Update test requirements to CUDA 12.8

[22quinn]

This is a follow-up fix to #16859
By default, it will install torch with CUDA 12.6

[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.

🚀

[houseroad]

@huydhn could you take a look?

[houseroad]

Also lint?

[huydhn]

I don't think test.txt file can be updated like this directly. You will need to update test.in and let precommit does the job or generate the file manually with uv pip compile requirements.in -o requirements.txt --index-strategy unsafe-best-match

However, the bigger issue here is that while vLLM CI uses CUDA 12.8. PyTorch default is still 12.6, causing this discrepancy. I have been working around this by using uv pip install -r requirements/dev.txt --extra-index-url https://download.pytorch.org/whl/cu128 instead, i.e. https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile#L177

[22quinn]

@huydhn I tried using extra-index-url, but always got torch 2.7 cu126 installed. Also there were a bunch of CUDA dependencies in test.txt not matching CUDA 12.8, e.g. nvidia-cuda-runtime-cu12==12.6.77

I think need to modify the hook by adding --torch-backend cu128 unless there's a way to specify this in in file

[huydhn]

From what I read in their doc https://docs.astral.sh/uv/guides/integration/pytorch/#automatic-backend-selection, does --torch-backend auto work for you (I haven't tried that before)? If it does, it would look nicer as you don't need to hardcode cu128 in test.txt. Maybe we want to try it out. If it doesn't work, then the current revision LGTM

[22quinn]

discussed with Huy offline, this PR is needed as it unifies the version of other cuda dependencies like nvidia-cuda-runtime-cu12