invokeai-xpu

InvokeAI running on Intel Arc GPUs using native PyTorch XPU support, deployed headlessly on Ubuntu LXC/VM with systemd.

InvokeAI on Intel GPU (XPU, no CUDA)

This is a single-script installer for running InvokeAI on Intel GPUs using native XPU support (introduced in InvokeAI 6.14).

Just: run InvokeAI on Intel GPU with one script.

Disclaimer: purely vibecoded.

Why this script exists

I have an Intel Arc B50 Pro and wanted InvokeAI to work on it. Since InvokeAI 6.14 added native torch.xpu device support, all the custom patches, workarounds, and manual torch pinning are no longer needed.

This script captures the essentials in one reproducible installer:

• ✅ Native torch.xpu backend
• ✅ Headless, systemd-managed, browser UI accessible
• ✅ Pre-flight checks for Ubuntu version and kernel
• ✅ Installs Intel compute runtime from Intel's repo (25.40+ for Battlemage stability)
• ✅ Works on Proxmox LXC and VM (GPU passthrough)
• ✅ Uses uv with --torch-backend=xpu — gets correct torch+xpu automatically

---

What this script does

The installer (install-invokeai-xpu-proxmox.sh) performs:

Pre-flight

• Checks Ubuntu version (24.04, 25.04, 25.10, 26.04)
• Checks kernel version (≥ 6.14 required for Battlemage)
• Verifies /dev/dri/renderD* access inside LXC/VM

System & GPU

• Adds Intel GPU repo
• Removes stale Ubuntu compute runtime (24.x — known BMG crash)
• Installs Intel GPU userspace from Intel's repo:
  • Level Zero (libze1, libze-intel-gpu1)
  • OpenCL ICD (intel-opencl-icd)
  • Compiler/runtime (intel-igc-cm, intel-gsc)

Python & InvokeAI

• Installs uv (Astral's fast Python package manager)
• Creates relocatable venv at /opt/invokeai-xpu with uv-managed Python 3.12
• Installs invokeai[xpu]==6.14.0 via uv pip install --torch-backend=xpu
• Verifies torch.xpu.is_available() and device detection

Runtime

• Writes minimal invokeai.yaml (device: xpu, bfloat16, lazy_offload)
• Creates optional VRAM override at /etc/invokeai/invokeai-xpu.env
• Creates systemd service with Intel XPU env vars
• Includes BMG stability workarounds (NEOReadDebugKeys=1, RenderCompressedBuffersEnabled=0)
• Starts service on boot, logs to journal

End result

InvokeAI runs on Intel GPU, images generate successfully, UI accessible at http://<IP>:9090.

---

Requirements

Proxmox Host

• Intel GPU (Arc Alchemist A-series, Battlemage B-series)
• i915/Xe kernel driver loaded
• GPU passed through to LXC or VM

LXC Container Config (/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

⚠️ Device numbers may change. Verify with ls -l /dev/dri on host.

VM (PCI Passthrough)

• Add GPU as PCI device in VM hardware config
• Ensure vfio-pci driver binds on host

---

LXC Container / VM

• Ubuntu 24.04, 25.04, 25.10, or 26.04
• Kernel 6.14+ (Ubuntu 24.04: install linux-generic-hwe-24.04)
• Fresh install recommended
• Internet access
• Privileged container (LXC) or standard VM

---

Quick Start

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

That's it. Wait for completion, then open http://<VM-IP>:9090.

---

Configuration

All settings are at the top of install-invokeai-xpu-proxmox.sh:

VENV_DIR="/opt/invokeai-xpu"        # Virtual environment location
INVOKE_ROOT="/data/invokeai"        # InvokeAI root directory (models, outputs)
PORT="9090"                         # Web UI port
SERVICE_NAME="invokeai.service"     # systemd unit name
INVOKEAI_XPU_VRAM_TOTAL_GB="16"     # VRAM override (for passthrough VMs)
SERVICE_USER="root"                 # Service user (change to dedicated user if desired)
INVOKE_VER="6.14.0"                 # InvokeAI version

VRAM Override

For GPU passthrough VMs where driver-global VRAM query (Sysman) may not work, set:

INVOKEAI_XPU_VRAM_TOTAL_GB="16"  # Match your GPU's VRAM

Non-root Service User

# Before running installer:
useradd -r -s /bin/false invokeai
# Then set in script:
SERVICE_USER="invokeai"

---

Troubleshooting

Issue	Check
/dev/dri missing	GPU not passed through to LXC/VM
Kernel too old	Install HWE kernel or upgrade Ubuntu
torch.xpu.is_available() false	Level Zero packages missing
OOM on generation	Increase INVOKEAI_XPU_VRAM_TOTAL_GB
Service won't start	journalctl -u invokeai.service -f
SIGABRT in command_encoder	Stale compute runtime — re-run installer to upgrade from Intel repo
patchmatch warning	Non-fatal, does not affect generation

---

Thanks / Credits

• InvokeAI team — Native XPU support (PR #9401)
• LexiconCode — Primary author of native XPU backend
• MordragT — Original XPU patches that paved the way
• Astral — uv package manager

---

License

MIT