[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>
Gabriel Wu committed
Aug 19, 2026 at 21:25 UTC
2b7fcbf52782f8729fd6ce6c9ab803617d72897b
1 file changed
+1
-1
csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh
+1
-1
@@ -250,7 +250,7 @@ void cutlass_gemm_blockwise_sm120_fp8_dispatch(torch::stable::Tensor& out,
250
torch::stable::Tensor const& b_scales) {
251
int M = a.size(0);
252
// more heuristic tuning can be done here by checking N/K dimensions as well
253
- bool swap_ab = (M <= 64) || (M % 4 != 0);
253
+ bool swap_ab = (M <= 64);
254
255
if (!swap_ab) {
256
if (M <= 256) {