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

Revert "[AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1641)"

Browse Source 
This reverts commit 00c1eb098c.
This commit is contained in:
PedramAlizadeh
2025-05-21 08:50:55 -07:00
committed by Arm Patinyasakdikul
parent 66ef428714
commit 7f878baef0
7 changed files with 15 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/tuning.cc
+4 -4
View File
@@ -329,11 +329,11 @@ static struct tuningModel tuning_model_5 {
// Follow order in RcclTunableColls
.llProtoRanges = {
/*ReduceScatter*/
{/*LL (min/max/factor/thread_threshold)*/ {0, 655360, 1, 16}, /*LL64/128 (min/max/factor/thread_threshold)*/ {131072, 4793500, 1, 64}},
{/*LL (min/max/factor)*/ {0, 655360, 1}, /*LL64/128 (min/max/factor)*/ {131072, 3211264, 1}},
/*AllGather*/
{/*LL (min/max/factor/thread_threshold)*/ {0, 98304, 1, 16}, /*LL64/128 (min/max/factor/thread_threshold)*/ {98304, 5592500, 1, 64}},
{/*LL (min/max/factor)*/ {0, 98304, 1}, /*LL64/128 (min/max/factor)*/ {98304, 5046272, 1}},
/*AllReduce*/
{/*LL (min/max/factor/thread_threshold)*/ {0, 1048576, 1, 0},/*LL64/128 (min/max/factor/thread_threshold)*/ {1048576, 144217728, 3145728, 0}},
{/*LL (min/max/factor)*/ {0, 1048576, 1},/*LL64/128 (min/max/factor)*/ {1048576, 9437184, 3145728}},
},
};
@@ -722,7 +722,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
}
}
}
// Set per-thread amount of work before we increase nThreads and nChannels
for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) {
comm->threadThresholds[a][NCCL_PROTO_LL] = NCCL_LL_THREAD_THRESHOLD;