[GRAPH] Increase default nChannels to 112 for gfx950 (#1596)

Signed-off-by: nileshnegi <Nilesh.Negi@amd.com>

[ROCm/rccl commit: 1df73e209e]
This commit is contained in:
Nilesh M Negi
2025-03-14 14:47:03 -07:00
committed by GitHub
parent afd04a5117
commit cf17cff5b6
+4 -1
View File
@@ -1338,7 +1338,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
allGather3Data[rank].nc = std::max(allGather3Data[rank].nc, 4/ringGraph->nChannels);
if (ringGraph->nChannels > MAXCHANNELS/2)
allGather3Data[rank].nc = 1;
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx94") || IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")) {
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx94")) {
// Multi-node MI300A
int managed = 0;
CUDACHECK(hipDeviceGetAttribute(&managed, hipDeviceAttributeDirectManagedMemAccessFromHost, 0));
@@ -1355,6 +1355,9 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
allGather3Data[rank].nc = 4;
}
}
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")) {
allGather3Data[rank].nc = 4;
}
allGather3Data[rank].pivotA2AEnabled = comm->topo->pivotA2AEnabled && rcclParamPivotAlltoallEnable();
comm->topo->ll128Enabled = comm->topo->ll128Enabled || rcclParamLL128ForceEnable();