[Bugfix] Fix for issue 17396

[frreiss]

This PR contains a patch for the bug reported in issue #17396. The LoRA weights embedding layers can have a different amount of padding than the weights of the base model, depending on what hardware the model and LoRA adapters are running on. The existing code in the bgmv_expand() kernel compensates for this difference when the LoRA weights have more padding, but it does not compensate for the difference when the LoRA weights have less padding. When LoRA adapters are used on CPU or MPS hardware with default parameters, the LoRA weights have less padding and the server crashes with a PyTorch tensor size mismatch error.

This patch updates the logic in bgmv_expand() so that it works regardless of which set of embeddings, LoRA or base, contain more padding dimensions.

FIX #17396

(anything written below this line will be removed by GitHub Actions)

[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.

🚀

[aldettinger]

Tested locally and the fix looks good to me. We are past this issue now, many thanks @frreiss 👍