Support fine-tuning Qwen3.5 models in qwen-vl-finetune

[yossibiton]

I'm trying to fine-tune Qwen3.5 models (e.g. Qwen/Qwen3.5-9B) using qwen-vl-finetune.

Model loading works via AutoModelForImageTextToText.from_pretrained() (resolves to Qwen3_5ForConditionalGeneration). However, the data processor computes position IDs using get_rope_index_3 when model_type="qwen3vl".

However, Qwen3.5 has a different architecture (Gated DeltaNet + partial RoPE with partial_rotary_factor=0.25) compared to Qwen3-VL's 3D mRoPE with temporal/spatial sections. Is get_rope_index_3 correct for Qwen3.5, or does it need a dedicated position ID function?

[dmhdmhdmh]

I also ran into the same problem. Did you manage to solve it later?

[yossibiton]

Yes, I went into huggingface source code (transformers) and took it from there.
I'm still not 100% sure it's correct, since the rope index implementation on huggingface for qwen3 was slightly different then this repo.

[dmhdmhdmh]

May I ask if fine-tuning the Qwen3.5-VL model with qwen-vl-finetune yields good results? Can it run without issues?

[yossibiton]

I still don't have answers for that

[dmhdmhdmh]

OK~thank you, I will also try it out