[jbellis]

Because you need kv proportional to context length during inference of a single token to avoid quadratic recomputation. So compressing the kv lets you handle longer contexts in the same amount of vram.