Search papers, labs, and topics across Lattice.
1
0
3
Get 2.2x memory savings in your video VLMs without sacrificing accuracy by proactively pruning the KV-cache using optical flow and saliency, *before* computing attention.
