In the context of large language model (LLM) inference, the difference between **raw KV rows** and **comp