perf(metal): faster moe prefill gather below 64 rows per expert - #33

Merged

asher merged 3 commits into

mainfrom

gather-qmm-prefill

Aug 9, 2026

Commits on Aug 9, 2026

(metal): iq2_xxs nax gather loader dequants via vectorized chunk16s, bit-identical
asher
committed
f6577ab
perf(metal): gather rhs nax simdgroups skip mma work for expert segments outside their row band
asher
committed
b8ee030
perf(metal): bm32 gather rhs nax tile dispatched below 64 rows per expert
asher
committed
25164da