Merge remote-tracking branch 'rccl/develop' into 2.19.4

[ROCm/rccl commit: 1a134b283b]

projects/rccl/src/graph/paths.cc

@@ -863,6 +863,7 @@ ncclResult_t ncclTopoComputeP2pChannels(struct ncclComm* comm) {
   } else {
     // Round to next pow2 nChannelsPerPeer and nChannels
     comm->p2pnChannelsPerPeer = (ncclParamNChannelsPerPeer() == -2 ? nextPow2(minChannels) : ncclParamNChannelsPerPeer());
+    comm->p2pnChannelsPerPeer *= 2;
     comm->p2pnChannels = nextPow2(comm->p2pnChannels);
   }
 

projects/rccl/src/graph/tuning.cc

@@ -362,6 +362,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
       if (coll == ncclFuncAllGather && a != NCCL_ALGO_RING && a != NCCL_ALGO_NVLS) continue;
 
       for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
+        if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_SIMPLE && IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx94") && comm->topo->nodes[GPU].count == comm->topo->nRanks) continue;
         if ((a == NCCL_ALGO_NVLS || a == NCCL_ALGO_NVLS_TREE) && p != NCCL_PROTO_SIMPLE) continue;
         int collnet = (a == NCCL_ALGO_COLLNET_DIRECT || a == NCCL_ALGO_COLLNET_CHAIN) ? 1 : 0;
         float bw = nNodes <= 2 || collnet ? graphs[a]->bwIntra : graphs[a]->bwInter;