diff --git a/paddle/fluid/eager/accumulation/accumulation_node.cc b/paddle/fluid/eager/accumulation/accumulation_node.cc
index 9c4089af092e418d6845864671124917c6498cf1..10696dbacd35bb2b4895de5a1efa0a8a249fc508 100644
--- a/paddle/fluid/eager/accumulation/accumulation_node.cc
+++ b/paddle/fluid/eager/accumulation/accumulation_node.cc
@@ -39,8 +39,9 @@ static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
 }
 
 std::vector<std::vector<paddle::experimental::Tensor>> GradNodeAccumulation::
-operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
-           bool create_graph) {
+operator()(
+    std::vector<std::vector<paddle::experimental::Tensor>>& grads,  // NOLINT
+    bool create_graph) {
   VLOG(3) << "Running Eager Backward Node: GradNodeAccumulation";
   PADDLE_ENFORCE(grads.size() == 1,
                  paddle::platform::errors::Fatal(
diff --git a/paddle/fluid/eager/accumulation/accumulation_node.h b/paddle/fluid/eager/accumulation/accumulation_node.h
index a91a0b6e34c0d9440e3645d1a6982748c4315962..2e38d7e9e91e227681ec8dfbc252ca59465ccf09 100644
--- a/paddle/fluid/eager/accumulation/accumulation_node.h
+++ b/paddle/fluid/eager/accumulation/accumulation_node.h
@@ -35,7 +35,7 @@ class GradNodeAccumulation : public GradNodeBase {
 
   // Functor: perform backward computations
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
+      std::vector<std::vector<paddle::experimental::Tensor>>& grads,  // NOLINT
       bool create_graph = false) override;
 
   void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
diff --git a/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc b/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc
index 0bc998a03a80b7b8a1e486ad68f1575c130d2c1b..d9f5447a88e9bf2c8f95c55911066b76078ea25f 100644
--- a/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc
+++ b/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc
@@ -145,8 +145,9 @@ void GradNodeScale::SetTensorWrappers_X(
 void GradNodeScale::SetAttributes_scale(float scale) { scale_ = scale; }
 
 std::vector<std::vector<paddle::experimental::Tensor>> GradNodeScale::
-operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
-           bool create_graph) {
+operator()(
+    std::vector<std::vector<paddle::experimental::Tensor>>& grads,  // NOLINT
+    bool create_graph) {
   // 1. Check Output Size
   PADDLE_ENFORCE(
       ((grads.size() == 1) && (grads[0].size() == 1)),
diff --git a/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h b/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h
index e263f73a6b8a4a1f9ce23d9b5ca383fd6828016b..0b942d2a06707817e58af6d42441c7fa7620f7a0 100644
--- a/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h
+++ b/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h
@@ -39,7 +39,7 @@ class GradNodeScale : public GradNodeBase {
 
   // Functor: perform backward computations
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
+      std::vector<std::vector<paddle::experimental::Tensor>>& grads,  // NOLINT
       bool create_graph = false) override;
 
   void ClearTensorWrappers() override { VLOG(6) << "Do nothing here now"; }
diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
index df2cdc35626a8aa27899f7340fa14285299a11d1..229817596423cd46c29db3f0dae589d0655b8485 100644
--- a/paddle/fluid/eager/auto_code_generator/eager_generator.cc
+++ b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
@@ -47,6 +47,9 @@ std::unordered_map<std::string, std::vector<std::string>>
 static std::unordered_map<std::string, paddle::framework::AttributeMap>
     operators_with_attrs = {};
 
+static std::unordered_set<std::string> ops_to_fill_zero_for_empty_grads = {
+    "split"};
+
 /* --- Black Ops list that's NO NEED to apply code generation --- */
 static std::unordered_set<std::string> black_ops_list = {"run_program"};
 
@@ -2243,11 +2246,21 @@ static std::string GenerateGradNodeCCContents(
   // [Generation] Get Full Grad Function
   const char* GRAD_FUNCTION_TEMPLATE =
       "std::vector<std::vector<paddle::experimental::Tensor>> "
-      "GradNode%s::operator()(const "
-      "std::vector<std::vector<paddle::experimental::Tensor>>& grads, "
-      "bool create_graph) {\n%s\n}";
-  std::string grad_function_str = paddle::string::Sprintf(
-      GRAD_FUNCTION_TEMPLATE, fwd_op_type, generated_grad_function_body);
+      "GradNode%s::operator()("
+      "std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool "
+      "create_graph) {\n"
+      "%s"
+      "%s"
+      "\n}";
+  std::string fill_zero_str = "";
+  if (ops_to_fill_zero_for_empty_grads.count(fwd_op_type)) {
+    fill_zero_str =
+        "egr::EagerUtils::FillZeroForEmptyGradInputs(&grads, "
+        "this->InputMeta());\n";
+  }
+  std::string grad_function_str =
+      paddle::string::Sprintf(GRAD_FUNCTION_TEMPLATE, fwd_op_type,
+                              fill_zero_str, generated_grad_function_body);
 
   VLOG(6) << "Generated returns";
 
@@ -2279,9 +2292,9 @@ static std::string GenerateGradNodeHeaderContents(
       "  ~GradNode%s() override { VLOG(6) << \" Destruct GradNode%s \"; }\n"
       "\n"
       "  virtual std::vector<std::vector<paddle::experimental::Tensor>> "
-      "operator()(const "
-      "std::vector<std::vector<paddle::experimental::Tensor>>& grads, const "
-      "bool create_graph = false) "
+      "operator()("
+      "std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool "
+      "create_graph = false) "
       "override;\n"
       "\n"
       "  void ClearTensorWrappers() override { \n"
diff --git a/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
index 92cee056d52a76fb4731f10166f49646dcc97bda..1de050d1230f8cc4a72032161e1e00cf71a1ad3b 100644
--- a/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
+++ b/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
@@ -17,6 +17,8 @@ import re
 import argparse
 import os
 
+ops_to_fill_zero_for_empty_grads = set(list("split"))
+
 # For API dispatch used at python-level
 # { op_name : [arg_name, ...] }
 core_ops_returns_info = {}
@@ -598,7 +600,8 @@ class {} : public egr::GradNodeBase {{
   ~{}() override = default;
 
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph = false) override;
+      std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph = false) override;
+  
   std::string name() override {{ return \" {} \"; }}
   
   void ClearTensorWrappers() override {{
@@ -656,10 +659,11 @@ def GenerateNodeDefinition(fwd_api_name, bwd_api_name, backward_fwd_input_map,
     for _, (ttype, fwd_position,
             grad_api_position) in backward_grad_input_map.items():
         if IsPlainTensorType(ttype):
-            grad_api_args[grad_api_position] = f"grads[{fwd_position}][0]"
+            grad_api_args[
+                grad_api_position] = f"hooked_grads[{fwd_position}][0]"
         else:
             assert IsVectorTensorType(ttype)
-            grad_api_args[grad_api_position] = f"grads[{fwd_position}]"
+            grad_api_args[grad_api_position] = f"hooked_grads[{fwd_position}]"
 
     for name, _, _, grad_api_position in backward_attrs_list:
         saved_attribute_name = GetSavedName(name)
@@ -687,23 +691,30 @@ def GenerateNodeDefinition(fwd_api_name, bwd_api_name, backward_fwd_input_map,
 
     grad_node_name = GetGradNodeName(fwd_api_name)
 
+    fill_zero_str = ""
+    if fwd_api_name in ops_to_fill_zero_for_empty_grads:
+        fill_zero_str = "egr::EagerUtils::FillZeroForEmptyGradInputs(&grads, this->InputMeta());\n"
+
     if len(namespace) > 0:
         grad_api_namespace = f"paddle::experimental::{namespace}"
     else:
         grad_api_namespace = f"paddle::experimental"
 
     FUNCTION_TEMPLATE = """
-std::vector<std::vector<paddle::experimental::Tensor>> {}::operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph) {{
+std::vector<std::vector<paddle::experimental::Tensor>> {}::operator()(std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph) {{
+    {}
+    auto hooked_grads = ApplyGradientHooks(grads);
+    
     // Call grad_api function
-    VLOG(3) << \"Finally State Running: \" << \"{}\"; 
+    VLOG(3) << \"Final State Running: \" << \"{}\"; 
     auto grad_api_returns = {}::{}({});
     {}
 }}
   """
 
     node_definition_str = FUNCTION_TEMPLATE.format(
-        grad_node_name, grad_node_name, grad_api_namespace, bwd_api_name,
-        grad_api_args_str, returns_str)
+        grad_node_name, fill_zero_str, grad_node_name, grad_api_namespace,
+        bwd_api_name, grad_api_args_str, returns_str)
 
     return node_definition_str
 
diff --git a/paddle/fluid/eager/custom_operator/custom_operator_node.cc b/paddle/fluid/eager/custom_operator/custom_operator_node.cc
index 72af1cc4b068679e72ae6bdc5e09fab8f56bac04..08ca3bed5a653637b77ba512a812a0b863fc80c7 100644
--- a/paddle/fluid/eager/custom_operator/custom_operator_node.cc
+++ b/paddle/fluid/eager/custom_operator/custom_operator_node.cc
@@ -20,8 +20,8 @@
 
 namespace egr {
 std::vector<std::vector<paddle::experimental::Tensor>> RunCustomOpNode::
-operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
-           bool create_graph) {
+operator()(std::vector<std::vector<paddle::experimental::Tensor>>& grads,
+           bool create_graph) {  // NOLINT
   paddle::CustomOpKernelContext ctx;
   auto grad_inputs_name = paddle::framework::OpMetaInfoHelper::GetInputs(
       egr::Controller::Instance().GetOpMetaInfoMap().at(op_type_)[1]);
diff --git a/paddle/fluid/eager/custom_operator/custom_operator_node.h b/paddle/fluid/eager/custom_operator/custom_operator_node.h
index 6ece2658575c795856438904c2716d61f0985879..33b56fc8c863ac23dc29f2468198f2610beef164 100644
--- a/paddle/fluid/eager/custom_operator/custom_operator_node.h
+++ b/paddle/fluid/eager/custom_operator/custom_operator_node.h
@@ -37,8 +37,9 @@ class RunCustomOpNode : public GradNodeBase {
 
   // Functor: perform backward computations
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
-      bool create_graph) override;
+      std::vector<std::vector<paddle::experimental::Tensor>>& grads,
+      bool create_graph = false)  // NOLINT
+      override;
 
   std::string name() {
     return paddle::string::Sprintf("RunCustomOpNode: %s_grad", op_type_);
diff --git a/paddle/fluid/eager/grad_node_info.cc b/paddle/fluid/eager/grad_node_info.cc
index 1d44d842b0825aa96380c947c67082fbcb5e1642..25610a3f95fe5d969ffafa8379842b1ef2333b54 100644
--- a/paddle/fluid/eager/grad_node_info.cc
+++ b/paddle/fluid/eager/grad_node_info.cc
@@ -102,6 +102,7 @@ const std::vector<std::vector<GradSlotMeta>>& GradNodeBase::OutputMeta() const {
 
 void GradNodeBase::SetGradInMeta(const paddle::experimental::Tensor& fwd_out,
                                  size_t slot_rank) {
+  VLOG(6) << "Set GradSlotMeta for Grad Inputs";
   auto* fwd_out_meta = egr::EagerUtils::nullable_autograd_meta(fwd_out);
   PADDLE_ENFORCE_LE(
       slot_rank, (bwd_in_meta_.size() - 1),
@@ -117,6 +118,12 @@ void GradNodeBase::SetGradInMeta(const paddle::experimental::Tensor& fwd_out,
   auto& meta = metas[0];
   meta.SetStopGradient(fwd_out_meta->StopGradient());
 
+  if (!fwd_out.is_initialized()) {
+    VLOG(6)
+        << "Skip Configuring GradSlotMeta for uninitialized GradInput Tensor";
+    return;
+  }
+
   // Record TensorMeta
   if (phi::DenseTensor::classof(fwd_out.impl().get())) {
     // Only Copy Meta
@@ -128,7 +135,9 @@ void GradNodeBase::SetGradInMeta(const paddle::experimental::Tensor& fwd_out,
         paddle::platform::errors::Fatal(
             "Attempting to copy DenseTensorMeta with phi::DataType::UNDEFINED,"
             "which is illegal."));
+
     meta.SetTensorMeta(dense_tensor->meta());
+    meta.SetPlace(fwd_out.inner_place());
 
     if (paddle::framework::IsComplexType(
             paddle::framework::TransToProtoVarType(dense_tensor->type()))) {
@@ -143,6 +152,7 @@ void GradNodeBase::SetGradInMeta(const paddle::experimental::Tensor& fwd_out,
 void GradNodeBase::SetGradInMeta(
     const std::vector<paddle::experimental::Tensor>& fwd_out,
     size_t slot_rank) {
+  VLOG(6) << "Set GradSlotMeta for Grad Inputs";
   size_t slot_size = fwd_out.size();
   PADDLE_ENFORCE_LE(
       slot_rank, (bwd_in_meta_.size() - 1),
@@ -172,6 +182,12 @@ void GradNodeBase::SetGradInMeta(
       meta.SetStopGradient(fwd_out_meta->StopGradient());
     }
 
+    if (!fwd_out_tensor.is_initialized()) {
+      VLOG(6)
+          << "Skip Configuring GradSlotMeta for uninitialized GradInput Tensor";
+      return;
+    }
+
     // Record TensorMeta
     if (phi::DenseTensor::classof(fwd_out_tensor.impl().get())) {
       // Only Copy Meta
@@ -184,6 +200,8 @@ void GradNodeBase::SetGradInMeta(
                                           "with phi::DataType::UNDEFINED,"
                                           "which is illegal."));
       meta.SetTensorMeta(dense_tensor->meta());
+      meta.SetPlace(fwd_out_tensor.inner_place());
+
       if (paddle::framework::IsComplexType(
               paddle::framework::TransToProtoVarType(dense_tensor->type()))) {
         need_complex_to_real_ = true;
@@ -228,6 +246,7 @@ void GradNodeBase::SetGradOutMeta(const paddle::experimental::Tensor& fwd_in,
                                           "with phi::DataType::UNDEFINED,"
                                           "which is illegal."));
       meta.SetTensorMeta(dense_tensor->meta());
+      meta.SetPlace(fwd_in.inner_place());
     }
   } else {
     VLOG(6) << "Unable to initialize the DenseTensorMeta of GradSlotMeta with "
@@ -272,6 +291,7 @@ void GradNodeBase::SetGradOutMeta(
                               "phi::DataType::UNDEFINED,"
                               "which is illegal."));
         meta.SetTensorMeta(dense_tensor->meta());
+        meta.SetPlace(fwd_in_tensor.inner_place());
       }
     } else {
       VLOG(6) << "Unable to initialize the DenseTensorMeta of GradSlotMeta "
diff --git a/paddle/fluid/eager/grad_node_info.h b/paddle/fluid/eager/grad_node_info.h
index 28c12717a24b0c89b8a3b6544124ad6533d6c70d..4dec1c1f9f4e5c0088fc05a8e581bb637117b4a4 100644
--- a/paddle/fluid/eager/grad_node_info.h
+++ b/paddle/fluid/eager/grad_node_info.h
@@ -76,8 +76,12 @@ class GradSlotMeta {
     return *meta_.get();
   }
 
+  void SetPlace(const phi::Place& place) { place_ = place; }
+  const phi::Place& GetPlace() const { return place_; }
+
  private:
   bool stop_gradient_{false};
+  phi::Place place_;
   std::shared_ptr<phi::DenseTensorMeta> meta_ = nullptr;
 };
 
@@ -102,7 +106,7 @@ class GradNodeBase {
    * is better choice to fit this format.
    * **/
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
+      std::vector<std::vector<paddle::experimental::Tensor>>& grads,  // NOLINT
       bool create_graph = false) = 0;
 
   virtual void ClearTensorWrappers() = 0;
diff --git a/paddle/fluid/eager/grad_tensor_holder.h b/paddle/fluid/eager/grad_tensor_holder.h
index 8c00f9161b629f7a3f093a1225d3d5b0b9bcca8b..db03789ea7632b0baf6e72e6da7624c04284d316 100644
--- a/paddle/fluid/eager/grad_tensor_holder.h
+++ b/paddle/fluid/eager/grad_tensor_holder.h
@@ -53,7 +53,7 @@ class GradTensorHolder {
     return buffer_[pos];
   }
 
-  const std::vector<std::vector<paddle::experimental::Tensor>>& Buffers() {
+  std::vector<std::vector<paddle::experimental::Tensor>>& Buffers() {
     return buffer_;
   }
 
diff --git a/paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc b/paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc
index 28682ab0fe094df6d27eb27e9118e6576685c95a..6c6c7fd25e5e53c7bd96d649fcab8b694b697d7b 100644
--- a/paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc
+++ b/paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc
@@ -80,13 +80,15 @@ TEST(AccumulationNode, Tensor) {
   grad_meta->SetStopGradient(false);
 
   // operator()
-  paddle::experimental::Tensor ret_et0 = node->operator()({{et0}})[0][0];
+  std::vector<std::vector<paddle::experimental::Tensor>> et0_vec = {{et0}};
+  paddle::experimental::Tensor ret_et0 = node->operator()(et0_vec)[0][0];
   auto* ret_et0_ptr =
       std::dynamic_pointer_cast<phi::DenseTensor>(ret_et0.impl())
           ->data<paddle::platform::float16>();
   CHECK_EQ(ret_et0_ptr[0], paddle::platform::float16(10.0f));
 
-  paddle::experimental::Tensor ret_et1 = node->operator()({{et1}})[0][0];
+  std::vector<std::vector<paddle::experimental::Tensor>> et1_vec = {{et1}};
+  paddle::experimental::Tensor ret_et1 = node->operator()(et1_vec)[0][0];
 
   auto* ret_et1_ptr =
       std::dynamic_pointer_cast<phi::DenseTensor>(ret_et1.impl())
@@ -121,7 +123,7 @@ TEST(AccumulationNode, Tensor) {
       std::make_shared<egr::CppTensorVoidHook>(reduce_hook_1));
 
   // operator()
-  paddle::experimental::Tensor _ret = node->operator()({{et0}})[0][0];
+  paddle::experimental::Tensor _ret = node->operator()(et0_vec)[0][0];
 
   // Check operator() result, should be 36.0
   auto* _ret_ptr = std::dynamic_pointer_cast<phi::DenseTensor>(_ret.impl())
diff --git a/paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h b/paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h
index 0b167203735d65683b0f978fa34fe7f457aae4f2..dff12fdfc34a1330cf7be300a99ee95c0086f668 100644
--- a/paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h
+++ b/paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h
@@ -32,7 +32,7 @@ class GradTestNode : public egr::GradNodeBase {
   GradTestNode() : GradNodeBase() { val_ = 1.0; }
   std::string name() override { return "GradTestNode"; }
   std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>>& grads,
+      std::vector<std::vector<paddle::experimental::Tensor>>& grads,
       bool create_graph = false) override {
     val_ = std::dynamic_pointer_cast<phi::DenseTensor>(grads[0][0].impl())
                ->data<float>()[0];
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 217055e4e9e4a19e695f42bf57c2331c9b98e2bd..7486e711641fc9ae4a02d8e66dbcd1099c548abf 100644
--- a/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc
+++ b/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc
@@ -247,4 +247,20 @@ TEST(EagerUtils, GetGradAccumulationNode) {
   ASSERT_ANY_THROW(egr::EagerUtils::GetGradAccumulationNode(t0));
 }
 
+TEST(EagerUtils, FillZeroForEmptyGradInputs) {
+  std::vector<std::vector<paddle::experimental::Tensor>> grads = {
+      std::vector<paddle::experimental::Tensor>(1)};
+  std::vector<std::vector<GradSlotMeta>> slot_metas = {
+      std::vector<GradSlotMeta>(1)};
+
+  phi::DenseTensorMeta tensor_meta;
+  tensor_meta.dtype = paddle::experimental::DataType::FLOAT32;
+  tensor_meta.dims = {2, 4};
+  slot_metas[0][0].SetTensorMeta(tensor_meta);
+  slot_metas[0][0].SetPlace(phi::CPUPlace());
+
+  EagerUtils::FillZeroForEmptyGradInputs(&grads, slot_metas);
+  eager_test::CompareTensorWithValue<float>(grads[0][0], 0.0);
+}
+
 }  // namespace egr
diff --git a/paddle/fluid/eager/to_static/run_program_op_node.h b/paddle/fluid/eager/to_static/run_program_op_node.h
index 4eaa64d3ac659ca0ec76083b70855d8b6b241556..c83e16e9a1ec21b3e7303834ac35b55fed60b2a6 100644
--- a/paddle/fluid/eager/to_static/run_program_op_node.h
+++ b/paddle/fluid/eager/to_static/run_program_op_node.h
@@ -370,7 +370,7 @@ class GradNodeRunProgram : public egr::GradNodeBase {
   ~GradNodeRunProgram() override = default;
   // Functor: perform backward computations
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
-      const std::vector<std::vector<paddle::experimental::Tensor>> &grads,
+      std::vector<std::vector<paddle::experimental::Tensor>> &grads,  // NOLINT
       bool create_graph) override {
     VLOG(3) << "Running Eager Backward Node: GradNodeRunProgram";
     PADDLE_ENFORCE_EQ(
diff --git a/paddle/fluid/eager/utils.cc b/paddle/fluid/eager/utils.cc
index 048087903a47c1699a7d7f32199c313146bd37ab..20faae95281db87ad4896b19e63857cf4b7e5e02 100644
--- a/paddle/fluid/eager/utils.cc
+++ b/paddle/fluid/eager/utils.cc
@@ -20,6 +20,7 @@
 
 #include "paddle/phi/api/all.h"
 #include "paddle/phi/common/layout.h"
+#include "paddle/phi/core/compat/convert_utils.h"
 #include "paddle/phi/core/tensor_meta.h"
 
 #include "paddle/fluid/framework/data_layout.h"
@@ -392,4 +393,28 @@ std::shared_ptr<egr::GradNodeBase> EagerUtils::GetGradAccumulationNode(
   }
 }
 
+void EagerUtils::FillZeroForEmptyGradInputs(
+    std::vector<std::vector<paddle::experimental::Tensor>>* in_grads,
+    const std::vector<std::vector<GradSlotMeta>>& grad_in_metas) {
+  for (size_t i = 0; i < in_grads->size(); i++) {
+    for (size_t j = 0; j < (*in_grads)[0].size(); j++) {
+      paddle::experimental::Tensor& grad = (*in_grads)[i][j];
+      if (!grad.is_initialized()) {
+        const GradSlotMeta& grad_in_meta = grad_in_metas[i][j];
+        PADDLE_ENFORCE(
+            grad_in_meta.HasTensorMeta(),
+            paddle::platform::errors::Fatal(
+                "Unable to fill empty grad inputs due to empty GradSlotMeta"));
+
+        const auto& tensor_meta = grad_in_meta.GetTensorMeta();
+        phi::Place place = grad_in_meta.GetPlace();
+
+        auto tensor_with_zero = paddle::experimental::full(
+            phi::vectorize(tensor_meta.dims), 0.0, tensor_meta.dtype, place);
+        grad.set_impl(tensor_with_zero.impl());
+      }
+    }
+  }
+}
+
 }  // namespace egr
diff --git a/paddle/fluid/eager/utils.h b/paddle/fluid/eager/utils.h
index fbd080ef70e25408abcb979360610ad08d752f96..396837f101c6518d32a908a80ff4782bf45ea090 100644
--- a/paddle/fluid/eager/utils.h
+++ b/paddle/fluid/eager/utils.h
@@ -217,6 +217,13 @@ class EagerUtils {
       const std::vector<paddle::experimental::Tensor>& tensors);
   static std::shared_ptr<egr::GradNodeBase> GetGradAccumulationNode(
       const paddle::experimental::Tensor& tensor);
+
+  /**
+    * Fill Zero
+    * **/
+  static void FillZeroForEmptyGradInputs(
+      std::vector<std::vector<paddle::experimental::Tensor>>* out_grads,
+      const std::vector<std::vector<GradSlotMeta>>& grad_out_metas);
 };
 
 }  // namespace egr
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_auto_prune.py b/python/paddle/fluid/tests/unittests/test_imperative_auto_prune.py
index 44d73612b1cb5e0bce1e561ab25884355b083d77..39b79dd4ba26b6b375404d452789758dddeba89d 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_auto_prune.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_auto_prune.py
@@ -182,7 +182,7 @@ class TestImperativeAutoPrune(unittest.TestCase):
         self.func_auto_prune2()
 
     # TODO(jiabin): Support this when we support better split tensor
-    def test_auto_prune3(self):
+    def func_auto_prune3(self):
         with fluid.dygraph.guard():
             case3 = AutoPruneLayer3(input_size=784)
             value1 = np.arange(784).reshape(1, 784).astype("float32")
@@ -194,7 +194,12 @@ class TestImperativeAutoPrune(unittest.TestCase):
             self.assertTrue(case3.linear.weight._grad_ivar() is not None)
             self.assertTrue((part2.gradient() == 0).all())
 
-    def test_auto_prune4(self):
+    def test_auto_prune3(self):
+        with _test_eager_guard():
+            self.func_auto_prune3()
+        self.func_auto_prune3()
+
+    def func_auto_prune4(self):
         with fluid.dygraph.guard():
             case4 = AutoPruneLayer3(input_size=784)
             value1 = np.arange(784).reshape(1, 784).astype("float32")
@@ -206,7 +211,12 @@ class TestImperativeAutoPrune(unittest.TestCase):
             self.assertTrue(case4.linear.weight._grad_ivar() is not None)
             self.assertTrue((part2.gradient() == 1).all())
 
-    def test_auto_prune5(self):
+    def test_auto_prune4(self):
+        with _test_eager_guard():
+            self.func_auto_prune4()
+        self.func_auto_prune4()
+
+    def func_auto_prune5(self):
         with fluid.dygraph.guard():
             case4 = AutoPruneLayer3(input_size=784)
             value1 = np.arange(784).reshape(1, 784).astype("float32")
@@ -218,6 +228,11 @@ class TestImperativeAutoPrune(unittest.TestCase):
             self.assertTrue(case4.linear.weight._grad_ivar() is not None)
             self.assertTrue((part2.gradient() == 0).all())
 
+    def test_auto_prune5(self):
+        with _test_eager_guard():
+            self.func_auto_prune5()
+        self.func_auto_prune5()
+
     def func_auto_prune6(self):
         with fluid.dygraph.guard():
             value0 = np.arange(26).reshape(2, 13).astype("float32")