Head-level attention fusion trims compute

Merging full‑attention and linear‑attention at the head granularity slashes transformer FLOPs without appreciably hurting downstream quality. The trick is to keep the expensive quadratic path only whe...