OneLA Shares Linear-Attention States Across Beams for 1.54-2.46x Faster Generative-Rec Decoding

_reachsumit · x · 2026-09-14

Original post →

More from Infra

Infra channel →