diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/fused_softmax_mask_grad_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/fused_softmax_mask_grad_kernel_register.cu
index 0bb64afd52e..800a8c3f23d 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/fused_softmax_mask_grad_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/fused_softmax_mask_grad_kernel_register.cu
@@ -14,7 +14,7 @@
 
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/full_kernel.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_softmax_mask_grad_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/fused_softmax_mask_grad_kernel.h"
 #include "paddle/phi/kernels/fusion/gpu/fused_softmax_mask_utils.h"
 
 PD_CUSTOM_KERNEL_REGISTER(fused_softmax_mask_grad,
diff --git a/backends/metax_gpu/kernels/fusion/fused_softmax_mask_grad_kernel_register.cu b/backends/metax_gpu/kernels/fusion/fused_softmax_mask_grad_kernel_register.cu
index 55883496a72..2c596c6a4eb 100644
--- a/backends/metax_gpu/kernels/fusion/fused_softmax_mask_grad_kernel_register.cu
+++ b/backends/metax_gpu/kernels/fusion/fused_softmax_mask_grad_kernel_register.cu
@@ -14,7 +14,7 @@
 
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/full_kernel.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_softmax_mask_grad_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/fused_softmax_mask_grad_kernel.h"
 #include "paddle/phi/kernels/fusion/gpu/fused_softmax_mask_utils.h"
 
 PD_CUSTOM_KERNEL_REGISTER(fused_softmax_mask_grad,