Persistent KV cache for vLLM that reduces repeated pre-fill work and improves time to first token.