Flux Loader Debug (All-In-One)

Flux Loader Debug (All-In-One) is the diagnostic version of the pack's headliner - Flux Loader (All-In-One) - with the same Flux defaults (auto-detected clip_l/t5 text encoders, ae.safetensors VAE, CLIP on CPU, dual-block Flux transformer sharded across your GPUs) plus the debug toolkit. And since the README openly calls the Flux loaders "early/less tested," this is the version you should actually start with.

What the debug mode adds

The test_mode toggle (or the test_mode entry in the ckpt_name dropdown) runs the pack's synthetic smoke test instead of loading real models. It creates a tensor on each selected GPU, confirms each tensor is genuinely resident on the card it claims, and checks memory allocation. All cards respond → a hot-pink "✓ MultiGPU Success!" test_image; any dead card → a gray image naming the failure. You also get dummy CLIP/VAE objects so a test graph can run end to end without a real model.

Why do this before a Flux load specifically? Because the checkpoint you're about to load is huge, and there's no better way to learn "GPU 2 can't allocate" than from a 30-second synthetic test instead of a five-minute load that dies at 90%.

log_vram_snapshot is the other big one: it records per-GPU VRAM usage in GB before and after the shard layout is applied, straight into the status string. When a Flux load claims multi-GPU but only card 0's VRAM moved, that snapshot is your proof.

Inputs and outputs

Same inputs as the production Flux loader - ckpt_name, num_gpus, gpu_ids, clip_name1/clip_name2, clip_loader_type (fixed flux), vae_name, safety_ok - plus test_mode and log_vram_snapshot. Outputs are model, clip, vae, test_image (IMAGE), and status (STRING with the full diagnostic report).

Model files

Same set as the production loader: Flux checkpoint in models/checkpoints/, clip_l.safetensors + t5xxl_fp8_e4m3fn_scaled.safetensors in models/text_encoders/, ae.safetensors in models/vae/. Missing pieces get called out, not silently skipped.

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, look under MultiGPU/Debug.

Reality check

Keep expectations calibrated: the author's verified win is Flux Dev Full fp32 (~22GB) across 4×RTX 3070s over PCIe, and Flux Dev 2 (~60GB) is still unverified. The project is sunset - as-is, GPL-3.0, no active support, diagnostics that can under-report on later runs. This node's job is to tell you whether your cards are up to that bet before you commit the disk time and VRAM to finding out.