Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Affiliation:
1
0
3
Speculative decoding drafters can achieve over 2x speedup on 256K context by treating distant tokens as shared far memory with sharpened attention.