[simone-chen]

Measurements

• Add 141,773 fresh B200 / vLLM 0.31.0 operation rows:113,557 GEMM,636 MLA BMM,3,256 paired quantization,10,332 GDN and13,992 MSA context measurements. Keep this PR draft and main-targeted; no defaults are promoted.
• GEMM includes37,518 BF16,36,408 block FP8,2,113 tensor FP8 and37,518 NVFP4 rows. Preserve1,110 explained historical block-FP8 gaps. The35,405 separately qualified FP8 reuse keys and deferred GLM shape expansion are not inserted as new measurements.
• BMM covers300 pre-BMM and336 post-BMM cases. Quantization covers1,628 paired callbacks, including four legitimate clipped-zero compute-scale estimates with positive static companions.
• GDN covers84 callbacks /5,166 paired points with zero failures; all four historical failed callbacks recovered. Fix the0.31 logical context alias to find its measured FlashInfer recurrence; preserve packed-decode normalization and existing guards.
• MSA continuation76851457 adds7,478 new cases with zero new failures and preserves6,514 measurements. Original17,568-case scope =13,992 measured +1,400 explained failures +2,176 source-blocked unattempted cases. Preserve per-case source groups184/6,330/7,478; aggregate finalizer metadata does not relabel seed measurements.
• Preserve immutable image/source/runtime pins, checkpoint and case inventories, raw hashes and failure evidence. The MSA Parquet is unchanged; its sidecar is marked partial. Only affected backend facts and reuse-manifest counters change; tiers and hardware/framework membership stay unchanged.

Validation

• Check exact identities, readable Parquets, provenance, duplicates and valid latencies. Preserve all prior accepted data byte-for-byte when adding each scope. Existing GEMM SOL floors affect39 BF16 rows; raw timings remain unchanged.
• Python/native consumers pass with shared-layer reuse disabled:113,557 GEMM,636 BMM,3,256 quantization identities,15,036 GDN queries and13,992 MSA queries per interface. Both interfaces share the compiled engine; retired legacy APIs are not claimed as independent validation.
• Retain bounded native eager/graph numerical probes, timing comparisons and raw review reports under benchmarks/evidence/vllm031-b200-refresh. Later compute-scale component probes do not replace missing components from the original full run.
• All seven collector-data rules and2,452 backend-fact slices pass. Latest MSA addition passes40 focused data/MSA tests, Ruff, exact consumer checks and independent review. Earlier GDN routing validation passed22 Rust tests,21 Python tests and389 parity tests with unchanged goldens.
• Auto collection and personal numerical checks used different host drivers. Sustained clock lock, full-grid arithmetic and end-to-end prediction accuracy are not established.
• Signed/DCO head 0794dc832f982193397174be2f55577096b216a6 is verified by GitHub. Full CI is requested for this exact head; its result remains pending.

Remaining work

• Other B200/vLLM campaign scopes remain open. Encoder measurements are not included in this commit. No merge or default promotion is authorized.
• Existing shared-tier resolution can use new data for gap filling under its current policy. This PR grants no new cross-version or cross-framework performance-equivalence claim.

[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 06ccfdc

[simone-chen]

/ok to test 7228466

[simone-chen]

/ok to test ddc5eb7

[simone-chen]

/ok to test ffd7c2c

[simone-chen]

/ok to test 65e61e4

[simone-chen]

/ok to test 4dbc51c

[simone-chen]

/ok to test 7530c32

[simone-chen]

/ok to test 8ab494f

[simone-chen]

/ok to test 0794dc8

[simone-chen]

Closing our data publication PR because the user selected Tianhao's #517 as the data update path. Retain only collector review PRs #604, #605, and #591. This does not merge these measurements or claim #517 covers vLLM 0.31. The branch, artifacts, provenance, and history remain available; collection remains stopped.