1 paper · 2 filters
Bingchen Yao, Haobo Xu, Haokun Lin +6
Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Direc…