Fix topo explorer's nccl 2.23 compatibility (#1623)

* Fix compiler issues due to broken compatibility 

* Fix segfault and pass rank instead of busid and add a pointer to cover a new algorithm

[ROCm/rccl commit: aace4e27f8]
Dieser Commit ist enthalten in:
Mustafa Abduljabbar
2025-04-02 09:47:29 -04:00
committet von GitHub
Ursprung 25dafc0c82
Commit 0a81478bd9
2 geänderte Dateien mit 12 neuen und 10 gelöschten Zeilen
+5 -3
Datei anzeigen
@@ -353,7 +353,7 @@ static ncclResult_t selectTransport(struct ncclComm* comm, struct ncclTopoGraph*
struct ncclTransport *transport = ncclTransports[t];
struct ncclTransportComm* transportComm = type == 1 ? &transport->send : &transport->recv;
int ret = 0;
NCCLCHECK(transport->canConnect(&ret, comm->topo, graph, myInfo, peerInfo));
NCCLCHECK(transport->canConnect(&ret, comm, graph, myInfo, peerInfo));
if (ret) {
connector->transportComm = transportComm;
NCCLCHECK(transportComm->setup(comm, graph, myInfo, peerInfo, connect, connector, channelId, connIndex));
@@ -721,7 +721,7 @@ ncclResult_t initTransportsRank_1(struct ncclComm* comm, struct allGatherInfo *a
//struct ncclTopoGraph treeGraph;
//struct ncclTopoGraph collNetGraph;
//struct ncclTopoGraph nvlsGraph;
struct ncclTopoGraph* graphs[] = { &treeGraph, &ringGraph, &collNetGraph, &collNetGraph, &nvlsGraph, &nvlsGraph };
struct ncclTopoGraph* graphs[] = { &treeGraph, &ringGraph, &collNetGraph, &collNetGraph, &nvlsGraph, &nvlsGraph, &nvlsGraph};
int nChannelsOrig;
struct ncclTopoRanks** allTopoRanks = NULL;
@@ -963,6 +963,7 @@ ncclResult_t initTransportsRank_1(struct ncclComm* comm, struct allGatherInfo *a
}
comm->nChannels = std::min(treeGraph.nChannels, ringGraph.nChannels);
TRACE(NCCL_INIT,"treeGraph.nChannels: %d , ringGraph.nChannels: %d", treeGraph.nChannels, ringGraph.nChannels);
NCCLCHECKGOTO(ncclTopoPreset(comm, graphs, &allGather3Data[rank].topoRanks), ret, fail);
fail:
return ret;
@@ -975,7 +976,7 @@ ncclResult_t initTransportsRank_3(struct ncclComm* comm, struct allGatherInfo *a
int nranks = comm->nRanks;
cpu_set_t affinitySave;
struct ncclTopoGraph* graphs[] = { &treeGraph, &ringGraph, &collNetGraph, &collNetGraph, &nvlsGraph, &nvlsGraph };
struct ncclTopoGraph* graphs[] = { &treeGraph, &ringGraph, &collNetGraph, &collNetGraph, &nvlsGraph, &nvlsGraph, &nvlsGraph};
int nChannelsOrig;
struct ncclTopoRanks** allTopoRanks = NULL;
@@ -1090,6 +1091,7 @@ ncclResult_t initTransportsRank_3(struct ncclComm* comm, struct allGatherInfo *a
}
NCCLCHECKGOTO(ncclCalloc(&rings, nranks*MAXCHANNELS), ret, fail);
TRACE(NCCL_INIT, "rank %d nranks %d nchannels %d", rank, nranks, comm->nChannels);
NCCLCHECKGOTO(ncclTopoPostset(comm, nodesFirstRank, nodesTreePatterns, allTopoRanks, rings, graphs, NULL, nc), ret, fail);
if (comm->topo->treeDefined) NCCLCHECK(ncclTreeBasePostset(comm, &treeGraph));