https://blog.mumong.cn/posts/local-llm-inference-optimization/