qwen 3.5 vision finetuning

[dcrey7]

Hi Abdur,

I’m trying the Qwen3.5 vision example from mlx-tune on March 13, 2026 on a Mac M4 Pro.

mlx_vlm imports fine, but mlx_tune.vlm still sets HAS_MLX_VLM = False.

My env:

• mlx-tune==0.4.2
• mlx-vlm==0.4.0

What I found:

• mlx_tune.vlm imports from mlx_vlm.trainer.trainer import Trainer, Dataset, save_adapter
• but in my installed mlx_vlm, mlx_vlm.trainer has datasets, lora, orpo_trainer, sft_trainer, utils and no trainer module

Which exact versions / commit SHAs of mlx-tune and mlx-vlm are known to work for the Qwen3.5 vision finetuning example?

[ARahim3]

Hi @dcrey7, thanks for the report!

You're right — mlx-vlm 0.4.0 restructured their trainer subpackage. The mlx_vlm.trainer.trainer module was removed and split into sft_trainer.py and datasets.py. Our import silently catches the ImportError and sets HAS_MLX_VLM = False, which is why it appears to not detect mlx-vlm even though it's installed.

We developed and tested against mlx-vlm 0.3.12 (which was the latest at the time), and didn't catch the breaking API change when 0.4.0 came out. Fixed in mlx-tune v0.4.3 — now supports both mlx-vlm 0.3.x and 0.4.0+.

uv pip install --upgrade mlx-tune
# or: pip install --upgrade mlx-tune

Let me know if you hit any other issues!