Add 8P6L multi-node models (#239)
This commit is contained in:
@@ -329,6 +329,12 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int64_t busId, int
|
||||
|
||||
// Check if we are close enough that it makes sense to enable GDR
|
||||
int netGdrLevel = PATH_PXB;
|
||||
#ifdef TOPO_EXPL
|
||||
int arch, vendor, model;
|
||||
NCCLCHECK(ncclTopoCpuType(system, &arch, &vendor, &model));
|
||||
if (arch == NCCL_TOPO_CPU_ARCH_X86 && vendor == NCCL_TOPO_CPU_VENDOR_AMD && model == NCCL_TOPO_CPU_TYPE_ROME)
|
||||
netGdrLevel = PATH_PHB;
|
||||
#endif
|
||||
NCCLCHECK(ncclGetLevel(&ncclTopoUserGdrLevel, NULL, "NCCL_NET_GDR_LEVEL"));
|
||||
if (ncclTopoUserGdrLevel != -2) netGdrLevel = ncclTopoUserGdrLevel;
|
||||
int distance = gpu->paths[NET][n].type;
|
||||
|
||||
Reference in New Issue
Block a user