ggml-cuda: add flash-attn support for DKQ=320/DV=256 with ncols2=32 (…#22286

Merged

JohannesGaessler merged 7 commits into

ggml-org:masterfrom

lnigam:fattn_320_256_GQA32

Apr 28, 2026

Commits on Apr 22, 2026

ggml-cuda: add flash-attn support for DKQ=320/DV=256 with ncols2=32 (GQA=32)
lnigam
committed

Commits on Apr 24, 2026

Adding check to return BEST_FATTN_KERNEL_NONE in case GQA!=32
lnigam
committed

Commits on Apr 28, 2026