pottokao/H3-RotNVFP4-ComfyUI-Loader

H3RotNVFP4Loader — ComfyUI loader for rotated-NVFP4 MiniMax-H3

The ComfyUI custom node that loads our rotated-NVFP4 MiniMax-H3 video+audio models — a single self-contained file that runs fully resident on one 16 GB Blackwell card, zero CPU offload, on the native nunchaku W4A4 fp4 kernel at bf16-class sharpness.

Weights (download the tier you want):

• MiniMax-H3-NVFP4-rotated — base (8-step, quality line)
• MiniMax-H3-FastH3-NVFP4-rotated — FastH3 (4-step, faster)

Why this node exists (short version)

MiniMax-H3 isn't supported by nunchaku's stock ComfyUI loader (Flux/Z-Image/Sana only), and its quality depends on a block-256 Hadamard rotation applied at runtime (it can't be folded through H3's timestep-dependent AdaLN). The config-only alternatives don't give both native speed and sharpness on Blackwell:

• plain NVFP4 (stock loader) = un-rotated → mushy;
• ComfyUI-native convrot_w4a4 = dequantizes to bf16 on Blackwell (no sm_120 kernel) and quantizes coarser → degrades.

So, like every other community H3 quant, this ships its own loader — the one that gets genuinely-native fp4 and rotation quality. Full reasoning + measurements in METHODOLOGY.md.

Nodes

node	use
H3RotNVFP4Loader	Load a single self-contained *.safetensors (base/fast tier) → MODEL. This is the one you want.
H3EmbedsTEFree	Optional: inject precomputed prompt embeds so the text encoder runs off-card.
H3RotNVFP4Patch	Legacy: patch a pre-loaded H3 model from an external pack + bf16 shell.

Install

Linux x86_64 + Blackwell GPU (RTX 50-series / GB10), CUDA-13 driver. Windows currently blocked (no nunchaku cu13 Windows wheel).

1. nunchaku (the fp4 kernel) — install the prebuilt wheel matching your python/torch/CUDA from nunchaku releases. Verified: torch 2.12.1+cu130 + nunchaku 1.3.0.dev20260306+cu13.0torch2.12 (see requirements.txt).
2. Copy this folder into ComfyUI/custom_nodes/h3_rotnvfp4/; restart ComfyUI.
3. Download a tier file from a weights repo above into ComfyUI/models/diffusion_models/, add H3RotNVFP4Loader, set ckpt_name. Log success marker: [rot] wrapped 200 linears.

A ready workflow.json ships in each weights repo.

Credits

MiniMax-H3 (MiniMaxAI) · nunchaku / SVDQuant (NVFP4 W4A4 kernel) · FastVideo (FastH3 distillation) · ComfyUI. Node © the author, Apache-2.0. Weights carry the MiniMax-H3 Community License (see the weights repos).