1 paper
Yuechuan Chen, Yanda Guo, Xiaoxi Liu +4
Some observation-window-based KV cache compressors use prompt attention to identify important tokens but allocate uniform or predefined budgets across heads, overlooking input-depe…