[ROCm/rccl commit: e5d2dfed34]
This commit is contained in:
Edgar Gabriel
2022-09-13 17:29:32 -05:00
zatwierdzone przez GitHub
rodzic 4c17f4dcc1
commit 4a86adcaba
+4 -4
Wyświetl plik
@@ -1098,12 +1098,12 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, ncclUniqueId* comm
for (int c=0; c<comm->nChannels; c++) {
struct ncclChannel* channel = comm->channels+c;
if (comm->nRanks == 1) continue;
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, channel, NCCL_MAX_TREE_ARITY, channel->tree.down, 1, &channel->tree.up, 0), ret, affinity_restore);
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, channel, 1, &channel->tree.up, NCCL_MAX_TREE_ARITY, channel->tree.down, 0), ret, affinity_restore);
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, c, NCCL_MAX_TREE_ARITY, channel->tree.down, 1, &channel->tree.up, 0), ret, affinity_restore);
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, c, 1, &channel->tree.up, NCCL_MAX_TREE_ARITY, channel->tree.down, 0), ret, affinity_restore);
// RCCL: need to connect binTree as well
if (comm->topo->pivotA2ANumBiRings == 3) {
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, channel, NCCL_MAX_TREE_ARITY, channel->binTree.down, 1, &channel->binTree.up, 0), ret, affinity_restore);
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, channel, 1, &channel->binTree.up, NCCL_MAX_TREE_ARITY, channel->binTree.down, 0), ret, affinity_restore);
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, c, NCCL_MAX_TREE_ARITY, channel->binTree.down, 1, &channel->binTree.up, 0), ret, affinity_restore);
NCCLCHECKGOTO(ncclTransportP2pConnect(comm, c, 1, &channel->binTree.up, NCCL_MAX_TREE_ARITY, channel->binTree.down, 0), ret, affinity_restore);
}
}
NCCLCHECKGOTO(ncclTransportP2pSetup(comm, &treeGraph, 0), ret, affinity_restore);