Load Checkpoint (MultiGPU Debug)

Load Checkpoint (MultiGPU Debug) is the diagnostic twin of Load Checkpoint (MultiGPU) - same loader, same sharding logic, but with three extra tools bolted on: test_mode, log_vram_snapshot, and two extra outputs (test_image, status). This is the version the pack's USER_GUIDE tells you to use first, and for once that advice is right: it de-risks the setup before you point it at a 20GB+ model.

The smoke test

Flip test_mode on (or pick the test_mode entry in the ckpt_name dropdown - same thing) and the loader skips loading entirely. Instead it runs a synthetic test that creates a tensor on every selected GPU, verifies each tensor is actually resident on the card it claims, and checks basic allocation. All GPUs pass → you get a hot-pink "✓ MultiGPU Success!" image and a status line listing each card's memory usage. Any failure → a gray image naming the dead cards.

This is genuinely the right way to first-run a multi-GPU box. It tells you whether CUDA sees all your cards and whether they can all allocate memory before you burn five minutes loading a checkpoint that then dies on a bad card.

The VRAM snapshot

log_vram_snapshot appends per-GPU VRAM readings (in GB) before and after the shard layout is applied, straight into the status. That's the before/after evidence you want when a load says "multi-GPU active" but only one card's VRAM actually moved.

Outputs

model, clip, vae like the production loader, plus test_image (IMAGE - the pink/gray smoke-test result or a colored load-status image) and status (STRING - the full diagnostic report). In test mode you also get real dummy CLIP/VAE objects, so a downstream graph won't error even though nothing real was loaded.

Inputs

Same as the production loader (ckpt_name, num_gpus, gpu_ids, clip_name1/2, clip_loader_type, vae_name, safety_ok), plus test_mode and log_vram_snapshot. When safety_ok is false the loader refuses to run - that's the Hardware Validator gate doing its job.

Installing it

Part of Comfy-MultiGPU-Loader. 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.

The workflow the author actually wants you to run

Use the debug loader with log_vram_snapshot, pair it with the debug sampler, and feed statuses to a display node to get everything into comfyui.log. Once the graph behaves, swap to the non-debug loader for production. Remember the state of play: sunset project, GPL-3.0, as-is, no active support - and only Flux Dev Full fp32 was verified on multi-GPU, so treat this as "promising experiment," not a product.