Deal with shared memory scenarios

[muellerzr]

What does this PR do?

We need to deal with shared memory scenarios when saving with safetensors and a model. Since the safetensor version relies on passing the entire model instead of the state dict directly, see here, to deal with this, for now a raw copy/paste of the code is done instead to remove duplicate names.

(All state_dict in torch are OrderedDict, hence the check)

Fixes # (issue)
Two failures on the transformers CI

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@SunMarc

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[SunMarc]

LGTM ! I know that in the save_model in accelerate and in save_pretrained in transformers, we remove the shared tensors in a different way. I don't know how comparable it is to this method in safetensors but my guess it that it is pretty similar.

[muellerzr]

Good call, I’ll refactor this into a function so that it only warns, doesn’t actually raise an err :)

I think logically it should work the same, bar the bnb explicit part

[SunMarc]

Thx for iterating ! LGTM !