invokeai-xpu

InvokeAI running on Intel Arc GPUs using PyTorch XPU (Level Zero), deployed headlessly on Ubuntu 24.04 LXC with systemd

InvokeAI on Intel GPU (XPU, no CUDA)

This repo is basically a “what finally worked” script for running InvokeAI on Intel GPUs (Arc or iGPU) using PyTorch XPU on Ubuntu 24.04, inside a Proxmox LXC.

No CUDA.
No Docker.
No fancy setup.

Just: make InvokeAI run on Intel GPU without fighting it for days.

Disclaimer: this is purely vibecoded

---

Why this repo exists

I have an intel Arc B50 and too much free time. I like how easy InvokeAI is to use and wanted to see if i can make it work with the Arc.

After a long debugging session with ChatGPT, this script captures everything that was needed, in the order to actually get it working:

• PyTorch XPU
• InvokeAI 6.10
• Intel Arc / iGPU
• Headless Ubuntu 24.04
• Proxmox LXC

So this is:

• ✅ A reproducible install that runs InvokeAI on Intel XPU
• ✅ Headless, systemd-managed, browser UI accessible
• ✅ Tested on Intel Arc (B-series)
• ❌ Not optimized
• ❌ Not officially supported
• ❌ Not guaranteed to survive future InvokeAI releases

---

What this script does

The install script (install-invoke-xpu.sh) performs the following:

System & GPU

• Installs Intel GPU userspace:
  • Level Zero
  • OpenCL ICD
  • Media drivers
• Verifies /dev/dri/renderD* access inside LXC

Python & PyTorch

• Creates a clean Python virtualenv at: /opt/invokeai-xpu
• Forces PyTorch XPU wheels
• Prevents CUDA wheels from being pulled in

InvokeAI

• Installs InvokeAI 6.10.0
• Writes a minimal, XPU-safe InvokeAI config
• Applies required runtime patches, including:
  • Making intel_extension_for_pytorch (IPEX) optional
  • Guarding against missing torch.xpu.mem_get_info()
  • Fixing invocation stats crashes when VRAM info is unavailable

Runtime

• Creates a systemd service
• Verifies XPU availability at startup
• Runs InvokeAI fully headless
• Exposes the web UI over HTTP

End result

InvokeAI runs on Intel GPU, images generate successfully, and the UI is accessible from a browser.

Known issues

• InvokeAI cannot accurately detect available VRAM on Intel XPU
  (workarounds are applied; generation still works)
• Generation preview may not update in real time
• UI may require a refresh after generation finishes
• Model downloads may not show progress until refresh
• GPU may stay “awake” after generation unless the service is stopped. This keeps fans spinning quite agressively. On Proxomox current workaround is to run intel_gpu_top -l on host after you want the fans to spin down OR constantly while running Invoke.

---

Requirements

Proxmox host

• Intel GPU
• XE driver
• GPU passed through to the LXC
• This is what I added to the container config file on the Proxmox host (/etc/pve/lxc/<id>.conf):

lxc.cgroup2.devices.allow: c 226:0 rwm
lxc.cgroup2.devices.allow: c 226:129 rwm
lxc.mount.entry: /dev/dri/card0 dev/dri/card0 none bind,optional,create=file
lxc.mount.entry: /dev/dri/renderD129 dev/dri/renderD129 none bind,optional,create=file

---

LXC container

• Ubuntu 24.04 LTS
• Fresh install recommended
• Internet access
• Privileged, nesting

---

Quick start

On a clean Ubuntu 24.04 LXC:

git clone https://github.com/Raasu2/invokeai-xpu.git
cd invokeai-xpu
chmod +x install-invoke-xpu.sh
sudo bash install-invoke-xpu.sh

Thanks / Credits

Huge thanks to MordragT for the original InvokeAI XPU patches. This setup would not exist without that work.