Unofficial ComfyUI reference implementation of HSWQ (calibration + V1 FP8 conversion)

[Shiba-2-shiba]

以下、GPTが出してくれた英語ですが、ComfyUIで実装を試してみたカスタムノード案のリポジトリを作成しました。
Z-Imageにも使えるか分かりませんが、実装の参考になればと思います。

Hi, and thank you for releasing the HSWQ code and documentation.
I’ve been experimenting with your quantization method and found it very effective.

I recently implemented an unofficial reference integration of HSWQ for ComfyUI,
with the goal of making the original algorithm usable in practical SDXL workflows
without modifying its core logic.

What I implemented:

• A ComfyUI custom-node implementation of HSWQ:
  • a calibration node that collects HSWQ statistics during normal image generation
  • a conversion node that applies V1-compatible FP8 quantization
• The quantization logic itself follows your original design
  (layer sensitivity ranking + weighted histogram MSE for amax optimization)
• The focus is strictly on workflow-level integration
  (UI, persistence, safety), not on changing or extending the algorithm

Repository:
https://github.com/Shiba-2-shiba/ComfyUI-HSWQ-Quantizer

This is not intended as a replacement or fork, but purely as a practical
reference implementation for users who work inside ComfyUI.
If you find it useful, feel free to reference it (for example as a community
implementation), or extract any ideas that might be helpful upstream.

If you’d like, I can also isolate specific parts (such as calibration handling
or persistence) as minimal patches or design notes.

Thanks again for sharing your work.

[ussoewwin]

この度は、サンプラーとスケジューラーに関する極めて重要かつ有用なご提案を頂き、かつComfyUIにおける実装プラン迄提示して頂き、ありがとうございました。
Z Image Turboのfp8化については、壮絶に苦戦しましたが昨日何とか基礎理論部分は完成しました。後は、どこまでスコアを上げられるかの調整です。
Thank you for your extremely important and useful proposal regarding the sampler and scheduler, and for presenting an implementation plan. The transition of Z Image turbo to fp8 proved an immense struggle, but I managed to complete the fundamental theoretical part yesterday. The remaining task is to adjust how far we can improve the score.

[Shiba-2-shiba]

作成したComfyUIの実装案で検討したのですが、キャリブレーションについて、サンプラーとスケジューラーで比較したところ、量子化しないレイヤーに差が出ない結果でした。
また、キャリブレーション数についても、想定より早く収束しているようです。Note記事にそのあたりのデータを記載した記事を作成しています。
I ran some tests using my ComfyUI implementation draft. For calibration, comparing the sampler and scheduler didn't show any difference in the non-quantized layers.
It also looks like it converges faster than I thought, so we don't need as many samples. I'm writing up a Note article with the details.