If you anything with a GPU anywhere, you can essentially forget it. Or at least this was the case a few years ago when I briefly toyed with using qubes seriously.
The situation has indeed improved and things have gotten smoother over the past few years IME! (AMD)
If you run multiple GPUs (so one for your GUI and the rest for whatever else you want to do), PCIe passthrough for the latter is pretty straightforward these days.
You can set up a dedicated gpu qube for the former - recommended but optional.
Edit: Seems like someone has managed to get CUDA to work, with some effort.
https://forum.qubes-os.org/t/nvidia-gpu-passthrough-into-lin...