Reducing Transformer Key-Value Cache Size with Cross-Layer Attention | Read Paper on Bytez