GPU passthrough to an appVM as opposed to standalone?

User Support General
1

I want to use my gpu run run AI models. All the guides I’ve seen regarding gpu passthrough say to make a standalone qube. Is there a known way to do this but with a typical template+appVM setup as opposed to using a standalone qube?

I have managed to attach my gpu to my appVM but installing the drivers into the template is where I’m having issues. I plan on running Ollama/open webui. Not running agents or anything like that.

2

Yes, there’s no reason why it shouldn’t work with an AppVM. Just install the drivers in the TemplateVM and then create an AppVM HVM with the PCI devices attached.

3

I would argue if its even worth to create a separate template for it.

I think the base template for all the others vms should also be expanded to allow for hvm/gpu boots. Same as locker docker installs, let all vm’s have it …

4

Sure, but you have to keep two things in mind:

  1. In certain situations (all in my case), the Template/AppVM’s GUI would just simply not work at all without the GPU when GPU drivers are installed: Cannot launch application in TemplateVM after installing `nvidia-driver-full` (unless GPU is passed through) · Issue #10485 · QubesOS/qubes-issues · GitHub . Maybe it’s possible to make some funky X11 config that would fix that.
  2. GPU driver installation (at least for Nvidia) requires you to use the in-VM kernel, which necessitates using a HVM. Using PVH qubes is better unless you really have to use HVM, so I wouldn’t do that for all my Qubes.