Extensions/ComfyUI Intel Arc XPU (OmniXPU Windows)
ComfyUI Extension

ComfyUI Intel Arc XPU (OmniXPU Windows)

Native Intel Arc Battlemage (B70 / XPU) acceleration for ComfyUI on Windows (ComfyUI-OmniXPU adapters and diagnostics). Comfy Registry installs require one extra setup step for native XPU binaries plus --enable-dynamic-vram --reserve-vram 4; see repository README. Benchmarks in the repository.

By wildersa·Created 2 days ago·Updated about 4 hours ago· 0
wildersa/ComfyUI-Intel-Arc-XPU
Nodes—
On cloudLocal install
Stars0
Updatedabout 4 hours ago
Readme

ComfyUI-Intel-Arc-XPU

Native Intel Arc Battlemage (bmg / Arc Pro B70) acceleration for ComfyUI on Windows, combining ComfyUI-OmniXPU model adapters and diagnostics with Windows XPU compute kernels, comfy-kitchen XPU backend, and DynamicVRAM support — zero core ComfyUI files modified.


1. Prerequisite (Required)

Your Windows ComfyUI environment must already have Windows 11 x64, Python 3.13, and torch==2.13.0+xpu installed:

"<ComfyUI_Python>\python.exe" -m pip install --upgrade torch==2.13.0+xpu torchvision==0.28.0+xpu --index-url https://download.pytorch.org/whl/xpu

[!IMPORTANT] If Python or PyTorch does not match Python 3.13 + torch==2.13.0+xpu (or if Intel XPU is unavailable), the extension prints a CONFLICT: message and disables itself cleanly instead of crashing ComfyUI. If an upstream ComfyUI update changes an internal hook, conflicting patches fall back to stock ComfyUI behavior.


2. Installation

Option A — ComfyUI Manager / Comfy Registry (+ 1 extra binary step on Windows)

Per Comfy Registry security standards, Registry packages exclude prebuilt .whl binaries and automated pip installers. Installing ComfyUI Intel Arc XPU (OmniXPU Windows) (comfyui-intel-arc-xpu) via the Manager Registry installs the Python adapters, conflict checks, and the OmniXPU Status diagnostic node — then finish setup with this one extra PowerShell step:

git clone https://github.com/wildersa/ComfyUI-Intel-Arc-XPU.git
cd ComfyUI-Intel-Arc-XPU
.\install.ps1

(If you already have the extension folder via Registry, running .\install.ps1 updates/installs the XPU runtime packages in your active ComfyUI .venv cleanly.)

Option B — Direct GitHub Install (All-in-One via Git Clone or Git URL)

If installing directly from GitHub (git clone into ComfyUI\custom_nodes\ or ComfyUI Manager → Install via Git URL, which requires allow_git_url_install = True in user\__manager\config.ini), the bundled Windows XPU runtimes are included in the repository and installed automatically by install.py on restart.

  • Remove any older comfyui-xpu-backend folder in custom_nodes/ if present so patches do not conflict.

3. Required / Recommended Flags & Environment Setup

After installing, configure these two Windows settings for full performance:

1. ComfyUI Launch Flags (Explicitly Recommended)

Add these launch arguments to your ComfyUI startup command or Comfy Desktop → Settings → Launch Arguments (launchArgs):

--enable-dynamic-vram --reserve-vram 4

What works with vs. without --enable-dynamic-vram --reserve-vram 4:

  • Without --enable-dynamic-vram: Compute acceleration still works — native Intel XPU kernels (omni_xpu_kernel), comfy-kitchen's xpu backend (INT8, ConvRot, FP8, W4A8), and model adapters still accelerate workflows that fit inside VRAM.
  • With --enable-dynamic-vram --reserve-vram 4: Unlocks Windows DynamicVRAM (comfy-aimdo XPU overlay) and host RAM weight staging, preventing XPU out of memory on large/video workflows (WAN_Animate2, MiniMax H3).

2. Attention Backend (OMNI_ATTN_BACKEND)

Windows defaults to torch (PyTorch SDPA) for stability. To enable CUTLASS-SYCL CUTE FMHA attention on Intel Arc BMG (or revert to torch when isolating experimental routes):

[Environment]::SetEnvironmentVariable("OMNI_ATTN_BACKEND", "cute", "User")   # enable CUTE FMHA
[Environment]::SetEnvironmentVariable("OMNI_ATTN_BACKEND", "torch", "User")  # revert to Windows default

Restart ComfyUI / Comfy Desktop after changing environment variables or launchArgs.


4. Verify Installation

Add the OmniXPU Status node to any workflow, or check <ComfyUI>\user\comfyui.log in PowerShell:

Select-String -Path "<ComfyUI>\user\comfyui.log" -Pattern "OmniXPU|comfy_kitchen backend xpu|DynamicVRAM|aimdo|Native ops|CUTE|OMNI_ATTN_BACKEND|CONFLICT|XPU out of memory"

Expected when healthy:

[OmniXPU] omni_xpu_kernel 0.2.0b3+torch213.bmg - available: sdp, norm, rotary, linear_fp8, int8, layout
Found comfy_kitchen backend xpu: {'available': True, 'disabled': False, ...}
DynamicVRAM support detected and enabled
[OmniXPU] XPU RAM weight staging enabled: budget ... MiB

| Log Signature | Meaning / Fix | | --- | --- | | [OmniXPU] CONFLICT: ... requires ... torch==2.13.0+xpu | Incompatible Python/PyTorch build; install torch==2.13.0+xpu on Python 3.13. | | Native Windows XPU runtime binaries (omni_xpu_kernel) are not installed | Installed via Registry without running the extra binary setup step (.\install.ps1 / install.py). | | XPU out of memory (without DynamicVRAM support detected and enabled) | Add --enable-dynamic-vram --reserve-vram 4 to your ComfyUI launch flags. | | experimental capability-driven BMG D128 CUTE route enabled | OMNI_ATTN_BACKEND=cute hit an experimental attention route; switch OMNI_ATTN_BACKEND back to torch to isolate artifacts. | | Attention backend 'comfy kitchen attention' is unavailable; using PyTorch attention. | CUDA-origin workflow; switch ModelAttentionBackend to "pytorch attention" (and bypass CUDA-only sol-attn sparse nodes). | | Found comfy_kitchen backend triton: {'available': False, 'disabled': True...} | Expected on Windows (not an error); XPU uses the xpu backend, not triton. |


5. Benchmarks & Telemetry

Measured on Windows 11 (Intel Arc Pro B70, 32 GB, torch==2.13.0+xpu) comparing Stock ComfyUI (OOTB) (out-of-the-box, no extension, DynamicVRAM off) against ComfyUI-Intel-Arc-XPU (--enable-dynamic-vram):

| Model | Config | Stock ComfyUI (OOTB) | With Extension (torch) | With Extension (cute) | | --- | --- | --- | --- | --- | | MOSS-Image | int8_convrot + W4A8, 25 steps | 375.28s cold · 50.75s warm | 103.45s cold · 20.01s warm | 29.53s cold (12.7×) · 18.27s warm (2.78×) | | MiniMax H3 Ref2VA | int8_convrot + NVFP4, 20 steps | 399.91s cold · 1582.00s warm | 209.29s cold · 179.84s warm | 186.08s cold (2.15×) · 160.25s warm (9.87×) | | Qwen Image 2.1 T2I | int8_convrot, 1024×1024, 25 steps | — (not tested) | 22.48s cold · 12.58s warm | 21.51s cold · 11.38s warm | | Z-Image Turbo | bf16, 1024×1024, 8 steps | — (not tested) | — | 33.41s cold · 5.19s–5.27s warm | | WAN 2.2 T2V 14B | fp8_scaled, 2 steps (Turbo) | — (not tested) | 104.37s cold | 93.09s cold · 82.62s warm |

<sub>cold = first run (loading models from disk + XPU/kernel warmup) · warm (hot) = subsequent runs with models/caches already loaded. Full tables and sample outputs in benchmarks/README.md.</sub>


6. Licenses, Upstream Sources & Advanced Variables

<details> <summary>Advanced Environment Variables &amp; Manual CLI Verification</summary>

| Variable | Default (Windows) | Purpose | | --- | --- | --- | | OMNI_ATTN_BACKEND | torch | torch (PyTorch SDPA), cute (CUTLASS-SYCL FMHA), auto, or esimd. | | OMNIXPU_PROVIDER_BOOTSTRAP | auto | Provider routing mode: auto, off, or required. | | OMNIXPU_XPU_MEMORY_FRACTION | 0.99 | Caching-allocator fraction applied on prestartup (0 / off disables). | | OMNIXPU_ENABLE | 1 | Set to 0 to disable all OmniXPU providers and patches. | | OMNIXPU_DEBUG / OMNIXPU_DEBUG_VERBOSE | 0 | Set to 1 for kernel or dispatch tracing. |

To inspect or rerun bundled wheel alignment from CLI (--no-deps + SHA-256 verified against SHA256SUMS.txt):

"<ComfyUI_Python>\python.exe" install.py --dry-run
"<ComfyUI_Python>\python.exe" install.py
</details>