Gemma 3 270M support, please

[msluszniak]

Discussed in #566

^{Originally posted by rmarioni August 28, 2025}
Google DeepMind juts dropped the small language model w/270M parameters which is very mobile/device friendly. Aware that Gemma 3 nano E1B also is availble but this model is still a big stretch for mobile apps. Kindly ask for some support on the RN bindings for Gemma 3 270M. This is going to be huge.

Has anyone started or looking into support for the model?

[mkopcins]

We had a look into this model, but this is mostly model intended for further finetuning usecases by users. AFAIK @NorbertKlockiewicz tried exporting this, but I am not completely sure about the results. Inference results were not amazing, but that is expected from 270m model where I think around 170m are the embeddings parameters and only 100m are used for actual inference

[msluszniak]

@mkopcins This issue is created from the discussion and I think you've missed quite a few messages. Click on a link that redirects to the actual discussion in #566. Here you'll see some additional information like that this model is exportable via executorch optimum. Also, I think that people who explicitly requested for this model in discussion section are aware of the inference results because they probably tried this model earlier and want to give it try via executorch.

[mkopcins]

I don't think torch ao kernels are needed for any exports, they are just quantized versions of some kernels and add some funtionality like shared embeddings. Currently included kernels should be able to handle those operations, but with possible quality degradarion.

[msluszniak]

Ok, so then there is no more blockers, I'm removing #609 as a blocker