Comment by sercand

I have one of my own. My reason for this is that you can get up to 2x performance compared to llama.cpp on the same model weights. It is optimized for single model to serve it well.