Adjust nChannels for mi350 for single host use cases

Summary:
MI350 uses less channels than mi300x. Need more channels for single node communicators that do not uses all the xgmi links per gpu (i.e., nranks  <8). For such cases, we inflate the nChannels a bit to achieve higher b/w.

Keep nChannels for multi-node or single-node but nranks=8 cases the same.

update:

Upstream PR merged: ROCm/rccl#2027

Reviewed By: haoyuz

Differential Revision: D86027843

fbshipit-source-id: 6d8e8024a013d63db251b9de4e69d5f0027914a1

Author: dmwu

comms/rcclx/develop/src/init.cc

@@ -1340,7 +1340,13 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
     }
   }
   if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")) {
-    allGather3Data[rank].nc = 4;
+    if (nranks == 2 && nNodes == 1){
+      allGather3Data[rank].nc = 16;
+    } else if (nranks == 4 && nNodes == 1){
+      allGather3Data[rank].nc = 8;
+    } else {
+      allGather3Data[rank].nc = 4;
+    }
   }
 
   allGather3Data[rank].pivotA2AEnabled = comm->topo->pivotA2AEnabled && rcclParamPivotAlltoallEnable();