MSCCL++ fix split path null deref (#1959)
* Add speriaswamy-amd to CODEOWNERS
* MSCCL++: fix split path null deref; key maps by parent ncclUniqueId
* removed no-op
[ROCm/rccl commit: 5bd5079de1]
Этот коммит содержится в:
коммит произвёл
GitHub
родитель
8e5016ebfd
Коммит
014fae1b51
@@ -2043,10 +2043,10 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
|
||||
comm->mscclppCompatible = true;
|
||||
comm->mscclpp_threshold = job->parent->mscclpp_threshold;
|
||||
comm->mscclpp_comm = job->parent->mscclpp_comm;
|
||||
auto& mscclppUniqueId = mscclpp_uniqueIdMap[*job->commId];
|
||||
mscclpp_uniqueIdMap[*job->commId] = mscclppUniqueId;
|
||||
mscclpp_uniqueIdReverseMap[mscclppUniqueId].insert(*job->commId);
|
||||
ncclCommToUniqueIdMap[comm] = *job->commId;
|
||||
const ncclUniqueId& parentUniqueId = ncclCommToUniqueIdMap[job->parent];
|
||||
auto& mscclppUniqueId = mscclpp_uniqueIdMap[parentUniqueId];
|
||||
mscclpp_uniqueIdReverseMap[mscclppUniqueId].insert(parentUniqueId);
|
||||
ncclCommToUniqueIdMap[comm] = parentUniqueId;
|
||||
}
|
||||
}
|
||||
else
|
||||
|
||||
Ссылка в новой задаче
Block a user