add reduce/broadcast algo/proto selection table for multi-node gfx940 (#1889)

[ROCm/rccl commit: 9c36439354]
This commit is contained in:
isaki001
2025-09-10 14:25:23 -05:00
gecommit door GitHub
bovenliggende 2577b33de8
commit 9fa7a738da
4 gewijzigde bestanden met toevoegingen van 9 en 3 verwijderingen
@@ -79,7 +79,7 @@ inline size_t rcclGetSizePerRank(ncclFunc_t const& func, size_t const& nBytes, i
// For AG, this is the send size per rank
// For RS, this is the recv size per rank
// For AR, this is the send/recv size per rank
return (func == ncclFuncReduceScatter || func == ncclFuncAllGather || func == ncclFuncBroadcast) ? nBytes / nRanks : nBytes;
return (func == ncclFuncReduceScatter || func == ncclFuncAllGather || func == ncclFuncBroadcast || func == ncclFuncReduce) ? nBytes / nRanks : nBytes;
}
ncclResult_t rcclGetAlgoProtoIndex(const char *envStr, const char* algoProtoString[], int nEntries, int& result);
ncclResult_t rcclOverrideProtocol(const char* ncclProtoStr[], float table[][NCCL_NUM_PROTOCOLS], struct ncclTaskColl* info);