HW Supporter (ANY passthrough)

ComfyUI has global hardware knobs that most people never touch: reserved VRAM, TF32, fp16 accumulation. They're global because they're torch backend settings, not per-node ones - but the normal way to set them is to edit config or use a dedicated node that needs MODEL/CLIP/VAE inputs. HwSupporterAny is the version that doesn't care what you feed it. ANY input in, the same value out, and while it passes through it applies the global tuning settings. It's the "just set up my machine" node you drop anywhere in a graph.

How it works

Unlike the main HW Supporter, this variant takes a single input of ANY type and applies global knobs without needing model wiring:

• Reserved VRAM - apply_reserved_vram + reserved_vram_gb (default 1.25GB) sets ComfyUI's reserved VRAM for model weights. reserved_vram_mode offers manual or auto_used_plus (used VRAM plus a headroom, capped by reserved_vram_auto_max_gb).
• fp16 accumulation and tf32 - each with auto/on/off. These flip the torch backend flags that affect matmul precision and speed; auto lets the node decide from your hardware (TF32 on Ampere+, for example).
• clean_gpu_before - purges GPU memory before applying settings.
• Profile - auto, 12GB_VRAM_32GB_RAM, low_vram, balanced, max_speed presets that bundle sensible defaults for common setups.

Outputs are output (your input, passed through untouched) and report_json describing what it applied. include_hardware_report appends a full hardware report to the JSON when you want to see the reasoning. console_log mirrors the action to the console.

Inputs that matter

• input - anything; the node's whole trick is that it works with any type. The output passes it through.
• profile - pick your tier, or leave auto and let it guess.
• reserved_vram_gb and the fp16_accumulation/tf32 toggles - the levers you'll actually tune if you're chasing OOMs or speed.

Install

Part of IAMCCS-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Or ComfyUI Manager → "IAMCCS" → install → restart. No extra dependencies.

Gotchas

Because these are global torch settings, the node has an effect everywhere in the session, not just in the subgraph around it. That's powerful and also a footgun: set tf32 on globally and every model in that session inherits it, whether or not that's what you wanted. Also, reserved-VRAM settings interact with ComfyUI's own low-VRAM logic - set it too high and models offload more aggressively than necessary; 1.25GB is a sane default, and the HW Probe node can give you a data-backed number. And remember the passthrough: it's the same value out, so don't expect it to transform anything.