Add support for Gemma 3?

[alexgusevski]

As far as I can tell all Gemma 3 models are multimodal except maybe 1B ones? Not sure but it says all of them on their HF.

[alexgusevski]

Okay saw now that you already support it! Nice!

The 1B one seems to not be able to be converted through this mlx_vlm library, but it is not able to be converted with mlx_lm library either right now.

[Blaizzy]

I already made a PR to MLX-LM to support 1B.

ml-explore/mlx-examples#1336

[vlbosch]

I already made a PR to MLX-LM to support 1B.

ml-explore/mlx-examples#1336

You're very quick! Great work!

Would that PR also work for text only use of larger models, like 27B?

[Blaizzy]

Yes, it will :)

[Blaizzy]

Thanks!

[vlbosch]

I just tried the converted model in "--chat"-mode, but as response to a text-only query I get only "< pad >" as output

[psm-2]

@Blaizzy The 12b model does not seem to work for me, it just outputs a lot of nonsense... But, when you do add the compatibility, would it be possible to add fine-tuning?

[alexgusevski]

@psm-2 Same, this is what I get when running the 4B model, got it from mlx community, havent tried other versions of it.

Oh and seems @vlbosch also gets this

[BlackLuny]

@psm-2 Same, this is what I get when running the 4B model, got it from mlx community, havent tried other versions of it.

Oh and seems [@vlbosch](https://github.com/vlbosch) also gets this

same on 27B 8bit

[alexgusevski]

More people having same issue as above: lmstudio-ai/lmstudio-bug-tracker#513

[FL33TW00D]

This should be fixed!