[Kernel] SM120: stop routing misaligned-M blockwise FP8 GEMMs to the small-M swapAB config (#52775)

Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
This commit is contained in:
Gabriel Wu
2026-08-19 21:25:29 +08:00
committed by GitHub
parent c2e7242ab7
commit 2b7fcbf527
@@ -250,7 +250,7 @@ void cutlass_gemm_blockwise_sm120_fp8_dispatch(torch::stable::Tensor& out,
torch::stable::Tensor const& b_scales) {
int M = a.size(0);
// more heuristic tuning can be done here by checking N/K dimensions as well
bool swap_ab = (M <= 64) || (M % 4 != 0);
bool swap_ab = (M <= 64);
if (!swap_ab) {
if (M <= 256) {