Making quants by using the GPU

[Nexesenex]

Is it already, or would it be reasonably possible to use the compute capabilities of our GPUs to quantize models with Llama.CPP?
This considering notably the additional calculation brought by the new iMatrix feature when making guided quants, and notably the IQ ones (which are out of my i7-6700k league for quantizing 70b models in a reasonable time).