Update Rome model matching (#461)

* Update Rome model matching

* Add another Rome model

* Automatically setup NET GDR level from model
This commit is contained in:
Wenkai Du
2021-11-05 08:53:47 -07:00
committed by GitHub
parent 04c10a6025
commit 0331e39f81
8 changed files with 276 additions and 52 deletions
+2
View File
@@ -531,6 +531,8 @@ ncclResult_t initTransportsRank_1(struct ncclComm* comm, struct allGather1Data_t
//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