Fix trainer and Qwen2-VL

[Blaizzy]

Closes #176
Closes #173

[qinxuye]

Any plan to release a new version? mlx-lm 0.21.0 has pinned mlx>=0.22.0.

Once mlx-vlm released a new version, we could loose the pin of mlx<0.22.0 in Xinference.

[Blaizzy]

Yes, there will be a new release this week :)

[Blaizzy]

Once mlx-vlm released a new version, we could loose the pin of mlx<0.22.0 in Xinference

Not sure I follow. Could you elaborate?

[Blaizzy]

I will pin v0.22.0 as well. There are some recent models and changes that require the latest MLX release.

[qinxuye]

Once mlx-vlm released a new version, we could loose the pin of mlx<0.22.0 in Xinference

Not sure I follow. Could you elaborate?

In which part do you need help?

[Blaizzy]

I need more context...

For instance, what is Xinference ?

[qinxuye]

I need more context...

For instance, what is Xinference ?

Oh, I misunderstood, https://github.com/xorbitsai/inference is our project, and we have adopted your project for the Mac VLM engine.

[qinxuye]

https://github.com/xorbitsai/inference/blob/7c6249a3383b2841a9e96243c6a900ce19a7f1d7/setup.cfg#L101

We have pinned mlx version, but the better option is that mlx-vlm can be compatible with mlx>=0.22.0

[Blaizzy]

Oh, I misunderstood, https://github.com/xorbitsai/inference is our project, and we have adopted your project for the Mac VLM engine.

No worries,

So awesome, really happy to see this!🔥

Sure thing, I will update the dependency to use >=0.22.0 on the next release

[qinxuye]

@Blaizzy when will the next release come? Our CI is broken, and really appreciate you can release new version as soon as possible.

[Blaizzy]

Hey,

It' done v0.1.11 is out :)