[feat] Support MRoPE + YaRN

[JJJYmmm]

Support MRoPE + YaRN for Qwen3-VL, tested with internal ckpt. 🙏

Enable YaRN with modified config:

...
    "max_position_embeddings": 1000000,
    "rope_scaling": {
      "rope_type": "yarn",
      "mrope_section": [
        24,
        20,
        20
      ],
      "mrope_interleaved": true,
      "factor": 3.0, # MRoPE has a slower factor compared to vanilla rope
      "original_max_position_embeddings": 256000
    },
...

[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 adds support for MRoPE with YaRN scaling. The changes in vllm/model_executor/layers/rotary_embedding/__init__.py correctly dispatch to MRotaryEmbedding when YaRN scaling is used with mrope_section. The changes in vllm/model_executor/layers/rotary_embedding/mrope.py add YaRN parameters and reuse logic from YaRNScalingRotaryEmbedding. However, I've found a critical issue in how MRotaryEmbedding interacts with the YaRN scaling logic, which will lead to incorrect rotary embeddings. Please see my detailed comment.

[ywang96]

Thanks for the contribution! This looks reasonable to me!