Skip to content

[Issue]: NCCL will not find next team #2016

@Qizhi697

Description

@Qizhi697

nccl/src/dev_runtime.cc

Lines 252 to 271 in 25368a7

while (true) {
if (t == nullptr) {
teamIsNew = true;
t = (struct ncclDevrTeam*)malloc(sizeof(struct ncclDevrTeam) + team.nRanks*sizeof(int));
t->team = team;
t->mcHandle = 0x0;
t->mcBasePtr = nullptr;
for (int i=0; i < team.nRanks; i++) {
t->worldRankList[i] = comm->rank + (i - team.rank)*team.stride;
}
break;
} else if (t->team.rank == team.rank && t->team.nRanks == team.nRanks && t->team.stride == team.stride) {
if (!multimem || t->mcBasePtr != nullptr) {
// Matching team is sufficient
if (outTeam) *outTeam = t;
return ncclSuccess;
}
break; // Need to enable multimem
}
}

NCCL Version

2.29.3

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions