[Dong1017]

Governing Issue and delivery slice

Refs #297

Downstream engine contribution

vLLM-Omni Draft PR #9745
publishes only the C1 BF16 prefix-attention implementation, tests and recipe
against main: base b84bade9b9be8fdad0a1a9b3e8dd3ab4273dc4d9, head
7a20cf9b987b22fe3ce00221cb30cc3c6fb90e63. Its new-base focused native suite
passed 12 tests with actual Ascend fused attention, source binding and cleanup.
Portable native record. New-base full-model 5 s/15 s quality and
clean HTTP A/B are not run. Historical C1 strict/latency results below retain
their original revision; the 50% campaign target remains unmet.

Publish the requested tracking snapshot for MiniMax-H3 Ref2VA on
four logical Ascend 910C devices, covering 5 s/15 s × lossless/lossy.
Campaign verdict: INCONCLUSIVE. No accepted 50% complete-HTTP reduction.
This is the issue owner's evidence delivery slice; final goal acceptance remains
with the requester.

The structure follows PR #294:
frozen plans, source attribution, results, rejected attempts, validators,
integrity hashes and replay limits. Its eight-device 910B1 campaign and accepted
result remain separate from this four-logical-device 910C campaign.

• Accepted scope / delegation: the requester authorized the isolated
  campaign and this tracking PR. The target remains at least 50% matched-B0
  complete-HTTP reduction for each track, strict bit-exact lossless checks and
  frozen automatic lossy gates without human media inspection.
• What this PR delivers: 96 campaign files, a four-track status/report,
  C0–C23 direction ledger, 28 automatic quality comparisons, 40 C12–C23 run
  projections, experimental engine diffs and source identities, eight unchanged
  frozen snapshots, plus inventory registration and count assertions.
• What it does not deliver: a selected fast quality-qualified configuration,
  final four-track acceptance, a production engine contribution or a
  self-contained device replay environment.

Current results

Evidence cutoff: 2026-11-02 15:52 +08.

Track	Numerical / quality evidence	Clean complete-HTTP evidence	Target status
5 s lossless	C1 direct-B0 strict 439 PASS; C13 fixed-request strict 439 PASS	One C1 screening pair: 291.247763→286.666760 s, 1.572888689% lower	Performance FAIL at 50%; final confirmation incomplete
15 s lossless	C1 direct-B0 strict 439 PASS	Not run	INCONCLUSIVE
5 s lossy	Tested fast profiles have observed frozen quality failures	No qualified candidate's clean pair	No accepted candidate
15 s lossy	Tested profiles have independent visual/audio failures at this duration	No qualified candidate's clean pair	No accepted candidate

The clean 5 s pair used two warmups and three measured complete MP4 responses
per fresh server, with audit/profiling disabled. Source integrity, cleanup and
strict paired checks passed. It is one screening pair, not the final three-pair
AB/BA/AB protocol. The delta includes migration+C1; no isolated migration-only
clean timing arm exists. Statistical significance is not claimed.

Recent completed diagnostic 5s-00 requests:

Candidate	Audited HTTP	Frozen quality disposition
C21 pre-gate activation cache	163.806 s	FAIL: 7 observed violations
C22 sigma-secant pre-gate cache	166.859 s	FAIL: 9 observed violations
C23 fresh conditioning / cached video context	199.065 s	FAIL: 8 observed violations

Each passed 134 input checks and cleanup; none proceeds to 15 s or clean timing.
C23 V2 also passed 44 native tests, including three executed four-rank suites,
and reduced its static 50-call attention cycle by 66.46%. That primitive
measurement excludes FFN, changing-input quality, complete-model timing and HTTP;
the complete model's quality failure prevents promotion.

Changed artifacts and evidence

Requirement	Artifact	Evidence and remaining gap
Workload, revisions and budget	Execution checkpoint, plans	Original model/50 steps, isolated stack, current requester ceiling24 directions/32 h/128 logical-device-hours; original start/deadline retained
Independent quality contracts	Lossless, lossy, frozen hashes	Exact validator/scorer bytes preserved; quality corpus and heldout execution incomplete
Candidate results and origins	Report and direction ledger, run projections, quality comparisons	Includes negative/inconclusive outcomes, actual dispatch counters, C13 baseline-reproduced regression failure and retained C21/C22 controller-alias repair
Engine sources	Patch notes and identities	Experimental patches, including the chunked-DLO adaptation of upstream PR9351; private engine commits remain unmerged and are not adopted production support
Availability and integrity	Replay notes, publication digests	Available bytes are verifiable; original weights/image/media/tensors/logs and private bootstrap/runtime remain external

Baseline engine: df1c1b05936381dbaa3891c1ae4c4b21ab375a5d.
Migration source: 922f1e9f1cb85b9f2b66d86e1eb500e6872b2dcb.
Executed InferRSI: 50c19da43e0b358e6f990c00719891ad308a8ac1.
Latest rejected engine candidate C23: 6eb7bb094effb40853431208d1631effaf452a3f.
The publication commit is not a new experiment runtime revision.

Checks and dependencies

• python tools/validate_evidence.py: PASS, 23 bundles / 533 hash checks,
  zero errors. Its 18 external/unavailable records across the inventory remain
  explicit; successful byte checks do not replay devices or change verdicts.
• ruff check src tests tools: PASS.
• pytest tests/test_evidence_inventory.py tests/test_trial_checks.py tests/test_smoke.py tests/test_claim.py:
  47 passed, 1 skipped at 70500b3d. The symlink-escape fixture skipped because symlink
  creation is unavailable in this Windows environment.
• All 96 staged evidence files matched their working bytes; eight original frozen
  snapshots remained exact. Relative Markdown links and the pending artifact
  scan passed. Private host identifiers, paths, credentials, raw media and weights
  are excluded. These checks do not substitute for independent review.
• Device results are carried from their pinned executions; no model/device run
  was started for this publication. Original MindIE-SD source/binaries, shared
  checkouts and global SDK/driver environments were not changed.
• Hosted checks apply to this PR head. Full repository/hardware qualification
  and independent review are not claimed by the local focused tests.

Knowledge authorship declaration

Not applicable under the current protected-path policy: this changes campaign
evidence, its inventory entry and inventory count assertions only.

Agent-assisted-by: OpenAI Codex (author)
Agent-assisted-by: Claude Code / deepseek-v4.1-flash-1017, claude-opus-5-5 (author)

Remaining work and Issue closure

All 24 direction families are charged. The original active deadline is
2026-11-03 01:19 +08, with no extra final-confirmation budget. Further new
mechanism search requires a requester extension of the direction ceiling;
this tracking PR grants no execution budget or new device run.

Four-track final confirmation, full lossy quality coverage and independent review
remain incomplete. The observed failed configurations stay FAIL; missing samples
are not PASS. Goal acceptance belongs to the requester and is not claimed.
Keep the acceptance Issue open.

[Dong1017]

Static evidence review at a0677a03: one historical-validator limitation should be documented before reuse. The reported single-pair 1.57% result, rejected candidates and incomplete qualification remain appropriately bounded; this is not campaign acceptance or an unconditional LGTM.

• Please preserve frozen validator bytes/receipts and document the mixed FAIL/missing-evidence precedence limitation below. A future qualified version should preserve observed FAIL while also reporting missing evidence.
• Reuse the existing cross-PR inventory finding when integrating both bundles; it remains unresolved.
• Software CI passed, but no device/model execution or independent raw-artifact replay was performed in this review. New-base full-model and HTTP qualification remain NOT_RUN.

Inspected evidence

Head: a0677a0; base: 2baebf6. Static inspection covered archived validators/scorers, frozen contracts, principal reports/projections, inventory and upstream receipt. The 28 quality projections retain 24 FAIL / 4 INCONCLUSIVE; 40 run projections agree with the report. Current-head software run: https://github.com/InferMatrixLab/InferRSI/actions/runs/41418369337 . No hardware efficacy inferred.

[amy-why-3459]

Review at head a0677a03 (draft): not approving yet.

1. Merge conflict: the PR is currently not mergeable with main (dirty). Please rebase.
2. Open validator finding (evidence/h3-ref2va-910c-4logical-20261101-297/validator/strict_compare.py vs validator/README.md:19): an observed cleanup failure or tensor mismatch combined with a missing rank configuration yields INCONCLUSIVE, while the README says observed gate violations remain FAIL. No commit has addressed or documented this since the review on 2026-11-02. Expected: document this as a known historical-validator limitation in the README (or fix precedence if the validator is not frozen).
3. Size: +27,509 lines / 98 files of evidence; please confirm this fits the evidence byte-budget discussion in [Investigation] Agent-led workflow gaps: closure, decision visibility, review provenance and focus #273 or note which artifacts should move to external storage.

[amy-why-3459]

Re-review at head a0677a03a7fc26ac0625924328997bba1ba159ed (range 2baebf6ebb815ebf07d21c41e3b3754858c0798e...a0677a03a7fc26ac0625924328997bba1ba159ed). Draft; not approving.

Prior short-SHA note still applies; no new commits since then. Concrete blockers unchanged:

1. Merge conflict / dirty: mergeable_state=dirty against current main (1246348f252dbc7e473efb1752c2d81da15e23fd). Please rebase onto current main.
2. Validator precedence vs README (evidence/h3-ref2va-910c-4logical-20261101-297/validator/strict_compare.py): verdict = 'INCONCLUSIVE' if missing else 'FAIL' if failures else 'PASS' — any missing evidence forces INCONCLUSIVE even when failures is non-empty. That contradicts validator/README.md ("an observed gate violation remains FAIL even when other samples are missing"). Document this as a known historical-validator limitation, or fix precedence if the validator is not frozen.
3. Size / budget: still +27,509 / 98 files of evidence; please confirm fit against the evidence byte-budget discussion in [Investigation] Agent-led workflow gaps: closure, decision visibility, review provenance and focus #273 or note which artifacts should move to external storage.

Reviewed commit: a0677a03a7fc26ac0625924328997bba1ba159ed.

[amy-why-3459]

增量 review：上次看的是 a0677a03a7fc26ac0625924328997bba1ba159ed，这次是 rebase 到 main 9f61a83f030bfa2ffe3b54eac1781593309d6380 之后的 head 4684e82c2423b8b2123f4dd6f1d1005ff62b4457。新增的提交只有 4684e82c2423b8b2123f4dd6f1d1005ff62b4457（evidence: document the frozen validator verdict precedence）。PR 仍是 draft，这次先不 Approve。

上次意见的处理情况

1. 合并冲突：已解决，现在 mergeable=true。
2. 校验器判定优先级：已按建议写成已知限制。我核对了 validator/strict_compare.py:109 确实是 verdict = 'INCONCLUSIVE' if missing else 'FAIL' if failures else 'PASS'，这个文件的字节没有变（sha256 仍是 b83f2bff…）。新 README 的 sha256 9136399b… 也和 results/publication-integrity.json 里的记录一致。“本 bundle 中没有同时含 failures 和 missing 的判定”这句，我抽查了 results/c1-15s-strict-comparison.json、c1-5s-strict-comparison.json、c13-5s-strict-comparison.json，三个都是 PASS，failures 和 missing_evidence 都为空，和描述一致。
3. 证据体积：还没有回应。仍是 98 个文件。请对照 [Investigation] Agent-led workflow gaps: closure, decision visibility, review provenance and focus #273 的字节预算确认一下，或者列出哪些产物应该挪到外部存储。

新提交引入的问题

4. 换行符被改成了 CRLF：evidence/h3-ref2va-910c-4logical-20261101-297/validator/README.md（25 行）和 results/publication-integrity.json（97 行）在 a0677a03 里都是 LF，这次每一行都变成了 CRLF，所以 diff 里是整文件重写。publication-integrity.json 实际只有 validator/README.md 那一条 hash 变了。因为 .gitattributes 对 /evidence/** 设了 -text，这些 CR 会原样进入归档字节，而同目录其他文件（比如 strict_compare.py）都是 LF。这应该是编辑器无意中改的。期望修改：把这两个文件恢复成 LF，然后按 LF 字节重新计算 validator/README.md 的 hash，让 diff 只剩新增的那几行说明和一条 hash 变化。如果有其他清单记录了 publication-integrity.json 自身的 hash，也请一并确认。

CI：我看的时候 full-check (3.10) 还在跑，其余检查都已通过。

[amy-why-3459]

Incremental review of 4684e82c2423b8b2123f4dd6f1d1005ff62b4457..979c9301cd8b3ab392e108472fd985177c93facf (1 commit).

Thanks for the quick fix. Verified:

• evidence/h3-ref2va-910c-4logical-20261101-297/validator/README.md and results/publication-integrity.json are LF again (no CR bytes in either file at 979c930).
• Compared with the pre-PR base a0677a03, the README diff is now only the limitation paragraph (lines 19-25), and publication-integrity.json differs only in line 95. The new hash 7588ed18…e135 matches the sha256 of the committed README bytes.
• All checks are green on this head (full-check 3.10/3.11, installed-package 3.10/3.11, pr-references, declarations).

One item is still open from my earlier comment: the size/budget question raised in #273 hasn't been answered in this thread yet. Could you confirm whether that's resolved, or link where it was decided? Once it's clarified I'm happy to approve.

[Dong1017]

Answering the size question at head 979c9301cd8b3ab392e108472fd985177c93facf (base 9f61a83f).

Size against the #273 R6 budget

Measured over blobs in the PR tree, in the units the R6 decision defines them (bytes under evidence/, not diff lines):

Measure	Value	R6 threshold	Result
Net addition under evidence/	1,258,808 B (1.200 MiB) — 96 new files, 0 removals, 0 modified	2.0 MiB per PR	under
Largest file, patch/c23-candidate.patch	110,042 B (107.5 KiB)	512 KiB per file	under
Next two largest	101,925 B (99.5 KiB), 99,044 B (96.7 KiB)	512 KiB per file	under

So +27,514 / 98 files is diff lines and file count; the byte figures above are the ones the budget is stated in. No artifact needs to move to external storage under the adopted thresholds. Reproduce with git ls-tree -r --long 979c9301 evidence/.

Where it was decided, and what is not yet enforced

• Thresholds and scope: owner decision on R6, 2026-11-04 — 512 KiB per file, 2.0 MiB net addition per PR.
• Status: adopted as reported thresholds, not a merge gate. Enforcement review is deferred to on or after 2026-12-02.
• Nothing enforces or reports it on main yet. tools/check_evidence_budget.py exists only in #328, which is open and dirty, and no repository document on main states the two numbers — only the Issue comment does. The figures above were measured locally, not read from a check run.

Also confirmed from your last point

publication-integrity.json is not hashed by any manifest: validator/frozen-snapshots.json pins 8 paths (2 contracts, 4 validator .py, 2 result JSONs), while tools/evidence_inventory.json declares publication-integrity.json as a bundle-sha256 manifest, i.e. as a hash source. Same relocation pattern as the evidence/indexer-255-* bundles.

Re-verified at this head

tools.validate_evidence.validate_inventory → 23 bundles, 533 checked hashes, 0 errors, 18 unavailable (external sources, unchanged). pytest tests/test_evidence_inventory.py tests/test_trial_checks.py → 40 passed, 1 skipped.

The PR is still draft; nothing here is an approval.

[zengchuang-hw]

PR review verdict: DO NOT PASS

Non-voting evidence-publication review. Reviewed base 9f61a83f030bfa2ffe3b54eac1781593309d6380 and head 979c9301cd8b3ab392e108472fd985177c93facf.

P2: Correct the reconstruction chain for C17–C20.

plan/reproduction.md:38–41 and patch/README.md:33–36 describe all C12/C13/C15–C23 patches as cumulative over C1 and tell readers to apply C1 followed by exactly one candidate patch.

However, the C17, C18, C19 and C20 source-identity records each declare patch_base=e89b1d40d37f7f72767892a56fa7980ea572ba8e, the recorded C13 head. Each corresponding candidate patch contains only two H3 implementation/test paths and omits C13's ten-path DLO/collective changes. Following the documented recipe therefore omits part of the recorded tested source configuration, even if the selected patch applies.

Please distinguish these C13-based candidates from the C1-based cumulative patches. Document the migration → C1 → C13 → selected C17/C18/C19/C20 chain, follow each other candidate's recorded patch base, and refresh the affected publication hashes. Preserve the historical patch and frozen-validator bytes. Re-review should verify the corrected chains against the identity records and patch contents, then rerun inventory/hash checks.

Earlier findings

• The historical strict-validator missing-first limitation is now explicit, with frozen bytes unchanged. The original inline thread remains unresolved.
• The two previously reported CRLF files now contain no CR bytes; the validator README hash matches its manifest.
• The 96 added evidence files total 1,258,808 bytes, with a largest file of 110,042 bytes. These are below the report-only R6 thresholds.

Independent checks

• All 98 changed files from the fixed-head archive matched their Git blob identities.
• Inventory validation: 23 bundles, 533 hash checks, zero errors, with 18 unavailable dependencies explicitly retained.
• All eight frozen snapshots and all 95 publication-manifest entries matched.
• pytest tests/test_evidence_inventory.py tests/test_trial_checks.py tests/test_smoke.py tests/test_claim.py: 48 passed, zero skipped.
• ruff check --no-cache src tests tools: passed.
• All 13 check runs returned for this exact head were completed/success.
• The retained projections contain 24 FAIL and four INCONCLUSIVE quality comparisons. The clean screening samples reproduce the stated 1.572888689% reduction, not the 50% campaign goal.

I did not run the complete local test suite, reconstruct/apply the external engine trees, replay private raw artifacts, or run NPU/model/media tests. This review covers the evidence publication and its reconstruction instructions; it does not certify device measurements or production correctness of every experimental patch. Campaign acceptance remains INCONCLUSIVE, and #297 should remain open.

[zengchuang-hw]

LGTM

PR review verdict: PASS

Incremental re-review of 979c9301cd8b3ab392e108472fd985177c93facf..70500b3d68a3158f63be7083e206342a3b5abdaa (one commit). Reviewed base 9f61a83f030bfa2ffe3b54eac1781593309d6380 and exact head 70500b3d68a3158f63be7083e206342a3b5abdaa. This supersedes the reconstruction-chain blocker in my previous review.

Resolved

• Both replay documents now distinguish the C1-based candidates from C17–C20, whose recorded base is C13 e89b1d40d37f7f72767892a56fa7980ea572ba8e. The documented migration → C1 → C13 → one selected C17/C18/C19/C20 chain agrees with the source identities and patch contents. I checked every C12/C13/C15–C23 identity; C13 carries the ten DLO/collective paths absent from each two-path C17–C20 patch.
• Only patch/README.md, plan/reproduction.md, and their two publication-manifest entries changed. All 93 other campaign files, including patches, identities and frozen validators/results, are byte-identical to the previously reviewed head. The three changed files retain LF endings.
• The earlier frozen-validator limitation remains explicitly documented. Its original inline thread is still open, but the requested annotation is present; it is not an outstanding defect in this publication.

Independent verification

• Isolated review tree verified against all 2,793 Git blobs at this head.
• Inventory validation: 23 bundles, 533 hash checks, zero errors; the same 18 unavailable external dependencies remain explicit.
• All 95 publication-manifest entries and all eight frozen snapshots match.
• pytest tests/test_evidence_inventory.py tests/test_trial_checks.py tests/test_smoke.py tests/test_claim.py: 48 passed, zero skipped.
• ruff check --no-cache src tests tools: passed.
• All six check runs for this exact head completed successfully, including full-check and installed-package on Python 3.10/3.11. Required knowledge-model-policy status is also successful. The Python 3.11 full-check log records 3,582 passed, 14 skipped and 286 subtests passed.

No blocking correction remains for the evidence-publication scope. Nonblocking cleanup: the owner can resolve the addressed historical inline thread and refresh stale Draft/head wording in the tracking descriptions.

I did not locally rerun the complete repository suite, reconstruct/apply the external engine trees, replay private raw artifacts, or run NPU/model/media tests. This approval covers publication of the bounded historical evidence and corrected reconstruction instructions. It does not certify experimental patches for production or accept the campaign: the 50% target remains unmet, campaign acceptance remains INCONCLUSIVE, and #297 remains open.

Independent reviewer: OpenAI Codex acting through zengchuang-hw, under the direct review request recorded on this PR and the account owner's standing delegation to review, reply and approve qualifying changes. No authoring, merging, deployment or campaign execution was performed.

[amy-why-3459]

LGTM

Incremental review of 979c9301cd8b3ab392e108472fd985177c93facf..70500b3d68a3158f63be7083e206342a3b5abdaa (1 commit: correct the C17–C20 reconstruction chain).

Verified: the corrected chain's hashes match the recorded artifacts, the patch bases match their identities, no CR bytes were reintroduced, and the size question against the #273 R6 budget was answered in the thread. All checks (pr-references, declarations, full-check 3.10/3.11, installed-package 3.10/3.11) are green on this head.

[Dong1017]

@zengchuang-hw @amy-why-3459 The C17–C20 P2 is fixed at 70500b3d68a3158f63be7083e206342a3b5abdaa, one commit over 979c9301. Please re-review.

Fix for the reconstruction chain (review)

patch/README.md and plan/reproduction.md now group the candidates by their recorded patch_base:

Candidates	patch_base	Chain
C12, C13, C15, C16, C21, C22, C23	C1 be319692	922f1e9f → C1 → one selected patch
C17, C18, C19, C20	C13 e89b1d40	922f1e9f → C1 → C13 → one selected patch

In results/publication-integrity.json, only the hashes of those two files changed: d3ebf520… → 477741c1… and 5afacd7b… → 02e769c6…. Patch bytes, identity bytes and the eight frozen snapshots are unchanged.

How the chains were checked

This was a static check of the published patch and identity bytes:

• In all eleven C12–C23 identities, patch_base is be319692, except C17–C20, which record e89b1d40.
• C17–C20 each change two paths: flash_h3.py and one test. The flash_h3.py pre-image blob 6ebfb10 equals C1's post-image, and C13 changes neither path. The order C1 → C13 → C17–C20 is therefore consistent at blob level. None of C17–C20 contains any of C13's ten paths.
• C21, C22 and C23 already contain all ten C13 paths, with post-image blobs identical to C13's. C1 plus one patch is therefore complete for them.
• Not done: applying each chain to a 922f1e9f checkout and comparing the result with the identity tree.

Re-run at 70500b3d

• python tools/validate_evidence.py: 23 bundles, 533 hash checks, 0 errors, 18 unavailable.
• pytest tests/test_evidence_inventory.py tests/test_trial_checks.py tests/test_smoke.py tests/test_claim.py: 47 passed, 1 skipped. The skipped test is the symlink fixture, because this Windows host cannot create symlinks. Your Linux run had 48 passed.
• ruff check --no-cache src tests tools: passed.

Size question (amy's open item)

The full answer is in this earlier comment. At this head the 96 bundle files total 1,259,483 B (1.201 MiB); this commit adds 675 B. The largest file is unchanged at 110,042 B. Both figures are under the report-only R6 thresholds: 2.0 MiB per PR and 512 KiB per file.

Other updates

• The inline thread on strict_compare.py:109 is answered and resolved. validator/README.md lines 19–25 document the limitation, and the frozen bytes are unchanged.
• The PR body is refreshed: inventory counts, the pytest scope, links to this head, and the outdated Draft wording.