Serving GLM-5.3 at 607.8 tok/s across two Nvidia DGX stations with vLLM