Search papers, labs, and topics across Lattice.
1
0
2
Forget balanced trees: exploiting the quality gap between token sources in speculative decoding unlocks 33% faster inference without sacrificing accuracy.
