Support for ONNX export of SeamlessM4TModel

[AlArgente]

Feature request

Add SeamlessM4Tv2 Model support to onnx_export_from_model.

Motivation

Being able to deploy SeamlessM4Tv2 models to production using onnx.

Your contribution

I got the speech-to-text model to ONNX, but I'm not able to generate the audio as expected, even though I'm trying to give the tgt_lang_token_ids as decoder_input_ids. I could help with by submitting a PR, but I might start creating the model_config/model_patcher first if it is needed.

EDIT: I got the speech-to-text model, not the speech-to-speech model. I'd like to export the t2u_model and the vocoder to onnx, but it seems that is giving problems, any advice on how to do it?

[LDelPinoNT]

Hello!

I'm also interested in exporting SeamlessM4Tv2 to ONNX. I'm pretty new to ONNX so any clue or guideline with that is also appreciated.

[sgonzalezsilot]

I'm having exactly the same issue.

Seems like its not possible to export SeamlessM4TModel to ONNX or even export SeamlessM4Tv2ForSpeechToSpeech.