HQH-XPU Info

Let's be honest about what this node is: HQH-XPU Info doesn't generate anything, doesn't process a pixel, and doesn't even check anything. It's a static label that ships at the root of the HQH-XPU pack, and its entire job is to sit on your canvas and say "the bundle is here, and here's what's in it." That sounds useless until you realize what HQH-XPU actually is - a bundle that deploys seven separate custom node packs at once, aimed at one very specific machine: an Intel Arc A770 on Windows running the XPU build of PyTorch.

If you own an Arc card, you already know the pain this pack answers. ComfyUI on Intel XPU is real but fiddly: oneAPI setup, torch-xpu builds from Intel's own index, and the classics the README names outright - OOM, DEVICE_LOST, driver kernel hangs on long video runs. HQH-XPU wraps a pile of patches and nodes (INT4 quantization, XPU kernel integration, a VRAM cache cleaner, a MiniMax H3 video pipeline) into one install so you can run 22B-class video models on 16GB of VRAM. This Info node is the front door of that bundle.

How it works

The mechanism is one line. The node declares no inputs at all, and its run() method returns a hardcoded string:

HQH-XPU bundle installed. Bundled nodes: TINT4, OmniXPU, XPU-CacheClean,
ClipProj (XPU), UniversalIO, H3-Latent-Upscaler, HQH-H3-Pipeline.

That's it. No introspection, no version check, no probe of whether the XPU kernels actually hooked in. Worth saying plainly: the name oversells it - "status" is really a roster. The string is baked in, so this node can't tell you if the bundle is broken, only that it loaded. If you want to confirm XPU acceleration is actually live, that's the OmniXPU diagnostic nodes, not this one. Treat this as a map, not a health check.

The inputs and outputs that matter

There are none to set. The required and optional input lists are both empty, so you drop the node on the canvas and it runs with zero wiring. Its single output is a status string (the roster above), and you don't need to connect it to anything - that's the point. Its best use is the same as any plumbing node from the ComfyUI-node-plumbing playbook: it makes the graph legible to whoever opens it next.

Installing it

This is where you need to slow down, because HQH-XPU isn't a normal clone-and-restart pack. The repo tracks source only; the actual node packs live in a nodes/ folder that gets deployed into custom_nodes/. Two paths:

Path A - ComfyUI Manager. Search for HQH-XPU in Manager and install. The pack ships a prestartup hook (install.py) that auto-copies the nodes/ sub-packs into custom_nodes/ on the next restart - no PowerShell involved.

Path B - the documented one (Windows):

git clone https://github.com/hqh330/HQH-XPU && cd HQH-XPU
powershell -ExecutionPolicy Bypass -File install.ps1 -ComfyRoot "<path-to-ComfyUI>"
# omit -ComfyRoot if the repo already sits in <ComfyUI>/custom_nodes/HQH-XPU

install.ps1 junction-links each sub-pack into custom_nodes/ - same folder, zero copying. Either way, restart ComfyUI afterward; the frontend extension then registers HQH-XPU as a search alias, so typing it in the canvas search box lists the whole bundle at once.

Common issues

The deploy is the thing that bites. Existing directories are never overwritten. If you already had ComfyUI-TINT4 or any sub-pack installed from upstream, the installer skips it - meaning you get the stock version, not the oneDNN-patched one the bundle depends on, and nothing tells you. If your Info node lists TINT4 but the TINT4 nodes behave stock, remove the old directory from custom_nodes/ first, then re-run install.

The other classic: the Info node appears but the sub-pack nodes don't show up in search at all. That means the deploy step failed or you skipped the restart. Check that custom_nodes/ actually has the junctioned/copied folders, restart, and hard-refresh the browser tab (the search-alias JS extension can be cached). And before you go further - this pack is XPU-only, aimed at an Arc A770 on Windows with the Intel PyTorch build. On an NVIDIA box, most of it is just dead weight; this Info node will load and tell you so cheerfully, but it's not for you.

Also worth knowing before you invest: the README lists honest driver-level limits - a single-graph A+B prompt causes a kernel hang, SDPA is Xe2-only, and TeaCache drifts on quantized models. This is a working snapshot of one person's machine, published source-only, not a polished upstream product. Use it for what it is: a fast path to a working Arc video setup, with the Info node as your confirmation that everything mounted.