Repository navigation

cuda: use the MMA flash attention kernel for GQA above 4 with quantized K/V on Ada - #307

Open

sb32445 wants to merge 3 commits into

PrismML-Eng:prismfrom

sb32445:pr/fattn-gqa-mma

Commits on Oct 4, 2026

cuda: use the MMA flash attention kernel for GQA above 4 with quantized K/V on Ada

sb32445
and
claude
committed
431509d
cuda: remove the GQA/MMA and vector-kernel query-limit switches

sb32445
and
claude
committed
069cd48

Commits on Oct 7, 2026

cuda : keep the vector FA kernel when MMA cannot read quantized K/V in place
sb32445
committed
d01570f