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

minor latency tuning (#591)

Browse Source 
* minor tuning for tree ll
This commit is contained in:
akolliasAMD
2022-08-03 15:07:44 -06:00
committed by GitHub
parent 9089e68a99
commit 4cecdc9be5
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/enqueue.cc
+3 -3
View File
@@ -464,11 +464,11 @@ static ncclResult_t getAlgoInfo(struct ncclInfo* info, int collNetTypeSupport, i
info->nChannels = nc;
if (!userTuneInput) {
// always respect user settings
if (info->nBytes <= 196608) {
if (info->nBytes <= 900000) {
info->protocol = NCCL_PROTO_LL;
info->algorithm = NCCL_ALGO_TREE;
info->nChannels = std::min(comm->nChannels, info->nBytes <= 65536? 4 : 12);
} else if (info->nBytes <= 1048576) {
info->nChannels = std::min(comm->nChannels, info->nBytes <= 16384? 4 : 24);
} else if (info->nBytes <= 2200008) {
info->protocol = NCCL_PROTO_LL;
info->algorithm = NCCL_ALGO_RING;
} else {