From 678366f5e243d03c32e3b04993fb0ecdf9d9883a Mon Sep 17 00:00:00 2001 From: isaki001 <36317038+isaki001@users.noreply.github.com> Date: Thu, 30 Oct 2025 12:08:12 -0500 Subject: [PATCH] gx950 multi-node tuning for LL/LL128 (#1953) * increased LL threshold for gfx950 AR to 256KB * AG/RS proto threshold update [ROCm/rccl commit: 72996e4d9f04ae18c7110a8d7718e622f20a9552] --- projects/rccl/src/graph/tuning.cc | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/projects/rccl/src/graph/tuning.cc b/projects/rccl/src/graph/tuning.cc index fe6ac8ec78..973b61b458 100644 --- a/projects/rccl/src/graph/tuning.cc +++ b/projects/rccl/src/graph/tuning.cc @@ -396,11 +396,11 @@ static struct tuningModel tuning_model_6 { // Follow order in RcclTunableColls .llProtoRanges = { /*ReduceScatter*/ - {/*LL (min/max/factor/thread_threshold)*/ {0, 131071, 1, 16}, /*LL64/128 (min/max/factor/thread_threshold)*/ {131071, 4194304, 1, 64}}, + {/*LL (min/max/factor/thread_threshold)*/ {0, 65536, 1, 16}, /*LL64/128 (min/max/factor/thread_threshold)*/ {65536, 4194304, 1, 64}}, /*AllGather*/ - {/*LL (min/max/factor/thread_threshold)*/ {0, 7, 1, 16}, /*LL64/128 (min/max/factor/thread_threshold)*/ {7, 8388608, 1, 64}}, + {/*LL (min/max/factor/thread_threshold)*/ {0, 32768, 1, 16}, /*LL64/128 (min/max/factor/thread_threshold)*/ {32768, 8388608, 1, 64}}, /*AllReduce*/ - {/*LL (min/max/factor/thread_threshold)*/ {0, 131071, 1, 0},/*LL64/128 (min/max/factor/thread_threshold)*/ {131071, 17660227, 3145728, 0}}, + {/*LL (min/max/factor/thread_threshold)*/ {0, 262144, 1, 0},/*LL64/128 (min/max/factor/thread_threshold)*/ {262144, 17660227, 3145728, 0}}, /*Reduce*/ {/*LL (min/max/factor/thread_threshold)*/ {0, 16383, 1, 0},/*LL64/128 (min/max/factor/thread_threshold)*/ {16383, 16777216, 1, 0}}, /*Broadcast*/