Understanding KV Cache: The Hidden Memory Cost of Serving LLMs | Dark Hacker News