GPU Status Display

The whole pitch of Comfy-MultiGPU-Loader is that your second GPU should stop being a paperweight. So the first question you'll ask is the one every multi-GPU setup dies on: is the second card actually doing anything? GPU Status Display is the pack's answer. It's an output node in MultiGPU/Diagnostics that reads every CUDA device and turns it into a text report plus a colored bar-chart image.

The honest part up front: this is a pure diagnostics node, and it's the one piece of this pack I'd genuinely reach for even if I'd given up on the rest of the sharding. It costs nothing, works with zero inputs, and it's the fastest way to see whether a sharded load actually spread across cards or quietly fell back to one.

How it works

When it runs, the node queries torch.cuda for every GPU, grabs used vs. total VRAM per card (mem_get_info), and prints a line per GPU plus a total. Then it draws that into a 512×256 image - one bar per GPU, filled green proportional to usage, labeled with percentage and GB. That status_image is a normal IMAGE, so you wire it into a Preview Image node to see it.

The second half is what makes it multi-GPU-aware. If you feed it the model output from this pack's loader, it reads a _multi_gpu_device_map attribute that the loader stamps onto the model and prints the shard layout: which block lives on which device (input_blocks.0 → cuda:0, etc.). That's the "is the sharding real" check.

Inputs and outputs

Everything is optional, so the node works naked:

• model - plug in the MODEL output of a loader to get the shard-layout report.
• status - a STRING you can feed in to override the auto-generated status text.
• layout - a multiline STRING appended to the report as the shard layout.

Outputs are status_text (STRING - hook it into a Show Text / console logger) and status_image (IMAGE - preview it). It's flagged as an output node, so it runs at the end of the graph like a Preview.

Installing it

The node ships inside the Comfy-MultiGPU-Loader pack. Easiest path is ComfyUI Manager (search "Comfy-MultiGPU-Loader"); otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/AngelCookiesLab/Comfy-MultiGPU-Loader
pip install -r requirements.txt   # just accelerate; torch is already there

Restart ComfyUI and it appears under MultiGPU/Diagnostics. Fair warning: this project is explicitly sunset/"as-is" - no active support, use at your own risk, GPL-3.0.

Where people get burned

The layout report only shows up when the model came from this pack's loader. Feed it a model from a stock Checkpoint Loader and you just get VRAM bars - which is still useful, but don't go hunting for a bug that isn't there. Also, VRAM "used" includes everything else on the box, so a card at 90% might be ComfyUI, not your workflow. And one known issue the author flags: logging/diagnostics can under-report on later runs. If the bars look wrong, that's a known quirk, not you.