462ca82..e2214532008952fe4ad11e09d9001a65a376b90a 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_mnist.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_mnist.py
@@ -27,6 +27,7 @@ from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 from utils import DyGraphProgramDescTracerTestHelper, is_equal_program
+from paddle.fluid.framework import _test_eager_guard, _in_eager_mode
 
 
 class SimpleImgConvPool(fluid.dygraph.Layer):
@@ -114,7 +115,7 @@ class TestImperativeMnist(unittest.TestCase):
 
         return _reader_imple
 
-    def test_mnist_float32(self):
+    def func_test_mnist_float32(self):
         seed = 90
         epoch_num = 1
         batch_size = 128
@@ -152,7 +153,7 @@ class TestImperativeMnist(unittest.TestCase):
                     label = data[1]
                     label.stop_gradient = True
 
-                    if batch_id % 10 == 0:
+                    if batch_id % 10 == 0 and not _in_eager_mode():
                         cost, traced_layer = paddle.jit.TracedLayer.trace(
                             mnist, inputs=img)
                         if program is not None:
@@ -257,6 +258,11 @@ class TestImperativeMnist(unittest.TestCase):
         for key, value in six.iteritems(static_param_value):
             self.assertTrue(np.allclose(value, dy_param_value[key], atol=1e-5))
 
+    def test_mnist_float32(self):
+        with _test_eager_guard():
+            self.func_test_mnist_float32()
+        self.func_test_mnist_float32()
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_mnist_sorted_gradient.py b/python/paddle/fluid/tests/unittests/test_imperative_mnist_sorted_gradient.py
index bda1958c0f3544bef51e51cf418ae6c07bdd7056..8e3cbaf9488bd787547f1d634c5c9e02db5382fc 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_mnist_sorted_gradient.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_mnist_sorted_gradient.py
@@ -26,10 +26,11 @@ from paddle.fluid.optimizer import SGDOptimizer
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 from test_imperative_mnist import MNIST
+from paddle.fluid.framework import _test_eager_guard
 
 
 class TestImperativeMnistSortGradient(unittest.TestCase):
-    def test_mnist_sort_gradient_float32(self):
+    def func_test_mnist_sort_gradient_float32(self):
         seed = 90
         epoch_num = 1
 
@@ -144,6 +145,11 @@ class TestImperativeMnistSortGradient(unittest.TestCase):
         for key, value in six.iteritems(static_param_value):
             self.assertTrue(np.allclose(value, dy_param_value2[key], atol=1e-5))
 
+    def test_mnist_sort_gradient_float32(self):
+        with _test_eager_guard():
+            self.func_test_mnist_sort_gradient_float32()
+        self.func_test_mnist_sort_gradient_float32()
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py
index 1c183a8c2b74aab3539607f845f89df28af19df1..7f2ce131a05d776687f9cb9f5e8d3cc533dfac25 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py
@@ -27,6 +27,7 @@ from test_imperative_base import new_program_scope
 import numpy as np
 import six
 from utils import DyGraphProgramDescTracerTestHelper, is_equal_program
+from paddle.fluid.framework import _test_eager_guard, _in_eager_mode
 
 
 class SimpleLSTMRNN(fluid.Layer):
@@ -210,10 +211,15 @@ class PtbModel(fluid.Layer):
 
 
 class TestDygraphPtbRnn(unittest.TestCase):
-    def test_ptb_rnn(self):
+    def func_test_ptb_rnn(self):
         for is_sparse in [True, False]:
             self.ptb_rnn_cpu_float32(is_sparse)
 
+    def test_ptb_rnn(self):
+        with _test_eager_guard():
+            self.func_test_ptb_rnn()
+        self.func_test_ptb_rnn()
+
     def ptb_rnn_cpu_float32(self, is_sparse):
         seed = 90
         hidden_size = 10
@@ -260,7 +266,7 @@ class TestDygraphPtbRnn(unittest.TestCase):
                 y = to_variable(y_data)
                 init_hidden = to_variable(init_hidden_data)
                 init_cell = to_variable(init_cell_data)
-                if i % 5 == 0:
+                if i % 5 == 0 and (not _in_eager_mode()):
                     outs, traced_layer = TracedLayer.trace(
                         ptb_model, [x, y, init_hidden, init_cell])
                     outs_static = traced_layer([x, y, init_hidden, init_cell])
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py
index 2d67af82de87a61f9149c2d09a0ad1cb743b8138..32e4aacf880e9f4e27c852925a0d397b74abbbcd 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py
@@ -26,6 +26,7 @@ from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 from utils import DyGraphProgramDescTracerTestHelper, is_equal_program
 from paddle.fluid.dygraph import TracedLayer
+from paddle.fluid.framework import _test_eager_guard, _in_eager_mode
 
 #NOTE(zhiqiu): run with FLAGS_cudnn_deterministic=1
 
@@ -242,7 +243,7 @@ class TestDygraphResnet(unittest.TestCase):
 
         return _reader_imple
 
-    def test_resnet_float32(self):
+    def func_test_resnet_float32(self):
         seed = 90
 
         batch_size = train_parameters["batch_size"]
@@ -259,14 +260,9 @@ class TestDygraphResnet(unittest.TestCase):
                 train_parameters, parameter_list=resnet.parameters())
             np.random.seed(seed)
 
-            batch_py_reader = fluid.io.PyReader(capacity=1)
-            batch_py_reader.decorate_sample_list_generator(
-                paddle.batch(
-                    self.reader_decorator(
-                        paddle.dataset.flowers.train(use_xmap=False)),
-                    batch_size=batch_size,
-                    drop_last=True),
-                places=fluid.CPUPlace())
+            train_reader = paddle.batch(
+                paddle.dataset.flowers.train(use_xmap=False),
+                batch_size=batch_size)
 
             dy_param_init_value = {}
             for param in resnet.parameters():
@@ -275,16 +271,21 @@ class TestDygraphResnet(unittest.TestCase):
             helper = DyGraphProgramDescTracerTestHelper(self)
             program = None
 
-            for batch_id, data in enumerate(batch_py_reader()):
+            for batch_id, data in enumerate(train_reader()):
                 if batch_id >= batch_num:
                     break
 
-                img = data[0]
-                label = data[1]
+                dy_x_data = np.array(
+                    [x[0].reshape(3, 224, 224) for x in data]).astype('float32')
+                y_data = np.array([x[1] for x in data]).astype('int64').reshape(
+                    batch_size, 1)
+
+                img = to_variable(dy_x_data)
+                label = to_variable(y_data)
                 label.stop_gradient = True
 
                 out = None
-                if batch_id % 5 == 0:
+                if batch_id % 5 == 0 and not _in_eager_mode():
                     out, traced_layer = TracedLayer.trace(resnet, img)
                     if program is not None:
                         self.assertTrue(
@@ -430,6 +431,11 @@ class TestDygraphResnet(unittest.TestCase):
             self.assertTrue(np.isfinite(value.all()))
             self.assertFalse(np.isnan(value.any()))
 
+    def test_resnet_float32(self):
+        with _test_eager_guard():
+            self.func_test_resnet_float32()
+        self.func_test_resnet_float32()
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py b/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py
index 13570d1bf71a56b6c074bc36e310cfebe92d97ba..daa778288ddf5488151a3adeed7d088a7240ded8 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py
@@ -22,6 +22,7 @@ from paddle.fluid import core
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 from test_imperative_resnet import ResNet
+from paddle.fluid.framework import _test_eager_guard
 
 batch_size = 8
 train_parameters = {
@@ -71,7 +72,7 @@ def optimizer_setting(params, parameter_list=None):
 
 
 class TestDygraphResnetSortGradient(unittest.TestCase):
-    def test_resnet_sort_gradient_float32(self):
+    def func_test_resnet_sort_gradient_float32(self):
         seed = 90
 
         batch_size = train_parameters["batch_size"]
@@ -230,6 +231,11 @@ class TestDygraphResnetSortGradient(unittest.TestCase):
             self.assertTrue(np.isfinite(value.all()))
             self.assertFalse(np.isnan(value.any()))
 
+    def test_resnet_sort_gradient_float32(self):
+        with _test_eager_guard():
+            self.func_test_resnet_sort_gradient_float32()
+        self.func_test_resnet_sort_gradient_float32()
+
 
 if __name__ == '__main__':
     unittest.main()