Several attention heads side by side, each with its own pattern, then concatenated and mixed. Grouped-query and multi-query attention share keys and values among heads, and the KV cache shrinks.
This interactive visualization needs JavaScript.