mlx-community

/

Qwen3.8-27B-4bit

Like 192

Follow

MLX Community 17.6k

Hi @prince-canuma , I tested Raushan's PR and it works for me, with it you can load the processor in either mlx-vlm or transformers:

>>> from transformers import AutoProcessor
>>> p = AutoProcessor.from_pretrained("./qwen3.8-27b-4bit-raushan")
>>> type(p)
<class 'transformers.models.qwen3_vl.processing_qwen3_vl.Qwen3VLProcessor'>
>>> type(p.image_processor)
<class 'transformers.models.qwen2_vl.image_processing_qwen2_vl.Qwen2VLImageProcessor'>

>>> from mlx_vlm import load
>>> _, processor = load("./qwen3.8-27b-4bit-raushan")
>>> type(processor)
<class 'mlx_vlm.models.qwen3_vl.processing_qwen3_vl.Qwen3VLProcessor'>
>>> type(processor.image_processor)
<class 'mlx_vlm.models.qwen3_vl.processing_qwen3_vl.Qwen3VLImageProcessor'>
>>>

However, the original version in this repo is not compatible with transformers.

Note that Qwen3VLImageProcessor exists in mlx-vlm, of course, but not in transformers.

Do you think it's safe to merge this PR?

Thanks!