Use CUDA 12.6 as default for release and nightly wheels

[huydhn]

This is a follow-up of #16859 as PyTorch is now released on CUDA 12.8, 12.6 (default), and 11.8 https://pytorch.org/get-started/locally.

To be rebased after #16859 merges.

[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.

🚀

[cliffwoolley]

Moving up is good, but we should move to 12.8 rather than to 12.6, because 12.8 is where Blackwell support is added.

[huydhn]

Moving up is good, but we should move to 12.8 rather than to 12.6, because 12.8 is where Blackwell support is added.

Yeah, I will move this to 12.8. Per comment from @simon-mo, I have added this as part of https://github.com/vllm-project/vllm/pull/16859/files#diff-60f61ab7a8d1910d86d9fda2261620314edcae5894d5aaa236b821c7256badd7L57 so that they can be landed together

[cliffwoolley]

Yeah, I will move this to 12.8. Per comment from @simon-mo, I have added this as part of https://github.com/vllm-project/vllm/pull/16859/files#diff-60f61ab7a8d1910d86d9fda2261620314edcae5894d5aaa236b821c7256badd7L57 so that they can be landed together

Thanks. Though with https://github.com/vllm-project/vllm/pull/16859/files#diff-60f61ab7a8d1910d86d9fda2261620314edcae5894d5aaa236b821c7256badd7L57 , shouldn't https://github.com/vllm-project/vllm/pull/16859/files#diff-7269b1b148c1b82edce5234e9cefabaaa80c1f2f6599db2d007856d9582b1e92 have also been moved forward to 12.8 package versions?

[huydhn]

Thanks. Though with https://github.com/vllm-project/vllm/pull/16859/files#diff-60f61ab7a8d1910d86d9fda2261620314edcae5894d5aaa236b821c7256badd7L57 , shouldn't https://github.com/vllm-project/vllm/pull/16859/files#diff-7269b1b148c1b82edce5234e9cefabaaa80c1f2f6599db2d007856d9582b1e92 have also been moved forward to 12.8 package versions?

The problem here is that I couldn't hard code --extra-index-url https://download.pytorch.org/whl/cu128 to test.txt because that file is also used by 12.6 and 11.8 (#16859 (comment)). So, CI is currently using a command like pip install -r requirements/test.txt --extra-index-url https://download.pytorch.org/whl/cu[128,126,118] depending on the CUDA_VERSION.

I think the underlying issue here is that while PyTorch default is 12.6, vLLM needs to get the 12.8 build from https://download.pytorch.org/whl/cu128