gx950 multi-node tuning for LL/LL128 (#1953)

* increased LL threshold for gfx950 AR to 256KB

* AG/RS proto threshold update

[ROCm/rccl commit: 72996e4d9f]
이 커밋은 다음에 포함됨:
isaki001 작성하고 GitHub가 2025-10-30 12:08:12 -05:00를 커밋함
1 부모 40f3faead0
커밋 678366f5e2
1 파일이 변경됨
+3 -3
+3 -3
파일 보기
@@ -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*/