AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[GRAPH] Match maxChannels for gfx942 CUs (#1302)

Browse Source
This commit is contained in:
Nilesh M Negi
2025-07-16 09:07:02 -05:00
committed by GitHub
parent 106024b0db
commit 6632183efe
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/connect.cc
+4 -4
View File
@@ -804,10 +804,10 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePa
} }
} }
// Only use full MAXCHANNELS for gfx94x and gfx950 // Only use full MAXCHANNELS for gfx942 (MI300X) and gfx950
maxChannels = (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx942") || IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950")) ? maxChannels = (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx942") ||
((comm->topo->nodes[GPU].nodes[0].gpu.cu == 80 || comm->topo->nodes[GPU].nodes[0].gpu.cu == 20 || comm->topo->nodes[GPU].nodes[0].gpu.cu == 38) IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950"))
? comm->topo->nodes[GPU].nodes[0].gpu.cu : MAXCHANNELS) : 2*CHANNEL_LIMIT; ? std::min(comm->topo->nodes[GPU].nodes[0].gpu.cu, MAXCHANNELS) : 2*CHANNEL_LIMIT;
if (graphs[NCCL_ALGO_RING]->nIntraChannels > 0 || comm->nNodes > 1) { if (graphs[NCCL_ALGO_RING]->nIntraChannels > 0 || comm->nNodes > 1) {
maxChannels = std::min(64, maxChannels); maxChannels = std::min(64, maxChannels);