Fix weights loading for Apertus

[nathanrchn]

Purpose

This PR fixes the weight loading mechanism for the Apertus models. The model uses a custom XIELU activation function that needs custom parameters for each layers. These parameters need to be stored alongside the model's weights. So we need to load the buffers too.

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 ask your reviewers to trigger select CI tests on top of fastcheck CI.

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 either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[gemini-code-assist]

Code Review

This pull request correctly addresses the need to load buffers for the XIELU activation function in the Apertus model. The implementation adds the necessary buffers to the params_dict for weight loading. However, the method used to identify these buffers by their name suffix is slightly brittle and could lead to issues in the future. I've suggested a more robust approach that directly targets XIELU modules, which will be more resilient to future changes in the model architecture.

[gemini-code-assist]

The current approach of identifying buffers to load by checking if their names end with .beta or .eps is a bit brittle. It could unintentionally match buffers from other parts of the model if they happen to use the same suffix. This could lead to incorrect weights being loaded or errors if shapes don't match.

A more robust approach would be to explicitly iterate over XIELU modules and add their buffers to params_dict. This ensures that you are only targeting the intended buffers and makes the code more resilient to future changes.

Suggested change

	for name, buffer in self.named_buffers():
	if name.endswith(".beta") or name.endswith(".eps"):
	params_dict[name] = buffer
	for module_name, module in self.named_modules():
	if isinstance(module, XIELU):
	for buffer_name, buffer in module.named_buffers(recurse=False):
	full_name = f'{module_name}.{buffer_name}'
	params_dict[full_name] = buffer

[malte-aws]

Hi,
What is the ETA for getting this merged?

[DarkLight1337]

Retrying the failing test