Ok, I need to try that. I'm getting 45tok/s with vLLM on my 6000. >600tok/s concurrent, but 45tok/s single request.