Search papers, labs, and topics across Lattice.
Affiliation:
1
0
2
LLM value caches are surprisingly redundant, allowing for significant memory compression through simple symmetry exploitation without hurting inference speed.
