2 ms·Serving Open Source Models 4x faster than vLLM by quantizing with ~no tradeoffs3 points by georgehill 3y ago