[Qwen] Remove cuda hard-code in qwen3 next

[wxsIcey]

Purpose

[Qwen] Changed device specification from hard-coded CUDA devices to platform-independent device selectors

Test Plan

Test through the exsiting tests

---

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.

[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.

🚀

[gemini-code-assist]

Code Review

This pull request refactors the device specification in the Qwen3-Next model to be platform-independent. The change replaces a hard-coded CUDA device reference (torch.cuda.current_device()) with a platform-agnostic device selector (current_platform.current_device()). This is a good improvement for code portability across different hardware backends like CUDA, ROCm, and others supported by vLLM. The implementation is correct and follows the established pattern for platform abstraction in the codebase. I have no further comments.

[MengqingCao]

LGTM, this make it easier for oot platforms to support Qwen3 Next, thx!