Add skip_post_init flag

[Aravind-11]

What does this PR do?

This PR adds a skip_post_init parameter to PreTrainedModel.from_pretrained() to allow users to skip the post-initialization step that reinitializes model weights. This is essential for users who subclass models with custom parameters and want to preserve their initialization.

When users subclass HuggingFace models (e.g., Qwen2VLForConditionalGeneration) and add custom parameters initialized in their __init__(), the initialization gets silently overwritten by post_init() which calls _init_weights() .

Fixes #42418

Before submitting

• 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? Let me know if additional tests are needed.

Who can review?

@ArthurZucker @Cyrilvallez

[ArthurZucker]

Ty tho I think we want to be a bit more careful in designing this!
We need tests that's for sure, but yes we need a good api to make sure users are not supprised.

For example -> based on the classes that are in the public init only.
cc @Cyrilvallez wdyt

[Cyrilvallez]

Hey @Aravind-11 ! The thing is that post_init is responsible for more than weight initialization, so we cannot simply skip it. Also, adding yet another flag is not a good idea IMO. Custom models should instead do

def _init_weights(self, module): ...

to skip weight initialization, instead of

self.skip_post_init = True

It's not more verbose, and avoid adding a flag and missing the rest of the post_init operations.

However, we could indeed think of something so that it's the default from remote code models. But in general we cannot be sure if people will rely on it or not, so a bit hard to decide. What's sure it that this used to be the default, so could make sense to go back to it for remote code

[Aravind-11]

That makes a lot of sense @Cyrilvallez ! Thank you for clarifying!!