[MTP][Runtime] Reuse draft attention metadata across MTP steps

[voipmonitor]

Summary

This stacks on top of the model-side GLM patch in upstream PR vllm-project#39550.

It avoids rebuilding draft attention metadata on every multi-step MTP iteration by reusing and updating metadata in-place for builders that explicitly support drafting updates.

Files:

• vllm/v1/spec_decode/eagle.py
• vllm/model_executor/layers/attention/mla_attention.py
• vllm/v1/attention/backends/mla/indexer.py

Tracking issue:

• [SM120][GLM-5.1] NVFP4 DCP/MTP stack tracker vllm-project/vllm#37113

Dependency:

• Depends on upstream PR [GLM-5] Fix loading of appended NextN/MTP draft layers vllm-project/vllm#39550 for the GLM model-side MTP path.
• This PR is intentionally opened as a stacked PR against glm5-nextn-model-parity in the fork so the diff stays runtime-only.

Validation

Long single-request benchmark (ignore_eos, max_tokens=4096, global_warmup_requests=2, temperature=0):

• Unpatched vLLM GLM-5.1 MTP=4 + probabilistic: 114.75 tok/s
• Hotfix runtime image GLM-5.1 MTP=5 + probabilistic: 139.76 tok/s median
• Git-backed rebuilt image from this branch, after full warm state: 156.67 tok/s median
• SGLang GLM-5.1 reference on the same long benchmark: 138.82 tok/s median

Notes:

• The first full benchmark run immediately after server startup is not a stable throughput reference for this path; after startup, one long warm run should be discarded before comparing medians.
• This patch is specific to the MLA/indexer drafting path. It is not yet evidence of the same gain for Qwen3.5 MTP.

[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.

🚀