Hunyuan 3 GPU Info

Hunyuan 3 is the most memory-hungry image model most people will ever run in ComfyUI, and when it fails, the first question is always "where did my VRAM go?" Hunyuan 3 GPU Info (class HunyuanImage3GPUInfo) is the pack's answer: a zero-config diagnostic that dumps your GPU state into a single string so you can actually see what's happening.

It's not a workflow node in the generation sense - you don't wire it into a pipeline, and it doesn't affect output. It's a debugging instrument, the same way a multimeter isn't part of your stereo. You add it to a graph when something's wrong, read the report, and delete it. That's its entire job, and it does it well.

How it works

It queries the CUDA runtime directly - total and free VRAM, GPU name, and (the part that matters on big rigs) detection of every GPU, not just GPU 0. Multi-GPU setups are where this earns its keep: if your second card isn't being used, this node is the first thing the README's troubleshooting section tells you to run, because it reveals exactly what torch.cuda sees.

The inputs and outputs

Zero inputs. One output:

• gpu_info (STRING) - the full report. It's an output node, so just drop it in and read the string on the node after queueing.

How to install it

Part of the Comfy_HunyuanImage3 pack:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
cd Comfy_HunyuanImage3
pip install -r requirements.txt

Restart ComfyUI and hard-refresh the browser. Manager users: search HunyuanImage in Install Custom Nodes.

Common issues & troubleshooting

The README's multi-GPU section is basically the user manual for this node. If ComfyUI only sees one of your GPUs:

1. Run Hunyuan 3 GPU Info and read what it reports.
2. Check the CUDA_VISIBLE_DEVICES environment variable - it's the most common way a card gets hidden.
3. Confirm torch.cuda.device_count() matches your hardware.

# Clear any GPU visibility restrictions, then restart ComfyUI
unset CUDA_VISIBLE_DEVICES

If the report shows a card you didn't expect, that's usually a driver or CUDA_VISIBLE_DEVICES leftover, not a node problem. And one honest caveat: GPU Info reports the CUDA-visible state at execution time. If another tab has grabbed VRAM, this node will show you the total and free numbers - but it won't tell you which model ate it. For that, the Force Unload node's memory_report output is the companion tool.

For a memory tool that looks the other direction - system RAM, where Hunyuan spills its offloaded weights - the pack's Hunyuan 3 RAM Diagnostic is the one to reach for. GPU Info covers the card; RAM Diagnostic covers the desk it's sitting on.