AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.12.10-1

Browse Source 
Fix bug with CollNet
Fix bug with zero-bytes send/recv operations
Fix NCCL_PARAM implementation to avoid taking a lock on every call
Fix bug when setting NCCL_IB_QPS_PER_CONNECTION to more than one.
Improve error reporting for network errors.


[ROCm/rccl commit: 353e8ba446]
This commit is contained in:
Sylvain Jeaugey
2022-03-30 02:25:49 -07:00
parent 7707479804
commit 27130280b2
16 changed files with 203 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/group.cc
+2 -2
View File
@@ -274,8 +274,8 @@ sched_delta:
if (sendbytes > sendChunkSize) { sendbytes = sendChunkSize; } else { sendRemaining = 0; }
// 0-bytes send/recv are considered as syncs. Make sure we only add syncs when requested
// (total size == 0), otherwise set size to -1.
if (sendbytes <= 0 && totSendBytes != 0) send = NULL;
if (recvbytes <= 0 && totRecvBytes != 0) recv = NULL;
if (sendbytes < 0 || (sendbytes == 0 && totSendBytes != 0)) send = NULL;
if (recvbytes < 0 || (recvbytes == 0 && totRecvBytes != 0)) recv = NULL;
if (recv) {
NCCLCHECKGOTO(scheduleRecv(comm, recvPeer, channelId, recvbytes, ((char*)recvBuff)+recvOffset), ret, group_cleanup);
}