How stupid would It be to install LM Studio in dom0?

General Discussion
1

I have exactly 1 GPU. I do like using models often. How bad would it be to have an AppVM to download things into LM Studio, but also just install vulkan drivers into dom0 (which may not be possible) and move all the models and runtime components etc into dom0 after the download?

I either I use dual-boot so I can run models quicker or I try to somehow get LM Studio into dom0.

I know both options break the Qubes security model. I would like to know which is worse. I also use models frequently enough that I find myself using Qubes less.

1 Like
2

I’d say your best option is to get GPU passthrough into a qube for your needs.

2 Likes
3

It makes everything so much less portable. And I only need a GPU when I use models.

1 Like
4

It depends on your threat model.

1 Like
5

You should probably wait until this gets implemented: Support for opt-in GPU acceleration via virtio-GPU native contexts · Issue #8552 · QubesOS/qubes-issues · GitHub

1 Like
6

It makes everything so much less portable.

How so?

And I only need a GPU when I use models.

Perfect opportunity to just have an LLM Qube with GPU passthrough for all this.

7

Looks like it’s been discussed. I wonder if it’s months away or years away from being implemented. No one is assigned but 3 of 14 items are done. Last update in Jan 2025.

1 Like
8

Because I would need an external GPU that I attach just for LLMs.

9

So you only have an integrated GPU inside your PC?