I'm not sure it could be for this complex quantization scheme. It's not a simple round to nearest quantization. Even if it is possible, it would require someone

I'm not sure it could be for this complex quantization scheme. It's not a simple round to nearest quantization. Even if it is possible, it would require someone to develop that functionality. Currently it does load the entire unquantized model into RAM first.