ARM64 CUDA 12.9 wheels built and uploaded to index incorrectly

[Gregory-Pereira]

Fix: arm wheels tagged incorrectly

Related to: https://vllm-dev.slack.com/archives/C07R5PAL2L9/p1760721161081999

Issue

The vLLM wheels index contains both arm64 and amd64 wheels for CUDA, however the wheel for arm does not contain the cu<CUDA_VERSION> identifier, while the amd64 one is.

Cause

I believe this is happening because in the arm wheel pipeline we pass VLLM_MAIN_CUDA_VERSION as 12.9 which it understands to be the same Major and Minor version as CUDA_VERSION of 12.9.1 for this pipeline - as a result it omits the cu129 identifier from the build. This should fix our release pipeline to ensure wheels match the index generated for them and are consistent with the x86_64 ones.

cc @wseaton @simon-mo @khluu

[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.

🚀