AVX512+AVXVNNI GEMM implementation for quants using Q8_K for activations #710

Merged

ikawrakow merged 16 commits intomainfrom

Aug 22, 2025

Commits on Aug 20, 2025

q8_k_r16: basics
Iwan Kawrakow
committed
q8_k_r16: iq4_xs now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq2_xxs now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq2_xs now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq2_s now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq3_xxs now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq3_s now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq1_s and iq1_m now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: q2_K and q3_K now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq2_ks and iq2_k now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq2_kl now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq3_ks and iq3_k now uses q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq4_kss, iq4_ks, and iq4_k now use q8_k_r16 on Zen4+
Iwan Kawrakow
committed
q8_k_r16: iq5_ks, iq5_k, and iq6_k now use q8_k_r16 on Zen4+
Iwan Kawrakow
committed
Fix AVX2
Iwan Kawrakow
committed

Commits on Aug 21, 2025

Just always set num_rows to 16
Iwan Kawrakow
committed