There’s a merge request !21658 on the mesa gitlab repository on freedesktop which to my understanding would allow you to run AMD drivers in a guest VM while passing through (parts of) the GPU with virtio.
According to the comments on a Phoronix article “VirtIO Native Context Being Worked On For AMD Drivers To Enhance VM Performance” about this merge request, this would allow you to have GPU acceleration for multiple VMs with a single GPU without having something like mediated passthrough and I see someone from the QubesOS team is interested in this for use in GPU acceleration.
I am not entirely familiar with Mesa or anything low level regarding GPUs, but would this actually allow for running multiple VMs at the same time off a single GPU or is this just an easier way to assign a single GPU to a single VM?