diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/fused_stack_transpose_quant_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/fused_stack_transpose_quant_kernel_register.cu
index f05323d6e7..749eb2d482 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/fused_stack_transpose_quant_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/fused_stack_transpose_quant_kernel_register.cu
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_stack_transpose_quant_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/fusion/gpu/fused_stack_transpose_quant_kernel.h"
 #include "paddle/phi/kernels/fusion/gpu/quant_utils.h"
 #include "paddle/phi/kernels/primitive/datamover_primitives.h"
 
diff --git a/backends/metax_gpu/kernels/fusion/fused_stack_transpose_quant_kernel_register.cu b/backends/metax_gpu/kernels/fusion/fused_stack_transpose_quant_kernel_register.cu
index 02c24f6d90..683918e172 100644
--- a/backends/metax_gpu/kernels/fusion/fused_stack_transpose_quant_kernel_register.cu
+++ b/backends/metax_gpu/kernels/fusion/fused_stack_transpose_quant_kernel_register.cu
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_stack_transpose_quant_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/fusion/gpu/fused_stack_transpose_quant_kernel.h"
 #include "paddle/phi/kernels/fusion/gpu/quant_utils.h"
 #include "paddle/phi/kernels/primitive/datamover_primitives.h"