Query-Key-Value (QKV) attention models can be designed using single-head attention or multiple attention heads paired with various sharing mechanisms. While a multi-head model performs attention on a group of feature sub-spaces in parallel, its Key-Value (KV) cache must retain the key and value representations for all of these parallel heads, denoted mathematically as $\left\{(\mathbf{K}_{\le i}^{[1]},\mathbf{V}_{\le i}^{[1]}),\dots,(\mathbf{K}_{\le i}^{[\tau]},\mathbf{V}_{\le i}^{[\tau]})\right\}$. To manage these representations efficiently, different sharing mechanisms dictate how keys and values are organized and shared across the multiple attention heads.