KV cache cut by ~45% with near‑same accuracy

Grouped Value Attention slashes transformer KV memory by roughly 45 % without hurting benchmark scores. By storing only grouped values and reconstructing keys on the fly, it eliminates the need to mat...