From b8fa212c95ebfe463ae5b568f1fc7080bf060600 Mon Sep 17 00:00:00 2001
From: Binyang Li <binyli@microsoft.com>
Date: Wed, 21 Feb 2024 10:07:37 +0000
Subject: [PATCH] make it work for rocm6.0

---
 apps/nccl/src/nccl.cu   | 4 ++--
 include/mscclpp/gpu.hpp | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/apps/nccl/src/nccl.cu b/apps/nccl/src/nccl.cu
index 0ce6f8bab..ec47362cf 100644
--- a/apps/nccl/src/nccl.cu
+++ b/apps/nccl/src/nccl.cu
@@ -511,8 +511,8 @@ NCCL_API ncclResult_t ncclCommCount(const ncclComm_t comm, int* count) {
 }
 
 NCCL_API ncclResult_t ncclCommCuDevice(const ncclComm_t comm, int* device) {
-  // TODO: implement this function
-  return ncclInternalError;
+  *device = comm->comm->bootstrap()->getRank();
+  return ncclSuccess;
 }
 
 NCCL_API ncclResult_t ncclCommUserRank(const ncclComm_t comm, int* rank) {
diff --git a/include/mscclpp/gpu.hpp b/include/mscclpp/gpu.hpp
index d25f9ce63..188f04966 100644
--- a/include/mscclpp/gpu.hpp
+++ b/include/mscclpp/gpu.hpp
@@ -8,7 +8,7 @@
 
 #include <hip/hip_runtime.h>
 #include <hip/hip_fp16.h>
-#include <hip/hip_bf16.h>
+// #include <hip/hip_bf16.h>
 
 using cudaError_t = hipError_t;
 using cudaGraph_t = hipGraph_t;