diff --git a/src/enqueue.cc b/src/enqueue.cc index 98a3d78491..1714aa1158 100644 --- a/src/enqueue.cc +++ b/src/enqueue.cc @@ -1783,7 +1783,14 @@ static ncclResult_t computeCollChunkInfo(struct ncclInfo* collInfo, size_t nByte if (collInfo->protocol == NCCL_PROTO_LL) chunkSize /= 2; if (collInfo->protocol == NCCL_PROTO_LL128) chunkSize = (chunkSize / NCCL_LL128_LINEELEMS) * NCCL_LL128_DATAELEMS; - if (collInfo->algorithm == NCCL_ALGO_COLLNET_DIRECT) { + if (collInfo->algorithm == NCCL_ALGO_TREE && collInfo->protocol == NCCL_PROTO_SIMPLE) { + if (collInfo->pattern == ncclPatternTreeUpDown) { + // Optimize chunkSize / nSteps + while (collInfo->nBytes / (collInfo->nChannels*chunkSize) < collInfo->comm->channels[0].tree.depth*8 && chunkSize > 131072) chunkSize /= 2; + while (collInfo->nBytes / (collInfo->nChannels*chunkSize) < collInfo->comm->channels[0].tree.depth*4 && chunkSize > 65536) chunkSize /= 2; + while (collInfo->nBytes / (collInfo->nChannels*chunkSize) < collInfo->comm->channels[0].tree.depth && chunkSize > 32768) chunkSize /= 2; + } + } else if (collInfo->algorithm == NCCL_ALGO_COLLNET_DIRECT) { // Optimize chunkSize / nSteps while (nBytes / (nChannels * collInfo->comm->channels[0].collnetDirect.nHeads * chunkSize) < collInfo->comm->channels[0].collnetDirect.depth * 64 && chunkSize > 131072) chunkSize /= 2; while (nBytes / (nChannels * collInfo->comm->channels[0].collnetDirect.nHeads * chunkSize) < collInfo->comm->channels[0].collnetDirect.depth * 8 && chunkSize > 65536) chunkSize /= 2;