Search papers, labs, and topics across Lattice.
1
0
2
SparDA achieves up to 5.3× higher decode throughput by cleverly decoupling attention mechanisms and optimizing KV block selection.
