diff --git a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
index 2e720f88004513ab57e7049bf760706bde9b2d73..8c48e40aea6228cfef7e14b0b681e47e68e88e28 100644
--- a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
+++ b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
@@ -191,6 +191,7 @@ paddle::small_vector<std::vector<paddle::experimental::Tensor>, egr::kSlotSmallV
 
 FORWARD_FUNCTION_TEMPLATE = """
 {} {}({}) {{
+  FLAGS_tensor_operants_mode = "eager";
   VLOG(3) << \"Running AD API: \" << \"{}\";
   // Dygraph Record Event
 {}
@@ -246,6 +247,7 @@ BEFORE_LOG_PRINT_TEMPLATE = """
 
 FORWARD_ONLY_FUNCTION_TEMPLATE = """
 {} {}({}) {{
+  FLAGS_tensor_operants_mode = "eager";
   VLOG(3) << \"Running AD API: \" << \"{}\";
   // Dygraph Record Event
 {}
@@ -364,6 +366,7 @@ FORWARD_CC_FILE_TEMPLATE = """
 #include "paddle/fluid/eager/nan_inf_utils.h"
 #include "paddle/fluid/eager/api/manual/eager_manual/dygraph_forward_api.h"
 DECLARE_bool(check_nan_inf);
+DECLARE_string(tensor_operants_mode);
 {}
 {}
 """
diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index 2400a81958cf5a372f0e010af9ed17bb6dc370f8..3041c72890a5e32d2126ba89089403444b0d9f5b 100755
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -1203,7 +1203,9 @@ cc_library(
        string_helper
        phi_tensor
        op_meta_info
-       phi_api)
+       phi_api
+       phi_tensor_operants
+       operants_manager)
 
 set(FLUID_FRAMEWORK_MODULES
     proto_desc
diff --git a/paddle/fluid/framework/custom_operator.cc b/paddle/fluid/framework/custom_operator.cc
index e0a6fbd37dafffc6fb3d334355b09127be368ee3..98d07da65e8540eb039693e2fd72ddc21aa67aa9 100644
--- a/paddle/fluid/framework/custom_operator.cc
+++ b/paddle/fluid/framework/custom_operator.cc
@@ -45,6 +45,12 @@ limitations under the License. */
 #include "paddle/phi/backends/device_manager.h"
 #endif
 
+#include "gflags/gflags.h"
+#include "paddle/phi/api/include/tensor_operants.h"
+#include "paddle/phi/core/operants_manager.h"
+
+DECLARE_string(tensor_operants_mode);
+
 namespace paddle {
 namespace framework {
 
@@ -270,6 +276,15 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
 
   try {
     VLOG(3) << "Custom Operator: Run ComputeFunc.";
+
+    FLAGS_tensor_operants_mode = "phi";
+    if (paddle::operants::OperantsManager::Instance().phi_operants.get() ==
+        nullptr) {
+      paddle::operants::OperantsManager::Instance().phi_operants.reset(
+          new paddle::operants::PhiTensorOperants());
+      VLOG(4) << "Initialize phi tensor operants successfully";
+    }
+
     func(&kernel_ctx);
 
     // sync output tensor data into original output
diff --git a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
index e782d6b65bba62ac47615e5f4de4b339575781ee..c231ab9b292e51c462c1469cd8fdc0f318fc7750 100644
--- a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
+++ b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
@@ -29,7 +29,7 @@ void tanh_grad(const Tensor& out, const Tensor& grad_out, Tensor* grad_x) {
   if (!grad_x) return;
   auto tmp = pow<T>(out, 2.0);
   tmp = scale<T>(tmp, -1.0, 1.0, true);
-  auto grad_x_tmp = multiply<T>(grad_out, tmp);
+  auto grad_x_tmp = grad_out * tmp;
   set_output<T>(grad_x_tmp, grad_x);
 }
 
@@ -172,7 +172,7 @@ void divide_grad(const Tensor& x,
     auto tmp0 = pow<T>(y, 2.0);
     auto tmp1 = divide<T>(x, tmp0);
     auto tmp2 = scale<T>(tmp1, -1.0, 0.0, true);
-    auto dy_res = multiply<T>(tmp2, out_grad);
+    auto dy_res = tmp2 * out_grad;
     if (x.dims() != y.dims()) {
       // Maybe need reduce here
       phi::DDim reduce_dim = get_reduce_dims(y.dims(), x.dims());
@@ -192,7 +192,7 @@ void divide_grad(const Tensor& x,
     // dx = (1/y) * dout
     auto one_tensor = full<T>(phi::vectorize(y.dims()), 1.0, y.dtype());
     auto tmp0 = divide<T>(one_tensor, y);
-    auto dx_res = multiply<T>(tmp0, out_grad);
+    auto dx_res = tmp0 * out_grad;
     if (y.dims() != x.dims()) {
       // Maybe need reduce here
       auto reduce_dim = get_reduce_dims(x.dims(), y.dims());
@@ -216,7 +216,7 @@ void sqrt_grad(const Tensor& out, const Tensor& out_grad, Tensor* x_grad) {
   if (x_grad) {
     auto div_x = full<T>(phi::vectorize(out.dims()), 0.5);
     auto tmp = divide<T>(div_x, out);
-    auto x_grad_tmp = multiply<T>(out_grad, tmp);
+    auto x_grad_tmp = out_grad * tmp;
     set_output<T>(x_grad_tmp, x_grad);
   }
 }
@@ -229,7 +229,7 @@ void multiply_grad(const Tensor& x,
                    Tensor* x_grad,
                    Tensor* y_grad) {
   if (x_grad) {
-    auto x_grad_unreduce = multiply<T>(out_grad, y);
+    auto x_grad_unreduce = out_grad * y;
     if (x_grad_unreduce.dims() != x.dims()) {
       auto axes = get_reduce_dims_from_out(x_grad_unreduce.dims(), x.dims());
       if (!axes.size()) {
@@ -249,7 +249,7 @@ void multiply_grad(const Tensor& x,
     }
   }
   if (y_grad) {
-    auto y_grad_unreduce = multiply<T>(out_grad, x);
+    auto y_grad_unreduce = out_grad * x;
     if (y_grad_unreduce.dims() != y.dims()) {
       auto axes = get_reduce_dims_from_out(y_grad_unreduce.dims(), y.dims());
       if (!axes.size()) {
@@ -297,7 +297,7 @@ void expand_grad(const Tensor& x,
 template <typename T>
 void exp_grad(const Tensor& out, const Tensor& out_grad, Tensor* x_grad) {
   if (x_grad) {
-    set_output<T>(multiply<T>(out_grad, out), x_grad);
+    set_output<T>(out_grad * out, x_grad);
   }
 }
 
diff --git a/paddle/fluid/prim/tests/CMakeLists.txt b/paddle/fluid/prim/tests/CMakeLists.txt
index 92b34352a70f22ee1ba808be334e5d6e074ea403..f42eb8db5682ae38bf42a0dfc046c9085e4bc58a 100644
--- a/paddle/fluid/prim/tests/CMakeLists.txt
+++ b/paddle/fluid/prim/tests/CMakeLists.txt
@@ -33,9 +33,16 @@ cc_test_old(
   activation_op
   phi_api
   phi_dygraph_api
-  static_global_utils)
+  static_global_utils
+  static_tensor_operants
+  operants_manager)
 
 if(NOT (NOT WITH_PYTHON AND ON_INFER))
+  cc_library(
+    init_env_utils
+    SRCS init_env_utils.cc
+    DEPS operants_manager eager_tensor_operants static_tensor_operants)
+
   cc_test_old(
     test_comp_eager
     SRCS
@@ -44,5 +51,6 @@ if(NOT (NOT WITH_PYTHON AND ON_INFER))
     ${prim_eager_deps}
     ${prim_generated_deps}
     prim_utils
-    static_global_utils)
+    static_global_utils
+    init_env_utils)
 endif()
diff --git a/paddle/fluid/prim/tests/init_env_utils.cc b/paddle/fluid/prim/tests/init_env_utils.cc
new file mode 100644
index 0000000000000000000000000000000000000000..708c187e678070f8e3c62e5dc85089d323b1ae9a
--- /dev/null
+++ b/paddle/fluid/prim/tests/init_env_utils.cc
@@ -0,0 +1,31 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/prim/tests/init_env_utils.h"
+#include "paddle/fluid/prim/utils/eager/eager_tensor_operants.h"
+#include "paddle/fluid/prim/utils/static/static_tensor_operants.h"
+#include "paddle/phi/core/operants_manager.h"
+
+namespace paddle {
+namespace prim {
+
+void InitTensorOperants() {
+  paddle::operants::OperantsManager::Instance().eager_operants.reset(
+      new paddle::operants::EagerTensorOperants());
+  paddle::operants::OperantsManager::Instance().static_operants.reset(
+      new paddle::operants::StaticTensorOperants());
+}
+
+}  // namespace prim
+}  // namespace paddle
diff --git a/paddle/fluid/prim/tests/init_env_utils.h b/paddle/fluid/prim/tests/init_env_utils.h
new file mode 100644
index 0000000000000000000000000000000000000000..4b6c7f40d4cce566e100e647f6aefe430a758e42
--- /dev/null
+++ b/paddle/fluid/prim/tests/init_env_utils.h
@@ -0,0 +1,23 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+namespace paddle {
+namespace prim {
+
+void InitTensorOperants();
+
+}  // namespace prim
+}  // namespace paddle
diff --git a/paddle/fluid/prim/tests/test_eager_prim.cc b/paddle/fluid/prim/tests/test_eager_prim.cc
index 35902797ea24517d834715711a670f6ece4b899d..a4ab38a6ad137b98e52fddc2c6d399b43d9a4ca3 100644
--- a/paddle/fluid/prim/tests/test_eager_prim.cc
+++ b/paddle/fluid/prim/tests/test_eager_prim.cc
@@ -14,17 +14,21 @@
 
 #include <sstream>
 
+#include "gflags/gflags.h"
 #include "glog/logging.h"
 #include "gtest/gtest.h"
 #include "paddle/fluid/eager/api/generated/eager_generated/forwards/dygraph_functions.h"
 #include "paddle/fluid/eager/api/utils/hook_utils.h"
 #include "paddle/fluid/eager/backward.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
+#include "paddle/fluid/prim/tests/init_env_utils.h"
 #include "paddle/fluid/prim/utils/utils.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/core/tensor_meta.h"
 
+DECLARE_string(tensor_operants_mode);
+
 PD_DECLARE_KERNEL(full, CPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(tanh, CPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(tanh_grad, CPU, ALL_LAYOUT);
@@ -46,6 +50,8 @@ namespace prim {
 TEST(EagerPrim, TanhBackwardTest) {
   // 1. Initialized
   eager_test::InitEnv(paddle::platform::CPUPlace());
+  FLAGS_tensor_operants_mode = "eager";
+  paddle::prim::InitTensorOperants();
   // 2. pre
   paddle::framework::DDim ddim = phi::make_ddim({4, 16, 16, 32});
   paddle::experimental::Tensor tensor0 =
diff --git a/paddle/fluid/prim/tests/test_static_prim.cc b/paddle/fluid/prim/tests/test_static_prim.cc
index 5a53101ab13bb01c8336cf462873517a48596dcc..b93e16b507f99c74414303102739cc8191c6074d 100644
--- a/paddle/fluid/prim/tests/test_static_prim.cc
+++ b/paddle/fluid/prim/tests/test_static_prim.cc
@@ -21,11 +21,15 @@
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/prim/api/manual_prim/utils/utils.h"
 #include "paddle/fluid/prim/utils/static/desc_tensor.h"
+#include "paddle/fluid/prim/utils/static/static_tensor_operants.h"
 #include "paddle/fluid/prim/utils/utils.h"
 #include "paddle/phi/core/enforce.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/core/operants_manager.h"
 
 DECLARE_bool(prim_enabled);
+DECLARE_string(tensor_operants_mode);
+
 PD_DECLARE_KERNEL(full, CPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(tanh, CPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(tanh_grad, CPU, ALL_LAYOUT);
@@ -142,6 +146,11 @@ class TestCompositeGradMaker : public CompositeGradOpMakerBase {
 };
 
 TEST(StaticPrim, TanhBackwardComposite) {
+  // Initialized environment
+  FLAGS_tensor_operants_mode = "static";
+  paddle::operants::OperantsManager::Instance().static_operants.reset(
+      new paddle::operants::StaticTensorOperants());
+
   TestBaseProgram base_program = TestBaseProgram();
   auto* target_block = base_program.GetBlock(0);
   // Prepare for forward tanh
@@ -223,6 +232,11 @@ TEST(StaticPrim, TanhBackwardComposite) {
 }
 
 TEST(StaticCompositeGradMaker, TestMutiInputMethod) {
+  // Initialized environment
+  FLAGS_tensor_operants_mode = "static";
+  paddle::operants::OperantsManager::Instance().static_operants.reset(
+      new paddle::operants::StaticTensorOperants());
+
   TestBaseProgram base_program = TestBaseProgram();
   auto* target_block = base_program.GetBlock(0);
   std::vector<int64_t> shape = {2, 2};
@@ -285,6 +299,11 @@ TEST(StaticCompositeGradMaker, TestMutiInputMethod) {
 }
 
 TEST(StaticCompositeGradMaker, TestMutiOutputMethod) {
+  // Initialized environment
+  FLAGS_tensor_operants_mode = "static";
+  paddle::operants::OperantsManager::Instance().static_operants.reset(
+      new paddle::operants::StaticTensorOperants());
+
   TestBaseProgram base_program = TestBaseProgram();
   auto* target_block = base_program.GetBlock(0);
   std::vector<int64_t> shape = {4, 2};
diff --git a/paddle/fluid/prim/utils/eager/CMakeLists.txt b/paddle/fluid/prim/utils/eager/CMakeLists.txt
index 8b137891791fe96927ad78e64b0aad7bded08bdc..2551a011015188ea65e899e431fb9f5a8dee26e3 100644
--- a/paddle/fluid/prim/utils/eager/CMakeLists.txt
+++ b/paddle/fluid/prim/utils/eager/CMakeLists.txt
@@ -1 +1,6 @@
-
+if(NOT (NOT WITH_PYTHON AND ON_INFER))
+  cc_library(
+    eager_tensor_operants
+    SRCS eager_tensor_operants.cc
+    DEPS final_dygraph_function)
+endif()
diff --git a/paddle/fluid/prim/utils/eager/eager_tensor_operants.cc b/paddle/fluid/prim/utils/eager/eager_tensor_operants.cc
new file mode 100644
index 0000000000000000000000000000000000000000..e42a0d6dd9135001505e64c79e5c435b671847ec
--- /dev/null
+++ b/paddle/fluid/prim/utils/eager/eager_tensor_operants.cc
@@ -0,0 +1,29 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/prim/utils/eager/eager_tensor_operants.h"
+
+#include "glog/logging.h"
+#include "paddle/fluid/eager/api/generated/eager_generated/forwards/dygraph_functions.h"
+
+namespace paddle {
+
+namespace operants {
+
+Tensor EagerTensorOperants::multiply(const Tensor& x, const Tensor& y) {
+  return ::multiply_ad_func(x, y);
+}
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/fluid/prim/utils/eager/eager_tensor_operants.h b/paddle/fluid/prim/utils/eager/eager_tensor_operants.h
new file mode 100644
index 0000000000000000000000000000000000000000..b7571de350fb03d9233e9357ae8ff5c645898b1d
--- /dev/null
+++ b/paddle/fluid/prim/utils/eager/eager_tensor_operants.h
@@ -0,0 +1,36 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/core/macros.h"
+#include "paddle/phi/core/operants_base.h"
+
+namespace paddle {
+
+namespace operants {
+
+class EagerTensorOperants : public TensorOperantsBase {
+ public:
+  EagerTensorOperants() = default;
+
+  Tensor multiply(const Tensor& x, const Tensor& y) override;
+
+ private:
+  DISABLE_COPY_AND_ASSIGN(EagerTensorOperants);
+};
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/fluid/prim/utils/static/CMakeLists.txt b/paddle/fluid/prim/utils/static/CMakeLists.txt
index db310e2ab61d4e89a0204a9186654c8974c1f44c..aa72fadb591a6d9c8ce403cd311f65c6b7726b0d 100644
--- a/paddle/fluid/prim/utils/static/CMakeLists.txt
+++ b/paddle/fluid/prim/utils/static/CMakeLists.txt
@@ -2,3 +2,8 @@ cc_library(
   static_global_utils
   SRCS static_global_utils.cc
   DEPS proto_desc)
+
+cc_library(
+  static_tensor_operants
+  SRCS static_tensor_operants.cc
+  DEPS static_prim_api)
diff --git a/paddle/fluid/prim/utils/static/composite_grad_desc_maker.h b/paddle/fluid/prim/utils/static/composite_grad_desc_maker.h
index efb2479e4051444fb51a3ecf10f9a9b83598498a..6de5c1ae41c0308f9d87bfbde1390114ffce5e99 100644
--- a/paddle/fluid/prim/utils/static/composite_grad_desc_maker.h
+++ b/paddle/fluid/prim/utils/static/composite_grad_desc_maker.h
@@ -29,6 +29,10 @@
 #include "paddle/fluid/prim/utils/static/desc_tensor.h"
 #include "paddle/fluid/prim/utils/static/static_global_utils.h"
 #include "paddle/phi/core/enforce.h"
+#include "paddle/phi/core/flags.h"
+
+DECLARE_string(tensor_operants_mode);
+
 namespace paddle {
 namespace prim {
 
@@ -59,6 +63,7 @@ class CompositeGradOpMakerBase {
     // TODO(jiabin): This should always execute by one thread...
     VLOG(6) << "Constructing Composite Grad func for " << fwd_op_.Type()
             << "_grad ";
+    FLAGS_tensor_operants_mode = "static";
     StaticCompositeContext::Instance().SetBlock(
         acting_program_.MutableBlock(0));
   }
diff --git a/paddle/fluid/prim/utils/static/static_tensor_operants.cc b/paddle/fluid/prim/utils/static/static_tensor_operants.cc
new file mode 100644
index 0000000000000000000000000000000000000000..ada4cca58a92c20d020df8add123a89eda5ead42
--- /dev/null
+++ b/paddle/fluid/prim/utils/static/static_tensor_operants.cc
@@ -0,0 +1,31 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/prim/utils/static/static_tensor_operants.h"
+
+#include "glog/logging.h"
+#include "paddle/fluid/prim/api/generated_prim/prim_generated_api.h"
+#include "paddle/fluid/prim/utils/static/desc_tensor.h"
+
+namespace paddle {
+
+namespace operants {
+using DescTensor = paddle::prim::DescTensor;
+
+Tensor StaticTensorOperants::multiply(const Tensor& x, const Tensor& y) {
+  return paddle::prim::multiply<DescTensor>(x, y);
+}
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/fluid/prim/utils/static/static_tensor_operants.h b/paddle/fluid/prim/utils/static/static_tensor_operants.h
new file mode 100644
index 0000000000000000000000000000000000000000..26365fcf233ee48685ee0347208e9aafb94cf281
--- /dev/null
+++ b/paddle/fluid/prim/utils/static/static_tensor_operants.h
@@ -0,0 +1,36 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/core/macros.h"
+#include "paddle/phi/core/operants_base.h"
+
+namespace paddle {
+
+namespace operants {
+
+class StaticTensorOperants : public TensorOperantsBase {
+ public:
+  StaticTensorOperants() = default;
+
+  Tensor multiply(const Tensor& x, const Tensor& y) override;
+
+ private:
+  DISABLE_COPY_AND_ASSIGN(StaticTensorOperants);
+};
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
index 283b305d71806facbe647a00310e12f0edec66c0..f06efbb211a0140b8febd3fd5006f8b34d84c4ce 100755
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -497,6 +497,10 @@ if(WITH_PYTHON)
     list(APPEND PYBIND_DEPS python)
     list(APPEND PYBIND_DEPS custom_operator)
     list(APPEND PYBIND_DEPS custom_operator_node)
+    list(APPEND PYBIND_DEPS operants_manager)
+    list(APPEND PYBIND_DEPS eager_tensor_operants)
+    list(APPEND PYBIND_DEPS static_tensor_operants)
+    list(APPEND PYBIND_DEPS phi_tensor_operants)
   endif()
 
   # On Linux, cc_library(paddle SHARED ..) will generate the libpaddle.so,
diff --git a/paddle/fluid/pybind/eager_functions.cc b/paddle/fluid/pybind/eager_functions.cc
index cc5a8d64e1234e7819118b6ac8a0d2ea71c4f6ea..e1a050178585f9f0d817938efce60fb9f3116a74 100644
--- a/paddle/fluid/pybind/eager_functions.cc
+++ b/paddle/fluid/pybind/eager_functions.cc
@@ -35,6 +35,8 @@ typedef SSIZE_T ssize_t;
 #include "paddle/fluid/platform/device/gpu/gpu_info.h"
 #include "paddle/fluid/platform/dynload/dynamic_loader.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/prim/utils/eager/eager_tensor_operants.h"
+#include "paddle/fluid/prim/utils/static/static_tensor_operants.h"
 #include "paddle/fluid/pybind/eager.h"
 #include "paddle/fluid/pybind/eager_utils.h"
 #include "paddle/fluid/pybind/exception.h"
@@ -54,6 +56,12 @@ typedef SSIZE_T ssize_t;
 #include "paddle/fluid/pybind/cuda_streams_py.h"
 #endif
 
+#include "gflags/gflags.h"
+#include "paddle/phi/api/include/tensor_operants.h"
+#include "paddle/phi/core/operants_manager.h"
+
+DECLARE_string(tensor_operants_mode);
+
 namespace paddle {
 namespace pybind {
 
@@ -487,10 +495,32 @@ static PyObject* eager_api_jit_function_call(PyObject* self,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
+static PyObject* eager_api_init_eager_and_static_tensor_operants(
+    PyObject* self, PyObject* args, PyObject* kwargs) {
+  EAGER_TRY
+
+  paddle::operants::OperantsManager::Instance().eager_operants.reset(
+      new paddle::operants::EagerTensorOperants());
+  paddle::operants::OperantsManager::Instance().static_operants.reset(
+      new paddle::operants::StaticTensorOperants());
+  VLOG(4) << "Initialize eager and static tensor operants successfully";
+
+  RETURN_PY_NONE
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 static PyObject* eager_api_run_custom_op(PyObject* self,
                                          PyObject* args,
                                          PyObject* kwargs) {
   EAGER_TRY
+  FLAGS_tensor_operants_mode = "phi";
+  if (paddle::operants::OperantsManager::Instance().phi_operants.get() ==
+      nullptr) {
+    paddle::operants::OperantsManager::Instance().phi_operants.reset(
+        new paddle::operants::PhiTensorOperants());
+    VLOG(4) << "Initialize phi tensor operants successfully";
+  }
+
   paddle::CustomOpKernelContext ctx =
       CastPyArg2CustomOpKernelContext(PyTuple_GET_ITEM(args, 0), 0);
   std::string op_type = CastPyArg2AttrString(PyTuple_GET_ITEM(args, 1), 1);
@@ -1090,6 +1120,11 @@ PyMethodDef variable_functions[] = {
      (PyCFunction)(void (*)(void))eager_api_run_custom_op,
      METH_VARARGS | METH_KEYWORDS,
      NULL},
+    {"_init_eager_and_static_tensor_operants",
+     (PyCFunction)(void (*)(
+         void))eager_api_init_eager_and_static_tensor_operants,
+     METH_VARARGS | METH_KEYWORDS,
+     NULL},
     {"tensor_copy",
      (PyCFunction)(void (*)(void))eager_api_tensor_copy,
      METH_VARARGS | METH_KEYWORDS,
diff --git a/paddle/phi/api/include/tensor.h b/paddle/phi/api/include/tensor.h
index d3e1a21eb929d62d9109fb4d2d2ed0fb64d3f132..35f7e88943291007fde13f8f46e9f42b08edf774 100644
--- a/paddle/phi/api/include/tensor.h
+++ b/paddle/phi/api/include/tensor.h
@@ -615,5 +615,7 @@ class PADDLE_API Tensor final {
   std::string name_{""};
 };
 
+PADDLE_API Tensor operator*(const Tensor& x, const Tensor& y);
+
 }  // namespace experimental
 }  // namespace paddle
diff --git a/paddle/phi/api/include/tensor_operants.h b/paddle/phi/api/include/tensor_operants.h
new file mode 100644
index 0000000000000000000000000000000000000000..89ac46a1953c2909dc14b2a6a50e18a01db99bd3
--- /dev/null
+++ b/paddle/phi/api/include/tensor_operants.h
@@ -0,0 +1,36 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/core/macros.h"
+#include "paddle/phi/core/operants_base.h"
+
+namespace paddle {
+
+namespace operants {
+
+class PhiTensorOperants : public TensorOperantsBase {
+ public:
+  PhiTensorOperants() = default;
+
+  Tensor multiply(const Tensor& x, const Tensor& y) override;
+
+ private:
+  DISABLE_COPY_AND_ASSIGN(PhiTensorOperants);
+};
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/phi/api/lib/CMakeLists.txt b/paddle/phi/api/lib/CMakeLists.txt
index 05499be590377f056799cfa081f1e32534ff2fd8..f6ade16361ef0280461ca5f15a4cc30e1a364293 100644
--- a/paddle/phi/api/lib/CMakeLists.txt
+++ b/paddle/phi/api/lib/CMakeLists.txt
@@ -4,17 +4,20 @@ if(WITH_GPU)
   nv_library(
     phi_tensor_raw
     SRCS tensor.cc
-    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool)
+    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool
+         operants_manager)
 elseif(WITH_ROCM)
   hip_library(
     phi_tensor_raw
     SRCS tensor.cc
-    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool)
+    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool
+         operants_manager)
 else()
   cc_library(
     phi_tensor_raw
     SRCS tensor.cc
-    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool)
+    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool
+         operants_manager)
 endif()
 
 set(api_gen_base ${CMAKE_SOURCE_DIR}/paddle/phi/api/yaml/generator/api_base.py)
@@ -308,3 +311,8 @@ cc_library(
   api_int_array
   SRCS int_array.cc
   DEPS tensor_copy)
+
+cc_library(
+  phi_tensor_operants
+  SRCS tensor_operants.cc
+  DEPS phi_function_api)
diff --git a/paddle/phi/api/lib/tensor.cc b/paddle/phi/api/lib/tensor.cc
index 8985383fd2cba156f64962a1dda43094c8e5e00f..118be82654964b75c52a504030b1d0efba406c82 100644
--- a/paddle/phi/api/lib/tensor.cc
+++ b/paddle/phi/api/lib/tensor.cc
@@ -27,6 +27,7 @@ limitations under the License. */
 #include "paddle/phi/core/ddim.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/enforce.h"
+#include "paddle/phi/core/operants_manager.h"
 #include "paddle/phi/core/selected_rows.h"
 #include "paddle/phi/core/sparse_coo_tensor.h"
 #include "paddle/phi/core/sparse_csr_tensor.h"
@@ -413,5 +414,9 @@ void Tensor::reset_inplace_version(bool set_to_zero) {
   }
 }
 
+PADDLE_API Tensor operator*(const Tensor &x, const Tensor &y) {
+  return paddle::operants::OperantsManager::Instance().multiply(x, y);
+}
+
 }  // namespace experimental
 }  // namespace paddle
diff --git a/paddle/phi/api/lib/tensor_operants.cc b/paddle/phi/api/lib/tensor_operants.cc
new file mode 100644
index 0000000000000000000000000000000000000000..22ab35c47860e7d9ee719b0924b4c5d2b16d8444
--- /dev/null
+++ b/paddle/phi/api/lib/tensor_operants.cc
@@ -0,0 +1,28 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/api/include/tensor_operants.h"
+#include "glog/logging.h"
+#include "paddle/phi/api/include/api.h"
+
+namespace paddle {
+
+namespace operants {
+
+Tensor PhiTensorOperants::multiply(const Tensor& x, const Tensor& y) {
+  return paddle::experimental::multiply(x, y);
+}
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/phi/core/CMakeLists.txt b/paddle/phi/core/CMakeLists.txt
index 85df40cc47c5c2fe8e090477c3cba8b9220405ec..caa571aff8d8284972441ef3ceb059027e793a24 100644
--- a/paddle/phi/core/CMakeLists.txt
+++ b/paddle/phi/core/CMakeLists.txt
@@ -114,6 +114,11 @@ cc_library(
   SRCS custom_kernel.cc
   DEPS kernel_factory)
 
+cc_library(
+  operants_manager
+  SRCS operants_manager.cc
+  DEPS flags)
+
 cc_library(
   mixed_vector
   SRCS mixed_vector.cc
diff --git a/paddle/phi/core/flags.cc b/paddle/phi/core/flags.cc
index 680661c890519a076af746bf34084ffe4d655a7c..53fd2a59d397dfd3edc27a8c8f17426b73a2a8e6 100644
--- a/paddle/phi/core/flags.cc
+++ b/paddle/phi/core/flags.cc
@@ -1206,3 +1206,19 @@ PADDLE_DEFINE_EXPORTED_bool(trt_ibuilder_cache,
 PADDLE_DEFINE_EXPORTED_bool(use_shm_cache,
                             false,
                             "Use shm cache in mmap_allocator.");
+
+/**
+ * Tensor operants related FLAG
+ * Name: tensor_operants_mode
+ * Since Version: 2.5.0
+ * Value Range: string, {eager, phi, static}
+ * default=eager
+ * Example:
+ * Note: For switching tensor operants mode of PaddlePaddle.
+ *       - eager mode: tensor operants with dygraph autograd;
+ *       - phi mode: tensor operants with only phi forward API;
+ *       - static mode: tensor operants within static graph.
+ */
+PADDLE_DEFINE_EXPORTED_string(tensor_operants_mode,
+                              "eager",
+                              "Tensor operants mode");
diff --git a/paddle/phi/core/operants_base.h b/paddle/phi/core/operants_base.h
new file mode 100644
index 0000000000000000000000000000000000000000..ca993e7d719f4b02d095e2e5189f4c70a14b708b
--- /dev/null
+++ b/paddle/phi/core/operants_base.h
@@ -0,0 +1,33 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/api/include/tensor.h"
+
+namespace paddle {
+
+namespace operants {
+
+using Tensor = paddle::experimental::Tensor;
+
+class TensorOperantsBase {
+ public:
+  virtual ~TensorOperantsBase() = default;
+
+  virtual Tensor multiply(const Tensor& x, const Tensor& y) = 0;
+};
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/phi/core/operants_manager.cc b/paddle/phi/core/operants_manager.cc
new file mode 100644
index 0000000000000000000000000000000000000000..14db7320de720d8d0fdbe57e90104c6200c5abac
--- /dev/null
+++ b/paddle/phi/core/operants_manager.cc
@@ -0,0 +1,67 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/core/operants_manager.h"
+
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "paddle/phi/core/enforce.h"
+#include "paddle/phi/core/errors.h"
+
+DECLARE_string(tensor_operants_mode);
+
+namespace paddle {
+
+namespace operants {
+
+OperantsManager& OperantsManager::Instance() {
+  static OperantsManager g_op_manager;
+  return g_op_manager;
+}
+
+Tensor OperantsManager::multiply(const Tensor& x, const Tensor& y) {
+  if (FLAGS_tensor_operants_mode == "eager") {
+    PADDLE_ENFORCE_NE(
+        this->eager_operants.get(),
+        nullptr,
+        phi::errors::Unavailable("The eager_operants pointer of "
+                                 "OperantsManager is not initialized"));
+    VLOG(4) << "OperantsManager reaches eager mode";
+    return this->eager_operants->multiply(x, y);
+  } else if (FLAGS_tensor_operants_mode == "static") {
+    PADDLE_ENFORCE_NE(
+        this->static_operants.get(),
+        nullptr,
+        phi::errors::Unavailable("The static_operants pointer of "
+                                 "OperantsManager is not initialized"));
+    VLOG(4) << "OperantsManager reaches static mode";
+    return this->static_operants->multiply(x, y);
+  } else if (FLAGS_tensor_operants_mode == "phi") {
+    PADDLE_ENFORCE_NE(
+        this->phi_operants.get(),
+        nullptr,
+        phi::errors::Unavailable(
+            "The phi_operants pointer of OperantsManager is not initialized"));
+    VLOG(4) << "OperantsManager reaches phi mode";
+    return this->phi_operants->multiply(x, y);
+  } else {
+    PADDLE_THROW(phi::errors::Unimplemented(
+        "FLAGS_tensor_operants_mode is not nitialized, please set "
+        "FLAGS_tensor_operants_mode first, which currently supports eager, "
+        "phi, and static mode"));
+  }
+}
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/paddle/phi/core/operants_manager.h b/paddle/phi/core/operants_manager.h
new file mode 100644
index 0000000000000000000000000000000000000000..5bed59d6591529014aade994e48d6f9413310c65
--- /dev/null
+++ b/paddle/phi/core/operants_manager.h
@@ -0,0 +1,77 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/api/include/tensor.h"
+#include "paddle/phi/core/macros.h"
+#include "paddle/phi/core/operants_base.h"
+
+namespace paddle {
+
+namespace operants {
+
+using Tensor = paddle::experimental::Tensor;
+
+/**
+ * [ Why need OperantsManager? ]
+ *
+ * Ideally, overloading tensor operators should call Tensor API directly.
+ * However, we faced two problems:
+ *
+ * 1. Support multiple modes: Tensor operator overloading needs to support
+ * [static mode / autograd mode / custom operator mode] at the same time.
+ *
+ * 2. Decouple phi and fluid: Tensor belongs to the phi library, but it relies
+ * upon functions in fluid when overloading Tensor operators.
+ *
+ * We design OperantsManager to solve these two problems:
+ *
+ * 1. use `FLAGS_tensor_operants_mode` to handle overloading mode, set this flag
+ * at the entry point of each mode:
+ *
+ * - FLAGS_tensor_operants_mode = "static": at the construction function of
+ * `CompositeGradOpMakerBase`.
+ * - FLAGS_tensor_operants_mode = "eager": at the beginning of dygraph_function.
+ * - FLAGS_tensor_operants_mode = "phi": at the beginning of the
+ * `eager_api_run_custom_op` function in eager mode and at the beginning of
+ * calling kernels in static mode.
+ *
+ * In order to guarantee the performance, OperantsManager holds three pointers
+ * to identify each mode respectively.
+ *
+ * 2. Decouple phi with the help of the polymorphism mechanism,
+ * TensorOperantsBase derives three child classes: PhiTensorOperants,
+ * EagerTensorOperants, and StaticTensorOperants. We set eager and static tensor
+ * operants at the fluid library and set phi operants at the phi library.
+ *
+ */
+class OperantsManager {
+ public:
+  static OperantsManager& Instance();
+
+  Tensor multiply(const Tensor& x, const Tensor& y);
+
+ public:
+  std::unique_ptr<TensorOperantsBase> eager_operants{nullptr};
+  std::unique_ptr<TensorOperantsBase> static_operants{nullptr};
+  std::unique_ptr<TensorOperantsBase> phi_operants{nullptr};
+
+ private:
+  OperantsManager() = default;
+  DISABLE_COPY_AND_ASSIGN(OperantsManager);
+};
+
+}  // namespace operants
+}  // namespace paddle
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index 1bd64072d14021621b4a4304981465ba61bf66be..acc7e55218598e8de9dd1ae5d52ca09d58457dfa 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -239,6 +239,7 @@ def __bootstrap__():
         core.init_glog(sys.argv[0])
     # don't init_p2p when in unittest to save time.
     core.init_devices()
+    core.eager._init_eager_and_static_tensor_operants()
     core.init_default_kernel_signatures()
 
 
diff --git a/python/paddle/fluid/tests/custom_op/CMakeLists.txt b/python/paddle/fluid/tests/custom_op/CMakeLists.txt
index b939ee43883a9eb5952384d30c8382cd5acbf5fe..955dff3b08bf89d8cf1361fff97b031ace56c06c 100644
--- a/python/paddle/fluid/tests/custom_op/CMakeLists.txt
+++ b/python/paddle/fluid/tests/custom_op/CMakeLists.txt
@@ -27,6 +27,8 @@ endif()
 
 py_test(test_custom_raw_op_kernel_op SRCS test_custom_raw_op_kernel_op.py)
 set_tests_properties(test_custom_raw_op_kernel_op PROPERTIES TIMEOUT 180)
+py_test(test_custom_power_jit SRCS test_custom_power_jit.py)
+set_tests_properties(test_custom_power_jit PROPERTIES TIMEOUT 180)
 
 # CPU custom op tests: only compile .cc file
 py_test(test_dispatch_jit SRCS test_dispatch_jit.py)
diff --git a/python/paddle/fluid/tests/custom_op/custom_power.cc b/python/paddle/fluid/tests/custom_op/custom_power.cc
new file mode 100644
index 0000000000000000000000000000000000000000..2c7bd53115ef8595b330ade0d1631a9c171d6ddb
--- /dev/null
+++ b/python/paddle/fluid/tests/custom_op/custom_power.cc
@@ -0,0 +1,47 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <iostream>
+#include <vector>
+
+#include "paddle/extension.h"
+
+std::vector<paddle::Tensor> PowerForward(const paddle::Tensor& x) {
+  if (x.is_cpu() || x.is_gpu()) {
+    return {x * x};
+  } else {
+    PD_THROW("Not implemented.");
+  }
+}
+
+std::vector<paddle::Tensor> PowerBackward(const paddle::Tensor& x,
+                                          const paddle::Tensor& out,
+                                          const paddle::Tensor& grad_out) {
+  if (x.is_cpu() || x.is_gpu()) {
+    paddle::Tensor middle_result = grad_out * x;
+    return {paddle::add(middle_result, middle_result)};
+  } else {
+    PD_THROW("Not implemented.");
+  }
+}
+
+PD_BUILD_OP(custom_power)
+    .Inputs({"X"})
+    .Outputs({"Out"})
+    .SetKernelFn(PD_KERNEL(PowerForward));
+
+PD_BUILD_GRAD_OP(custom_power)
+    .Inputs({"X", "Out", paddle::Grad("Out")})
+    .Outputs({paddle::Grad("X")})
+    .SetKernelFn(PD_KERNEL(PowerBackward));
diff --git a/python/paddle/fluid/tests/custom_op/custom_relu_op_xpu.cc b/python/paddle/fluid/tests/custom_op/custom_relu_op_xpu.cc
index 4bcd5af1c4aed90ce925f500da54bce3b3fc9df9..0fb63266c8ac2c0b6ab2a03df567b9a3ef56bef9 100644
--- a/python/paddle/fluid/tests/custom_op/custom_relu_op_xpu.cc
+++ b/python/paddle/fluid/tests/custom_op/custom_relu_op_xpu.cc
@@ -118,7 +118,7 @@ std::vector<paddle::Tensor> relu_xpu_backward(const paddle::Tensor& x,
   auto zeros = paddle::experimental::full_like(x, 0.0, x.dtype(), x.place());
   auto condition = paddle::experimental::greater_than(x, zeros);
 
-  grad_x = paddle::multiply(grad_out, paddle::where(condition, ones, zeros));
+  grad_x = grad_out * paddle::where(condition, ones, zeros);
 
   return {grad_x};
 }
diff --git a/python/paddle/fluid/tests/custom_op/test_custom_power_jit.py b/python/paddle/fluid/tests/custom_op/test_custom_power_jit.py
new file mode 100644
index 0000000000000000000000000000000000000000..ab0d3edc11c4af19ee47203254b2a74f149674ae
--- /dev/null
+++ b/python/paddle/fluid/tests/custom_op/test_custom_power_jit.py
@@ -0,0 +1,126 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import unittest
+
+import numpy as np
+from utils import extra_cc_args, paddle_includes
+
+import paddle
+import paddle.static as static
+from paddle.utils.cpp_extension import get_build_directory, load
+from paddle.utils.cpp_extension.extension_utils import run_cmd
+
+# Because Windows don't use docker, the shared lib already exists in the
+# cache dir, it will not be compiled again unless the shared lib is removed.
+file = '{}\\custom_power_jit\\custom_power_jit.pyd'.format(
+    get_build_directory()
+)
+if os.name == 'nt' and os.path.isfile(file):
+    cmd = 'del {}'.format(file)
+    run_cmd(cmd, True)
+
+custom_module = load(
+    name='custom_power_jit',
+    sources=['custom_power.cc'],
+    extra_include_paths=paddle_includes,  # add for Coverage CI
+    extra_cxx_cflags=extra_cc_args,  # test for cc flags
+    verbose=True,
+)
+
+
+def custom_power_dynamic(func, device, dtype, np_x, use_func=True):
+    paddle.set_device(device)
+
+    t = paddle.to_tensor(np_x, dtype=dtype)
+    t.stop_gradient = False
+
+    out = func(t) if use_func else paddle.pow(t, 2)
+    out.stop_gradient = False
+
+    out.backward()
+    if t.grad is None:
+        return out.numpy(), t.grad
+    else:
+        return out.numpy(), t.grad.numpy()
+
+
+def custom_power_static(func, device, dtype, np_x, use_func=True):
+    paddle.enable_static()
+    paddle.set_device(device)
+
+    with static.scope_guard(static.Scope()):
+        with static.program_guard(static.Program()):
+            x = static.data(name='X', shape=[None, 8], dtype=dtype)
+            x.stop_gradient = False
+            out = func(x) if use_func else paddle.pow(x, 2)
+            static.append_backward(out)
+
+            exe = static.Executor()
+            exe.run(static.default_startup_program())
+            # in static graph mode, x data has been covered by out
+            out_v = exe.run(
+                static.default_main_program(),
+                feed={'X': np_x},
+                fetch_list=[out.name],
+            )
+
+    paddle.disable_static()
+    return out_v
+
+
+class TestJITLoad(unittest.TestCase):
+    def setUp(self):
+        self.custom_op = custom_module.custom_power
+        self.devices = ['cpu']
+        self.dtypes = ['float32', 'float64']
+        if paddle.is_compiled_with_cuda():
+            self.devices.append('gpu')
+            self.dtypes.append('float16')
+
+    def test_static(self):
+        for device in self.devices:
+            for dtype in self.dtypes:
+                if device == 'cpu' and dtype == 'float16':
+                    continue
+                x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
+
+                out = custom_power_static(self.custom_op, device, dtype, x)
+                pd_out = custom_power_static(
+                    self.custom_op, device, dtype, x, False
+                )
+                np.testing.assert_allclose(out, pd_out, rtol=1e-5, atol=1e-8)
+
+    def test_dynamic(self):
+        for device in self.devices:
+            for dtype in self.dtypes:
+                if device == 'cpu' and dtype == 'float16':
+                    continue
+                x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
+
+                out, x_grad = custom_power_dynamic(
+                    self.custom_op, device, dtype, x
+                )
+                pd_out, pd_x_grad = custom_power_dynamic(
+                    self.custom_op, device, dtype, x, False
+                )
+                np.testing.assert_allclose(out, pd_out, rtol=1e-5, atol=1e-8)
+                np.testing.assert_allclose(
+                    x_grad, pd_x_grad, rtol=1e-5, atol=1e-8
+                )
+
+
+if __name__ == '__main__':
+    unittest.main()