Search papers, labs, and topics across Lattice.
1
0
2
Prefill latency can be slashed by 80% without sacrificing long-context retrieval accuracy by pairing concatenation-aware fine-tuning with selective KV recomputation.
