I was looking into trying the (crappy) NPU included in my AMD laptop, and it turns out that there is no support at this time for PCI passthrough of the device (just like with the iGPU, woe is me…).
I wonder if that is something that could be adapted to QubesOS. That would let user with available AMD NPUs to do locally accelerated AI even if they cannot pass-through their GPU (which in the case of laptops with just an iGPU can be tricky).
The repository you mentioned is under an excessively restrictive nonfree license. I reported this as a bug, as I suspect AMD incorrectly applied the license used for firmware. In general, it might be possible but is probably roughly as hard as GPU acceleration.