Installing HA in a Proxmox LXC (because I need to share iGPU)

Configuration
1

So I have Proxmox which is currently has the following installed…
iSpy (LXC)
JellyFin (LXC)
Home Assistant (VM)

I have the iGPU passed thru to iSpy and JellyFin. I would also like to pass the GPU to Home Assistant for use with voice & LLM.

From what I understand, you can pass the GPU to the HA VM but is will disable use in the other LXC’s.

So I was thinking about putting HA in an LXC so i can access the GPU without knocking off the other apps.

I know HA in an LXC has its cons but I’m not sure if there’s any other way to get to the iGPU. I suppose I could install a dedicated GPU for HA but wanted to get some feedback from the forum before continuing.

BTW: HA is also using 2 USB devices for Bluetooth and Matter antennas. I realize that if I go with LXC, these can also be accessed by other containers.

2

Personally, its opening up a world of potential problems with addons, but if this is not an issue for you, then an LXC could be the way to go.

3

You don’t need to pass the GPU to HA. There’s no GPU driver in haos and it (HA) connects to LLMs through OpenAI compatible API calls… The piper / whisper addons shipped with haos are even CPU optimized. (And quite good) so you can give HAOS your GPU but it won’t do anything with it.

Setup a VM running ollama optimized fo your GPU with the correct drivers and pass the GPU to that…

4

A little off topic: have you looked at Frigate

5

Yes, i used Frigate for a while but switched to ispy. Both are good but I preferred ispy’s UI and configuration options.

6

Thanks for the information. My thoughts were to set up a local LLM eventually. So…can something like ollama be installed in a separate LXC container (using GPU) and have HA access it? If so, then that will be the best solution going forward.

7

Thats exactly what these two little boxes are doing:

image
image437×536 28.6 KB

The NUC14 is hosting:

  • Iona, a vm which runs a bunch of base home lab containers that don’t need heavy compute or any GPU it also has an NVidia 5700ti punched through to it to ollama
  • Taran, another vm - this time with an onboard iGPU (Intel a770xe) punched through to a SECOND instance of ollama (effectively, different serve) this one has more vram but slower silicon,

The Nuc 15 is hosting my HA instance.

The VMs easily could be on any of the hosts, and I have temporarily run homeassistant on the same box running taran and iona when I had a hw failure on NUC10 (the predecessor for NUC15, and yes i had to punch through the USBs to the host as well and move my dongles)

8

That sounds like a plan for me but all on one box

  1. Keep HA in the VM
  2. Get voice working without local LLM
  3. Eventually spin up an LXC with ollama, using the shared iGPU
  4. If iGPU can’t handle the load, go with a dedicated GPU.
1 Like
9

I’ve tested ollama on a local LXC and whilst it works fine, it really struggles without some decent horsepower behind it, which for me was a deal breaker given that was going to dramatically increase my idle power usage for the limited amount of use I was going to give it.

10

Correct if you don’t have at LEAST an Nvidia 3xxx card or better with 12G or BETTER vram I wouldn’t even attempt it.

11

thanks for the input