🧼 remove v4.44 deprecations

[gante]

What does this PR do?

See title :)

@zucchini-nlp -> video llava changes
@SunMarc -> all other changes

[gante]

(followed this instruction whenever shard_checkpoint was used)

[SunMarc]

Unfortunately, this is not the exact function. We can't just replace it directly. You need to build the shard and the index again using the output of this function. Check the example here: https://huggingface.co/docs/huggingface_hub/main/en/package_reference/serialization#huggingface_hub.split_torch_state_dict_into_shards.example

[zucchini-nlp]

Thanks for fixing but the processing logic refactor wen longer than I expected. We are still doing the last bits in #33424 and the checkpoints on the hub are not updated yet (tracker #33374). So we can't be raising ValueError yet

Maybe we can change the deprecation version to be after a few more minor releases. We have also same warnings in all llava and blip models afair. Also now I think we can even do more than just a few releases, since the changes will invalidate many models on the hub that are not owned by us. WDYT, which version we can put in the message?

[gante]

since the changes will invalidate many models on the hub that are not owned by us.

We should then support it for as long as possible :) If it is critical to any popular model, I would not even set a deprecation version target at all. Leave the exception without target, and set a target when it blocks you from doing something OR if it is causing many issues. An example of this is supporting model.config modifications to parameterize generate -- it is super inconvenient for maintenance, but many examples/models rely on it.

If it is only used in some niche models, and applying the deprecation (raising exception) is a significant win for transformers, then I would set at least 6 minor versions.

LMK which one is it, and I will update the warning.

[zucchini-nlp]

I hope you meant warning, not exception. I don't think we should be raising exception unless at least the official checkpoint is updated accordingly.

I would say this particular one is a niche model, given download stats from the hub. And most older llava and BLIP models may be outdated fast as we get newer and better releases each month. But I think we still can't enforce users to do something by elevating this to an Exception. My preference is to keep it for 6 minor versions as a warning, and try to update official models on the hub until that. We have write access to all of them afaik except for Video-LLaVA so updating will be straightforward

[SunMarc]

Thanks for the update ! I left a comment. Also, I check that we need to enforce huggingface-hub to 0.24.0 because of this commit. we accept max_shard_size with string format only from this version.

[SunMarc]

Unfortunately, this is not the exact function. We can't just replace it directly. You need to build the shard and the index again using the output of this function. Check the example here: https://huggingface.co/docs/huggingface_hub/main/en/package_reference/serialization#huggingface_hub.split_torch_state_dict_into_shards.example

[gante]

I hope you meant warning, not exception

@zucchini-nlp haha yes, warning! In essence, allowing the behavior.

[gante]

@zucchini-nlp @SunMarc PR comments addressed, LMK if they look good to you :)

[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.