KSampler (MultiGPU Debug)

Straight answer: KSamplerMultiGPUDebug is a subclass of KSampler (MultiGPU) with the category flipped to MultiGPU/Debug. Same inputs, same outputs, same sampling engine. The "Debug" in the name is a menu placement, not extra features - the base node already does all the shard reporting. So why does this variant exist? Because the pack's USER_GUIDE deliberately points you at the debug nodes for first runs, and keeping the whole diagnostic chain in one menu makes it easier to build that graph.

It's the node to use when your goal is figuring out why a multi-GPU load isn't spreading. Everything KSamplerMultiGPU reports - shard layout, per-GPU block counts, VRAM deltas across the run, peak VRAM, idle-GPU warnings - comes out of the status STRING output. Wire that into a display node (the guide suggests feeding it into a GPU Status Display) and the full report lands in comfyui.log after each run.

Same inputs, same defaults

All the standard KSampler fields: model, seed, steps (20), cfg (8 - remember to drop to 1 for Flux), sampler_name (all of ComfyUI's samplers), scheduler (all 9), positive, negative, latent_image, denoise (1.0). Outputs: latent and status.

A realistic debugging loop looks like this:

1. Load with a debug loader, log_vram_snapshot on.
2. Sample with this node.
3. Read status - did every GPU execute blocks? Any zero-usage warnings? Where did peak VRAM land?
4. Fix the GPU selection, re-run, compare.

Installing it

Ships with Comfy-MultiGPU-Loader. Via ComfyUI Manager (search "Comfy-MultiGPU-Loader") or:

cd ComfyUI/custom_nodes
git clone https://github.com/AngelCookiesLab/Comfy-MultiGPU-Loader
pip install -r requirements.txt

Restart, find it under MultiGPU/Debug.

One caveat to carry

The author flags a known issue: diagnostics can under-report activity on later runs. So when you're iterating on the same loaded model, a "zero executions" warning on run three might be the logger lying, not your sharding collapsing. If a warning looks wrong, reload the model before you trust it. And remember the project is sunset - as-is, GPL-3.0, no active support.