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

Fix min_nchannels bug for gfx94* nranks=4 (#1202)

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

[ROCm/rccl commit: d9661c17e6]
This commit is contained in:
Nilesh M Negi
2024-06-07 14:31:28 -05:00
committed by GitHub
parent efe851cfff
commit 9b88e59ea4
1 changed files with 1 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/init.cc
+1 -4
View File
@@ -1396,10 +1396,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
allGather3Data[rank].nc = 16; allGather3Data[rank].nc = 16;
else if (nranks == 4) else if (nranks == 4)
// NCCL_MIN_NCHANNELS=24 // NCCL_MIN_NCHANNELS=24
allGather3Data[rank].nc = 6; allGather3Data[rank].nc = 4;
else if (nranks == 8)
// NCCL_MIN_NCHANNELS=56
allGather3Data[rank].nc = 2;
} }
allGather3Data[rank].pivotA2AEnabled = comm->topo->pivotA2AEnabled && rcclParamPivotAlltoallEnable(); allGather3Data[rank].pivotA2AEnabled = comm->topo->pivotA2AEnabled && rcclParamPivotAlltoallEnable();