diff --git a/paddle/fluid/eager/CMakeLists.txt b/paddle/fluid/eager/CMakeLists.txt
old mode 100644
new mode 100755
index fd02d247788e90c237db6e7b9e96eee73fe89821..d8a6ccc65bffa4f505684180abe2568d33a2201c
--- a/paddle/fluid/eager/CMakeLists.txt
+++ b/paddle/fluid/eager/CMakeLists.txt
@@ -13,9 +13,12 @@ set(eager_deps
     eager_nan_inf_utils
     grad_node_info
     grad_tensor_holder
-    accumulation_node
     custom_operator_node)
 
+if(NOT (NOT WITH_PYTHON AND ON_INFER))
+  set(eager_deps ${eager_deps} accumulation_node)
+endif()
+
 set(fluid_deps
     tracer
     layer
@@ -33,9 +36,9 @@ if(NOT ((NOT WITH_PYTHON) AND ON_INFER))
 endif()
 
 add_subdirectory(api)
-add_subdirectory(accumulation)
 add_subdirectory(custom_operator)
 if(NOT ((NOT WITH_PYTHON) AND ON_INFER))
+  add_subdirectory(accumulation)
   add_subdirectory(tests)
   add_subdirectory(pylayer)
   cc_library(
diff --git a/paddle/fluid/eager/accumulation/CMakeLists.txt b/paddle/fluid/eager/accumulation/CMakeLists.txt
old mode 100644
new mode 100755
index 297e853947dfbe4330a16678bced454a46ea982e..a924a9b106dd134be059851a9c194d900e345506
--- a/paddle/fluid/eager/accumulation/CMakeLists.txt
+++ b/paddle/fluid/eager/accumulation/CMakeLists.txt
@@ -1,4 +1,6 @@
-cc_library(
-  accumulation_node
-  SRCS accumulation_node.cc
-  DEPS gradient_accumulator phi_api grad_node_info)
+if(NOT (NOT WITH_PYTHON AND ON_INFER))
+  cc_library(
+    accumulation_node
+    SRCS accumulation_node.cc
+    DEPS gradient_accumulator phi_api grad_node_info)
+endif()
diff --git a/paddle/fluid/eager/accumulation/accumulation_node.cc b/paddle/fluid/eager/accumulation/accumulation_node.cc
old mode 100644
new mode 100755
index 4e5afca78ade0076db3d12f1f4e326bb6c6811f9..f06c4b63ae212fbd9a9e7d80baa1b8376b81c5ea
--- a/paddle/fluid/eager/accumulation/accumulation_node.cc
+++ b/paddle/fluid/eager/accumulation/accumulation_node.cc
@@ -15,6 +15,7 @@
 #include "paddle/fluid/eager/accumulation/accumulation_node.h"
 
 #include "glog/logging.h"
+#include "paddle/fluid/eager/api/generated/eager_generated/forwards/dygraph_functions.h"
 #include "paddle/fluid/eager/eager_tensor.h"
 #include "paddle/fluid/eager/utils.h"
 #include "paddle/fluid/imperative/gradient_accumulator.h"
@@ -44,8 +45,12 @@ static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
       // Accumulation
       if (LIKELY(t.is_dense_tensor())) {
         if (LIKELY(tensor->is_dense_tensor())) {
-          paddle::imperative::TensorAdd<paddle::experimental::Tensor>(t,
-                                                                      tensor);
+          if (t.is_custom_device()) {
+            *tensor = add_ad_func(t, *tensor);
+          } else {
+            paddle::imperative::TensorAdd<paddle::experimental::Tensor>(t,
+                                                                        tensor);
+          }
         } else {
           // TODO(jiabin): Support Other TensorBase later
           // TODO(zhanlve): Replace SelectedRowsAddTensor with
@@ -68,8 +73,12 @@ static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
           paddle::experimental::Tensor tensor_values(
               std::make_shared<phi::DenseTensor>(
                   tensor_sparse->non_zero_elements()));
-          paddle::imperative::TensorAdd<paddle::experimental::Tensor>(
-              t_values, &tensor_values);
+          if (t.is_custom_device()) {
+            tensor_values = add_ad_func(t_values, tensor_values);
+          } else {
+            paddle::imperative::TensorAdd<paddle::experimental::Tensor>(
+                t_values, &tensor_values);
+          }
         }
       } else {
         // TODO(jiabin): Support Other TensorBase later
diff --git a/paddle/fluid/eager/api/utils/CMakeLists.txt b/paddle/fluid/eager/api/utils/CMakeLists.txt
old mode 100644
new mode 100755
index 1fd4905605ea812fc73e8bd501510f1863fb4766..dbb59e5aae7266098e51205429785af374fee005
--- a/paddle/fluid/eager/api/utils/CMakeLists.txt
+++ b/paddle/fluid/eager/api/utils/CMakeLists.txt
@@ -1,12 +1,24 @@
-cc_library(
-  tensor_utils
-  SRCS tensor_utils.cc
-  DEPS phi_api autograd_meta grad_node_info accumulation_node)
-cc_library(
-  hook_utils
-  SRCS hook_utils.cc
-  DEPS phi tensor_utils autograd_meta grad_node_info utils accumulation_node)
 cc_library(
   global_utils
   SRCS global_utils.cc
   DEPS place tracer)
+
+if(NOT (NOT WITH_PYTHON AND ON_INFER))
+  cc_library(
+    tensor_utils
+    SRCS tensor_utils.cc
+    DEPS phi_api autograd_meta grad_node_info accumulation_node)
+  cc_library(
+    hook_utils
+    SRCS hook_utils.cc
+    DEPS phi tensor_utils autograd_meta grad_node_info utils accumulation_node)
+else()
+  cc_library(
+    tensor_utils
+    SRCS tensor_utils.cc
+    DEPS phi_api autograd_meta grad_node_info)
+  cc_library(
+    hook_utils
+    SRCS hook_utils.cc
+    DEPS phi tensor_utils autograd_meta grad_node_info utils)
+endif()
diff --git a/paddle/fluid/eager/tests/data_structure_tests/CMakeLists.txt b/paddle/fluid/eager/tests/data_structure_tests/CMakeLists.txt
old mode 100644
new mode 100755
index 3c695a5168cd3df1f05551adad8124e6fbdb661f..0d5876477c664684917b90ea8268022f96540430
--- a/paddle/fluid/eager/tests/data_structure_tests/CMakeLists.txt
+++ b/paddle/fluid/eager/tests/data_structure_tests/CMakeLists.txt
@@ -2,14 +2,14 @@ cc_test_old(test_egr_ds_eager_tensor SRCS eager_tensor_test.cc DEPS
             ${eager_deps})
 cc_test_old(test_egr_ds_auotgrad_meta SRCS autograd_meta_test.cc DEPS
             ${eager_deps})
-cc_test_old(test_egr_ds_grad_node_info SRCS grad_node_info_test.cc DEPS
-            ${eager_deps})
-cc_test_old(test_egr_ds_accumulation_node SRCS accumulation_node_test.cc DEPS
-            ${eager_deps})
-cc_test_old(test_egr_ds_tensor_wrapper SRCS tensor_wrapper_test.cc DEPS
-            ${eager_deps})
 
 if(NOT ((NOT WITH_PYTHON) AND ON_INFER))
   cc_test_old(test_egr_ds_grad_tensor_holder SRCS grad_tensor_holder_test.cc
               DEPS ${eager_deps} ${generated_deps})
+  cc_test_old(test_egr_ds_grad_node_info SRCS grad_node_info_test.cc DEPS
+              ${eager_deps} ${generated_deps})
+  cc_test_old(test_egr_ds_accumulation_node SRCS accumulation_node_test.cc DEPS
+              ${eager_deps} ${generated_deps})
+  cc_test_old(test_egr_ds_tensor_wrapper SRCS tensor_wrapper_test.cc DEPS
+              ${eager_deps} ${generated_deps})
 endif()
diff --git a/paddle/fluid/eager/tests/task_tests/CMakeLists.txt b/paddle/fluid/eager/tests/task_tests/CMakeLists.txt
old mode 100644
new mode 100755
index 2f57489999ff8118b7273737f30928fac07f5cd2..de963c86203c8f25806c6ce94c9e0279cd8c2e3d
--- a/paddle/fluid/eager/tests/task_tests/CMakeLists.txt
+++ b/paddle/fluid/eager/tests/task_tests/CMakeLists.txt
@@ -1,15 +1,3 @@
-cc_test(
-  test_egr_task_tensor_utils
-  SRCS tensor_utils_test.cc
-  DEPS ${eager_deps})
-cc_test(
-  test_egr_task_eager_utils
-  SRCS eager_utils_test.cc
-  DEPS ${eager_deps})
-cc_test(
-  test_egr_task_forward_autograd
-  SRCS forward_autograd_test.cc
-  DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
 cc_test(
   test_egr_task_nan_inf_utils
   SRCS nan_inf_utils_test.cc
@@ -44,4 +32,16 @@ if(NOT ((NOT WITH_PYTHON) AND ON_INFER))
     test_egr_task_autocodegen
     SRCS generated_test.cc
     DEPS ${eager_deps} ${fluid_deps} ${generated_deps})
+  cc_test(
+    test_egr_task_tensor_utils
+    SRCS tensor_utils_test.cc
+    DEPS ${eager_deps} ${generated_deps})
+  cc_test(
+    test_egr_task_eager_utils
+    SRCS eager_utils_test.cc
+    DEPS ${eager_deps} ${generated_deps})
+  cc_test(
+    test_egr_task_forward_autograd
+    SRCS forward_autograd_test.cc
+    DEPS ${eager_deps} ${fluid_deps} ${generated_deps} eager_scale scale_node)
 endif()
diff --git a/python/paddle/fluid/tests/custom_runtime/test_custom_cpu_plugin.py b/python/paddle/fluid/tests/custom_runtime/test_custom_cpu_plugin.py
old mode 100644
new mode 100755
index f6a05c07b4be7ab631d445ebe6fef8395feca1eb..0d087cab118e08302e9586c6c847d5ec9c4bbf04
--- a/python/paddle/fluid/tests/custom_runtime/test_custom_cpu_plugin.py
+++ b/python/paddle/fluid/tests/custom_runtime/test_custom_cpu_plugin.py
@@ -59,6 +59,7 @@ class TestCustomCPUPlugin(unittest.TestCase):
             self._test_eager_copy_to()
             self._test_fallback_kernel()
             self._test_scalar()
+            self._test_custom_device_gradient_accumulation()
         self._test_custom_device_dataloader()
         self._test_custom_device_mnist()
 
@@ -208,6 +209,60 @@ class TestCustomCPUPlugin(unittest.TestCase):
         k_t = paddle.to_tensor([3], dtype="int32")
         value_1, indices_1 = paddle.topk(data_1, k=k_t)
 
+    def _test_custom_device_gradient_accumulation(self):
+        import paddle
+
+        class MNIST(paddle.nn.Layer):
+            def __init__(self):
+                super().__init__()
+                self.shape = 1 * 28 * 28
+                self.size = 10
+                self.output_weight = self.create_parameter(
+                    [self.shape, self.size]
+                )
+                self.accuracy = paddle.metric.Accuracy()
+
+            def forward(self, inputs, label=None):
+                x = paddle.reshape(inputs, shape=[-1, self.shape])
+                x = paddle.matmul(x, self.output_weight)
+                x = paddle.nn.functional.softmax(x)
+                if label is not None:
+                    self.accuracy.reset()
+                    correct = self.accuracy.compute(x, label)
+                    self.accuracy.update(correct)
+                    acc = self.accuracy.accumulate()
+                    return x, acc
+                else:
+                    return x
+
+        paddle.set_device('custom_cpu')
+        dataset = paddle.vision.datasets.MNIST(
+            mode='train',
+            transform=paddle.vision.transforms.Compose(
+                [paddle.vision.transforms.ToTensor()]
+            ),
+        )
+        loader = paddle.io.DataLoader(
+            dataset, batch_size=64, num_workers=1, shuffle=True
+        )
+
+        mnist = MNIST()
+        sgd = paddle.optimizer.SGD(
+            learning_rate=0.01, parameters=mnist.parameters()
+        )
+
+        data = next(loader())
+        img = data[0]
+        label = data[1]
+        label_int32 = paddle.cast(label, 'int32')
+
+        pred, acc = mnist(img, label_int32)
+        avg_loss = paddle.nn.functional.cross_entropy(pred, label_int32)
+        avg_loss.backward(retain_graph=True)
+        avg_loss = paddle.nn.functional.cross_entropy(pred, label_int32)
+        avg_loss.backward()
+        sgd.step()
+
 
 if __name__ == '__main__':
     if os.name == 'nt' or sys.platform.startswith('darwin'):