4 ms·Serving Open Source Models 4x faster than vLLM by quantizing with ~no tradeoffs1 points by kmdupree 3y ago