diff --git a/paddle/phi/core/tensor_utils.cc b/paddle/phi/core/tensor_utils.cc
index 45f6c00affe052f940b6d2773abc2594a56368d0..8b3d4a1427340339e8250189faf723cd74af19f7 100644
--- a/paddle/phi/core/tensor_utils.cc
+++ b/paddle/phi/core/tensor_utils.cc
@@ -200,10 +200,9 @@ void Copy(const Context& dev_ctx,
     paddle::memory::Copy(
         dst_cuda_pinned_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
 #endif
-  }
 #ifdef PADDLE_WITH_XPU
-  else if (paddle::platform::is_xpu_place(src_place) &&  // NOLINT
-           paddle::platform::is_cpu_place(dst_place)) {
+  } else if (paddle::platform::is_xpu_place(src_place) &&  // NOLINT
+             paddle::platform::is_cpu_place(dst_place)) {
     paddle::memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
   } else if (paddle::platform::is_cpu_place(src_place) &&
              paddle::platform::is_xpu_place(dst_place)) {
@@ -216,11 +215,40 @@ void Copy(const Context& dev_ctx,
       return;
     }
     paddle::memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
+#endif
+#ifdef PADDLE_WITH_CUSTOM_DEVICE
+  } else if (paddle::platform::is_custom_place(src_place) &&  // NOLINT
+             paddle::platform::is_cpu_place(dst_place)) {
+    auto stream =
+        blocking
+            ? nullptr
+            : reinterpret_cast<const paddle::platform::CustomDeviceContext&>(
+                  dev_ctx)
+                  .stream();
+    paddle::memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, stream);
+  } else if (paddle::platform::is_cpu_place(src_place) &&  // NOLINT
+             paddle::platform::is_custom_place(dst_place)) {
+    auto stream =
+        blocking
+            ? nullptr
+            : reinterpret_cast<const paddle::platform::CustomDeviceContext&>(
+                  dev_ctx)
+                  .stream();
+    paddle::memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, stream);
+  } else if (paddle::platform::is_custom_place(src_place) &&  // NOLINT
+             paddle::platform::is_custom_place(dst_place)) {
+    auto stream =
+        blocking
+            ? nullptr
+            : reinterpret_cast<const paddle::platform::CustomDeviceContext&>(
+                  dev_ctx)
+                  .stream();
+    paddle::memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, stream);
+#endif
   } else {
     PADDLE_THROW(phi::errors::Unimplemented(
         "Copy from %s to %s is not supported.", src_place, dst_place));
   }
-#endif
 }
 
 template <typename Context>
@@ -363,4 +391,11 @@ template void Copy(const XPUContext& dev_ctx,
                    DenseTensor* dst);
 #endif
 
+#ifdef PADDLE_WITH_CUSTOM_DEVICE
+template void Copy(const CustomContext& dev_ctx,
+                   const DenseTensor& src,
+                   Place dst_place,
+                   bool blocking,
+                   DenseTensor* dst);
+#endif
 }  // namespace phi
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
index 00d7255a83f21c43ca85921d6b48c61f1a2ed6ac..02dabf899c6ef1e23a9c31296132503d849a5148 100644
--- a/python/paddle/fluid/tests/custom_runtime/test_custom_cpu_plugin.py
+++ b/python/paddle/fluid/tests/custom_runtime/test_custom_cpu_plugin.py
@@ -39,6 +39,7 @@ class TestCustomCPUPlugin(unittest.TestCase):
             self._test_custom_device_dataloader()
             self._test_custom_device_mnist()
             self._test_eager_backward_api()
+            self._test_eager_copy_to()
         self._test_custom_device_dataloader()
         self._test_custom_device_mnist()
 
@@ -133,6 +134,32 @@ class TestCustomCPUPlugin(unittest.TestCase):
 
         self.assertTrue(x_tensor.grad.place.is_custom_place())
 
+    def _test_eager_copy_to(self):
+        import paddle
+        x = np.random.random([2, 2]).astype("float32")
+        # cpu -> custom
+        cpu_tensor = paddle.to_tensor(x,
+                                      dtype='float32',
+                                      place=paddle.CPUPlace())
+        custom_cpu_tensor = cpu_tensor._copy_to(
+            paddle.CustomPlace('custom_cpu', 0), True)
+        self.assertTrue(np.array_equal(custom_cpu_tensor, x))
+        self.assertTrue(custom_cpu_tensor.place.is_custom_place())
+        # custom -> custom
+        another_custom_cpu_tensor = custom_cpu_tensor._copy_to(
+            paddle.CustomPlace('custom_cpu', 0), True)
+        self.assertTrue(np.array_equal(another_custom_cpu_tensor, x))
+        self.assertTrue(another_custom_cpu_tensor.place.is_custom_place())
+        # custom -> cpu
+        another_cpu_tensor = custom_cpu_tensor._copy_to(paddle.CPUPlace(), True)
+        self.assertTrue(np.array_equal(another_cpu_tensor, x))
+        self.assertTrue(another_cpu_tensor.place.is_cpu_place())
+        # custom -> custom self
+        another_custom_cpu_tensor = another_custom_cpu_tensor._copy_to(
+            paddle.CustomPlace('custom_cpu', 0), True)
+        self.assertTrue(np.array_equal(another_custom_cpu_tensor, x))
+        self.assertTrue(another_custom_cpu_tensor.place.is_custom_place())
+
     def tearDown(self):
         del os.environ['CUSTOM_DEVICE_ROOT']