diff --git a/backends/iluvatar_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu b/backends/iluvatar_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu
index 49e4eb84ad8..13516a5e0c9 100644
--- a/backends/iluvatar_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu
+++ b/backends/iluvatar_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu
@@ -16,7 +16,6 @@
 #include "paddle/phi/kernels/distributed_fused_lamb_init_kernel.h"
 #include "paddle/phi/kernels/funcs/algorithm.h"
 #include "paddle/phi/kernels/fusion/gpu/cast_with_ptr.h"
-#include "paddle/phi/kernels/fusion/gpu/distributed_fused_lamb_init_kernel.cu"  //NOLINT
 
 PD_CUSTOM_KERNEL_REGISTER(distributed_fused_lamb_init,
                           iluvatar_gpu,
diff --git a/backends/iluvatar_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cc b/backends/iluvatar_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cc
index b7920534ed1..24c7d3e1711 100644
--- a/backends/iluvatar_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cc
+++ b/backends/iluvatar_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cc
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_bias_act_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/fused_bias_act_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(fused_bias_act,
                           iluvatar_gpu,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu
index 9d4d57e9e5d..97ad0c25596 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/distributed_fused_lamb_init_kernel_register.cu
@@ -16,7 +16,6 @@
 #include "paddle/phi/kernels/distributed_fused_lamb_init_kernel.h"
 #include "paddle/phi/kernels/funcs/algorithm.h"
 #include "paddle/phi/kernels/fusion/gpu/cast_with_ptr.h"
-#include "paddle/phi/kernels/fusion/gpu/distributed_fused_lamb_init_kernel.cu"  //NOLINT
 
 PD_CUSTOM_KERNEL_REGISTER(distributed_fused_lamb_init,
                           metax_gpu,
diff --git a/backends/metax_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cu b/backends/metax_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cu
index 9345cb0a291..c49ceed2196 100644
--- a/backends/metax_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cu
+++ b/backends/metax_gpu/kernels/ernie_core/fused_bias_act_kernel_register.cu
@@ -14,7 +14,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_bias_act_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/fused_bias_act_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(fused_bias_act,
                           metax_gpu,