vLLM 0.34 sizes the KV cache from free memory, uses block size 64 and drops Model Runner V1. What changed, opt-out flags and a safe upgrade plan.