KV Cache (Key‑Value Cache)

Cached attention key/value tensors reused across decoding steps to avoid recomputation and reduce latency and cost.