fix nccl version (#53942)

paddle/fluid/operators/collective/alltoall_op.cu.cc

@@ -98,7 +98,7 @@ PD_REGISTER_STRUCT_KERNEL(alltoall,
                           ops::AllToAllOpCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/c_allgather_op.cu.cc

@@ -95,7 +95,7 @@ PD_REGISTER_STRUCT_KERNEL(c_allgather,
                           ops::CAllGatherOpCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc

@@ -28,7 +28,7 @@ PD_REGISTER_STRUCT_KERNEL(c_allreduce_max,
                           ALL_LAYOUT,
                           ops::CAllReduceMaxCUDAKernel,
                           float,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           double,

paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc

@@ -28,7 +28,7 @@ PD_REGISTER_STRUCT_KERNEL(c_allreduce_sum,
                           ALL_LAYOUT,
                           ops::CAllReduceSumCUDAKernel,
                           float,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           double,

paddle/fluid/operators/collective/c_broadcast_op.cu.cc

@@ -100,7 +100,7 @@ PD_REGISTER_STRUCT_KERNEL(c_broadcast,
                           int64_t,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {

paddle/fluid/operators/collective/c_concat_op.cu.cc

@@ -137,7 +137,7 @@ PD_REGISTER_STRUCT_KERNEL(c_concat,
                           double,
                           int,
                           int64_t,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {

paddle/fluid/operators/collective/c_embedding_op.cu

@@ -239,7 +239,7 @@ PD_REGISTER_STRUCT_KERNEL(c_embedding,
                           ops::CEmbeddingCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {
@@ -251,7 +251,7 @@ PD_REGISTER_STRUCT_KERNEL(c_embedding_grad,
                           ops::CEmbeddingGradCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {

paddle/fluid/operators/collective/c_identity_op.cu.cc

@@ -25,7 +25,7 @@ PD_REGISTER_STRUCT_KERNEL(c_identity,
                           double,
                           int,
                           int64_t,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {

paddle/fluid/operators/collective/c_reducescatter_op.cu.cc

@@ -87,7 +87,7 @@ PD_REGISTER_STRUCT_KERNEL(c_reducescatter,
                           ops::CReduceScatterOpCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/c_split_op.cu

@@ -123,7 +123,7 @@ PD_REGISTER_STRUCT_KERNEL(c_split,
                           double,
                           int,
                           int64_t,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {

paddle/fluid/operators/collective/mp_allreduce_sum_op.cu.cc

@@ -31,7 +31,7 @@ PD_REGISTER_STRUCT_KERNEL(mp_allreduce_sum,
                           double,
                           int,
                           int64_t,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           plat::float16) {

paddle/fluid/operators/collective/partial_allgather_op.cu.cc

@@ -108,7 +108,7 @@ PD_REGISTER_STRUCT_KERNEL(partial_allgather,
                           ops::PartialAllGatherOpCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/partial_recv_op.cu.cc

@@ -124,7 +124,7 @@ PD_REGISTER_STRUCT_KERNEL(partial_recv,
                           ops::PartialRecvOpCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/partial_send_op.cu.cc

@@ -123,7 +123,7 @@ PD_REGISTER_STRUCT_KERNEL(partial_send,
                           ops::PartialSendCUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/recv_v2_op.cu.cc

@@ -238,7 +238,7 @@ PD_REGISTER_STRUCT_KERNEL(recv_v2,
                           ops::RecvOpV2CUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/operators/collective/send_v2_op.cu.cc

@@ -223,7 +223,7 @@ PD_REGISTER_STRUCT_KERNEL(send_v2,
                           ops::SendOpV2CUDAKernel,
                           float,
                           double,
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
                           plat::bfloat16,
 #endif
                           int,

paddle/fluid/platform/device/gpu/nccl_helper.h

@@ -59,7 +59,7 @@ inline ncclDataType_t ToNCCLDataType(framework::proto::VarType::Type type) {
     return ncclUint8;
   } else if (type == framework::proto::VarType::BOOL) {
     return ncclUint8;
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
   } else if (type == framework::proto::VarType::BF16) {
     return ncclBfloat16;
 #endif
@@ -86,7 +86,7 @@ inline ncclDataType_t ToNCCLDataType(phi::DataType type) {
     return ncclInt8;
   } else if (type == phi::DataType::BOOL) {
     return ncclUint8;
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
   } else if (type == phi::DataType::BFLOAT16) {
     return ncclBfloat16;
 #endif

paddle/phi/core/utils/data_type.h

@@ -229,7 +229,7 @@ inline ncclDataType_t ToNCCLDataType(DataType type) {
     return ncclInt8;
   } else if (type == DataType::BOOL) {
     return ncclUint8;
-#if NCCL_VERSION_CODE >= 21000
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
   } else if (type == DataType::BFLOAT16) {
     return ncclBfloat16;
 #endif