[simone-chen]

Change

• Consolidate fix: adapt vLLM attention metadata to removed CPU fields #569 and fix: qualify vLLM 0.31 B200 attention collectors #587 for Tianhao's collector review alongside his data update PR. Preserve the original branches and campaign source SHAs.
• Adapt vLLM 0.31 attention metadata, pre/post MLA BMM, dense and encoder attention, MLA, DSA, and DSV4 to the pinned native serving APIs. Preserve older-version routes and explicit qualification gates.
• Retain runtime pins, backend selection, cache layouts, tests, and collector documentation. Leave campaign filtering, CI change detection, recovery infrastructure, and measurement data outside this PR.

Review order

• Review this attention PR first, then the recurrent/MoE compatibility PR fix: vLLM 0.31 recurrent and MoE compatibility #605 stacked on it.
• Communication qualification remains separate in fix: retain communication dispatch and failure evidence #591. Use Tianhao's data update data: 2026-10 campaign — vLLM 0.30.0 / TRT-LLM 1.3.0rc29 / SGLang 0.5.21 perf tables for 8 systems #517; this stack does not publish measurements or promote defaults. Its vLLM 0.30 measurements do not themselves qualify vLLM 0.31 timing reuse.

Validation

• Focused attention, registry, provenance, and routing checks: 308 passed, 24 skipped in the local environment. Ruff and formatting checks pass.
• The initial test run lacked the locally built native module; rerunning with the existing native build completed successfully.
• This consolidation does not rerun GPU qualification. Collection remains stopped. Full CI will be requested for the exact signed head.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[simone-chen]

/ok to test 2d0ed6c