AI/rocm-systems
2
0
Fork 0
Koodi Ongelmat Pull-pyynnöt Actions Paketit Projektit Julkaisut Wiki Toiminta

2.7.5-1

Selaa lähdekoodia 
Minor fixes for A100 platforms.
Add a WARN for invalid GroupEnd call.
This commit is contained in:
Sylvain Jeaugey
2020-06-22 09:36:20 -07:00
vanhempi 5949d96f36
commit 01afd20a77
4 muutettua tiedostoa jossa 8 lisäystä ja 5 poistoa
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/tuning.cc
+1 -1
Näytä tiedosto
@@ -107,9 +107,9 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
float speed = comm->nNodes <= 2 || a == NCCL_ALGO_COLLNET ? graphs[a]->speedIntra : graphs[a]->speedInter;
float busBw = graphs[a]->nChannels * speed;
if (compCap80) busBw *= 0.92;
// Various model refinements
if (compCap80) busBw = std::min(busBw, 235.0f);
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL) busBw *= (comm->nNodes > 1 || coll == ncclCollAllReduce || coll == ncclCollReduce) ? 1.0/4.0 : 1.0/3.0;
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (ppn < 2 ? 0.7 : 0.92 /*120.0/128.0*/), ll128MaxBwPerCh[coll]*graphs[a]->nChannels);
double maxTreeBw = comm->nNodes > 2 ?