Search papers, labs, and topics across Lattice.
1
0
Pushing KV cache quantization down to 2 bits without sacrificing downstream accuracy is now viable by rotating key and value matrices according to their joint second-order interaction statistics.
