This is somehow similar to this #678.

Support for quantized models. Currently finetuning TVM compiled models are not yet supported. The benefit of this is less GPU resource requirements which will make running models cheaper at the expense of performance.