From cf980e9b9cd00f8e72495302ee65b9173f27571d Mon Sep 17 00:00:00 2001 From: Ziyue Yang Date: Wed, 2 Oct 2024 04:32:41 +0800 Subject: [PATCH] Fix size matching in MSCCL (#1318) [ROCm/rccl commit: 7830af58440b6de18e80106c96e20278624e0020] --- projects/rccl/src/misc/msccl/msccl_lifecycle.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/projects/rccl/src/misc/msccl/msccl_lifecycle.cc b/projects/rccl/src/misc/msccl/msccl_lifecycle.cc index 39c4bd25f5..7492298762 100644 --- a/projects/rccl/src/misc/msccl/msccl_lifecycle.cc +++ b/projects/rccl/src/misc/msccl/msccl_lifecycle.cc @@ -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 &&