From 89da2f19c8831751bfe1288638060ecc1958f9e9 Mon Sep 17 00:00:00 2001
From: LiYuRio <63526175+LiYuRio@users.noreply.github.com>
Date: Tue, 23 May 2023 17:08:55 +0800
Subject: [PATCH] fix nccl version (#53942)

---
 paddle/fluid/operators/collective/alltoall_op.cu.cc          | 2 +-
 paddle/fluid/operators/collective/c_allgather_op.cu.cc       | 2 +-
 paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc   | 2 +-
 paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc   | 2 +-
 paddle/fluid/operators/collective/c_broadcast_op.cu.cc       | 2 +-
 paddle/fluid/operators/collective/c_concat_op.cu.cc          | 2 +-
 paddle/fluid/operators/collective/c_embedding_op.cu          | 4 ++--
 paddle/fluid/operators/collective/c_identity_op.cu.cc        | 2 +-
 paddle/fluid/operators/collective/c_reducescatter_op.cu.cc   | 2 +-
 paddle/fluid/operators/collective/c_split_op.cu              | 2 +-
 paddle/fluid/operators/collective/mp_allreduce_sum_op.cu.cc  | 2 +-
 paddle/fluid/operators/collective/partial_allgather_op.cu.cc | 2 +-
 paddle/fluid/operators/collective/partial_recv_op.cu.cc      | 2 +-
 paddle/fluid/operators/collective/partial_send_op.cu.cc      | 2 +-
 paddle/fluid/operators/collective/recv_v2_op.cu.cc           | 2 +-
 paddle/fluid/operators/collective/send_v2_op.cu.cc           | 2 +-
 paddle/fluid/platform/device/gpu/nccl_helper.h               | 4 ++--
 paddle/phi/core/utils/data_type.h                            | 2 +-
 18 files changed, 20 insertions(+), 20 deletions(-)

diff --git a/paddle/fluid/operators/collective/alltoall_op.cu.cc b/paddle/fluid/operators/collective/alltoall_op.cu.cc
index aacd76af4af..0ae338c745a 100644
--- a/paddle/fluid/operators/collective/alltoall_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/c_allgather_op.cu.cc b/paddle/fluid/operators/collective/c_allgather_op.cu.cc
index 70b7d70dc93..c3eff905851 100644
--- a/paddle/fluid/operators/collective/c_allgather_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc b/paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc
index 9be9674bb08..277988b5691 100644
--- a/paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc b/paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc
index 8e45b7e43b2..76d809cd234 100644
--- a/paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/c_broadcast_op.cu.cc b/paddle/fluid/operators/collective/c_broadcast_op.cu.cc
index d0d3ebb1394..e37657a1747 100644
--- a/paddle/fluid/operators/collective/c_broadcast_op.cu.cc
+++ b/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) {
diff --git a/paddle/fluid/operators/collective/c_concat_op.cu.cc b/paddle/fluid/operators/collective/c_concat_op.cu.cc
index dc47c184c22..1760b6ea390 100644
--- a/paddle/fluid/operators/collective/c_concat_op.cu.cc
+++ b/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) {
diff --git a/paddle/fluid/operators/collective/c_embedding_op.cu b/paddle/fluid/operators/collective/c_embedding_op.cu
index 4861b5d26ab..758734ada66 100644
--- a/paddle/fluid/operators/collective/c_embedding_op.cu
+++ b/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) {
diff --git a/paddle/fluid/operators/collective/c_identity_op.cu.cc b/paddle/fluid/operators/collective/c_identity_op.cu.cc
index 3d5f16c218c..9571168db15 100644
--- a/paddle/fluid/operators/collective/c_identity_op.cu.cc
+++ b/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) {
diff --git a/paddle/fluid/operators/collective/c_reducescatter_op.cu.cc b/paddle/fluid/operators/collective/c_reducescatter_op.cu.cc
index 35053b1511f..edae8feb612 100644
--- a/paddle/fluid/operators/collective/c_reducescatter_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/c_split_op.cu b/paddle/fluid/operators/collective/c_split_op.cu
index b01ed790e85..0b3e2aaf781 100644
--- a/paddle/fluid/operators/collective/c_split_op.cu
+++ b/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) {
diff --git a/paddle/fluid/operators/collective/mp_allreduce_sum_op.cu.cc b/paddle/fluid/operators/collective/mp_allreduce_sum_op.cu.cc
index b6af2dbd1c8..b4773a8eb54 100644
--- a/paddle/fluid/operators/collective/mp_allreduce_sum_op.cu.cc
+++ b/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) {
diff --git a/paddle/fluid/operators/collective/partial_allgather_op.cu.cc b/paddle/fluid/operators/collective/partial_allgather_op.cu.cc
index 2374f4a4aed..d22fd70bd0f 100644
--- a/paddle/fluid/operators/collective/partial_allgather_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/partial_recv_op.cu.cc b/paddle/fluid/operators/collective/partial_recv_op.cu.cc
index b0df94194e4..0c33ca7c25c 100644
--- a/paddle/fluid/operators/collective/partial_recv_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/partial_send_op.cu.cc b/paddle/fluid/operators/collective/partial_send_op.cu.cc
index dc24ea01fc9..4f9fc41bc4e 100644
--- a/paddle/fluid/operators/collective/partial_send_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/recv_v2_op.cu.cc b/paddle/fluid/operators/collective/recv_v2_op.cu.cc
index bfa12f91194..28058aa4868 100644
--- a/paddle/fluid/operators/collective/recv_v2_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/operators/collective/send_v2_op.cu.cc b/paddle/fluid/operators/collective/send_v2_op.cu.cc
index adea7db0b80..a80dc1f91e4 100644
--- a/paddle/fluid/operators/collective/send_v2_op.cu.cc
+++ b/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,
diff --git a/paddle/fluid/platform/device/gpu/nccl_helper.h b/paddle/fluid/platform/device/gpu/nccl_helper.h
index 8dd0639ce72..6afcd2eb7cd 100644
--- a/paddle/fluid/platform/device/gpu/nccl_helper.h
+++ b/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
diff --git a/paddle/phi/core/utils/data_type.h b/paddle/phi/core/utils/data_type.h
index 16b73e0f2ba..018672e45b5 100644
--- a/paddle/phi/core/utils/data_type.h
+++ b/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
-- 
GitLab