handle weight sharing with init_on_device

[aws-rhsoln]

What does this PR do?

This PR handles the case when we are using meta_device and have tied weights. During model init, when we have a case, where weight2 = weight1 , this would result in register_param getting called again. This call will re-init the param instead of using weight1. This will result in relationship between the weights getting lost.

Fixes # (issue)

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.

[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.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[SunMarc]

Hi @aws-rhsoln, thanks for the PR ! Could you add a few tests ? Also, could you explain in which context you needed that ?

[SunMarc]

Hi @aws-rhsoln, are you still planning to finish the PR ? Thanks !

[aws-rhsoln]

I'll add some tests soon. Context for the change: For our pipeline parallel implementation here, we first create the model and then trace it, use the traced model to partition for PP. Now we create the model using meta-device and then copy the weights into device after PP init. When copying weights, if we have shared weights such that self.output_layer.weight = self.embedding.weight , the init_on_device api would first initialize embedding weight and then when its about to initialize output.layer weight, since its the embedding weight, its already on device and hence would end up throwing an error. Hence to fix it, this PR was introduced.

[SunMarc]

Awesome ! Thanks for the update !