[REQUEST] Gemma 3n support

[StephennFernandes]

would be great to have support on inference and fine tuning the Gemma 3n series of models for both vision and audio tasks. (optionally a demo on multimodal fine tuning)

similar to: https://github.com/huggingface/huggingface-gemma-recipes/tree/main/scripts

[shralex]

We currently support SFT for multi-modal Gemma3 and Llama4 in MaxText.

To help us correctly prioritize the request, could you please include information about the use-case and any other relevant info you can share. Thank you!

[StephennFernandes]

Hi @shralex thanks a ton for considering this.

I am looking into ways to finetune the Gemma3n to possibly push the multilingual and multimodal capabiliites of gemma 3n

I am looking for ways to have gemma3n to be finetuned a mixture of multimodal tasks speech images texts, in some cases even finetune it where each input sample consists of audio + images and text and generate output tokens.

I am hoping that I could fientune gemma3n to have ASR, AST, audio language detection, text to text translation, OCR, OCR with translation for multiliple languages.

I also hope to generalize cross modal abilities in gemma 3n with given input images and audio file along with text instructions.