Serverless inference platform that continuously optimizes open-source LLMs and the full serving stack (kernels, engines, and hardware) for lower latency and higher efficiency