Improve model matching for GPUs with alltoall XGMI connection (#1372)

[ROCm/rccl commit: 5c367a21d0]
Tento commit je obsažen v:
Wenkai Du
2024-10-11 09:53:14 -07:00
odevzdal GitHub
rodič ef54dd7cbc
revize 4cd1b3a9f5
3 změnil soubory, kde provedl 205 přidání a 18 odebrání
+3 -1
Zobrazit soubor
@@ -971,10 +971,12 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
// try to match 8P6L
NCCLCHECK(parseChordalRing(system, graph));
if (graph->nChannels) return ncclSuccess;
// try to match 8P alltoall connected GPUs
NCCLCHECK(parseA2a8P(system, graph, nullptr));
if (graph->nChannels) return ncclSuccess;
// try to match Rome 4P2H
const char *remap_str = getenv("NCCL_RINGS_REMAP");
NCCLCHECK(parseRome4P2H(system, graph, remap_str));
if (graph->nChannels) return ncclSuccess;
// try to match 1H16P
NCCLCHECK(parse1H16P(system, graph));