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

2.17.1-1

Browse Source 
Add new NVLS algorithm for allreduce using NVLink SHARP (intra-node only).
Add new config options: cgaClusterSize, minCTAs, maxCTAs, netName.
Enable LL128 when we use PXN to close rings.
NVTX3 includes update.
Fix crash when one CollNet (SHARP) rail fails to initialize.
This commit is contained in:
Sylvain Jeaugey
2023-02-27 02:48:21 -08:00
parent f3d5166783
commit 5d3ab08b69
72 changed files with 4541 additions and 2391 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/connect.cc
+2 -2
View File
@@ -313,8 +313,8 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePa
// Honor NCCL_MIN_NRINGS/NCCL_MAX_NRINGS.
// We permit combining max, then min, to only use the first channels, then duplicate them.
nChannels = comm->nChannels = std::min((int)ncclMaxNchannels(), nChannels);
nChannels = comm->nChannels = copyChannels(comm, nChannels, ncclMinNchannels(), ringPrev, ringNext);
nChannels = comm->nChannels = std::min(std::min(ncclMaxNchannels(), nChannels), comm->maxCTAs);
nChannels = comm->nChannels = copyChannels(comm, nChannels, std::max(ncclMinNchannels(), comm->minCTAs), ringPrev, ringNext);
// Create rings array and check all is fine
NCCLCHECK(ncclBuildRings(nChannels, rings, comm->rank, comm->nRanks, ringPrev, ringNext));