Search papers, labs, and topics across Lattice.
University of California
1
0
2
Reading barely 20% of the KV cache is enough to match up to 99% of dense attention accuracy at 32K context, yielding an immediate 1.69× speedup over FlashAttention without any retraining.
