[NPU] fix npu op elementwise_mul_grad (#31592)

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>
 
@@ -58,18 +59,22 @@ 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();
 
+    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);
     }
+  }
 };
 
 }  // namespace operators
@@ -88,3 +93,4 @@ REGISTER_OP_NPU_KERNEL(
     ops::ElementwiseMulGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::ElementwiseMulGradNPUKernel<paddle::platform::NPUDeviceContext,
                                      paddle::platform::float16>);
+#endif

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)