[Misc] Add uninitialized params tracking for AutoWeightsLoader

[Isotr0py]

Currently, model have uninitialized parameters won't raise any error even a warning message. This makes some bugs like #10297 (comment) and #9422 (comment) caused by incorrect weights loading/model initialization difficult to be found.

• Add feature to track uninitialized weights for AutoWeightsLoader.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Isotr0py.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[DarkLight1337]

Thanks for the QoL improvement!

[DarkLight1337]

Hmm. not sure whether we're being too strict or is the weight loader buggy?

[Isotr0py]

Hmmm, seems that strict=True is too strict for sharded safetensors, because in that case the weights of a module will be sharded into multiple files. It will trigger strict check then raise error while the module is partially loaded from one of the shared files...

Perhaps we should add the loading check on top of DefaultModelLoader in loader.py instead of AutoWeightsLoader.

[DarkLight1337]

Agreed, let's move the weight loading check to a higher level.

[DarkLight1337]

Since we may skip loading some layers in vision models / PP case, we should update the check to error out when weights inside the initialized model aren't loaded, rather than when some weights from the checkpoint aren't loaded.