HW Probe Recommendations (JSON)

"How much VRAM do I actually have, and what should I set my tile sizes and reserved VRAM to?" That's the question every long video workflow asks, and most people answer it by trial and error - usually after the third OOM. HWProbeRecommendations turns the pack's hardware probe into a node: it inspects your machine, computes a profile, and hands back concrete numbers you can wire into VAE decode and HW supporter nodes. It's the "auto" button for settings that most people leave on defaults.

How it works

It wraps the pack's recommend_settings() logic, which does the probing:

• VRAM detection - tries pynvml first for total/used VRAM, falls back to torch.cuda.mem_get_info() if pynvml isn't available. If neither works, it degrades gracefully to an "auto" profile instead of crashing.
• Profile selection - conservative tiers: ≤8.5GB VRAM → low_vram, ≤16.5GB → balanced, above → max_speed. You'll see which one you got in the profile output.
• TF32 heuristic - on Ampere and newer (compute capability 8.x+), it recommends TF32 auto; older cards get off.
• Reserved VRAM - estimated as used VRAM + headroom, mirroring how ComfyUI's reserved-VRAM setting behaves, with an optional cap.
• VAE decode recommendations - tile size, overlap, temporal size, temporal overlap - tuned to your VRAM tier for long video VAE work.

You give it the width, height, frames, and fps of your job (0 means "unknown, don't factor in"). Outputs are report_json (the full recommendation document), profile, reserved_vram_effective_gb, and the four VAE numbers (vae_tile_size, vae_overlap, vae_temporal_size, vae_temporal_overlap). The whole point is that those seven outputs plug directly into other IAMCCS HW nodes.

Inputs that matter

• width / height / frames / fps - describe your job so the recommendations are sized to it; zeros skip those factors.
• pretty - pretty-print the JSON report; keep it on for humans, off for downstream parsing.

Install

Part of IAMCCS-nodes:

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

Or ComfyUI Manager → "IAMCCS" → install → restart. pynvml is optional - without it the probe falls back to torch's numbers, so no hard dependencies.

Gotchas

The recommendations are heuristics - they'll get you into the right ballpark, not perfection. If you have other apps using VRAM while you probe, the numbers reflect that moment's state, so probe when your GPU is otherwise idle. And the reserved-VRAM math is "used + headroom," which means it can be high right after a big generation; run it clean or the recommendations will over-reserve. It's a starting point you're meant to tune from, and honestly the report JSON is worth reading once just to see what your card is actually doing.