Update Rome model matching (#461)

* Update Rome model matching * Add another Rome model * Automatically setup NET GDR level from model [ROCm/rccl commit: 0331e39f81]
2021-11-05 08:53:47 -07:00
parent 709b3dc85b
commit fd98ee84b4
8 changed files with 276 additions and 52 deletions
@@ -828,6 +828,8 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, ncclUniqueId* comm
  NCCLCHECK(ncclTopoGetSystem(comm, &comm->topo));
  // save nRanks to ncclTopoSystem as indicator of multi-node
  comm->topo->nRanks = comm->nRanks;
+  // init netGdrLevel
+  comm->topo->netGdrLevel = -2;
  // Compute paths between GPUs and NICs
  NCCLCHECK(ncclTopoComputePaths(comm->topo, comm->peerInfo));
  // Remove inaccessible GPUs and unused NICs