diff --git a/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc b/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc
index cecb62d3580f9e84afa965424c6e8af20cda5dc5..08df6d4e27af0a79123f26ad2064ee0203cc1b28 100644
--- a/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc
+++ b/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc
@@ -12,6 +12,7 @@ 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. */
 
+#ifdef PADDLE_WITH_ASCEND_CL
 #include <memory>
 #include <string>
 
@@ -58,17 +59,21 @@ class ElementwiseMulGradNPUKernel : public framework::OpKernel<T> {
 
     auto place = ctx.GetPlace();
 
-    dx->mutable_data<T>(place);
-    dy->mutable_data<T>(place);
-
     auto stream =
         ctx.template device_context<paddle::platform::NPUDeviceContext>()
             .stream();
 
-    auto dx_runner = NpuOpRunner("Mul", {*dout, *y}, {*dx}, {});
-    dx_runner.Run(stream);
-    auto dy_runner = NpuOpRunner("Mul", {*x, *dout}, {*dy}, {});
-    dy_runner.Run(stream);
+    if (dx) {
+      dx->mutable_data<T>(place);
+      auto dx_runner = NpuOpRunner("Mul", {*dout, *y}, {*dx}, {});
+      dx_runner.Run(stream);
+    }
+
+    if (dy) {
+      dy->mutable_data<T>(place);
+      auto dy_runner = NpuOpRunner("Mul", {*x, *dout}, {*dy}, {});
+      dy_runner.Run(stream);
+    }
   }
 };
 
@@ -88,3 +93,4 @@ REGISTER_OP_NPU_KERNEL(
     ops::ElementwiseMulGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::ElementwiseMulGradNPUKernel<paddle::platform::NPUDeviceContext,
                                      paddle::platform::float16>);
+#endif
diff --git a/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py
index d0baa5877f224bbc8c0e505457b004918aaf6448..9bfb7e033e7ea454223c683877bb30f02506be75 100644
--- a/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py
+++ b/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py
@@ -122,6 +122,7 @@ class TestElementwiseMulNet(unittest.TestCase):
 
             e = paddle.multiply(a, b)
             f = paddle.multiply(c, d)
+            f.stop_gradient = True
             g = paddle.multiply(e, f)
 
             fc_1 = fluid.layers.fc(input=g, size=128)