cuda: Q4_K mmvq at vdr 4 on Pascal (+ Q4_K geometry) -- 121 -> 81 us at n=1 - #1

Open

mewsian wants to merge 3 commits into

Kmic-68:p100-optimizationsfrom

mewsian:q4k-pascal-pr

Commits on Sep 30, 2026

cuda: Q4_K mmvq at vdr 4 on Pascal -- 121 -> 87.6 us at n=1
mewsian
committed
4aa5235
cuda: Q4_K-specific Pascal mmvq geometry

mewsian
and
claude
committed
e595d2b
cuda: keep Q4_K mmvq vdr 4 to sm_60-only builds

mewsian
and
claude
committed
12aff2a