Dora_datacollector_updated

[shirinyamani]

This PR just updates the datacollector for fine-tuning using Dora

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Thanks for fixing this.

I think when creating the PR, something went wrong, as it shows 39 commits, even though there is only one change. But it's no big deal, I can still merge.

[shirinyamani]

Hi Benjamin @BenjaminBossan
Thanks for all the good you do. I just added the implementation SnapKV Cache paper to cache_utils.py file. To reflect the SnapKV approach, I added the implementation on llama_modeling under llama_snapkv.py for the user to see what changes has to be applied to flash_attention2 to reflect the snap_kv, however, I'm not entirely sure tht there is the best location for this implementation. I’m confident that the main SnapKV logic belongs in cache_utils.py, but I would really appreciate your insights on whether the llama_snapkv.py example is well-placed or if it should be integrated differently.
Much appreciated!

[BenjaminBossan]

@shirinyamani Thanks for working on this. I'm not a transformers maintainer, so I can't really tell you where best to implement this feature. It's best if you reach out to the transformers folks for this. What you could also do is create a draft PR with your implementation the way that you suggested, I'm sure they'll let you know if you have to change something.