[REQUEST]: request group scaled fp8 models for qwen3

[aabbccddwasd]

Has this been supported or requested before?

• I have checked the GitHub README.
• I have checked the Qwen documentation.
• I have checked the documentation of the related framework.
• I have searched the issues and there is not a similar one.

What is this feature about?

a new quantized model

Proposal

Introduction

I would like that ...

Rational

Implementation of this feature will help the following usecase:
in this vllm-project/vllm#17280 they said they supported fp8 on sm120, and they use this model for test https://huggingface.co/RedHatAI/Qwen2.5-3B-FP8-dynamic. but few days later someone says Qwen3-14B-FP8( https://huggingface.co/Qwen/Qwen3-14B-FP8 ) can't run on rtx pro 6000, after a few studies I found vllm only support group scaled fp8 but not block scaled fp8. but unfortunately all fp8 model provided by Qwen is block scaled. so I hope Qwen will provide group scaled models for deployment on sm120. thanks a lot!!!

Contributions are welcomed

• I am willing to help implement this feature.