From b8116456885af2c4152aa2f45ce7125604fd344e Mon Sep 17 00:00:00 2001
From: Dingming Wu <dmwu0506@gmail.com>
Date: Tue, 11 Nov 2025 07:46:51 -0800
Subject: [PATCH] Adjust nChannels on gfx950 based on ranks and nodes for
 better bandwidth (#2027)

---
 src/init.cc | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/src/init.cc b/src/init.cc
index eb936ac9bd..5ef4c40481 100644
--- a/src/init.cc
+++ b/src/init.cc
@@ -1453,8 +1453,16 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
         allGather3Data[rank].nc = 4;
     }
   }
+  // For single node communicators that do not uses the full xgmi links per gpu, i.e., nranks < 8
+  // Inflate the nChannels a bit to achieve higher b/w. 
   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();