Fix size matching in MSCCL (#1318)

This commit is contained in:
Ziyue Yang
2024-10-02 04:32:41 +08:00
committed by GitHub
parent 8b3ed8f104
commit 7830af5844
+1 -1
View File
@@ -330,7 +330,7 @@ static ncclResult_t mscclInternalSchedulerSelectAlgo(int rank, struct mscclSched
auto &m = status.algoMetas[i];
size_t nBytes = param->count * ncclTypeSize(param->dataType) * m.sizeMultiplier;
bool msgSizeIsValid =
param->count > 0 && (param->count % m.nChunksPerLoop) == 0 &&
param->count > 0 && ((param->count * m.sizeMultiplier) % m.nChunksPerLoop) == 0 &&
nBytes >= m.minBytes && (m.maxBytes == 0 || nBytes <= m.maxBytes);
if (msgSizeIsValid &&
m.nRanks == param->nRanks &&