KV cache

Runtime attention keys and values retained from prior tokens so decoding can reuse them instead of recomputing the full prefix.