From 6ab20a7c6bfa1d67c3b5b377af05c20e8d4bcda5 Mon Sep 17 00:00:00 2001 From: Nusrat Islam Date: Tue, 4 Jun 2024 15:22:38 -0500 Subject: [PATCH] graph: fix multi-node minChannel count --- src/graph/connect.cc | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-) diff --git a/src/graph/connect.cc b/src/graph/connect.cc index f00a34fd8c..d59c092981 100644 --- a/src/graph/connect.cc +++ b/src/graph/connect.cc @@ -668,11 +668,9 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePa nChannels = comm->nChannels = copyChannels(comm, nChannels, 2*nChannels, ringPrev, ringNext); } - int minNchannels = 64; - if (comm->nNodes == 1) { - minNchannels = ncclMinNchannels(); - } else { - minNchannels = std::min(64,ncclMinNchannels()); + int minNchannels = ncclMinNchannels(); + if (comm->nNodes > 1) { + minNchannels = std::min(64, maxChannels); } if (mscclEnabled() && (comm->topo->mscclEnabled || mscclForceEnabled())) {