cache past K/V
compute only new token states
reduce K/V storage with MQA, GQA, or MLA