Search papers, labs, and topics across Lattice.
1
0
0
2
A single attention module dynamically switches between linear prefilling and full decoding, achieving $21.9\times$ faster prefilling and Transformer-level long-context recall without the quadratic bottleneck.
