Some things are hard with QEMU ( GPU Passthrough ) which are easier with Docker/LXC(D)
Qemu can do everything docker can do and more. I'm not really interested in first-class support for docker.
I'd be happy to look into it -
I was looking at this: https://medium.com/@calerogers/gpu-virtualization-with-kvm-qemu-63ca98a6a172
and it seems pretty convoluted, if you suggest that GPU has gotten simpler (this suggest it has https://www.collabora.com/news-and-blog/blog/2018/02/12/virtualizing-gpu-access/)
I would also not see much need for containers
Well, on sr.ht upstream I don't have GPUs to share anyway, so this angle is never going to convince me. However, the code orchestrating the images is pretty simple and you could easily rig up GPU passthrough or docker yourself.
Most images have a genimg script which can be used to build them (though some of these require building from a similar host, which you should set up following that host's installation procedure).