diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu
index 3e6d41ab2b..c348464a92 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/label_smooth_grad_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/label_smooth_grad_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(label_smooth_grad,
                           iluvatar_gpu,
diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu
index 1d4b60374f..42dbe02d28 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/label_smooth_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/label_smooth_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(label_smooth,
                           iluvatar_gpu,
diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/lamb_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/lamb_kernel_register.cu
index 43f6245c9f..46a690a234 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/lamb_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/lamb_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/lamb_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/lamb_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(lamb,
                           iluvatar_gpu,
diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu
index 8e524c04fe..ea9acdcb41 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/lgamma_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/lgamma_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(lgamma,
                           iluvatar_gpu,
diff --git a/backends/metax_gpu/CMakeLists.txt b/backends/metax_gpu/CMakeLists.txt
index 1397340c2a..bc2f8eaf72 100755
--- a/backends/metax_gpu/CMakeLists.txt
+++ b/backends/metax_gpu/CMakeLists.txt
@@ -124,7 +124,11 @@ file(
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/funcs/math/*.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/funcs/eigen/*.cu
   # kernels/gpu
-  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/ctc_align_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/lgamma_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/lamb_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/label_smooth_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/label_smooth_grad_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/l1_norm_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/yolo_box_post_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/spectral_norm_grad_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/spectral_norm_kernel.cu
diff --git a/backends/metax_gpu/kernels/cuda_kernels/l1_norm_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/l1_norm_kernel_register.cu
index ae3c0ad97a..5905f6d44d 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/l1_norm_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/l1_norm_kernel_register.cu
@@ -14,6 +14,5 @@
 
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/gpu/l1_norm_kernel.cu"  // NOLINT
-
 PD_CUSTOM_KERNEL_REGISTER(
     l1_norm, metax_gpu, ALL_LAYOUT, phi::L1NormKernel, float) {}
diff --git a/backends/metax_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu
index 906efb6451..6aaf6ad38b 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/label_smooth_grad_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/label_smooth_grad_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/label_smooth_grad_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(label_smooth_grad,
                           metax_gpu,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu
index c2e73aab64..a7a60262bf 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/label_smooth_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/label_smooth_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/label_smooth_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(label_smooth,
                           metax_gpu,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/lamb_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/lamb_kernel_register.cu
index a8bd18a788..1d99df771d 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/lamb_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/lamb_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/lamb_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/lamb_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(lamb,
                           metax_gpu,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu
index 69c17c6df2..03346acb9c 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/lgamma_kernel_register.cu
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/lgamma_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/lgamma_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(lgamma,
                           metax_gpu,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/log_softmax_grad_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/log_softmax_grad_kernel_register.cu
index 99ea4e13dc..b07589a567 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/log_softmax_grad_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/log_softmax_grad_kernel_register.cu
@@ -14,7 +14,6 @@
 
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/log_softmax_grad_kernel.h"
-// #include "paddle/phi/kernels/gpu/log_softmax_grad_kernel.cu"
 
 PD_CUSTOM_KERNEL_REGISTER(log_softmax_grad,
                           metax_gpu,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/logsumexp_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/logsumexp_kernel_register.cu
index 41e9efa56f..92145aa9db 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/logsumexp_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/logsumexp_kernel_register.cu
@@ -14,8 +14,8 @@
 // limitations under the License.
 // clang-format off
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/logsumexp_kernel.cu" // NOLINT
 #include "paddle/phi/kernels/logsumexp_kernel.h"
+
 // clang-format on
 PD_CUSTOM_KERNEL_REGISTER(logsumexp,
                           metax_gpu,