Hardware Validator (MultiGPU)

This is the pack's own "safety rail," and it exists because the thing this pack does - sharding native fp32 safetensors checkpoints across GPUs - is exactly the kind of load that quietly eats a machine. The author's test case was Flux Dev 2 in FP32 at roughly 60GB. Before you ask six GPUs to swallow that, Hardware Validator (MultiGPU) checks your actual hardware against a profile and tells you, flatly, whether the load will fit. It's a MultiGPU/Safety node and it's the newest thing in the pack (added in 0.1.5).

How it works

It reads your real GPU inventory with torch.cuda.get_device_properties, sums total VRAM, checks per-GPU VRAM and GPU count, and compares against the selected profile. Four profiles are built in:

• generic - no enforced limits, passes anything with a GPU
• sdxl - 24GB total, 8GB min per GPU, 1+ GPU
• flux-dev - 32GB total, 10GB per GPU, 2+ GPUs
• flux-dev2-fp32 - 64GB total, 12GB per GPU, 6+ GPUs

The trap is hiding in plain sight: the default profile is flux-dev2-fp32, the most demanding one. On a 2×8GB rig that validator will fail by default - which is the point, but it reads as "node is broken" to a beginner. If you're running SDXL, switch the profile to sdxl.

The inputs that matter

• profile - which requirement set to check against (default flux-dev2-fp32, see above).
• num_gpus - how many GPUs to validate; Auto uses all available.
• gpu_ids - comma-separated allowed IDs (default 0,1,2,3,4,5,6,7); IDs not present on the machine get filtered out.
• allow_overcommit - when true, a failed check becomes a ⚠️ warning instead of a ❌, and ok still comes back true. Only flip this when you know what you're doing.
• status_image - returns a colored pass/fail image (blue = OK, orange = failed) so you can see the result at a glance.

What it outputs and how you wire it

Four outputs: gpu_ids (the filtered, validated list of GPU indices), status (a full text report per GPU), ok (BOOLEAN), and status_image (IMAGE). The intended wiring, straight from the README:

Hardware Validator
  gpu_ids  ──────────►  Loader.gpu_ids
  ok       ──────────►  Loader.safety_ok

The loaders (Load Checkpoint and Load Flux variants) accept a safety_ok input and abort the load if it's false - that's the gate. The gpu_ids string also tells the loader which cards to actually use, in order.

Installing it

Part of Comfy-MultiGPU-Loader: ComfyUI Manager search, or git clone https://github.com/AngelCookiesLab/Comfy-MultiGPU-Loader into custom_nodes/, pip install -r requirements.txt, restart. It shows up under MultiGPU/Safety.

Honest caveats

The author's own changelog admits the validator "has not been tested a lot so stuff might be broken," so treat a pass as a good sign, not a guarantee. The profiles are rough floor-guesses, not measured footprints - VRAM needs depend on resolution, batch, and the exact model. And remember this pack is sunset/"as-is" with no active support. Still, as a pre-flight check before a multi-GPU load, it's the right instinct, and it's the one node in this pack that can save you from a hard lockup.