KV Cache (Key-Value Cache)

Cached attention key/value tensors reused across tokens or turns to reduce latency and cost in long contexts.