Feature request

I’ve noticed that many newly added models do not include a ForTokenClassification implementation. Is this due to fundamental challenges in implementation (though I don’t perceive any major obstacles—perhaps I’ve overlooked something), or is it simply a matter of development priorities and time constraints?

Motivation

I am currently testing a prototype based on the Mamba series models, which requires token classification outputs.

Your contribution

If it’s merely a time issue preventing the implementation of ForTokenClassification in transformers, I’d be more than willing to contribute by adding this feature for Mamba/Mamba2. If time allows, I’d also be happy to extend the support to other models.

From my understanding, replicating the approach used in LlamaForTokenClassification should suffice to implement token classification model for most models. Any advice or guidance would be highly appreciated!