NCCL_TREES variable and rome model fixes (#856)

[ROCm/rccl commit: d33cd5a233]
This commit is contained in:
akolliasAMD
2023-08-21 10:35:37 -06:00
gecommit door GitHub
bovenliggende 47330a62a6
commit 56129830a6
8 gewijzigde bestanden met toevoegingen van 58 en 60 verwijderingen
+3 -1
Bestand weergeven
@@ -1110,6 +1110,8 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
comm->topo->ll128Enabled = false;
// Topology hint for MSCCL internal scheduler about whether to enable MSCCL
comm->topo->mscclEnabled = false;
// Topology hint if tree has been defined by model or User
comm->topo->treeDefined = false;
// Compute paths between GPUs and NICs
NCCLCHECKGOTO(ncclTopoComputePaths(comm->topo, comm), ret, fail);
// Remove inaccessible GPUs and unused NICs
@@ -1362,7 +1364,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
NCCLCHECKGOTO(ncclCalloc(&rings, nranks*MAXCHANNELS), ret, fail);
NCCLCHECKGOTO(ncclTopoPostset(comm, nodesFirstRank, nodesTreePatterns, allTopoRanks, rings, graphs, nc), ret, fail);
if (comm->topo->pivotA2ANumBiRings == 3) NCCLCHECK(ncclTreeBasePostset(comm, &treeGraph));
if (comm->topo->treeDefined) NCCLCHECK(ncclTreeBasePostset(comm, &treeGraph));
// AllGather3 - end