2.4.1-dtk23.04

paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu

@@ -884,7 +884,7 @@ static bool CreatePreMulScaleOpIfSupported(ncclDataType_t dtype,
                                            ncclComm_t comm,
                                            const void *scale,
                                            ncclRedOp_t *op) {
-#if NCCL_VERSION_CODE >= 21100
+#if (NCCL_VERSION_CODE >= 21100) && defined(PADDLE_WITH_NCCL)
   int ver;
   PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclGetVersion(&ver));
   if (ver >= 21100) {
@@ -965,7 +965,7 @@ static void NCCLSumWithScaleBase(const T *sendbuff,
         sendbuff, recvbuff, recvcount, dtype, op, comm, stream));
   }
 
-#if NCCL_VERSION_CODE >= 21100
+#if (NCCL_VERSION_CODE >= 21100) && defined(PADDLE_WITH_NCCL)
   if (should_destroy_op) {
     VLOG(10) << "ncclRedOpDestroy starts";
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclRedOpDestroy(op, comm));

paddle/phi/kernels/gpu/graph_send_recv_kernel.cu

@@ -16,6 +16,7 @@
 
 #include <thrust/device_vector.h>
 #include <thrust/fill.h>
+#include <thrust/execution_policy.h>
 
 #include <algorithm>
 #include <vector>

paddle/phi/kernels/gpu/graph_send_ue_recv_kernel.cu

@@ -19,6 +19,7 @@
 
 #include <thrust/device_vector.h>
 #include <thrust/fill.h>
+#include <thrust/execution_policy.h>
 #include <algorithm>
 #include <vector>
 

paddle/phi/kernels/gpu/unique_kernel.cu

@@ -21,6 +21,7 @@
 #include <thrust/scatter.h>
 #include <thrust/sequence.h>
 #include <thrust/unique.h>
+#include <thrust/sort.h>
 
 #include <iostream>
 #include <vector>

paddle/phi/kernels/sparse/gpu/coalesce_kernel.cu

@@ -22,6 +22,8 @@ limitations under the License. */
 #include "paddle/phi/kernels/funcs/sparse/flatten_indices.cu.h"
 #include "paddle/phi/kernels/funcs/sparse/scatter.cu.h"
 #include "paddle/phi/kernels/funcs/sparse/utils.cu.h"
+#include <thrust/sort.h>
+#include <thrust/unique.h>
 
 namespace phi {
 namespace sparse {