fix include for mp_allreduce_sum_kernel (PaddlePaddle#2218)

Author: fsylmxx

backends/iluvatar_gpu/kernels/cuda_kernels/mp_allreduce_sum_kernel_register.cu

@@ -14,7 +14,7 @@
 
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/all_reduce_kernel.h"
-#include "paddle/phi/kernels/gpu/mp_allreduce_sum_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/mp_allreduce_sum_kernel.h"  //NOLINT
 PD_CUSTOM_KERNEL_REGISTER(mp_allreduce_sum,
                           iluvatar_gpu,
                           ALL_LAYOUT,

backends/metax_gpu/CMakeLists.txt

@@ -252,6 +252,7 @@ file(
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/mean_all_grad_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/multiplex_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/multiplex_grad_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/mp_allreduce_sum_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/pow2_decay_with_linear_warmup_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/sigmoid_cross_entropy_with_logits_grad_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/top_k_kernel.cu

backends/metax_gpu/kernels/cuda_kernels/mp_allreduce_sum_kernel_register.cu

@@ -14,7 +14,7 @@
 
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/all_reduce_kernel.h"
-#include "paddle/phi/kernels/gpu/mp_allreduce_sum_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/mp_allreduce_sum_kernel.h"  //NOLINT
 PD_CUSTOM_KERNEL_REGISTER(mp_allreduce_sum,
                           metax_gpu,
                           ALL_LAYOUT,