Support for Qwen3-VL and up-to-date MLLM compression documentation

[FDH21]

I'm trying to apply auto-round to Qwen3-VL (or similar multimodal LLMs), but I noticed that the current documentation under auto_round/compressors/mllm/README.md seems outdated or incomplete.

Could you please:
Clarify whether Qwen3-VL (or Qwen-VL series) is supported or planned for support?
Update the MLLM README with current instructions, including how to add custom multimodal models?
Provide a minimal working example for quantizing a Qwen-VL model if possible?

[wenhuach21]

we have released quantized qwen3-vl model in intel hugging face space. You could refer to the model card to reproduce it.

The major issue is we have to modify the model architecture due to some specific constraints of this model so it could only inference on transformers but not vllm. we plan to fix it in the next release as shown in another issue.

Type on the phone, sorry for not attaching the related links

[FDH21]

I have fine-tuned my own version of Qwen3-vl-4B, how can I apply AutoRound to my own SFT-ed Qwen3 model?

[wenhuach21]

It’s not related to whether the model is fine-tuned. I think no change is needed for the 4B version, but I’m not completely sure. @n1ck-guo Could you take a look?

[n1ck-guo]

@FDH21 Thank you for your attention. As you mentioned, our readme is a bit outdated. We will update it soon.
You can use this command auto_round --model your_qwen3_vl_4B_model --iters 200 --output_dir tmp_dir to quantize your model. If there is any bugs or errors, please show me your logs and I will help to fix it.
Also, you can use auto_round --help to show the more details of using auto_round.

[FDH21]

What calibration datasets are currently supported for MLLMs, and can I use my own custom calibration dataset

[wenhuach21]

We currently support LLaVA for text–image paired datasets and refer here for text-only datasets. Based on our prior experiments, when only the language model is quantized, a text-only dataset works well