Granite and GraniteMoE models.

[janimo]

Rouge scores are good but logit values are very off.

[merrymercy]

let us know when it is ready to be merged.

[janimo]

I still need to debug why the logits in the ONLY_RUN test go so much over the thresholds (e.g. 60 instead of << 1 like in others) even if the rogue scores match. You can merge if you wish and I can continue debugging or I'll update the PR when I find it. I am not sure whether implementation details should cause such large differences.

[frreiss]

@janimo sorry, just noticed that this PR exists. I submitted a similar PR yesterday with Granite integration code that came out of our work on a NeurIPs booth demo. My PR is at #1818, and it appears to have been merged already.

@merrymercy would you mind also including @janimo in the credits for Granite support on the release notes of the next version?

[janimo]

@frreiss no problem. Apparently (I haven't tested, I had no GPU access lately) my missing the need for a granite specific chat template could have been the cause of the large logit value differences. Possibly the MoE model could pass now too.

[merrymercy]

Sure. I will do that and close this PR for now.