"KV快取" 与 "快取快取" 的术语口径问题"

** Phenomenon **: - §2.3.3 clearly defines: "KV Cache" is an internal optimization in the model, "caching key-value pairs of calculated tokens during a single inference process"; "Prompt Cache" is an optimization at the API service layer, "caching the results of the same prefix across multiple API requests." - However, in Figure 2-10, "Request 1" and "Request 2" are two independent API requests, and they share the prefix "System Prompt + Tools". However, the arrow in the figure indicates "KV reuse", and the title is "KV Cache Prefix Reuse Mechanism". ** Problem **: According to the definition in §2.3.3, prefix reuse across requests should belong to "Prompt Cache"; however, Figure 2-10 indicates this cross-request behavior as "KV Cache / KV Reuse". The terminology in the same book is inconsistent, and readers will misunderstand that the lifetime of KV Cache is across requests, which conflicts with §2.3.3 (and the intuitive understanding: KV Cache lifetime = single inference request). ** Suggestion **: My understanding is that the lifetime of KV Cache is a single inference request (prefill stage generates K and V, and decode stage reuses tokens one by one); prefix reuse across requests belongs to Prompt Cache. If we understand Figure 2-10 in this way, the part of cross-request reuse should be Prompt Cache, not KV Cache. Please let me know if I have misunderstood.