1 paper · 1 filter
Shuaiyi Li, Zhisong Zhang, Yan Wang +5
Block attention, which processes the input as separate blocks that cannot attend to one another, offers significant potential to improve KV cache reuse in long-context scenarios su…