update to cuda 12.9

[youkaichao]

Purpose

PyTorch 2.8 is only available for cuda 12.9 on arm64 platforms

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.

[gemini-code-assist]

Code Review

This pull request updates the CUDA version from 12.8.1 to 12.9.1 across the build configurations. The changes in .buildkite/release-pipeline.yaml and the CUDA_VERSION argument in docker/Dockerfile are consistent with this goal. However, I've identified a critical issue in the docker/Dockerfile where a hardcoded PyTorch version for arm64 builds is still pointing to a CUDA 12.8-specific build, which will likely cause the arm64 build to fail with the updated CUDA version. This needs to be addressed.

[gemini-code-assist]

While updating CUDA_VERSION to 12.9.1 is the goal of this PR, this change will likely break the arm64 build due to a hardcoded PyTorch version string.

In this file at line 349, the PyTorch version for arm64 is hardcoded as torch==2.8.0.dev20250318+cu128. The +cu128 suffix indicates it's built for CUDA 12.8.

With CUDA_VERSION set to 12.9.1, the build script will correctly look for packages in the CUDA 12.9 index (.../cu129), but it will try to install a package for CUDA 12.8, which will likely fail.

To fix this, the hardcoded torch version string needs to be updated to a version compatible with CUDA 12.9 (e.g., ...dev...+cu129). The torchvision version on the same line might also need to be updated accordingly. This is a critical issue that needs to be addressed to ensure the arm64 build succeeds.

Since this part of the file is not in the diff, I cannot provide a direct code suggestion, but the line to change is:

349:            "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319" ; \

[youkaichao]

The arm build is broken https://buildkite.com/vllm/release/builds/7808/steps/canvas?sid=01990371-cf21-43a1-bc1f-7b51b716a488 by the #20358 pytorch 2.8 update, because it finds a cpu built pytorch.

[Isotr0py]

Let's see if release pipeline can turn green then.

[nWEIdia]

it finds a cpu built pytorch.

Thank you! That explains why the build suddenly required numa.
So in this sense, installing libnuma-dev is not the right fix (in #23960), right?

[youkaichao]

So in this sense, installing libnuma-dev is not the right fix (in #23960), right?

I think so. Switching to the cuda 12.9 build looks more reasonable.

[youkaichao]

close as #23960 covers more aspects