vLLM uses Python 3.12 built-in hash() which leads to predictable hash collisions in prefix cache