specimba's picture
|
download
raw
2.06 kB
metadata
license: other
license_name: minimax-h3-community-license-agreement
license_link: >-
  https://huggingface.co/alibaba-pai/MiniMax-H3-Fun-Controlnet-Union/blob/main/LICENSE
base_model:
  - alibaba-pai/MiniMax-H3-Fun-Controlnet-Union
tags:
  - minimax-h3
  - controlnet
  - comfyui
  - w4a8
  - quantized

MiniMax-H3 Fun ControlNet-Union, pruned, W4A8

W4A8 quantization of the MiniMax-H3 Fun ControlNet-Union model patch for ComfyUI.

File minimax_h3_fun_controlnet_union_pruned_w4a8.safetensors
Size 1.45 GB (bf16 pruned build: 4.22 GB, int8_convrot build: 2.3 GB)
Source alibaba-pai/MiniMax-H3-Fun-Controlnet-Union, via the pruned bf16 build in Comfy-Org/MiniMax-H3
sha256 fe822947667d7625e0c622edc583952628f9ccf749a7444af11b2eb9747f3bdc

Quantization

  • Format: asym_w4a8_int8 (4-bit asymmetric weights, int8 activations), ConvRot rotation with group size 256, weight group size 16.
  • Applied to the 20 2D projection layers of the 5 control blocks. Norms, embeddings, biases and the AdaLN basis stay in bf16/fp32.
  • Calibration-free (reference format). Quantization metadata is stored in the safetensors header (_quantization_metadata, comfy_wxa8), the layout ComfyUI reads natively.

Usage (ComfyUI)

  1. Put the file in models/model_patches/.
  2. Load it with ModelPatchLoader, apply with MiniMaxH3FunControlNetApply (see the bundled template video_minimax_h3_fun_controlnet_union).
  3. Requires ComfyUI 0.35.0 or newer. This file carries native quantization metadata; the Comfy-Org int8_convrot build uses the legacy scale-tensor format and needs Comfy-Org/ComfyUI#16221 on 0.35.0.

Notes

  • Loads and builds all 5 control blocks. Output quality relative to the bf16 and int8_convrot builds has not been measured.
  • License follows the source model (MiniMax-H3 Community License Agreement).

Xet Storage Details

Size:
2.06 kB
·
Xet hash:
980f6688c49bb52cdce1d4a99b0c18033446044768efb18662848a3e235f4802

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.