[Bugfix] Fix Lora Name Parsing

[alex-jw-brooks]

Currently, we assume lora weights start with base_model.model. and split the prefix of lora weights accordingly. This is usually true, but in some situations, there may be lora adapters that don't have this prefix. An example of such a model is granite speech 3.3, where the audio specific lora weights are bundled directly with the model so that they can be automatically loaded together with from_pretrained using the transformers peft mixin, and the weight names align with the transformers (not peft) model tensor names.

This fix keeps the full prefix if the lora tensors don't start with base_model.model.. This is needed for #16246 to work correctly, since the lora is currently not applied due to erroneous slicing (e.g., language_model.model.layers.1.self_attn_q_proj vs layers.1.self_attn_q_proj)

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

🚀

[jeejeelee]

Thank you for your contribution, will look at this ASAP

[jeejeelee]

Thank you for your contribution, LGTM