diff --git a/paddle/fluid/operators/memcpy_op.cc b/paddle/fluid/operators/memcpy_op.cc
index ef430f8bfaf4368d34a0f447157864f4fb4c793f..9fb06c5968cbdaa2df471c831eac3a9e57884df5 100644
--- a/paddle/fluid/operators/memcpy_op.cc
+++ b/paddle/fluid/operators/memcpy_op.cc
@@ -113,7 +113,9 @@ class MemcpyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
                  "1: dst is on CUDAPlace. "
                  "2: dst is on CUDAPinnedPlace. "
                  "3: dst is on XPUPlace. "
-                 "4: dst is on NPUPlace. ");
+                 "4: dst is on NPUPlace. "
+                 "5: dst is on NPUPinnerPlace. "
+                 "6: dst is on CustomDevicePlace");
     AddComment(R"DOC(
     Memcpy Operator.
     By now, it ONLY supports the memcopy between CUDAPinnedPlace <-> CUDAPlace or 
diff --git a/paddle/fluid/operators/memcpy_op.h b/paddle/fluid/operators/memcpy_op.h
index a35fefa53ba3d4f206f26b5ee84afbb9573015f7..092d0cf368e6a6e019164adcf71536459f39f30f 100644
--- a/paddle/fluid/operators/memcpy_op.h
+++ b/paddle/fluid/operators/memcpy_op.h
@@ -41,6 +41,7 @@ class MemcpyFunctor {
     XPU = 3,
     NPU = 4,
     NPU_PINNED = 5,
+    CUSTOM_DEVICE = 6,
   };
 
  public:
@@ -67,6 +68,11 @@ class MemcpyFunctor {
     } else if (dst_place_type_ == DeviceType::NPU_PINNED) { /* npu->npu_pin */
       framework::TensorCopy(
           lod_tensor, platform::NPUPinnedPlace(), dev_ctx_, &out_tensor);
+#endif
+#ifdef PADDLE_WTIH_CUSTOM_DEVICE
+    } else if (dst_place_type_ == DeviceType::CUSTOM_DEVICE) {
+      framework::TensorCopy(
+          lod_tensor, dev_ctx_.GetPlace(), dev_ctx_, &out_tensor);
 #endif
     } else {
       PADDLE_THROW(platform::errors::Unimplemented(