vulkan: handle larger batch sizes (>4) efficiently for IQ3_S mat-vec … - #27449

Merged

ggerganov merged 2 commits into

ggml-org:masterfrom

LaurentZuijdwijk:vulkan/iq3s-mmv-register-spill

Sep 2, 2026

Commits on Aug 20, 2026

vulkan: handle larger batch sizes (>4) efficiently for IQ3_S mat-vec when NUM_COLS > 4. 5x perf at n=8
Marshall
committed
211abf6
adds 2 cases per quant type at `k=16*256` to the `all_types` mat-vec sweep
Marshall
committed
d19f920