[brockhargreaves-amd]

Motivation

Pinned GPU reference bundles are currently keyed by (operation, architecture): each pair has one baseline lock, one DVC pointer, and one archive. As more kernel authors add pinned coverage, this creates two problems:

• PR collisions. Two independent PRs that touch different kernel implementations on the same architecture both rewrite the same lock and pointer, so they always conflict and the second must requalify everything.
• One baseline per pair. Enrolling a new kernel family that does not exist at the pair's baseline revision forces re-anchoring every existing case on that architecture.

This PR keys references by (operation, architecture, variant), where a variant is a kernel implementation (for example convolution implicit_gemm_fwd, SDPA dense and wmma). Each variant owns its cases, baseline revision, lock, and archive, so independent kernel work lands in disjoint files and can pin its own baseline.

Stacked on #13651.

Technical Details

Layout (paths under rocke/library/tests/):

• Adapters and locks: <op>_reference/architectures/<arch>/<variant>/{__init__.py,baseline_lock.json}. <arch>/__init__.py only names the processor.
• Archives and DVC pointers: reference_bundles/<op>/<arch>/<variant>.tar.gz(.dvc), with one root .gitignore for archives and extracted bundles.
• Installed bundles: engines/test_arch_content/rocke/<op>/<arch>/<variant>/. The architecture directory stays the first processor-named level, so per-target packaging is unchanged.
• Initial variants: conv gfx942/gfx950 implicit_gemm_fwd; SDPA gfx942/gfx950 dense; SDPA gfx1151/gfx1200/gfx1201 wmma (gfx1200 enrolled, unpublished).

Migration without requalification:

• The six published bundles move byte for byte. Each .dvc pointer changes only its path: line, with the same md5 and size, so no new DVC objects are needed.
• Qualified manifests are untouched. Legacy manifests without a variant field are identified by their lock's path.

Commits:

1. Replay every published frozen runner against the current verification request (and one with an extra key), with a host stand-in runtime. This is a safety net for the request protocol.
2. Pin a cohort_sha256 digest of each qualified cohort in its baseline lock. Configuration checks it against the manifest, and a host test checks it against the adapter's CASES, so a cohort edit without requalification fails in every tree, even without bundle data.
3. Move the shared architecture registry into reference_common.
4. Key adapters and locks by variant. Replay resolves only the architecture; the source side resolves the variant adapter.
5. Store, stage, and install bundles per variant. CTest names stay one per (op, arch), running all its published variants with the timeout scaled by variant count.
6. Fail configuration on half-wired or pre-variant references: a lock without a pointer or adapter, an old-layout lock or pointer, an invalid variant name, or a nested ignore file. Host checks cover variant naming and adapter/directory consistency.
7. --variant for qualify/verify/cases/diff and --rocke-reference-variant for pytest, defaulting when an architecture has one variant. Test ids become <arch>-<variant>-<case>.
8. Record variant in newly qualified manifests and validate it when present.
9. status (one line per enrolled variant: publication, baseline, bundle state, shared cohorts) and artifact.py published [--json] for tooling.
10. Documentation: glossary, quickstart recipes (add a case, add a variant with its own baseline, add an architecture, un-share the WMMA cohort), workflow, guides, and a migration note.

Unchanged:

• CTest names and the provider category patterns.
• The set(ROCKE_PUBLISHED_<OP>_ARCHITECTURES ...) publication lists.
• The worker request's architecture key and the archive member root.

Where new kernels go:

• A family with the same math and input/output model (same Case, independent reference, input generator, and manifest operation) is a new variant.
• A family with different math or a different input/output model, such as backward convolution or attention decode/backward, is a new operation package with its own variants.

This keeps each contract single-purpose and independent kernel PRs in disjoint files. The docs include a table of what can still overlap between concurrent PRs: the same variant, the shared WMMA cohort, adding an architecture, and contract changes.

Developer and tooling notes:

• Old <op>/<arch>.tar.gz archives and extracted bundles in existing checkouts are ignored leftovers: delete them and dvc pull.
• Tools that build archive or install paths must add /<variant>, or use artifact.py published --json.

Test Plan

• Host suites (test_reference_common.py, conv/SDPA contract tests, test_sdpa_torch_reference.py, platform/tests/test_gpu_references.py with its CMake configure/install checks) on every commit in the stack. Also run at the head with a temporary path containing a space and $.
• Installed layout: the host reference CTests run from an installed tree. The status cohort test runs in both source and installed trees.
• Hardware, source mode: pinned pytest with --rocke-reference-operation/--rocke-reference-arch on gfx942 and gfx950 (conv and SDPA), gfx1201 (SDPA), and gfx1151 (SDPA). Each run includes the three failure controls: perturbed current output, perturbed baseline replay, and a missing launch.
• Hardware, installed mode: CTest on gfx942 and gfx950 (conv and SDPA) and gfx1151 (SDPA), with provider category selection checked for every category. Other processors' GPU entries must skip with exit 77.
• CLI on hardware: status, cases, and verify --case on gfx942.

Test Result

• Host: 779 passed at the head (also with a space-and-$ temporary path). Every commit passes.
• Source pinned suites, 0 skipped: conv gfx942 35/35, conv gfx950 35/35, SDPA gfx942 11/11, SDPA gfx950 10/10, SDPA gfx1201 11/11, SDPA gfx1151 11/11. All failure controls rejected their perturbations.
• Installed: rocke_reference_common_pytest 447 passed, 6 skipped. The 6 are the locked-baseline requalification checks, which need git history and skip in installed trees. The matching rocke_<op>_gpu_<arch>_pytest passes on gfx942, gfx950, and gfx1151; the other processors' entries skip; category selection is unchanged.
• CLI: status reports both convolution variants matching their bundles with a shared cohort, and verify --case passes one selected case with the subset caveat.
• Workflow demonstration gfx950, gfx120X, and gfx1151 passing: https://github.com/ROCm/rocm-libraries/actions/runs/40045375501

Submission Checklist

JIRA ID : AICK-2299

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[Copilot]

🔵 Needs a closer look

The cross-architecture bundle migration and frozen-runner compatibility require maintainer validation, alongside resolution of the retained findings.

2 open issues

• Create architecture directory before packing first publication · New
• Track published variant changes with CONFIGURE_DEPENDS · New

What changed in this PR

This PR splits rocKE’s pinned convolution and SDPA references by operation, architecture, and kernel variant, allowing kernel families to maintain independent baselines without rewriting shared reference files.

Changes:

• Moves adapters, locks, and DVC pointers into variant-specific paths while preserving published archive identities.
• Adds variant-aware discovery, selection, validation, and installation, retaining one GPU CTest per operation/architecture.
• Expands host checks and documents qualification, publication, and migration workflows.

Paths below are relative to dnn-providers/hip-kernel-provider/rocke/.

File	Description
TESTING.md	Documents variant-based coverage and installation.
platform/​tests/​test_gpu_references.py	Tests variant staging, publication, and configuration guards.
platform/​tests/​README.md	Updates installed reference-suite descriptions.
platform/​CMakeLists.txt	Scales GPU CTest timeouts by variant count.
platform/​cmake/​PublishedGpuReferences.cmake	Clarifies publication semantics in comments; lists remain unchanged.
platform/​cmake/​InstallGpuReferences.cmake	Validates reference wiring and installs variant bundles.
library/​tests/​test_sdpa_torch_reference.py	Checks variant identity and qualification cohort digests.
library/​tests/​test_sdpa_reference_contract.py	Adapts SDPA contract checks to variant adapters.
library/​tests/​test_conv_reference_contract.py	Adapts convolution contract and qualification checks.
library/​tests/​sdpa_reference/​worker.py	Separates source adapter selection from processor-only replay.
library/​tests/​sdpa_reference/​torch_reference.py	Uses the variant adapter lookup for oracle audits.
library/​tests/​sdpa_reference/​docs/​sdpa-test-reference.md	Documents SDPA variants and updated bundle paths.
library/​tests/​sdpa_reference/​architectures/​gfx950/​dense/​baseline_lock.json	Adds the qualified cohort digest.
library/​tests/​sdpa_reference/​architectures/​gfx950/​dense/​__init__.py	Houses the relocated gfx950 dense adapter.
library/​tests/​sdpa_reference/​architectures/​gfx950/​__init__.py	Retains processor identity without importing variants.
library/​tests/​sdpa_reference/​architectures/​gfx942/​dense/​baseline_lock.json	Adds the qualified cohort digest.
library/​tests/​sdpa_reference/​architectures/​gfx942/​dense/​__init__.py	Houses the relocated gfx942 dense adapter.
library/​tests/​sdpa_reference/​architectures/​gfx942/​__init__.py	Retains processor identity without importing variants.
library/​tests/​sdpa_reference/​architectures/​gfx1201/​wmma/​baseline_lock.json	Adds the qualified cohort digest.
library/​tests/​sdpa_reference/​architectures/​gfx1201/​wmma/​__init__.py	Houses the relocated gfx1201 WMMA adapter.
library/​tests/​sdpa_reference/​architectures/​gfx1201/​__init__.py	Retains processor identity without importing variants.
library/​tests/​sdpa_reference/​architectures/​gfx1200/​wmma/​__init__.py	Houses the unpublished gfx1200 WMMA adapter.
library/​tests/​sdpa_reference/​architectures/​gfx1200/​__init__.py	Retains processor identity without importing variants.
library/​tests/​sdpa_reference/​architectures/​gfx1151/​wmma/​baseline_lock.json	Adds the qualified cohort digest.
library/​tests/​sdpa_reference/​architectures/​gfx1151/​wmma/​__init__.py	Houses the relocated gfx1151 WMMA adapter.
library/​tests/​sdpa_reference/​architectures/​gfx1151/​__init__.py	Retains processor identity without importing variants.
library/​tests/​sdpa_reference/​architectures/​__init__.py	Delegates SDPA lookups to the shared registry.
library/​tests/​reference_common/​worker.py	Documents source and replay target roles.
library/​tests/​reference_common/​registry.py	Adds shared processor and variant discovery.
library/​tests/​reference_common/​pytest_support.py	Validates and filters explicit variant selections.
library/​tests/​reference_common/​pinned_tests.py	Runs comparisons and failure controls per variant.
library/​tests/​reference_common/​paths.py	Resolves variant-specific source and installed bundles.
library/​tests/​reference_common/​docs/​README.md	Adds variant terminology and migration guidance.
library/​tests/​reference_common/​docs/​quickstart.md	Adds variant authoring and execution recipes.
library/​tests/​reference_common/​docs/​gpu-reference-workflow.md	Updates qualification, publication, and troubleshooting procedures.
library/​tests/​reference_common/​docs/​gpu-ci-pinned-rocke-test-reference-plan.md	Updates reference architecture and storage documentation.
library/​tests/​reference_common/​artifact.py	Adds cohort validation, variant checks, and publication listing.
library/​tests/​reference_bundles/​sdpa/​gfx950/​dense.tar.gz.dvc	Updates the archive filename while preserving its hash and size.
library/​tests/​reference_bundles/​sdpa/​gfx942/​dense.tar.gz.dvc	Updates the archive filename while preserving its hash and size.
library/​tests/​reference_bundles/​sdpa/​gfx1201/​wmma.tar.gz.dvc	Updates the archive filename while preserving its hash and size.
library/​tests/​reference_bundles/​sdpa/​gfx1151/​wmma.tar.gz.dvc	Updates the archive filename while preserving its hash and size.
library/​tests/​reference_bundles/​sdpa/​.gitignore	Removes operation-local archive ignore rules.
library/​tests/​reference_bundles/​conv/​gfx950/​implicit_gemm_fwd.tar.gz.dvc	Updates the archive filename while preserving its hash and size.
library/​tests/​reference_bundles/​conv/​gfx942/​implicit_gemm_fwd.tar.gz.dvc	Updates the archive filename while preserving its hash and size.
library/​tests/​reference_bundles/​conv/​.gitignore	Removes operation-local archive ignore rules.
library/​tests/​reference_bundles/​.gitignore	Centralizes variant archive and extracted-bundle ignore rules.
library/​tests/​conv_reference/​worker.py	Separates source adapter selection from processor-only replay.
library/​tests/​conv_reference/​docs/​conv-test-reference.md	Documents convolution variant ownership and paths.
library/​tests/​conv_reference/​check_torch_reference.py	Uses variant adapter lookup in oracle checks.
library/​tests/​conv_reference/​architectures/​gfx950/​implicit_gemm_fwd/​baseline_lock.json	Adds the qualified cohort digest.
library/​tests/​conv_reference/​architectures/​gfx950/​implicit_gemm_fwd/​__init__.py	Houses the relocated gfx950 forward adapter.
library/​tests/​conv_reference/​architectures/​gfx950/​__init__.py	Retains processor identity without importing variants.
library/​tests/​conv_reference/​architectures/​gfx942/​implicit_gemm_fwd/​baseline_lock.json	Adds the qualified cohort digest.
library/​tests/​conv_reference/​architectures/​gfx942/​implicit_gemm_fwd/​__init__.py	Houses the relocated gfx942 forward adapter.
library/​tests/​conv_reference/​architectures/​gfx942/​__init__.py	Retains processor identity without importing variants.
library/​tests/​conv_reference/​architectures/​__init__.py	Delegates convolution lookups to the shared registry.
library/​tests/​conftest.py	Registers the pytest variant-selection option.

_{Review effort: Balanced}

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #14164 — split pinned GPU references by operation, architecture and variant — 🟢 First-round: no closer look needed — the six published bundles move byte for byte without requalification, every host check passes from a source checkout, and the moved gfx942 bundles pass on the GPU in CI

Prior item	Status
Copilot ×2 (10-23)	🟢 Both FIXED. The first-publication recipe now creates the architecture directory before pack (03fa4418). pack() itself still expects that directory, so creating it there would also cover direct callers (optional) (read). The variant glob now uses CONFIGURE_DEPENDS (d57e6c5a) (read)
Bots / CI	🟢 Green at head d57e6c5a: TheRock Multi-Arch, ASAN and the hipDNN superbuild. The provider build installs every published reference, and its configure step authenticates each moved archive against its moved lock, so the green builds cover all six (read). On both gfx94X shards, the moved gfx942 bundles pass on the GPU (conv 35 cases, SDPA 11). rocke_reference_common_pytest also runs the cohort and frozen-runner replay checks against them (435 passed). The gfx950 / gfx1151 / gfx1201 GPU tests skip in that run, which has no such GPUs. A second run with gfx950, gfx1151 and gfx120X stages started at 00:31 UTC and is still in progress (read)
Our checks	🟢 Every .dvc pointer changes only its path: line, with the same md5 and size (ran). On the merge, the three reference host suites pass 714 tests against develop's 598 (ran). That includes the six locked-baseline preparation checks, which CI's installed tree always skips. The 18 skips are bundle comparisons, because there is no DVC data here. The torch audit passes 5/5 (ran). The platform CMake and install tests, including the new configure guards, pass 41 against develop's 14 (ran). run_<op>_reference.py status lists the six published variants and the unpublished gfx1200 one, each with its dvc pull command (ran). The merge is clean (71 behind), and rule_scan is clean (ran)
Cross-PR	🟢 Open #13897 (RDNA SDPA catalog) edits the SDPA worker, WMMA implementation and contract test. It merges cleanly, and on develop + this PR + #13897 the reference host suites (562 passed) and the torch audit pass (ran). #14328 also edits test_reference_common.py and merges cleanly (ran). #13398 still edits the pre-variant conv locks and pointers, but it is 501 commits behind and already conflicts with develop. Once this lands, configuration rejects those old paths with a message that says where they go (read)

No net-new findings.

Utility: 🟢 Positive — independent kernel work on one architecture now lands in disjoint files with its own baseline, so pinning the next kernel family no longer requalifies every case on that architecture. Configuration also refuses half-wired or pre-variant references instead of silently dropping their coverage.

---

Legend: 🔴 blocking · 🟠 should-fix · 🟡 minor nit · 🟢 info/good · ⚪ retracted

First-round review: a local qwen3.6 model made the independent first pass. The orchestrator (Claude Opus 5.5) verified it. The checks were the reference host suites, the torch audit and the platform CMake tests on the merge and on develop, the DVC pointer diff, the status command and a three-way merge with #13897. They also covered scratch merges with the other open PRs that touch the reference layout, and the CI logs at head. The orchestrator owns this merge-worthiness call. It flags whether a closer look is needed; it is not an approval or a rejection.

[yraparti]

rocKE checklist re-review of the fixes since d57e6c5a382c

ROCm/rocm-libraries#14164 · eee5958744b0 · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 1, MEDIUM 0, LOW 0, NEEDS-INFO 13

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (d57e6c5a382c)

Resolved 0 · still open 1 · new 0

• ⏳ §D.1 still FAIL (HIGH): Put the knob on the KernelSpec and delete the environment read.

Findings

• HIGH §D · dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 · Put the knob on the KernelSpec and delete the environment read.

Needs info

• §B.4 At CLI, pytest, and CMake trust boundaries for the new variant parameter and lock files, invalid or ambiguous inputs are explicitly rejected rather than normalized into another valid variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:77). — no diff line supports this verdict
• §C.1 Re-review resolves the previous NEEDS-INFO: repeated per-operation architecture/variant lookup is now a named shared helper in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and both operation packages delegate to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:10 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:10. — no diff line supports this verdict
• §C.7 Re-review resolves the previous NEEDS-INFO: the bundle move is shown as a no-op archive identity move by preserving the DVC object hash while changing only the archive path, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:2 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:5. — no diff line supports this verdict
• §E.13 Several diffs (cli.py, test_reference_common.py, test_gpu_references.py, and docs/*.md) are truncated or omitted in the prompt, so full removal of old per-architecture references across all files cannot be verified from the provided diff alone.
• §E.15 Previously NEEDS-INFO; verified in the current diff that new variant parameters are read and used in validate_bundle (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), pack (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:203), unpack (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:293), and AdapterRegistry.resolve_variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65, read at…
• §F.6 Local pre-commit execution status before push is not visible from the diff alone.
• §F.7 Bot and Copilot review threads and their resolution status are not included in the prompt.
• §G.3 Missing current-head evidence that the required local non-GPU gate sweep was run, or a baseline-diff showing the affected gates are unchanged versus develop.

Section notes

• §D: This PR reorganizes pinned GPU reference test bundles and harnesses by (operation, architecture, variant); it does not touch dispatcher registration, routing, selectors, or KernelSpec definitions.
• §D: This PR only restructures pinned GPU test reference harnesses by (operation, architecture, variant) and does not touch the production dispatcher, candidates, or kernel selectors.
• §G: Items 18–34 of 34
• §H: No device-side masks, bounds, indirection tables, or kernel implementations are added or modified in this PR; it only reorganizes test reference bundle layouts, locks, DVC pointers, and Python/CMake test harness infrastructure.
• §J: Markdown present. Shapes: | Doc | Operands / inner dimensions | Result vs declared | |---|---|---| | library/tests/conv_reference/docs/conv-test-reference.md | Documents tensor layouts (NHWC, KYXC with C/groups, NHWK) but adds no equation using them. | No equation result shape is declared. | | TESTING.md, platform/tests/README.md, omitted docs | No visible operand shape annotations in the provided diff. | N/A. | Symbols: | Symbol | Definition / domain | |---|---| | architecture / arch | Processor-level GPU reference selection and package level. | | variant | Kernel-family reference adapter/bundle level under an architecture. | | cohort | Enrolled case set associated with a lock/bundle identity. | Novelty: | Claim | Check status | |---|---| | Conv now publishes the implicit_gemm_fwd variant and lists it via status. | Visible in the provided diff, but complete develop comparison is blocked by omitted Markdown hunks. | | Pinned gradients would be a separate…
• §K: This PR reorganizes pinned GPU test reference bundles by (operation, architecture, variant); it does not touch dispatch tuning, KnobSpace subclasses, knob axes, sweep entry points, or benchmark tuning replay.

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	N/A	This PR restructures pinned GPU reference harnesses and bundles, but does not add or modify kernel-body emitters.
2	N/A	This PR does not modify core IR or lowering-op handler paths such as core/ir or core/lower_*.
3	N/A	This PR changes reference-test adapters and publication plumbing, not production kernel-level selectors, spec builders, lowering, or codegen knobs.
4	N/A	No Python or C++ IR emitter code is changed, so emit-order parity does not apply.
5	N/A	The diff does not modify representative-IR golden files or intentional kernel IR output.
6	N/A	There is no engine or emitted-IR change in this PR requiring golden-vs-cross-engine parity proof.
7	N/A	The new logic is in pinned-reference test selection, packaging, and installation rather than a kernel or lowering hot path that emits IR.
8	N/A	The PR does not touch C++ engine, builder, selector, or lowering implementation code.
9	N/A	The PR contains no shared op-handler fix or lowerer-only architecture divergence.
10	N/A	The PR does not introduce a new kernel/compiler feature flag that changes emitted IR.
11	N/A	The PR does not add a new IR opcode.
12	N/A	The PR does not modify ROCKE_BACKEND behavior or core/backend.py::CPP_UNPORTED_ARCHES.
13	N/A	The PR does not change emitted .ll generation or ROCKE_DEBUG_LOC handling.
14	N/A	The PR does not modify classic helpers under platform/cpp/helpers or their C++ helper headers; helpers/tiling is not involved.

§B graded 6/6

#	Verdict	Evidence
1	N/A	This PR only touches test reference infrastructure and documentation, with no new kernel or dispatch rejection gates.
2	N/A	No preflight gates or kernel spec constructors are added or modified in this PR.
3	N/A	No model-zoo dispatch routing or kernel fallback logic is changed in this test-only PR.
4	NEEDS-INFO	At CLI, pytest, and CMake trust boundaries for the new variant parameter and lock files, invalid or ambiguous inputs are explicitly rejected rather than normalized into another valid variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:77). — no diff line supports this verdict
5	N/A	No kernel dispatch gates or supported configuration scopes are modified in this PR.
6	N/A	No is_valid_spec or supports_* functions are touched in this PR.

§C graded 10/10

#	Verdict	Evidence
1	NEEDS-INFO	Re-review resolves the previous NEEDS-INFO: repeated per-operation architecture/variant lookup is now a named shared helper in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and both operation packages delegate to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:10 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:10. — no diff line supports this verdict
2	N/A	This PR refactors pinned GPU reference test layout and staging, not kernel build_<family>(spec) builders or emitted IR.
3	PASS	Variant names are operation/algorithm family names such as implicit_gemm_fwd in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4, dense in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:4, and CMake rejects invalid or gfx-prefixed variant names in…
4	PASS	Arch-independent registry and lookup logic lives in the common helper dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, while processor packages now only name the enrolled processor, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/__init__.py:9, and algorithm-specific adapters live under variant subpackages such as…
5	PASS	The diff fan-out matches the inherent reference-key widening across tests, bundle paths, CMake staging, and docs: default_bundle_path gains a variant component in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, CMake installs variant payloads in dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:205, and TESTING documents the new installed path in…
6	PASS	Re-review resolves the previous NEEDS-INFO: the easy change is factored through the shared AdapterRegistry abstraction in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22 before operation packages bind get_adapter, variants, and baseline_lock to that abstraction in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:13.
7	NEEDS-INFO	Re-review resolves the previous NEEDS-INFO: the bundle move is shown as a no-op archive identity move by preserving the DVC object hash while changing only the archive path, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:2 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:5. — no diff line supports this verdict
8	PASS	New functionality extends the existing shared reference tooling rather than reimplementing separate operation tools: artifact.py adds the published command in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:316, and CLI cohort/variant handling is added to ReferenceOperation in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:170.
9	PASS	The variant-source/replay split is treated as a class across sibling workers: convolution dispatches source mode through get_adapter in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:45, and SDPA applies the same source/replay guard in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:64; CMake also rejects stale pre-variant locks and pointers repo-wide in…
10	PASS	The widened shared key (operation, architecture, variant) is propagated through option parsing in dnn-providers/hip-kernel-provider/rocke/library/tests/conftest.py:55, path construction in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, pytest parametrization in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:178, CMake staging in…

§D graded 21/21

#	Verdict	Evidence
1	FAIL	dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 reads the environment
2	N/A	This PR does not add or modify any dispatcher KernelCandidate or Capability declarations.
3	N/A	This PR does not add, modify, or call any dispatcher candidate _supports or admits predicates.
4	N/A	This PR does not touch dispatcher selectors or spec builders on the selection path.
5	N/A	This PR does not modify any KernelSpec identity, kernel_name, or cache-key logic.
6	N/A	This PR does not modify cohort routing, spec builders, launch metadata, or cache keys in the dispatcher.
7	N/A	This PR does not add or modify dispatcher heuristics, hand-pinned overrides, or resolvers.
8	N/A	This PR does not define or modify any dispatcher routing cohorts or spec override predicates.
9	N/A	This PR does not register any new kernel or architecture in the dispatcher registries.
10	N/A	This PR does not introduce any new KernelSpec fields or features.
11	N/A	This PR does not register, promote, or change precedence or opt-in status for any dispatcher candidates.
12	N/A	This PR only restructures pinned GPU test reference harnesses and does not add, rename, or reprioritize any dispatcher candidates.
13	N/A	This PR does not modify CandidateRegistry.select or add any dispatcher wrappers or entry points.
14	N/A	This PR does not touch kernel compilation caches, routing logs, tuning records, or kernarg layouts.
15	N/A	This PR does not add or modify any dispatcher candidates or arch modules in the production dispatch registry.
16	N/A	This PR does not introduce a new operator family in the dispatcher.
17	N/A	This PR does not add or modify dispatcher candidates or routing rules.
18	N/A	This PR does not modify CandidateRegistry settings, binding requirements, or exempt sets.
19	N/A	This PR makes no claims about unified attention auto routing or run_unified_attention_torch dispatch.
20	N/A	This PR makes no changes to kernel-level selectors, spec builders, or the Python registry/tuning layer.
21	N/A	This PR does not fix a kernel selection or emitter issue; it reorganizes test reference bundles by variant.

§E graded 17/17

#	Verdict	Evidence
1	PASS	library/tests imports rocke (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:57), while platform/cmake/InstallGpuReferences.cmake invokes artifact.py via an isolated subprocess (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:208) without importing library into platform.
2	PASS	Device queries in the reference test infrastructure use from rocke.runtime.hip_module import get_device_target_id (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:59) rather than torch.
3	PASS	Torch remains optional: dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/check_torch_reference.py:13 gates with pytest.importorskip("torch") and pinned reference tests verify assert not report["torch_imported"] (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:210).
4	N/A	This PR only modifies files under library/tests/, platform/, and TESTING.md, not production library/{kernels,dispatch,builders,benchmarks} layering or KNOWN_VIOLATIONS/KNOWN_BAD lists.
5	PASS	Platform CMake rules use relative directory variables (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:29) and Python helpers use tempfile and pathlib (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:18).
6	N/A	This PR does not touch emitter trees (core/, helpers/, instances/, library/kernels/).
7	PASS	Testing and workflow documentation are updated alongside the reference layout changes (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/docs/conv-test-reference.md:13).
8	PASS	No Jira/tracking IDs, reviewer handles, or PR numbers appear in code or comments in the diff (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:1).
9	PASS	All modified files are scoped to splitting pinned GPU reference bundles by operation, architecture, and variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	The PR explicitly accounts for concurrent branches and merge collisions by adding CMake guards against pre-variant locks and DVC pointers restored at legacy paths (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107).
11	PASS	Only deliverable test harness code, locks, DVC pointers, CMake rules, and required reference documentation are committed (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22).
12	PASS	All enrolled architectures (gfx942, gfx950, gfx1151, gfx1200, gfx1201) are migrated to the variant layout (dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:471, dnn-providers/hip-kernel-provider/rocke/platform/cmake/PublishedGpuReferences.cmake:11).
13	NEEDS-INFO	Several diffs (cli.py, test_reference_common.py, test_gpu_references.py, and docs/*.md) are truncated or omitted in the prompt, so full removal of old per-architecture references across all files cannot be verified from the provided diff alone.
14	PASS	Moved adapter files preserve upstream comments, historical API compatibility checks, and docstrings (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:43).
15	NEEDS-INFO	Previously NEEDS-INFO; verified in the current diff that new variant parameters are read and used in validate_bundle (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), pack (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:203), unpack (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:293), and AdapterRegistry.resolve_variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65, read at…
16	N/A	This item is a meta-instruction for review grading rather than a property of the pull request diff.
17	PASS	No performance numbers or superiority claims appear in the PR description, docs, or code (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/brockhargreaves-amd/pinned-reference-variants matches users/<user>/… (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).
2	PASS	title test(rocke): split pinned GPU references by operation, architecture, and variant matches type(scope): summary and is 80 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py is in the diff
5	PASS	no key, certificate, or secret filename in the diff
6	NEEDS-INFO	Local pre-commit execution status before push is not visible from the diff alone.
7	NEEDS-INFO	Bot and Copilot review threads and their resolution status are not included in the prompt.
8	PASS	Base branch is develop (though body notes Stacked on #13651), and the branch, title, body, tests, and files satisfy policy.yml checks (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312).
9	PASS	The PR keeps the full 10-commit migration and test suite together without artificial splitting for a size cap (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	Tests added in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:341 exercise real variant discovery, cohort digesting, and CMake staging behavior rather than throwaway checks.
11	PASS	The fetched PR metadata shows the populated Title and full Motivation/Technical Details/Test Plan Body are present (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§G graded 34/34

#	Verdict	Evidence
1	N/A	This PR refactors pinned GPU reference bundle layout and harness plumbing, not byte-identity parity as a correctness proof.
2	N/A	This PR does not add or modify a kernel implementation, so the kernel-work on-silicon done criterion does not apply.
3	NEEDS-INFO	Missing current-head evidence that the required local non-GPU gate sweep was run, or a baseline-diff showing the affected gates are unchanged versus develop.
4	PASS	New variant behavior is covered by falsifiable tests, e.g. selected declared variants are accepted/rejected in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:398 and missing/unpublished variant paths are exercised in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:807.
5	PASS	Tests exercise shipped surfaces rather than only unreachable internals, including the published artifact CLI via subprocess in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:309 and pytest session/selection behavior in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:807.
6	NEEDS-INFO	The diff is truncated and omits documentation hunks under reference_common docs, sdpa docs, platform CMakeLists, and platform/tests/test_gpu_references.py, so the load-bearing prose and its checks cannot be fully verified from the provided diff.
7	N/A	This PR does not introduce a new numerical oracle design doc or new fp32 parity-reference policy.
8	N/A	No dtype fallback behavior is added or modified in the shown harness changes.
9	N/A	This PR makes no LDS or occupancy resource claims.
10	N/A	This PR makes no performance-result or provenance claims that require device/script/ratio validation.
11	N/A	This PR does not add or optimize a kernel or dispatch selector whose reachability must be proven.
12	N/A	The changes are pinned-reference test infrastructure and do not introduce a two-engine emitted-IR or lowering parity gap.
13	N/A	No LLVM intrinsic or TableGen symbol is added, changed, or recommended.
14	N/A	No bottleneck classification or static-versus-dynamic performance analysis is introduced.
15	N/A	No wrong-branch risk is evident from the provided checklist scope for this infrastructure refactor.
16	NEEDS-INFO	Resolves the previous NEEDS-INFO: the shared registry defines and validates variant adapter lookup in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:81, and both operation architecture packages delegate their exported symbols to it, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/init.py:12 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/init.py:12. — no diff line supports this verdict
17	PASS	Shipped scripts are exercised end-to-end through subprocess tests, including artifact.py published in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:309 and common CLI round-tripping in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:207.
18	PASS	New configuration and cohort-digest guards are wired into CMake configuration and bundle validation (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:40) and are proven to fire on constructed fault conditions while passing valid states (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:588…
19	PASS	Test fixtures construct multi-variant, unpublished, half-wired, and mismatched-cohort shapes needed to express the failures (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1273, dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:338).
20	N/A	This PR does not add or modify a detector banning performance or numerical claims.
21	N/A	This PR refactors reference bundle layout by variant and does not introduce performance or statistical measurement code.
22	PASS	The harness explicitly fails rather than skipping when a selected operation, variant, or published architecture has no bundle or no published variants (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:96, dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1468), and the PR body confirms 0 skipped across all 6 hardware suites.
23	N/A	This PR makes no performance or routing speedup claims.
24	N/A	This PR does not re-bless IR goldens or change kernel codegen; the existing reference bundles are moved byte-for-byte with identical manifest_sha256…
25	N/A	This PR does not contain a codegen micro-optimization.
26	N/A	This PR does not benchmark kernels or report performance numbers.
27	N/A	This PR does not perform A/B performance comparisons or ship kernel optimization knobs.
28	N/A	This PR does not report performance deltas or counterfactual comparisons.
29	N/A	This PR does not swap math primitives in any kernel.
30	N/A	This PR does not make register pressure or zero-spill claims.
31	N/A	This PR does not claim to fix a kernel performance defect or bank conflict against a baseline.
32	N/A	This PR does not report measured performance or hardware resource figures.
33	N/A	This PR does not change any kernel dispatch resolver, heuristic, or spec builder decision function.
34	N/A	This PR does not add or modify an opt-in kernel output flag.

§H graded 5/5

#	Verdict	Evidence
1	N/A	This PR only reorganizes test reference bundles and CMake/Python test harness infrastructure without touching kernel specs, SRDs, or launch validation.
2	N/A	This PR does not add or modify paged/indirection kernels or device table lookups.
3	N/A	This PR does not add or modify any device kernel code or compile-time tile/page divisors.
4	N/A	This PR does not touch device-side masks, bounds, or runtime sequence-length inputs.
5	N/A	This PR does not introduce or modify kernel launch inputs feeding device-side bounds or masks.

§I graded 6/6

#	Verdict	Evidence
1	N/A	This PR refactors test reference bundle layout by variant and does not perform kernel routing or candidate space measurement.
2	PASS	Invalid or half-wired states are loudly rejected with actionable error messages and recovery commands rather than silently ignored or cryptically failing (e.g., dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:39-135 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143-173).
3	N/A	This PR splits existing pinned reference cohorts by variant without altering test case generation or hypothesis property tests.
4	N/A	This PR does not generate performance tables or benchmark reports comparing shipped vs achievable kernel selections.
5	PASS	Pinned reference bundles are explicitly tracked as frozen snapshots via baseline_lock.json (recording baseline_revision, cohort_sha256, and manifest_sha256) with an explicit qualification/regen CLI path and status/diff drift checks (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349-495, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:1261-1317).
6	N/A	This PR does not add or modify any DSL ops, hardware instructions, or lowering primitives.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The modified Markdown visible in the diff contains no mathematical equations or matrix products to type-check.
2	N/A	The PR does not introduce formula notation tables or multi-domain mathematical quantities such as decay versus log-decay.
3	N/A	The visible Markdown documents GPU reference packaging and workflows, not a kernel algorithm derivation owned by another doc.
4	NEEDS-INFO	Not resolved from the previous NEEDS-INFO: the diff still omits substantial Markdown hunks for library/tests/reference_common/docs/README.md, gpu-ci-pinned-rocke-test-reference-plan.md, gpu-reference-workflow.md, quickstart.md, sdpa-test-reference.md, and platform/tests/test_gpu_references.py-adjacent docs, so novelty claims there cannot be checked against develop.
5	PASS	TESTING.md:921-923 states the installed CTest timeout as a per-variant aggregate (five minutes per installed variant) rather than a bare per-test constant.
6	N/A	This PR reorganizes pinned test reference bundles and does not add model configuration tables, tuned dispatch band edges, or similar constants requiring deployment provenance.
7	N/A	No new Markdown design or algorithm document file is added in the provided diff; existing test documentation is updated.

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR does not add or modify any tuned dispatch candidates or KnobSpace subclasses.
2	N/A	This PR does not add or modify any kernel spec fields or KnobAxis declarations.
3	N/A	This PR does not change tuning_id, config_key, knob canonicalization rules, or tuned spec defaults.
4	N/A	This PR does not touch KnobSpace.defaults() or KnobSpace.recorded().
5	N/A	This PR does not modify knob dropping, refusal, inertness, or base_value() rules.
6	N/A	This PR does not modify KnobSpace.validate or policy-free tuning spec construction.
7	N/A	This PR does not touch KnobSpace.base(req) or request normalization for tuning.
8	N/A	This PR does not change normalize_knobs or knob value coercion.
9	N/A	This PR modifies GPU test reference bundle replay, not dispatch tuning_id/spec_hash benchmark replay.
10	N/A	This PR does not touch KnobSpace.find or tuning_id resolution in dispatch.
11	N/A	This PR does not touch sweep entry points, pin_to_spec, or sweep harnesses.
12	N/A	This PR does not change production stacks, accept_default, or base specs for auto tuning selection.
13	N/A	This PR does not touch benchmark or sweep harnesses or make claims about auto selection.

§L graded 10/10

#	Verdict	Evidence
1	N/A	This PR reorganizes pinned GPU test reference bundles by variant and does not add or modify any gfx hardware target or ISA backend.
2	N/A	This PR does not add or modify any MMA op_id or layout map.
3	N/A	This PR does not touch MMA selection, MfmaAtom catalogs, or kernel lowering.
4	N/A	This PR only relocates existing test adapter code and does not alter ArchTarget.wave_size or lane math.
5	N/A	This PR does not add or modify arch-specific kernel recipes or tile loaders.
6	N/A	This PR does not modify pipeline policies, ArchTarget, or rocke/ imports.
7	N/A	This PR does not add or modify logical dtypes, bit packing, instruction fragments, or FP8 dialects.
8	N/A	This PR does not touch scaled-MMA contracts or WMMA/MFMA scaled emission.
9	N/A	This PR does not introduce a new architecture target.
10	N/A	This PR does not modify intrinsic arch-domain artifacts or LLVM flavor membership lists.

§M graded 10/10

#	Verdict	Evidence
1	N/A	The diff does not touch DSL kernel bodies, SSA Values, or runtime predicate/control-flow emission.
2	N/A	The diff does not touch legalize_for_target or tile/MMA lowering.
3	N/A	The diff does not touch ck.dsl.ir serialization, print_ir, wire versions, attributes, or ABI parameter ordering.
4	N/A	The diff does not touch portable IR schemas, importers, magic division helpers, recipe VM, or recorder coverage.
5	N/A	The diff does not touch AsyncTileLoader, LDS layouts, or waitcnt placement.
6	N/A	The diff does not emit GPU barriers, divergent control flow, memory loads, or cross-lane shuffles.
7	N/A	The diff does not touch ISA backend buffer resource words or TemporalHint lowering.
8	N/A	The diff does not add LLVM intrinsic declarations or modify compiler backend selection precedence.
9	N/A	The diff does not touch public C entry points, cpp/include/rocke/ headers, or C++ codegen launch behavior.
10	N/A	The diff does not touch debug-info emission or DWARF consumers.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The diff does not add or modify helpers under platform/python/rocke/helpers/ or introduce a new helper promoted to helpers/.
2	N/A	The diff does not author or modify kernels or their use of shared helper mechanisms such as atoms, TensorView, loaders, epilogues, or transforms.
3	N/A	The diff does not touch SoftwarePipeline or any kernel pipeline code.
4	N/A	The diff does not add or modify fused kernels, epilogue fusion logic, fusion validation, or fusion legalizers.
5	N/A	The diff does not introduce or change any public tiling API under helpers/tiling/.
6	N/A	The diff does not add or modify front-door or toolbox tiling helpers or any LayoutStyle implementation.
7	N/A	The diff does not touch tiling layout overrides, accumulator descriptors, or MMA validation.
8	N/A	The diff does not touch tiling memory operations, LDS allocation, global load/store emission, clipping, padding, or wide-load selection.
9	N/A	The diff does not use or modify N-D tensor axis roles, batch reductions, flow-stage relabeling, or AB-swap routing in tiling.
10	N/A	The diff does not modify tiling emit, LDS staging, MMA issue logic, or tiling fail-fast raises.
11	N/A	The diff does not touch MMA traits, TileMma, lds_swizzle, scheduling group helpers, or related tiling policies.
12	N/A	The diff makes no LDS bank-conflict or coalescing claims and does not modify related simulator or coalescing analysis tools.
13	N/A	The diff does not construct or modify WarpDistributionEncoding, make_tile_desc, RegisterMapper, or tiling store casts.
14	N/A	The diff does not use tiling verbs, modify synchronization or barrier placement, or add a tiling pipeline variant.
15	N/A	The diff does not touch tiling allocation or bounds-checking logic for transformed address ranges.
16	N/A	The diff does not use or modify optimize_layout, recommend, or evaluate_transform.
17	N/A	The diff does not touch the tiling recorder, round-trip analysis, witness flags, or visualization modules.
18	N/A	The diff does not touch interleaved layouts, tiling demo kernels, dispatch registration of tiling demos, or tiling numeric gates.

§O graded 16/16

#	Verdict	Evidence
1	N/A	This PR does not touch core/verify.py, cpp/core/verify.cpp, or IR address spaces/attributes.
2	N/A	This PR does not modify scf.for, region structure, or IR block emission.
3	N/A	This PR does not modify IRBuilder, LLVM intrinsic declaration tables, or metadata emission.
4	N/A	This PR does not modify IR types, SmemType, or IR parsers.
5	N/A	This PR does not modify core/passes.py, cpp/core/passes.cpp, or op purity tables.
6	N/A	This PR does not touch constant folding in Python or C++ passes.
7	N/A	This PR does not modify kernel attributes, COMGR options, or cluster dimension checks.
8	N/A	This PR does not touch CodegenPolicy, SchedulerStrategy, or kernel compilation options.
9	N/A	This PR does not modify waitcnt encoding or barrier lowering.
10	N/A	This PR does not touch LLVM lowerers or parameter attribute/type lowering.
11	N/A	This PR does not touch ArchTarget or target catalog resolution.
12	N/A	This PR does not touch vector/LDS width caps or codegen arch defaults.
13	N/A	This PR does not touch public C headers under cpp/include/rocke/ or ABI versions.
14	N/A	This PR does not touch C++ engine sources, headers, bindings, or C++ compile flags in platform/CMakeLists.txt.
15	N/A	This PR does not touch C++ engine fallback handling or alternative lowerers.
16	N/A	This PR does not touch known_gaps.md or C++ library/builder capabilities.

§P graded 14/14

#	Verdict	Evidence
1	N/A	No kernel builders, specs, or REQUIRED_FIELDS are added or modified in this PR.
2	N/A	No kernel argument signatures, ABI orders, or packing routines are modified in this PR.
3	N/A	Launch geometry logic is unchanged; reference adapters only relocate existing prepare/launch wrappers into variant subpackages.
4	N/A	No LDS capacity gates or emitters in Python or C++ are touched by this PR.
5	N/A	No WMMA/RDNA GEMM, GemmMultiD, or GemmMultiAbd instance specs or validators are modified in this PR.
6	N/A	No caller-zeroed workspaces (Stream-K, MoE, split-K) or workspace sizing helpers are touched in this PR.
7	N/A	No Stream-K, persistent GEMM, GroupedGemmLauncher, or SUPPORT_MATRIX.md limits are modified in this PR.
8	N/A	No quantization helpers, MX GEMM scales, or quantize_scalar_f32 paths are touched in this PR.
9	N/A	No numeric emulation or decode paths are added or changed in this PR.
10	N/A	This PR updates test reference bundle manifests (schema 2 under reference_common), not ck.dsl.example.manifest/v1 runtime manifests in helpers/manifest.py.
11	N/A	No runtime launch lifecycle, stream synchronization, WorkspacePool, or timing utilities are modified in this PR.
12	N/A	No COMGR or HIP library loading code or hipError launch fixes are included in this PR.
13	N/A	No production host-side launch math, tile-count helpers, or kernel-cache admission guards are refactored in this PR.
14	N/A	No examples/expected.json files, instances/, or helpers/ kernel modules are added or modified in this PR.

§Q graded 7/7

#	Verdict	Evidence
1	NEEDS-INFO	The diffs for dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt (+7 -2 lines) and dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py (+485 -80 lines) remain omitted to fit the model budget, so CTest registration and installed test changes cannot be verified.
2	N/A	This PR refactors the GPU reference test bundle layout and does not add a new kernel family or variant to the platform IR or builders.
3	PASS	Tests and documentation grade the reference harness accurately, enforcing failure controls for perturbed outputs and missing launches in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:168-276 and noting subset verification caveats in dnn-providers/hip-kernel-provider/rocke/TESTING.md:154-156.
4	N/A	This PR does not re-bless IR goldens or widen the valid spec space in a lowerer or builder.
5	PASS	Tests remain portable across CPU/GPU and OS environments, skipping cleanly when hardware or bundles are absent (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:76-160) and querying the active device processor dynamically (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pytest_support.py:80-115).
6	PASS	The harness refactor preserves per-configuration semantics, scaling timeouts per installed variant (dnn-providers/hip-kernel-provider/rocke/TESTING.md:924-925), keeping at most one slot's workers alive at a time (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:82-163), and preserving error classes (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:227-232).
7	N/A	This PR does not modify engine backend rejection contracts or ROCKE_BACKEND / ROCKE_CPP_STRICT fallback behavior.

§R graded 11/11

#	Verdict	Evidence
1	N/A	This PR only restructures pinned GPU test reference harnesses and bundles and changes no kernel body, algorithm, tiling strategy, pipeline variant, or performance lever.
2	N/A	No performance lever or architecture-specific optimization recipe is modified in this test-reference refactoring PR.
3	N/A	No optimization lever covered by the utilities skills is touched by this test-infrastructure PR.
4	N/A	The diff does not alter kernel codegen, lowering, or performance primitives, so static ISA/occupancy probe evidence does not apply.
5	N/A	No global memory accesses in GPU kernels are added or modified by this diff.
6	N/A	No LDS or MMA layouts are added or changed in this pull request.
7	N/A	The PR makes no kernel bottleneck claims and implements no performance fixes.
8	N/A	No waits, barriers, or scheduling hints in GPU kernels are changed.
9	N/A	This PR is not a performance regression fix and changes no shape-dependent performance lever.
10	N/A	No compiler codegen changes, assembly parity claims, or performance benchmarks are included in this diff.
11	N/A	This PR introduces no new GPU optimization lessons, probes, or analysis tools.

§S graded 8/8

#	Verdict	Evidence
1	N/A	The PR does not add any public symbol, spec, builder, compile/launch parameter, or C entry point to the rocke package or C engine.
2	NEEDS-INFO	Existing public API test coverage in library/tests/ and platform/tests/ is preserved without deleting, skipping, or narrowing tests; the cohort preservation guards in dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py:393 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:67 are updated to resolve variant adapters while keeping all existing test cases intact. — no diff line supports this verdict
3	N/A	No public API in rocke or the C engine is removed or renamed.
4	N/A	No __all__ list or public rocke module is modified in this diff.
5	PASS	New and updated tests in library/tests/ assert concrete outputs, digests, and fail-fast error raises, such as cohort digest and lock validation at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:540 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:588.
6	PASS	CPU-checkable contracts (manifest validation, cohort digests, CLI commands, and CMake reference wiring) are tested on CPU without Torch or a GPU at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312 and dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61, while GPU launch and numeric comparisons are marked @​pytest.mark.gpu at…
7	N/A	No API with a C++ twin or C ABI binding is added or modified.
8	PASS	No rocke public API or rocke.run_manifest code is modified, the PR test results name each GPU architecture (gfx942, gfx950, gfx1201, gfx1151), and the reference installation and contract suites are exercised across dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312 and dnn-providers/hip-kernel-provider/rocke/TESTING.md:123.

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #14164 — split pinned GPU references by operation, architecture, and variant (brockhargreaves-amd, AICK-2299) — 🟡 Follow-up: fixes verified; minor items open

Prior item	Status
Copilot (10-23) ×2: create the archive directory before the first pack; CONFIGURE_DEPENDS on the variant glob. Ours (10-24), optional: pack() itself could create the directory	🟢 Still FIXED, and our optional note was adopted. pack() creates the directory once the bundle validates (a6450761, artifact.py:204), and test_pack_creates_the_archive_directory passes. Every glob now has CONFIGURE_DEPENDS (see the AnaghaRaoAMD row) (ran)
yraparti (10-24; raised again 10-30 03:13Z) HIGH §D.1: environment read at test_reference_common.py:892; NEEDS-INFO ×12, now ×13	🟢 Answered; not a Rule 3 case. • The line: :892 only builds the environment for a child pytest (dict(os.environ, AMDGPU_TARGETS=…)). The only AMDGPU_TARGETS read is pytest_support.py:85. It predates this PR (develop :72) and only chooses the GPU lane (exit 77), never a kernel. • The rule: Rule 3 says "Do not flag environment reads in tests … that only configure the harness" (.github/instructions/rocke-pr-review.instructions.md:34). That was the author's 10-24 reply, and that thread is resolved. • Now: the bot's re-review at head opened a new thread with the same finding. It is unanswered; the same reply applies. • NEEDS-INFO: the author answered all 12 on 10-27. The new §G.3 row asks for gate evidence, which is in Our checks (read)
CongMa13 (10-26): threads and comments without an author reply	🟢 Addressed. The two links were yraparti's review, answered on 10-27, and our 10-24 post, which had no findings to answer (read)
CongMa13 (10-27) ×2: a listed architecture missing from registry.json still installs; source verify / diff skip the lock's cohort_sha256	🟢 FIXED af45a8c8, bd2d28a6. • Configuration fails on an unenrolled listed architecture, now also when an existing build directory rebuilds. • The bundle loader checks the cohort digest, as staging does. • test_listed_architectures_need_enrollment and test_bundle_commands_check_the_cohort_that_the_lock_pins pass (ran). • Both threads are answered and resolved
CongMa13 (10-27 23:38Z) APPROVED	🟡 Stale: it predates the 10 commits pushed on 10-29/30 (353e8d92…eee59587), which change cli.py, artifact.py, the CMake checks and the tests
AnaghaRaoAMD (10-29) ×5 code items: status and diff disagree on inputs (cli.py:1598); a pre-variant manifest passes as any variant through --lock (:587); 5 globs lack CONFIGURE_DEPENDS; the Torch audits take no variant; host tests assume one variant per architecture	🟢 All five FIXED, but none of their threads has a reply. • Inputs 0de1b0d6: _entry_problem (cli.py:785) now compares the recorded input digests. Probe (ran): after an inputs-only change, diff exits 1 with "generated inputs changed: (k)", and verify stops before any worker. With the fix reverse-applied, diff reports "all 8 enrolled cases match" while status's digest differs, which reproduces their report. • Pre-variant manifests 1c37173e, by another route than they suggested. A manifest without variant belongs to the variant whose committed lock pins its digest (cli.py:610, artifact.py:150). test_committed_locks_pin_distinct_manifests keeps that owner unique. Probe (ran): we enrolled a second gfx942 variant with dense's cases and passed it a dense bundle with dense's lock. load_bundle, diff, verify (exit 2) and artifact.py's validate_bundle all refuse it. With the fix reverted, all four accept it, and verify reaches the workers. • Globs a010e8cf, 3ebff8db: all 7 file(GLOB) calls take CONFIGURE_DEPENDS. registry.json, the adapters, operation.json, the locks and the archives are configure dependencies, so a requalified bundle restages. The wiring tests also run as a rebuild (ran: checks). • Torch audits 4a4dc178: the SDPA audit takes --variant (torch_reference.py:125). The conv check needs no flag because it covers every distinct enrolled case (read; it needs Torch). An AST test now fails on any reference-tool call that relies on the default variant, and it passes, as does the SDPA audit test (ran). • Host tests 353e8d92, f516f414: we enrolled a second variant on gfx942 for both operations, SDPA with dense's cases and conv with a subset. The four host test files pass unchanged: 736 passed and 24 skipped, every skip for missing DVC data or git history (ran). The guide row and quickstart lines they asked to drop no longer tell a variant author to edit shared tests
AnaghaRaoAMD (10-29) ×3 docs items: one serial CTest per architecture; adding a Case field; paged KV and others as operation packages	🟢 Mostly, no replies. • CTests: the trade-off is recorded: at most two variants per entry, and per-variant CTests before a third (eee59587). • Case fields: the guide now says that a defaulted field changes every case's contract, so every published variant needs requalifying. Variant-constant knobs go on the adapter (93ee5b18). • Operation packages: paged K/V on SDPA cases is now a variant, and the guide states one case model per operation (0edf7060). Their scaffold request is still open: decode, varlen, LSE output and backward remain separate packages modelled on conv_reference (read)
AnaghaRaoAMD (10-29): split test_reference_common.py by concern ("A tracked follow-up is fine")	🟡 OPEN: no reply and no follow-up ticket linked. The file is now 5,356 lines (2,980 on develop). The open #14328 already conflicts with this head there, in one hunk at the end of the file where both add a test (ran: merge-tree) → file the follow-up and link it on the thread
Ours (10-24) cross-PR notes	🟢 #13897 merged on 10-26 and is part of the merge below. #13398 no longer touches the reference layout. For #14328, see the row above (ran: git)
Bots / CI	🟡 • PR CI at eee59587: these pass: TheRock Multi-Arch (Linux gfx94X ×2, Windows gfx110X ×2), ASAN, pre-commit, clang-tidy, Component CI, the hipDNN superbuild and the policy bot. • Moved bundles on hardware: conv 35 and SDPA 11 pass on gfx942 (both shards). A manual extra-architecture dispatch at head adds conv 35 and SDPA 10 on gfx950, and SDPA 11 on gfx1151 (Linux and Windows). gfx1201 has no GPU run: the gfx120X runner is a gfx1200, so its test skips. Its frozen runner still replays on that lane. • That dispatch failed only on Windows gfx1151, in 3 lanes outside this PR. ◦ Develop's ISA cache-policy and NT-bit tests skipped on the 10-24 runner. On this one they find Visual Studio's bundled llvm-objdump and fail. ◦ A COMGR compile in test_gdn_decode_spec.py timed out. ◦ rocke_pytest hit its CTest timeout. This PR's lanes pass on that runner (rocke_reference_common_pytest: 451 passed). • GitHub: blocked; auto-merge off (read: CI logs, gh)
Our checks	🟢 • The push (ran: git): 13 commits since d57e6c5a, plus the develop merge 964f9e9b. That merge resolves no conflict in rocke/. • Merge a7393a0d (head + develop 587f9942; 57 behind, 27 ahead, 0 conflicts) (ran): ◦ native build OK; ctest 33/33 (1 skip); ◦ byte identity GREEN for the gemm families at llvm20/22/23; ◦ representative golden PASS ×3; differential golden: 0 changed, 0 missing; ◦ the PR's 5 test files, plus 2 that its conftest.py governs: 943 passed, 32 skipped; ◦ clang 24 compile: 0 failures. • Bundles: each of the six .dvc pointers differs from develop's only in path:, with the same md5 and size (ran). • Develop since the base: no commit touches the reference trees. Only 4 touch a file of this PR (platform/CMakeLists.txt, outside the reference block). The merge has exactly the six variant locks and pointers and no old-layout file. Outside the PR's files, nothing uses the old paths or calls a lookup without a variant (ran: git grep). • Rules 1–3: the push adds no links, tracker IDs, perf numbers or environment reads (ran)

Net-new: none.

Utility: 🟢 Positive — independent per-variant baselines without requalification; review gaps closed with tests.

---

Legend: 🔴 blocking · 🟠 should-fix · 🟡 minor nit · 🟢 info/good · ⚪ retracted

Follow-up review: a local qwen3.6 model made the independent first pass, and the orchestrator (Claude Opus 5.5) verified it. It read the 13 commits since our first round (d57e6c5a). On a clean merge with develop, it ran the native build, ctest, byte identity, the representative and differential goldens, a ROCm clang 24 compile and the PR's test files. It compared the six DVC pointers with develop's and scanned develop's new commits for changes to the reference layout. Two probes exercised the input-digest and pre-variant fixes, each against a control with the fix reverse-applied. The host tests ran again with a second enrolled variant on gfx942. The CI logs at head and the GitHub activity since our first round were read. The orchestrator owns this merge-worthiness call. This review flags whether a closer look is needed; it is not an approval or a rejection.

[brockhargreaves-amd]

Doing another run on other architectures here: https://github.com/ROCm/rocm-libraries/actions/runs/41128621898

[brockhargreaves-amd]

Follow-ups from @AnaghaRaoAMD's review. They are owned by the rocKE team (@ROCm/rocke-lib,
@ROCm/rocke-platform), which decides whether and when to do them; none is part of this PR, and I
won't be picking them up. Each item says what triggers it. The rocKE team can copy any of them into an
issue as is.

1. One GPU CTest per (operation, architecture, variant). From the CTest thread.

• Register rocke_<op>_gpu_<arch>_<variant>_pytest for each installed variant, passing
  --rocke-reference-variant <variant>, with a timeout longer than the worker's 300 s request
  timeout so a hang still reports which worker hung.
• Widen the two rocke_[a-z0-9_]+_gpu_gfx[0-9a-z]+_pytest patterns in
  ROCKE_ENGINE_test_categories_external.yaml (owned by @ROCm/hipdnn-reviewers) to accept the
  trailing variant. Keep the processor before the variant in the name.
• Update _reference_ctests in platform/tests/test_gpu_references.py and the ctest -R
  selectors in the docs, then remove the "at most two variants" cap from
  gpu-reference-workflow.md (Enroll provider test selection, and the "Two on different
  variants" row) and from the quickstart's add-a-variant recipe.
• Trigger: before any architecture publishes a third variant of an operation, or as soon as
  one variant needs its own exclusion or category.

2. Add a Case field without requalifying every published variant. From the Case-field
thread.

• Today any new field, even one with a default, changes every case's contract, so every
  published variant of the operation is requalified. For SDPA that is four bundles on four GPUs.
• Prototyped design: a Case declares the fields added after qualification. case_contract(case)
  leaves each one out of the contract, cohort digest, manifest and request while it holds its
  default, and is used everywhere a contract is recorded, matched or sent. Declarations are
  append-only, a declared default must reproduce every earlier qualification, and Case.id must
  encode non-default values. With nothing declared it equals asdict(case), so no lock changes.
• Trigger: the first per-case knob added to an existing operation's Case, such as an SDPA
  sliding window or logit softcap.

3. Split test_reference_common.py by concern. From the split thread.

• A pure move into its own directory: a conftest.py with the operation and second_variant
  fixtures (not the root library/tests/conftest.py), a support module for the shared helpers,
  and one file per concern. rocke_reference_common_pytest keeps its name and points at the
  directory.
• Acceptance: the same collected test ids and pass/skip counts, in source and installed trees.
• Optional: an autouse guard in that conftest that fails a reference host test relying on the
  sole-variant default.
• Port any open edit of the file, such as feat(rocke): backend completeness for every production target and dense MMA instruction #14328's.
• Trigger: none; any time after this PR merges.

4. Optional: a pinned reference for #14552's paged K/V kernels. From the operation-or-variant
thread.

• Under the rule this PR now documents, uniform-length paged K/V on SDPA's cases is an SDPA
  variant.
• It needs a worker hook that binds and digest-checks read-only operands (block table, K/V
  lengths) in source and replay mode, and GPU qualification on gfx942 and gfx950.
• A split-KV variant also needs the worker to allow two launches per call.
• Trigger: if the rocKE team wants pinned coverage for paged K/V.

GPU status of this PR at 01e4921ab16:

• PR CI on gfx942 (both shards): conv 35 and SDPA 11 cases pass.
• Dispatch 40979046469 at
  eee5958744b:
  • gfx950: conv 35 and SDPA 10.
  • gfx1151 Linux and Windows: SDPA 11 each.
  • That run's Windows failure was outside this PR's tests.
• gfx1201 is pending in
  41128621898. Its
  gfx120X runner pool also has gfx1200 machines, where the gfx1201 test skips.

[brockhargreaves-amd]

Thanks @yraparti. The HIGH §D.1 is answered on its thread. As before, the 13 NEEDS-INFO rows concern hunks that the truncated diff left out. My 10-27 answers still hold at 01e4921ab16, with these updates from the review commits:

• §B.4. A manifest without a variant field is now accepted only as the variant whose committed lock pins its digest (1c37173). Every other variant rejects it, whichever lock is passed. test_a_pre_variant_bundle_belongs_to_the_variant_whose_lock_pins_it and test_committed_locks_pin_distinct_manifests cover this. The other rejections at the CLI, pytest and CMake boundaries are unchanged.

• §E.13, §G.16. Host tests and reference tools now name the variant they use (353e8d9, 4a4dc17). test_reference_tools_name_the_variant_they_use fails on any tool call that relies on the sole-variant default.

• §F.6. pre-commit passes at 01e4921ab16, both as the PR check and locally over the review commits.

• §F.7. Every thread from @AnaghaRaoAMD's review has a reply. The earlier Copilot and CongMa13 threads were fixed and resolved.

• §G.3. I ran the local non-GPU gates at 01e4921ab16:

  • reference host suites (conv and SDPA contracts, reference_common): 734 passed, 18 skipped;
  • test_sdpa_torch_reference.py: 5 passed;
  • platform/tests/test_gpu_references.py: 85 passed, 2 skipped.

  Each review commit also passes on its own. In PR CI at this head, both gfx94X hipkernelprovider shards pass, with 35 conv and 11 SDPA pinned GPU cases each.

• §G.6, §J.4. The doc hunks that the grader didn't see are in 93ee5b1, 0edf706 and eee5958, which are docs only, and in the doc parts of 353e8d9, 0de1b0d and 1c37173. They describe the behaviour those commits test, and they make no performance or novelty claims.

• §Q.1. platform/CMakeLists.txt gains a two-line comment pointing to the CTest trade-off in the workflow guide (eee5958). The registration itself is unchanged.

• §S.2. The four reference host test files collect 757 tests, against 742 at 964f9e9b949. There are 16 new ids, one of which replaces skips_supported_gpu_without_cohort[conv-gfx1151], now [conv-gfx1100]. No other id was removed.

[brockhargreaves-amd]

Thanks @cgmillette. On the open items:

• @AnaghaRaoAMD's threads: all ten now have replies.
• test_reference_common.py split: agreed, as a follow-up owned by the rocKE team. It's item 3 in the follow-ups comment, along with the other review follow-ups, and none of them is part of this PR.
• CongMa13's approval: it predates the review commits, so I've re-requested review.
• gfx1201: it has had no GPU run since the review commits, because the gfx120X job landed on a gfx1200 runner. 41128621898 reruns it.
• Windows gfx1151 failures in 40979046469: these are outside this PR, and they're for the rocKE team: the ISA tests that pick up Visual Studio's llvm-objdump, the GDN COMGR timeout, and rocke_pytest's CTest timeout. On that runner, rocke_reference_common_pytest (451 passed) and rocke_sdpa_gpu_gfx1151_pytest (11 passed) pass. The develop merge brings in fix(rocke): skip GDN/KDA tests on Windows CI #14710, which skips the GDN/KDA tests on Windows CI.

[yraparti]

rocKE checklist re-review of the fixes since eee5958744b0

ROCm/rocm-libraries#14164 · 01e4921ab16f · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 2, MEDIUM 0, LOW 0, NEEDS-INFO 60

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (eee5958744b0)

Resolved 0 · still open 1 · new 1

• ⏳ §D.1 still FAIL (HIGH): Put the knob on the KernelSpec and delete the environment read.
• 🆕 §B.4 new FAIL (HIGH): In both reference workers, require variant to be present for source-mode requests and pass request["variant"] to get_adapter, while leaving replay mode independent of the variant adapter.

Findings

• HIGH §B · dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46 · In both reference workers, require variant to be present for source-mode requests and pass request["variant"] to get_adapter, while leaving replay mode independent of the variant adapter.
• HIGH §D · dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 · Put the knob on the KernelSpec and delete the environment read.

Needs info

• §A.1 The changes-since list names new or modified kernel-body emitter files under library/kernels/common/ and C++ instance files under platform/cpp/instances/common/, but their hunks are omitted, so I cannot verify whether every non-exempt Python family has a synchronized C++ twin or a documented known_gaps entry.
• §A.2 The changes-since list names core IR and lowering files in both platform/python/rocke/core/ and platform/cpp/core/, but the relevant hunks are omitted, so I cannot verify mirrored op-handler/lowering behavior.
• §A.3 The changes-since list names attention_unified.py and attention_unified_selectors.cpp plus related headers, but their hunks are omitted, so I cannot verify selector/spec/lowering/codegen parity.
• §A.4 The changed Python and C++ emitter hunks are omitted, so I cannot verify exact emit order, operand binding order, or canonical IR equality.
• §A.5 The changes-since list names representative and library golden JSON changes, but their contents and the corresponding IR-changing hunks are omitted, so I cannot tell whether each re-bless is intentional and appropriate.
• §A.6 The diff shown does not provide the cross-engine byte-identity gate result or identify whether golden changes were only Python-drift goldens versus check_byte_identity coverage.
• §A.7 The hot-path kernel and lowering hunks that would show whether new logic is guarded or unconditional are omitted.
• §A.8 The current diff includes C++ engine/instance/lowering file changes in the changes-since list, but no check_byte_identity result or complete C++ hunks are provided to review by hand.

Section notes

• §A: The re-review diff is truncated and the changes-since list names production rocKE kernel, core, lowering, backend, helper, parity, and golden files whose hunks are omitted; items that depend on those hunks are marked NEEDS-INFO.
• §B: The provided diff is truncated; rows marked NEEDS-INFO identify the missing hunks needed to grade safely.
• §D: This PR reorganizes pinned GPU reference test bundles and harnesses by (operation, architecture, variant) and does not touch dispatcher registration, routing, or KernelSpec definitions.
• §D: This PR only restructures pinned GPU test reference bundles and harnesses by variant; it makes no changes to the production dispatcher, candidate registry, routing, or kernel selectors.
• §E: Re-review against 01e4921 (rebase onto develop). All PR-owned changes under reference_common, conv_reference, sdpa_reference, and platform/cmake maintain structure and hygiene invariants.
• §F: Re-evaluated against head 01e4921 (which rebases/merges upstream develop changes); all policy.yml checks remain satisfied while local pre-commit and review thread states remain external to the diff.

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	NEEDS-INFO	The changes-since list names new or modified kernel-body emitter files under library/kernels/common/ and C++ instance files under platform/cpp/instances/common/, but their hunks are omitted, so I cannot verify whether every non-exempt Python family has a synchronized C++ twin or a documented known_gaps entry.
2	NEEDS-INFO	The changes-since list names core IR and lowering files in both platform/python/rocke/core/ and platform/cpp/core/, but the relevant hunks are omitted, so I cannot verify mirrored op-handler/lowering behavior.
3	NEEDS-INFO	The changes-since list names attention_unified.py and attention_unified_selectors.cpp plus related headers, but their hunks are omitted, so I cannot verify selector/spec/lowering/codegen parity.
4	NEEDS-INFO	The changed Python and C++ emitter hunks are omitted, so I cannot verify exact emit order, operand binding order, or canonical IR equality.
5	NEEDS-INFO	The changes-since list names representative and library golden JSON changes, but their contents and the corresponding IR-changing hunks are omitted, so I cannot tell whether each re-bless is intentional and appropriate.
6	NEEDS-INFO	The diff shown does not provide the cross-engine byte-identity gate result or identify whether golden changes were only Python-drift goldens versus check_byte_identity coverage.
7	NEEDS-INFO	The hot-path kernel and lowering hunks that would show whether new logic is guarded or unconditional are omitted.
8	NEEDS-INFO	The current diff includes C++ engine/instance/lowering file changes in the changes-since list, but no check_byte_identity result or complete C++ hunks are provided to review by hand.
9	NEEDS-INFO	The changes-since list names parity mini-pair files and shared lowering/op-handler files, but the relevant hunks and test registration details are omitted, so I cannot verify the required cheap proof exists for each shared fix or arch divergence.
10	NEEDS-INFO	The changes-since list names new feature/kernel work and golden JSON changes, but the omitted hunks do not show which flag-on IR paths, LLVM flavors, and architectures are covered by golden hash entries.
11	NEEDS-INFO	Core IR, opcode documentation, C++ IR headers, verifier, and lowerers are listed as changed, but their hunks are omitted, so I cannot determine whether a new IR op was added or whether every required touchpoint was updated.
12	NEEDS-INFO	platform/python/rocke/core/backend.py is listed as changed but its hunk is omitted, so I cannot verify ROCKE_BACKEND=both behavior or that CPP_UNPORTED_ARCHES remains empty.
13	NEEDS-INFO	The changes-since list names build-id, LLVM profile, lowering, and golden changes, but the emitted .ll generation hunks and gate environment are omitted, so I cannot verify absence of stamps or debug locations.
14	NEEDS-INFO	The changes-since list names classic helper changes in both Python and C++ helper files, but most helper hunks are omitted, so I cannot verify that every changed classic helper updates its C++ twin and header consistently.

§B graded 6/6

#	Verdict	Evidence
1	NEEDS-INFO	Changes since the previous review include omitted rocKE kernel, dispatch, platform, and provider plan-builder hunks that may add new rejection gates; the full hunks are needed to verify every new rejection is classified as hardware limit versus implementation shortcut.
2	NEEDS-INFO	Changes since the previous review include omitted spec/validation and plan-builder hunks, so the full preflight and spec-constructor changes are needed to compare rejection scopes per architecture.
3	NEEDS-INFO	Changes since the previous review include omitted provider plan-builder and dispatch/routing hunks, so the full fallback-routing changes and their tests are needed to verify real model-zoo shapes are not silently demoted.
4	FAIL	The prior NEEDS-INFO is not resolved: the source-mode worker request is a trust boundary, but a missing meaningful variant field is normalized with request.get("variant"), which rewrites a malformed source request into the architecture's default/sole variant and runs that adapter instead of rejecting (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46).
5	NEEDS-INFO	The visible reference-variant gates are tested, but omitted kernel/dispatch/platform hunks in the changes since the previous review may open additional configurations; the full hunks are needed to verify each newly opened gate scope is tested or scoped down.
6	NEEDS-INFO	Changes since the previous review include omitted kernel/spec validation files that may touch is_valid_spec or supports_*; the full hunks are needed to verify those predicates remain functional-only and do not encode policy or coverage opinions.

§C graded 10/10

#	Verdict	Evidence
1	NEEDS-INFO	Re-review resolves the previous NEEDS-INFO: repeated per-operation architecture/variant lookup is factored into the shared AdapterRegistry helper in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and both operation packages delegate to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:12 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:12. — no diff line supports this verdict
2	N/A	This PR refactors pinned GPU reference test layout and staging, not kernel build_<family>(spec) builders or emitted IR.
3	PASS	Variant names are operation/algorithm family names such as implicit_gemm_fwd in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4 and dense in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:4, and CMake rejects invalid or gfx-prefixed variant names in…
4	PASS	Arch-independent registry and lookup logic lives in the common helper dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, while processor packages now only name the enrolled processor, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/__init__.py:9, and algorithm-specific adapters live under variant subpackages such as…
5	PASS	The diff fan-out matches the inherent reference-key widening across tests, bundle paths, CMake staging, and docs: default_bundle_path gains a variant component in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, CMake installs variant payloads in dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:218, and TESTING documents the new installed path in…
6	PASS	The change is factored through the shared AdapterRegistry abstraction in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22 before operation packages bind get_adapter, variants, and baseline_lock to that abstraction in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:15.
7	PASS	Re-review resolves the previous NEEDS-INFO: the bundle move preserves the DVC object hash and size while changing only the archive path, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:2 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:5, and the baseline lock retains manifest_sha256 in…
8	PASS	New functionality extends the existing shared reference tooling rather than reimplementing separate operation tools: artifact.py adds the published command in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:316, and CLI cohort/variant handling is added to ReferenceOperation in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:170.
9	PASS	The variant-source/replay split is treated as a class across sibling workers: convolution dispatches source mode through get_adapter in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:45, and SDPA applies the same source/replay guard in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:64; CMake also rejects stale pre-variant locks and pointers repo-wide in…
10	PASS	The widened shared key (operation, architecture, variant) is propagated through option parsing in dnn-providers/hip-kernel-provider/rocke/library/tests/conftest.py:55, path construction in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, pytest parametrization in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:178, CMake staging in…

§D graded 21/21

#	Verdict	Evidence
1	FAIL	dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 reads the environment
2	N/A	This PR does not add or modify any dispatcher KernelCandidate or Capability declarations.
3	N/A	This PR does not add, modify, or call any dispatcher candidate _supports or admits predicates.
4	N/A	This PR does not touch dispatcher selectors or spec builders on the selection path.
5	N/A	This PR does not modify any KernelSpec identity, kernel_name, or cache-key logic.
6	N/A	This PR does not modify cohort routing, spec builders, launch metadata, or cache keys in the dispatcher.
7	N/A	This PR does not add or modify dispatcher heuristics, hand-pinned overrides, or resolvers.
8	N/A	This PR does not define or modify any dispatcher routing cohorts or spec override predicates.
9	N/A	This PR does not register any new kernel or architecture in the dispatcher registries.
10	N/A	This PR does not introduce any new KernelSpec fields or features.
11	N/A	This PR does not register, promote, or change precedence or opt-in status for any dispatcher candidates.
12	N/A	This PR only restructures pinned GPU test reference harnesses and does not add, rename, or reprioritize any dispatcher candidates.
13	N/A	This PR does not modify CandidateRegistry.select or add any dispatcher wrappers or entry points.
14	N/A	This PR does not touch kernel compilation caches, routing logs, tuning records, or kernarg layouts.
15	N/A	This PR does not add or modify any dispatcher candidates or arch modules in the production dispatch registry.
16	N/A	This PR does not introduce a new operator family in the dispatcher.
17	N/A	This PR does not add or modify dispatcher candidates or routing rules.
18	N/A	This PR does not modify CandidateRegistry settings, binding requirements, or exempt sets.
19	N/A	This PR makes no claims about unified attention auto routing or run_unified_attention_torch dispatch.
20	N/A	This PR makes no changes to kernel-level selectors, spec builders, or the Python registry/tuning layer.
21	N/A	This PR does not fix a kernel selection or emitter issue; it reorganizes test reference bundles by variant.

§E graded 17/17

#	Verdict	Evidence
1	PASS	library/tests imports rocke (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:57), while platform/cmake/InstallGpuReferences.cmake invokes artifact.py via an isolated subprocess (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:208) without importing library into platform.
2	PASS	Device queries in the reference test infrastructure use from rocke.runtime.hip_module import get_device_target_id (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:59) rather than torch.
3	PASS	Torch remains optional: dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/check_torch_reference.py:13 gates with pytest.importorskip("torch") and pinned reference tests verify assert not report["torch_imported"] (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:210).
4	N/A	This PR only modifies files under library/tests/, platform/, and TESTING.md, not production library/{kernels,dispatch,builders,benchmarks} layering or KNOWN_VIOLATIONS/KNOWN_BAD lists.
5	PASS	Platform CMake rules use relative directory variables (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:20) and Python helpers use tempfile and pathlib (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:18).
6	N/A	This PR does not touch emitter trees (core/, helpers/, instances/, library/kernels/).
7	PASS	Testing and workflow documentation are updated alongside the reference layout changes (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/docs/conv-test-reference.md:13, dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md:321).
8	PASS	No Jira/tracking IDs, reviewer handles, or PR numbers appear in code or comments in the diff (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:1).
9	PASS	All PR-modified files are scoped to splitting pinned GPU reference bundles by operation, architecture, and variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	The PR was rebased onto current develop and explicitly guards against concurrent branches restoring pre-variant locks or DVC pointers at legacy paths (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107).
11	PASS	Only deliverable test harness code, locks, DVC pointers, CMake rules, and required reference documentation are committed (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22).
12	PASS	All enrolled architectures (gfx942, gfx950, gfx1151, gfx1200, gfx1201) are migrated to the variant layout (dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:471, dnn-providers/hip-kernel-provider/rocke/platform/cmake/PublishedGpuReferences.cmake:11).
13	NEEDS-INFO	Previously NEEDS-INFO due to diff budget truncation; cli.py, test_reference_common.py, test_gpu_references.py, and docs/*.md remain truncated or omitted in the prompt, although platform/cmake/InstallGpuReferences.cmake:61 and platform/cmake/InstallGpuReferences.cmake:107 enforce that no pre-variant locks or DVC pointers remain.
14	PASS	After rebasing onto develop, moved adapter files preserve upstream comments, historical API compatibility checks, and docstrings (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:43).
15	NEEDS-INFO	Previously NEEDS-INFO due to diff truncation; all visible new variant parameters in artifact.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), registry.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65), and the visible portion of cli.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349) are read and used, but the tail of cli.py and test_gpu_references.py are truncated.
16	N/A	This item is a meta-instruction for review grading rather than a property of the pull request diff.
17	PASS	No performance numbers or superiority claims appear in the PR description, docs, or code (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/brockhargreaves-amd/pinned-reference-variants matches users/<user>/… (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).
2	PASS	title test(rocke): split pinned GPU references by operation, architecture, and variant matches type(scope): summary and is 80 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py is in the diff
5	PASS	no key, certificate, or secret filename in the diff
6	NEEDS-INFO	The new commits since eee5958 do not include local pre-commit execution logs, which are not visible from the diff alone.
7	NEEDS-INFO	Bot and Copilot review threads and their resolution/ticket status are not included in the prompt or diff.
8	PASS	Base branch is develop (and body notes Stacked on #13651), and the branch, title, body, tests, and files satisfy policy.yml checks (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312).
9	PASS	The PR keeps the full 10-commit migration and test suite together without artificial splitting for a size cap (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	Tests added in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:341 exercise real variant discovery, cohort digesting, and CMake staging behavior rather than throwaway checks.
11	PASS	The fetched PR metadata shows the populated Title and full Motivation/Technical Details/Test Plan/Test Result Body are present (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§G graded 34/34

#	Verdict	Evidence
1	N/A	No shown change uses byte-identity or parity as a correctness proof.
2	NEEDS-INFO	Current commits introduce or wire kernel work, e.g. new convolution WMMA specs are imported in dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/direct_kernel_sweep.py:430, but the diff does not include current-head on-silicon numeric verification evidence with the required oracle/tolerance coverage.
3	NEEDS-INFO	Previous NEEDS-INFO is not resolved: missing current-head evidence that the rocKE root tools/run_checks.py sweep, the extra llvm23 byte-identity check for emission changes, and the relevant pytest/golden/IR gates were run or proven unchanged versus develop.
4	NEEDS-INFO	The visible reference-variant behavior has tests, but the current-head changes since the previous review include many omitted behavior and test hunks, so it is missing enough diff to verify every new behavior has a fail-before test rather than only substring or whitespace coverage.
5	NEEDS-INFO	The visible reference tests exercise shipped paths, but the current-head changes since the previous review include many omitted tests and internal surfaces, so it is missing enough diff to verify every new test is falsifiable and reachable through the shipped path or explicitly scoped as defense-in-depth.
6	NEEDS-INFO	Previous NEEDS-INFO is not fully resolved: multiple documentation and code-comment hunks are still omitted, including reference_common docs, sdpa docs, platform/tests/test_gpu_references.py, and large current-head platform/kernel/performance hunks, so load-bearing numeric and structural prose cannot be fully mapped to checks.
7	N/A	No shown design doc introduces a new fp32 parity-reference oracle policy to review.
8	N/A	No shown rocKE harness change adds or modifies an fp8-to-fp16 dtype fallback.
9	NEEDS-INFO	The current-head changes include omitted resource/optimization-related hunks, so it is missing enough diff to verify whether any load-bearing LDS or occupancy numbers are tool-sourced rather than hand-derived.
10	N/A	No shown rocKE diff makes a performance-result claim requiring device, ratio, shipped-script, and parity provenance validation.
11	NEEDS-INFO	Current commits introduce or wire optimized kernels, e.g. dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/direct_kernel_sweep.py:443 adds new WMMA spec classes to launch planning, but the relevant routing/selector and measurement-path hunks are omitted, so reachability on the shipped surface cannot be verified.
12	NEEDS-INFO	Current-head changes include omitted Python/C++ platform, parity, and lowering hunks, so the two-engine or cross-layer gap cannot be graded by grepping callers and confirming which engine builds the changed path.
13	N/A	No shown change recommends a mangled LLVM intrinsic symbol or overload class that would require checking a .td definition.
14	N/A	No shown rocKE diff classifies a performance bottleneck using static analysis or dynamic counters.
15	N/A	No wrong-branch risk is evident from the shown rocKE diff content for this checklist item.
16	NEEDS-INFO	Resolves the previous NEEDS-INFO: the shared registry is defined in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:21, and operation architecture packages delegate lookups to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/init.py:12 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/init.py:12. — no diff line supports this verdict
17	NEEDS-INFO	The visible artifact CLI has end-to-end subprocess coverage, but current-head changes since the previous review add or modify multiple shipped scripts in omitted hunks, so it is missing enough diff to verify every shipped script was run end-to-end rather than only parsed.
18	PASS	Variant and cohort guards are wired into bundle validation and packaging (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:136, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:163), and constructed negative tests prove they fire for mismatched variants and mixed cohort locks (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:398…
19	PASS	The fixtures explicitly create the multi-variant, unpublished, missing-bundle, candidate-bundle, and selection states needed for the new gates (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1269, dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1347).
20	N/A	This PR does not add or modify a detector banning performance or numerical claims.
21	N/A	This PR refactors pinned reference bundle layout and verification wiring, not performance or statistical measurement code.
22	PASS	The pinned harness fails selected unpublished or empty published-variant selections instead of passing without running (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:78), and CMake fails missing published archives during configuration (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:191).
23	N/A	This PR does not make performance speedup or routing A/B claims.
24	N/A	This PR does not re-bless IR goldens or use a golden re-bless as a correctness claim.
25	N/A	This PR does not contain a codegen micro-optimization.
26	N/A	This PR does not benchmark kernels or report performance measured on another target base.
27	N/A	This PR does not report an A/B result for a kernel optimization knob.
28	N/A	This PR does not report develop-to-PR performance deltas or counterfactual forced-off arms.
29	N/A	This PR does not swap guarded math primitives in a kernel.
30	N/A	This PR does not make no-spill, VGPR, scratch, or code-object resource claims.
31	N/A	This PR does not claim to fix a shipped kernel performance defect or bank conflict.
32	N/A	This PR does not report measured performance or hardware resource figures requiring provenance.
33	PASS	The changed variant resolver is explicitly enumerated by tests for enrolled, missing, ambiguous, override, and sole-variant selections (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1269), and downstream fixture outcomes are tabled across publication, bundle, option, and variant states (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1347).
34	N/A	This PR does not add or modify an opt-in kernel output flag.

§H graded 5/5

#	Verdict	Evidence
1	N/A	This PR only reorganizes test reference bundles and CMake/Python test harness infrastructure without touching kernel specs, SRDs, or launch validation.
2	N/A	This PR does not add or modify paged/indirection kernels or device table lookups.
3	N/A	This PR does not add or modify any device kernel code or compile-time tile/page divisors.
4	N/A	This PR does not touch device-side masks, bounds, or runtime sequence-length inputs.
5	N/A	This PR does not introduce or modify kernel launch inputs feeding device-side bounds or masks.

§I graded 6/6

#	Verdict	Evidence
1	N/A	This PR refactors test reference bundle layout by variant and does not perform kernel routing or candidate space measurement.
2	PASS	Invalid or half-wired states are loudly rejected with actionable error messages and recovery commands rather than silently ignored or cryptically failing (e.g., dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:39-135 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143-173).
3	N/A	This PR splits existing pinned reference cohorts by variant without altering test case generation or hypothesis property tests.
4	N/A	This PR does not generate performance tables or benchmark reports comparing shipped vs achievable kernel selections.
5	PASS	Pinned reference bundles are explicitly tracked as frozen snapshots via baseline_lock.json (recording baseline_revision, cohort_sha256, and manifest_sha256) with an explicit qualification/regen CLI path and status/diff drift checks (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349-495, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:1261-1317).
6	N/A	This PR does not add or modify any DSL ops, hardware instructions, or lowering primitives.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The modified Markdown visible in the diff (TESTING.md, conv-test-reference.md, platform/tests/README.md) contains no mathematical equations or matrix products to type-check.
2	N/A	The PR does not introduce mathematical formula notation tables or multi-domain quantities such as decay versus log-decay.
3	N/A	The visible Markdown documents GPU reference packaging and test workflows rather than restating a kernel algorithm derivation owned by another design document.
4	NEEDS-INFO	Not resolved from the previous NEEDS-INFO: the diff still omits the full Markdown hunks for dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/docs/README.md, gpu-ci-pinned-rocke-test-reference-plan.md, gpu-reference-workflow.md, quickstart.md, and sdpa_reference/docs/sdpa-test-reference.md, so novelty claims in those files cannot be verified.
5	PASS	dnn-providers/hip-kernel-provider/rocke/TESTING.md:949-950 and dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md:324-325 state the CTest timeout as a per-variant aggregate (five minutes per installed variant / five minutes per variant).
6	N/A	This PR reorganizes pinned test reference bundles and does not add model config.json tables, tuned dispatch band edges, or kernel constants requiring deployment provenance.
7	N/A	No new Markdown design or algorithm document file is added in the PR file list; existing test documentation files are updated in place.

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR does not add or modify any tuned dispatch candidates or KnobSpace subclasses.
2	N/A	This PR does not add or modify any kernel spec fields or KnobAxis declarations.
3	N/A	This PR does not change tuning_id, config_key, knob canonicalization rules, or tuned spec defaults.
4	N/A	This PR does not touch KnobSpace.defaults() or KnobSpace.recorded().
5	N/A	This PR does not modify knob dropping, refusal, inertness, or base_value() rules.
6	N/A	This PR does not modify KnobSpace.validate or policy-free tuning spec construction.
7	N/A	This PR does not touch KnobSpace.base(req) or request normalization for tuning.
8	N/A	This PR does not change normalize_knobs or knob value coercion.
9	N/A	This PR modifies GPU test reference bundle replay, not dispatch tuning_id/spec_hash benchmark replay.
10	N/A	This PR does not touch KnobSpace.find or tuning_id resolution in dispatch.
11	N/A	This PR does not touch sweep entry points, pin_to_spec, or sweep harnesses.
12	N/A	This PR does not change production stacks, accept_default, or base specs for auto tuning selection.
13	N/A	This PR does not touch benchmark or sweep harnesses or make claims about auto selection.

§L graded 10/10

#	Verdict	Evidence
1	N/A	This PR reorganizes pinned GPU test reference bundles by variant and does not add or modify any gfx hardware target or ISA backend.
2	N/A	This PR does not add or modify any MMA op_id or layout map.
3	N/A	This PR does not touch MMA selection, MfmaAtom catalogs, or kernel lowering.
4	N/A	This PR only relocates existing test adapter code and does not alter ArchTarget.wave_size or lane math.
5	N/A	This PR does not add or modify arch-specific kernel recipes or tile loaders.
6	N/A	This PR does not modify pipeline policies, ArchTarget, or rocke/ imports.
7	N/A	This PR does not add or modify logical dtypes, bit packing, instruction fragments, or FP8 dialects.
8	N/A	This PR does not touch scaled-MMA contracts or WMMA/MFMA scaled emission.
9	N/A	This PR does not introduce a new architecture target.
10	N/A	This PR does not modify intrinsic arch-domain artifacts or LLVM flavor membership lists.

§M graded 10/10

#	Verdict	Evidence
1	NEEDS-INFO	The current diff lists changes to IR/control-flow and kernel-emitting files, but the relevant hunks for Value truthiness, scf_if/scf_for result threading, and emitted runtime predicates are omitted; provide the full hunks for platform/python/rocke/core/ir.py, platform/cpp/core/ir/ir_flow.cpp, and affected kernel builder files.
2	NEEDS-INFO	The current diff lists changes to target/architecture data and lowering-related files, but the relevant hunks for legalize_for_target behavior are omitted; provide the full hunks for the legalizer/arch-data changes to verify it only binds or rejects.
3	N/A	The visible diff and listed changed files do not show changes to ck.dsl.ir serialization, print_ir interchange behavior, wire versions, attribute tagging, SSA ids, or ABI parameter ordering.
4	NEEDS-INFO	The current diff lists changes to portable IR importer, recipe VM, and recording-builder files, but their hunks are omitted; provide the full hunks for platform/cpp/portable_ir/ir_import_json.cpp, platform/cpp/portable_ir/recipe_vm.cpp, and platform/python/rocke/portable_ir/src/recording_builder.py.
5	NEEDS-INFO	The current diff lists memory-lowering and kernel-body changes, but the relevant hunks are omitted; provide the full hunks needed to check AsyncTileLoader dword limits, LDS layout validation, destination pointer swizzling, and waitcnt placement.
6	NEEDS-INFO	The current diff lists changes to control, memory, cross-lane lowering, and kernel bodies, but the relevant hunks are omitted; provide the full hunks needed to check barriers in divergent control flow, invalid-pointer loads before select, and lane_xor constraints.
7	NEEDS-INFO	The current diff lists ISA backend and buffer/OOB-related changes, but the relevant hunks are omitted; provide the full hunks for buffer resource word 3 ownership and TemporalHint validation.
8	NEEDS-INFO	The current diff lists intrinsic-domain, lowering, LLVM profile, and backend-selection related changes, but the relevant hunks are omitted; provide the full hunks for intrinsic declarations/override tables and backend precedence changes.
9	NEEDS-INFO	The current diff lists changes under cpp/include/rocke and C++ binding/core files, but the relevant hunks are omitted; provide the full hunks for public C entry points and headers to verify stable extern "C" signatures and no fabricated HIP launch behavior.
10	NEEDS-INFO	The current diff lists LLVM-profile/lowering changes but omits the relevant hunks; provide the full hunks that emit or consume debug info to verify Debug Info Version flags and DWARF consumption behavior.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The PR diff does not add or modify helpers under platform/python/rocke/helpers/.
2	N/A	The PR diff does not author or modify kernels or their use of shared helper mechanisms.
3	N/A	The PR diff does not touch SoftwarePipeline or any kernel pipeline code.
4	N/A	The PR diff does not add or modify fused kernels, epilogue fusion logic, fusion validation, or fusion legalizers.
5	N/A	The PR diff does not introduce or change any public tiling API under helpers/tiling/.
6	N/A	The PR diff does not add or modify front-door or toolbox tiling helpers or any LayoutStyle implementation.
7	N/A	The PR diff does not touch tiling layout overrides, accumulator descriptors, or MMA validation.
8	N/A	The PR diff does not touch tiling memory operations, LDS allocation, global load/store emission, clipping, padding, or wide-load selection.
9	N/A	The PR diff does not use or modify N-D tensor axis roles, batch reductions, flow-stage relabeling, or AB-swap routing in tiling.
10	N/A	The PR diff does not modify tiling emit, LDS staging, MMA issue logic, or tiling fail-fast raises.
11	N/A	The PR diff does not touch MMA traits, TileMma, lds_swizzle, scheduling group helpers, or related tiling policies.
12	N/A	The PR diff makes no LDS bank-conflict or coalescing claims and does not modify related simulator or coalescing analysis tools.
13	N/A	The PR diff does not construct or modify WarpDistributionEncoding, make_tile_desc, RegisterMapper, or tiling store casts.
14	N/A	The PR diff does not use tiling verbs, modify synchronization or barrier placement, or add a tiling pipeline variant.
15	N/A	The PR diff does not touch tiling allocation or bounds-checking logic for transformed address ranges.
16	N/A	The PR diff does not use or modify optimize_layout, recommend, or evaluate_transform.
17	N/A	The PR diff does not touch the tiling recorder, round-trip analysis, witness flags, or visualization modules.
18	N/A	The PR diff does not touch interleaved layouts, tiling demo kernels, dispatch registration of tiling demos, or tiling numeric gates.

§O graded 16/16

#	Verdict	Evidence
1	N/A	This PR does not touch core/verify.py, cpp/core/verify.cpp, or IR address spaces/attributes.
2	N/A	This PR does not modify scf.for, region structure, or IR block emission.
3	N/A	This PR does not modify IRBuilder, LLVM intrinsic declaration tables, or metadata emission.
4	N/A	This PR does not modify IR types, SmemType, or IR parsers.
5	N/A	This PR does not modify core/passes.py, cpp/core/passes.cpp, or op purity tables.
6	N/A	This PR does not touch constant folding in Python or C++ passes.
7	N/A	This PR does not modify kernel attributes, COMGR options, or cluster dimension checks.
8	N/A	This PR does not touch CodegenPolicy, SchedulerStrategy, or kernel compilation options.
9	N/A	This PR does not modify waitcnt encoding or barrier lowering.
10	N/A	This PR does not touch LLVM lowerers or parameter attribute/type lowering.
11	N/A	This PR does not touch ArchTarget or target catalog resolution.
12	N/A	This PR does not touch vector/LDS width caps or codegen arch defaults.
13	N/A	This PR does not touch public C headers under cpp/include/rocke/ or ABI versions.
14	N/A	This PR does not touch C++ engine sources, headers, bindings, or C++ compile flags in platform/CMakeLists.txt.
15	N/A	This PR does not touch C++ engine fallback handling or alternative lowerers.
16	N/A	This PR does not touch known_gaps.md or C++ library/builder capabilities.

§P graded 14/14

#	Verdict	Evidence
1	NEEDS-INFO	Missing hunks for the changed builder/instance files and the updated builder signature contract tests are needed to verify that all builders still take exactly (spec, *, arch) and that REQUIRED_FIELDS was not changed incompatibly.
2	NEEDS-INFO	Missing hunks for the changed conv ABI files, instance_conv_abi.h, packing/argument code, and parity emitters are needed to verify kernarg order and ABI-version parity.
3	NEEDS-INFO	Missing hunks for the changed launch-geometry and host harness code are needed to verify GEMM/conv grid single-sourcing and that harnesses do not recompute grids incorrectly.
4	NEEDS-INFO	Missing hunks for the changed LDS/admission gate files and C++ validator files are needed to verify the gate charges emitted allocations and uses ArchTarget capacity/fits_lds with C++ parity.
5	NEEDS-INFO	Missing hunks for the changed WMMA/RDNA GEMM, direct-conv WMMA, GemmMultiD/MultiAbd, and validator/spec files are needed to verify narrow-path restrictions and wave-size handling.
6	NEEDS-INFO	Missing hunks for the changed split-K/Stream-K/workspace-related kernels, helpers, launchers, and zeroing tests are needed to verify caller-zeroed workspace sizing and sentinel-zeroing coverage.
7	NEEDS-INFO	Missing hunks for SUPPORT_MATRIX.md, Stream-K sweep/launch code, grouped GEMM, and persistent/split-K admission guards are needed to verify known-rejected configurations remain rejected.
8	NEEDS-INFO	Missing hunks for the changed quantization, MX scale, MX GEMM, and decode helper files are needed to verify use of shared quant helpers, MX constraints, and edge decode behavior.
9	NEEDS-INFO	Missing hunks for the changed numeric emulation/decode code and edge-case tests are needed to verify finite overflow, Inf, NaN, denormal, and rounding-midpoint coverage in both engines.
10	N/A	This diff changes pinned GPU reference bundle metadata and descriptor-packaging plumbing, not rocKE runtime manifests under helpers/manifest.py or ck.dsl.example.manifest/v1.
11	NEEDS-INFO	Missing hunks for the changed runtime launcher, hip_module, stall_gate, and timing code are needed to verify buffer lifetime, stream synchronization, fences, WorkspacePool use, and HIP-event timing.
12	NEEDS-INFO	Missing hunks for the changed runtime/comgr.py and runtime/hip_module.py files are needed to verify COMGR/HIP loading does not map a second LLVM or import torch only to find ROCm libraries.
13	NEEDS-INFO	Missing hunks for the changed host-side launch math, kernel-cache, split-K, Stream-K, conv, and admission-guard code plus the corresponding test-plan documentation are needed to verify every consumer's numeric tests cover the refactor.
14	NEEDS-INFO	Missing hunks for the changed examples and any expected.json/performance-bound files are needed to verify correctness gates are preserved and no disallowed performance lower bound or misplaced kernel logic was added.

§Q graded 7/7

#	Verdict	Evidence
1	NEEDS-INFO	The diffs for dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt (+7 -2 lines) and dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py (+485 -80 lines) are still omitted to fit the model budget after the rebase, so CTest registration and installed test changes cannot be verified.
2	N/A	This PR refactors the GPU reference test bundle layout and does not add a new kernel family or variant to the platform IR or builders.
3	PASS	Tests and documentation grade the reference harness accurately, enforcing failure controls for perturbed outputs and missing launches in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:168-276 and noting subset verification caveats in dnn-providers/hip-kernel-provider/rocke/TESTING.md:154-156.
4	N/A	This PR does not re-bless IR goldens or widen the valid spec space in a lowerer or builder.
5	PASS	Tests remain portable across CPU/GPU and OS environments, skipping cleanly when hardware or bundles are absent (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:76-160) and querying the active device processor dynamically (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pytest_support.py:80-115).
6	PASS	The harness refactor preserves per-configuration semantics, scaling timeouts per installed variant (dnn-providers/hip-kernel-provider/rocke/TESTING.md:949-950), keeping at most one slot's workers alive at a time (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:82-163), and preserving error classes (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:227-232).
7	N/A	This PR does not modify engine backend rejection contracts or ROCKE_BACKEND / ROCKE_CPP_STRICT fallback behavior.

§R graded 11/11

#	Verdict	Evidence
1	N/A	This PR only restructures pinned GPU test reference harnesses and bundles and changes no kernel body, algorithm, tiling strategy, pipeline variant, or performance lever.
2	N/A	No performance lever or architecture-specific optimization recipe is modified in this test-reference refactoring PR.
3	N/A	No optimization lever covered by the utilities skills is touched by this test-infrastructure PR.
4	N/A	The diff does not alter kernel codegen, lowering, or performance primitives, so static ISA and occupancy probe evidence does not apply.
5	N/A	No global memory accesses in GPU kernels are added or modified by this diff.
6	N/A	No LDS or MMA layouts are added or changed in this pull request.
7	N/A	The PR makes no kernel bottleneck claims and implements no performance fixes.
8	N/A	No waits, barriers, or scheduling hints in GPU kernels are changed.
9	N/A	This PR is not a performance regression fix and changes no shape-dependent performance lever.
10	N/A	No compiler codegen changes, assembly parity claims, or performance benchmarks are included in this diff.
11	N/A	This PR introduces no new GPU optimization lessons, probes, or analysis tools.

§S graded 8/8

#	Verdict	Evidence
1	N/A	The diff does not add a public rocke package symbol, spec/builder/grid helper, compile_kernel/KernelLauncher/time_launches parameter, run_manifest option, or rocke_* C entry point.
2	NEEDS-INFO	The previous NEEDS-INFO is resolved: existing pinned-reference cohort guards are updated to use variant adapters without dropping coverage at dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py:406 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:64. — no diff line supports this verdict
3	N/A	No public rocke package or C engine API is removed or renamed; the moved reference adapters are library test-harness internals.
4	N/A	No __all__ list in rocke, rocke.core, rocke.runtime, rocke.analysis, or rocke.helpers.tiling is edited by this diff.
5	PASS	The updated tests assert contracts and failure behavior, including variant mismatch rejection at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:398 and cohort digest binding at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:537.
6	PASS	CPU-checkable reference contracts are tested in host pytest such as the standard-library published-listing test at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:309, while launch/numeric pinned checks are explicitly GPU-marked at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:184.
7	N/A	The diff does not add or modify a platform API with a C++ twin or a C ABI binding.
8	NEEDS-INFO	Missing diff-citable evidence that the README entry points from rocke/platform/ were run at the current head, including tests/test_rocke.py and tests/run_all.py, and any applicable manifest/runtime verification.

[yraparti]

rocKE checklist re-review of the fixes since 01e4921ab16f

ROCm/rocm-libraries#14164 · e49fb56d9f82 · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 2, MEDIUM 0, LOW 0, NEEDS-INFO 28

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (01e4921ab16f)

Resolved 0 · still open 2 · new 0

• ⏳ §B.4 still FAIL (HIGH): In dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:65, replace request.get("variant") with request["variant"] and require isinstance(request["variant"], str) in get_adapter or run for source-mode requests.
• ⏳ §D.1 still FAIL (HIGH): Put the knob on the KernelSpec and delete the environment read.

Findings

• HIGH §B · dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46 · In dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:65, replace request.get("variant") with request["variant"] and require isinstance(request["variant"], str) in get_adapter or run for source-mode requests.
• HIGH §D · dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 · Put the knob on the KernelSpec and delete the environment read.

Needs info

• §C.1 Re-review resolves the previous NEEDS-INFO: repeated per-operation architecture/variant lookup is factored into the shared AdapterRegistry helper in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and both operation packages delegate to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:10 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:10. — no diff line supports this verdict
• §E.13 The rebase since the previous review did not modify the PR's files, and cli.py (truncated at line 1445), test_reference_common.py (truncated at line 1676), test_gpu_references.py, and reference_common/docs/*.md are still truncated or omitted in the prompt, although dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61 and dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107 enforce that no pre-variant locks or DVC pointers remain.
• §E.15 The rebase since the previous review did not modify the PR's files; all visible new variant parameters in artifact.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), registry.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65), and the visible portion of cli.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349) are read and used, but the tail of cli.py (after line 1445) and test_gpu_references.py remain truncated/omitted.
• §F.6 The new commits since 01e4921 do not include local pre-commit execution logs, which cannot be verified from the diff alone.
• §F.7 The changes since 01e4921 do not include GitHub review thread metadata, so bot/Copilot thread resolution and deferred ticket IDs cannot be verified.
• §G.3 Previous NEEDS-INFO is not resolved: although the PR no longer modifies kernel emission, the PR test plan and diff still do not provide evidence that the rocKE root tools/run_checks.py gate sweep was run or proven unchanged versus develop.
• §G.4 The rebase removed the unrelated kernel files, and visible new behaviors have fail-before tests (e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312), but the tail of dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py, the tail of dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py, and all of dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py are omitted from the diff.
• §G.5 Visible tests in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:338 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:482 are falsifiable and exercise shipped entry points, but the truncated portion of dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py and the omitted dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py prevent verifying all added tests.

Section notes

• §A: The previous NEEDS-INFO grades were caused by changes-since history from rebasing onto develop; against the current PR diff (Files: list), this PR only reorganizes pinned GPU reference test harnesses, locks, DVC pointers, CMake staging, and docs, and does not touch any kernel emitters, selectors, core IR/lowering, helpers, or IR goldens.
• §D: This PR only restructures pinned GPU test reference harnesses by (operation, architecture, variant) and does not modify production dispatcher registration, routing, or kernel selection.
• §E: The branch was rebased onto develop; the PR-specific diff continues to pass all verifiable checks, while items 13 and 15 remain NEEDS-INFO due to prompt budget truncation of cli.py, test_reference_common.py, test_gpu_references.py, and docs/*.md.
• §G: Re-reviewed against e49fb56 after rebase onto develop; the PR diff now contains only pinned GPU reference variant harness changes.
• §G: Re-review includes a truncated changes-since-previous section with many omitted rocKE benchmark, dispatch, kernel, golden, and LSE hunks; items depending on those missing hunks are marked NEEDS-INFO.
• §H: No device-side masks, bounds, indirection tables, or kernel implementations are added or modified by this PR; all changes in the PR diff are confined to pinned GPU reference test harness layout, CMake staging, DVC pointers, and documentation.

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	N/A	Resolved from previous NEEDS-INFO: the PR diff against base (dnn-providers/hip-kernel-provider/rocke/library/tests/... and platform/cmake/InstallGpuReferences.cmake) touches only pinned GPU reference test…
2	N/A	Resolved from previous NEEDS-INFO: no core/lower_* or core/ir files are modified in this PR's file list.
3	N/A	Resolved from previous NEEDS-INFO: this PR does not modify any kernel selectors, spec builders, lowering, or codegen knobs.
4	N/A	Resolved from previous NEEDS-INFO: this PR modifies no Python or C++ IR emitters.
5	N/A	Resolved from previous NEEDS-INFO: no representative-IR or library IR golden files are touched in this PR (only test reference baseline_lock.json and .dvc files are moved/updated).
6	N/A	Resolved from previous NEEDS-INFO: this PR makes no IR or engine changes requiring parity or golden proof.
7	N/A	Resolved from previous NEEDS-INFO: no kernel or lowering logic is added or modified on any IR emission path.
8	N/A	Resolved from previous NEEDS-INFO: no C++ rocKE engine, lowering, helper, or instance code is touched in this PR.
9	N/A	Resolved from previous NEEDS-INFO: no shared op-handler fix or lowerer arch divergence is introduced in this PR.
10	N/A	Resolved from previous NEEDS-INFO: no new kernel or IR feature flag is introduced in this PR.
11	N/A	Resolved from previous NEEDS-INFO: no new IR op is added in this PR.
12	N/A	Resolved from previous NEEDS-INFO: core/backend.py and ROCKE_BACKEND handling are not touched by this PR.
13	N/A	Resolved from previous NEEDS-INFO: .ll emission and IR gate environments are not modified by this PR.
14	N/A	Resolved from previous NEEDS-INFO: no classic helper files under platform/python/rocke/helpers/ or platform/cpp/helpers/ are modified by this PR.

§B graded 6/6

#	Verdict	Evidence
1	N/A	The PR diff Against base (develop) only modifies test reference infrastructure, CMake staging, and docs, and does not add any kernel or dispatcher rejection gates (the broader commit range since the previous review is…
2	N/A	This PR does not touch any kernel spec constructors or preflight gates.
3	N/A	This PR only reorganizes pinned GPU reference test bundles by variant and does not change kernel routing or fallback behavior.
4	FAIL	The previous FAIL is not resolved: in source mode, conv_reference/worker.py (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46) and sdpa_reference/worker.py (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:65) still use request.get("variant"), which normalizes a malformed source request missing variant into None and causes get_adapter (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:72) to silently fall back to the architecture's sole variant instead of rejecting.
5	PASS	The previous NEEDS-INFO was caused by rebase commits from develop; in the actual PR diff, the new variant-selection, cohort-digest, and CMake reference-wiring gates (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:77) are covered by host tests…
6	N/A	This PR does not add or modify any is_valid_spec or supports_* predicates.

§C graded 10/10

#	Verdict	Evidence
1	NEEDS-INFO	Re-review resolves the previous NEEDS-INFO: repeated per-operation architecture/variant lookup is factored into the shared AdapterRegistry helper in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and both operation packages delegate to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:10 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:10. — no diff line supports this verdict
2	N/A	This PR refactors pinned GPU reference test layout and staging, not kernel build_<family>(spec) builders or emitted IR.
3	PASS	Variant names are operation/algorithm family names such as implicit_gemm_fwd in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4 and dense in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:4, and CMake rejects invalid or gfx-prefixed variant names in…
4	PASS	Arch-independent registry and lookup logic lives in the common helper dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, while processor packages now only name the enrolled processor, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/__init__.py:9, and algorithm-specific adapters live under variant subpackages such as…
5	PASS	The diff fan-out matches the inherent reference-key widening across tests, bundle paths, CMake staging, and docs: default_bundle_path gains a variant component in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, CMake installs variant payloads in dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:218, and TESTING documents the new installed path in…
6	PASS	The change is factored through the shared AdapterRegistry abstraction in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22 before operation packages bind get_adapter, variants, and baseline_lock to that abstraction in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:15.
7	PASS	The bundle move preserves the DVC object hash and size while changing only the archive path, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:2 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:5, and the baseline lock retains manifest_sha256 in…
8	PASS	New functionality extends the existing shared reference tooling rather than reimplementing separate operation tools: artifact.py adds the published command in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:316, and CLI cohort/variant handling is added to ReferenceOperation in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:170.
9	PASS	The variant-source/replay split is treated as a class across sibling workers: convolution dispatches source mode through get_adapter in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:45, and SDPA applies the same source/replay guard in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:64; CMake also rejects stale pre-variant locks and pointers repo-wide in…
10	PASS	The widened shared key (operation, architecture, variant) is propagated through option parsing in dnn-providers/hip-kernel-provider/rocke/library/tests/conftest.py:55, path construction in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, pytest parametrization in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:178, CMake staging in…

§D graded 21/21

#	Verdict	Evidence
1	FAIL	dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 reads the environment
2	N/A	This PR does not add or modify any dispatcher KernelCandidate or Capability declarations.
3	N/A	This PR does not add, modify, or call any dispatcher candidate _supports or admits predicates.
4	N/A	This PR does not touch dispatcher selectors or spec builders on the selection path.
5	N/A	This PR does not modify any KernelSpec identity, kernel_name, or cache-key logic.
6	N/A	This PR does not modify cohort routing, spec builders, launch metadata, or cache keys in the dispatcher.
7	N/A	This PR does not add or modify dispatcher heuristics, hand-pinned overrides, or resolvers.
8	N/A	This PR does not define or modify any dispatcher routing cohorts or spec override predicates.
9	N/A	This PR does not register any new kernel or architecture in the dispatcher registries.
10	N/A	This PR does not introduce any new KernelSpec fields or features.
11	N/A	This PR does not register, promote, or change precedence or opt-in status for any dispatcher candidates.
12	N/A	This PR only restructures pinned GPU test reference harnesses and does not add, rename, or reprioritize any dispatcher candidates.
13	N/A	This PR does not modify CandidateRegistry.select or add any dispatcher wrappers or entry points.
14	N/A	This PR does not touch kernel compilation caches, routing logs, tuning records, or kernarg layouts.
15	N/A	This PR does not add or modify any dispatcher candidates or arch modules in the production dispatch registry.
16	N/A	This PR does not introduce a new operator family in the dispatcher.
17	N/A	This PR does not add or modify dispatcher candidates or routing rules.
18	N/A	This PR does not modify CandidateRegistry settings, binding requirements, or exempt sets.
19	N/A	This PR makes no claims about unified attention auto routing or run_unified_attention_torch dispatch.
20	N/A	This PR makes no changes to kernel-level selectors, spec builders, or the Python registry/tuning layer.
21	N/A	This PR does not fix a kernel selection or emitter issue; it reorganizes test reference bundles by variant.

§E graded 17/17

#	Verdict	Evidence
1	PASS	library/tests imports rocke (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:57), while platform/cmake/InstallGpuReferences.cmake invokes artifact.py via an isolated subprocess (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:208) without importing library into platform.
2	PASS	Device queries in the reference test infrastructure use from rocke.runtime.hip_module import get_device_target_id (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:59) rather than torch.
3	PASS	Torch remains optional: dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/check_torch_reference.py:13 gates with pytest.importorskip("torch") and pinned reference tests verify assert not report["torch_imported"] (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:210).
4	N/A	This PR only modifies files under library/tests/, platform/, and TESTING.md, not production library/{kernels,dispatch,builders,benchmarks} layering or KNOWN_VIOLATIONS/KNOWN_BAD lists.
5	PASS	Platform CMake rules use relative directory variables (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:20) and Python helpers use tempfile and pathlib (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:18).
6	N/A	This PR does not touch emitter trees (core/, helpers/, instances/, library/kernels/).
7	PASS	Testing and workflow documentation are updated alongside the reference layout changes (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/docs/conv-test-reference.md:13, dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md:338).
8	PASS	No Jira/tracking IDs, reviewer handles, or PR numbers appear in code or comments in the diff (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:1).
9	PASS	All PR-modified files are scoped to splitting pinned GPU reference bundles by operation, architecture, and variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	The PR was rebased onto current develop and explicitly guards against concurrent branches restoring pre-variant locks or DVC pointers at legacy paths (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107).
11	PASS	Only deliverable test harness code, locks, DVC pointers, CMake rules, and required reference documentation are committed (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22).
12	PASS	All enrolled architectures (gfx942, gfx950, gfx1151, gfx1200, gfx1201) are migrated to the variant layout (dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:483, dnn-providers/hip-kernel-provider/rocke/platform/cmake/PublishedGpuReferences.cmake:11).
13	NEEDS-INFO	The rebase since the previous review did not modify the PR's files, and cli.py (truncated at line 1445), test_reference_common.py (truncated at line 1676), test_gpu_references.py, and reference_common/docs/*.md are still truncated or omitted in the prompt, although dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61 and dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107 enforce that no pre-variant locks or DVC pointers remain.
14	PASS	After rebasing onto develop, moved adapter files preserve upstream comments, historical API compatibility checks, and docstrings (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:43).
15	NEEDS-INFO	The rebase since the previous review did not modify the PR's files; all visible new variant parameters in artifact.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), registry.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65), and the visible portion of cli.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349) are read and used, but the tail of cli.py (after line 1445) and test_gpu_references.py remain truncated/omitted.
16	N/A	This item is a meta-instruction for review grading rather than a property of the pull request diff.
17	PASS	No performance numbers or superiority claims appear in the PR description, docs, or code (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/brockhargreaves-amd/pinned-reference-variants matches users/<user>/… (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).
2	PASS	title test(rocke): split pinned GPU references by operation, architecture, and variant matches type(scope): summary and is 80 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py is in the diff
5	PASS	no key, certificate, or secret filename in the diff
6	NEEDS-INFO	The new commits since 01e4921 do not include local pre-commit execution logs, which cannot be verified from the diff alone.
7	NEEDS-INFO	The changes since 01e4921 do not include GitHub review thread metadata, so bot/Copilot thread resolution and deferred ticket IDs cannot be verified.
8	PASS	Base branch is develop (body notes Stacked on #13651), and the branch prefix, Conventional Commit title, description with issue ref, unit tests, and absence of forbidden files satisfy policy.yml (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312).
9	PASS	The full variant-keyed reference migration and test suite remain together in one PR without artificial splitting for a size cap (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	Tests in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:353 exercise real variant discovery, cohort digesting, and CMake staging invariants rather than throwaway checks.
11	PASS	The fetched PR metadata shows the populated Title and complete Motivation/Technical Details/Test Plan/Test Result Body are intact (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§G graded 34/34

#	Verdict	Evidence
1	N/A	No change in this PR uses byte-identity or parity as a correctness proof.
2	N/A	Resolves the previous NEEDS-INFO: after rebasing onto develop, this PR only reorganizes pinned GPU reference test harnesses and bundles (e.g.…
3	NEEDS-INFO	Previous NEEDS-INFO is not resolved: although the PR no longer modifies kernel emission, the PR test plan and diff still do not provide evidence that the rocKE root tools/run_checks.py gate sweep was run or proven unchanged versus develop.
4	NEEDS-INFO	The rebase removed the unrelated kernel files, and visible new behaviors have fail-before tests (e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312), but the tail of dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py, the tail of dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py, and all of dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py are omitted from the diff.
5	NEEDS-INFO	Visible tests in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:338 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:482 are falsifiable and exercise shipped entry points, but the truncated portion of dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py and the omitted dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py prevent verifying all added tests.
6	NEEDS-INFO	Previous NEEDS-INFO is not resolved: multiple documentation and test hunks in this PR are omitted from the diff, including dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/docs/README.md, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/docs/gpu-reference-workflow.md, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/docs/quickstart.md, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/docs/sdpa-test-reference.md, and dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py.
7	N/A	This PR does not introduce a new fp32 parity-reference oracle policy in a design doc.
8	N/A	No harness change in this PR adds or modifies a dtype fallback such as fp8 to fp16.
9	N/A	Resolves the previous NEEDS-INFO: after rebasing onto develop, this PR contains only reference test harness changes and no LDS or occupancy numbers.
10	N/A	This PR makes no kernel performance claims requiring device, ratio, script, or parity provenance.
11	N/A	Resolves the previous NEEDS-INFO: after rebasing onto develop, this PR does not introduce or wire any optimized kernels or benchmark specs.
12	N/A	Resolves the previous NEEDS-INFO: after rebasing onto develop, this PR only touches the Python/CMake pinned reference test harness and introduces no two-engine or cross-layer kernel/platform changes.
13	N/A	This PR does not touch LLVM intrinsics or recommend mangled intrinsic symbols.
14	N/A	This PR does not classify any performance bottleneck.
15	N/A	No wrong-branch risk applies because the PR diff is provided directly from the head commit.
16	NEEDS-INFO	Resolves the previous NEEDS-INFO: the shared AdapterRegistry is defined in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22 and wired into both operation packages in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/init.py:12 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/init.py:12. — no diff line supports this verdict
17	PASS	Resolves the previous NEEDS-INFO: after rebasing onto develop, the only modified shipped CLI scripts are exercised end-to-end via subprocess in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:316 and CMake execute_process in dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:206.
18	PASS	Variant and cohort guards are wired into bundle authentication (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:136, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:157), and negative tests construct mismatched variant/cohort failures (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:398…
19	PASS	The fixtures now create multi-variant, ambiguous, override, sole-variant, published, unpublished, missing-bundle, and candidate-bundle states needed for the gates (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1269, dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1593).
20	NEEDS-INFO	The changes-since-previous section lists omitted claim/support-checking script hunks, so the diff is missing the detector logic needed to determine whether any claim ban sweeps words rather than only digits.
21	NEEDS-INFO	The changes-since-previous section lists omitted rocKE benchmark and performance-harness hunks, so the diff is missing the measurement-code changes needed to verify accepted inputs, measured/modelled labeling, false-positive behavior, and tolerance semantics.
22	PASS	The selected pinned-reference path fails unpublished or empty selected variants instead of silently passing (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:76), and CMake fails missing published variant archives during staging (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:194).
23	NEEDS-INFO	The changes-since-previous section lists omitted benchmark, routing, dispatch, and documentation hunks, so the diff is missing the perf/routing claims needed to verify configuration labeling and oracle comparisons.
24	NEEDS-INFO	The changes-since-previous section lists many omitted rocKE golden JSON hunks, so the diff is missing the re-blessed hash entries and validation evidence needed to assess golden-scope GPU coverage.
25	NEEDS-INFO	The changes-since-previous section lists omitted kernel/codegen and benchmark hunks, so the diff is missing the per-kernel resource deltas and corpus A/B evidence needed to assess any codegen micro-optimization.
26	NEEDS-INFO	The changes-since-previous section lists omitted benchmark and performance-documentation hunks, so the diff is missing the measurement provenance needed to verify the benchmark worktree was on the PR target base.
27	NEEDS-INFO	The changes-since-previous section lists omitted benchmark, dispatch, and kernel-tuning hunks, so the diff is missing the A/B toggle details needed to verify only the shipped knob was toggled.
28	NEEDS-INFO	The changes-since-previous section lists omitted performance and routing hunks, so the diff is missing any develop-to-PR or counterfactual tables needed to verify zero-delta and forced-off-arm handling.
29	NEEDS-INFO	The changes-since-previous section lists omitted kernel math hunks, so the diff is missing the primitive-use sites needed to determine whether any guarded fast-primitive swap respects its precondition.
30	NEEDS-INFO	The changes-since-previous section lists omitted benchmark/resource-reporting hunks, so the diff is missing any no-spill or resource claims and code-object evidence needed for assessment.
31	NEEDS-INFO	The changes-since-previous section lists omitted performance-defect, dispatch, kernel, and benchmark hunks, so the diff is missing the shipped-baseline A/B evidence needed to verify any defect was real before the PR.
32	NEEDS-INFO	The changes-since-previous section lists omitted benchmark and measurement hunks, so the diff is missing the reproducible provenance and compile/run target confirmation for measured figures.
33	NEEDS-INFO	The shown variant resolver is differentially exercised across ambiguous, override, and sole-variant states (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1269), but the changes-since-previous section lists many omitted dispatch/resolver hunks, so the diff is missing the full set of changed pure decision functions and their differential tables.
34	NEEDS-INFO	The changes-since-previous section lists omitted LSE/output-flag and attention-output hunks, so the diff is missing the flag-on/off output comparisons needed to verify pre-existing outputs are unchanged.

§H graded 5/5

#	Verdict	Evidence
1	N/A	This PR only reorganizes test reference bundles and CMake/Python test harness infrastructure without touching kernel specs, SRDs, or launch validation.
2	N/A	This PR does not add or modify paged/indirection kernels or device table lookups.
3	N/A	This PR does not add or modify any device kernel code or compile-time tile/page divisors.
4	N/A	This PR does not touch device-side masks, bounds, or runtime sequence-length inputs.
5	N/A	This PR does not introduce or modify kernel launch inputs feeding device-side bounds or masks.

§I graded 6/6

#	Verdict	Evidence
1	N/A	This PR refactors test reference bundle layout by variant and does not perform kernel routing or candidate space measurement.
2	PASS	Invalid or half-wired states are loudly rejected with actionable error messages and recovery commands rather than silently ignored or cryptically failing (e.g., dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:39-135 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143-173).
3	N/A	This PR splits existing pinned reference cohorts by variant without altering test case generation or hypothesis property tests.
4	N/A	This PR does not generate performance tables or benchmark reports comparing shipped vs achievable kernel selections.
5	PASS	Pinned reference bundles are explicitly tracked as frozen snapshots via baseline_lock.json (recording baseline_revision, cohort_sha256, and manifest_sha256) with an explicit qualification/regen CLI path and status/diff drift checks (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349-495, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:1261-1317).
6	N/A	This PR does not add or modify any DSL ops, hardware instructions, or lowering primitives.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The modified Markdown visible in the diff (TESTING.md, conv-test-reference.md, platform/tests/README.md) contains no mathematical equations or matrix products to type-check.
2	N/A	The PR does not introduce mathematical formula notation tables or multi-domain quantities such as decay versus log-decay.
3	N/A	The visible Markdown documents GPU reference packaging and test workflows rather than restating a kernel algorithm derivation owned by another design document.
4	NEEDS-INFO	Not resolved from the previous NEEDS-INFO: the rebase commit (01e4921..e49fb56) did not make the omitted PR Markdown hunks visible (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/docs/README.md, gpu-ci-pinned-rocke-test-reference-plan.md, gpu-reference-workflow.md, quickstart.md, and sdpa_reference/docs/sdpa-test-reference.md remain omitted to fit the model budget), so novelty claims in those files still cannot be verified.
5	PASS	dnn-providers/hip-kernel-provider/rocke/TESTING.md:1010-1011 and dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md:341-342 state the CTest timeout as a per-variant aggregate (five minutes per installed variant / five minutes per variant).
6	N/A	This PR reorganizes pinned test reference bundles and does not add model config.json tables, tuned dispatch band edges, or kernel constants requiring deployment provenance.
7	N/A	No new Markdown design or algorithm document file is added in the PR file list; existing test documentation files are updated in place.

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR does not add or modify any tuned dispatch candidates or KnobSpace subclasses.
2	N/A	This PR does not add or modify any kernel spec fields or KnobAxis declarations.
3	N/A	This PR does not change tuning_id, config_key, knob canonicalization rules, or tuned spec defaults.
4	N/A	This PR does not touch KnobSpace.defaults() or KnobSpace.recorded().
5	N/A	This PR does not modify knob dropping, refusal, inertness, or base_value() rules.
6	N/A	This PR does not modify KnobSpace.validate or policy-free tuning spec construction.
7	N/A	This PR does not touch KnobSpace.base(req) or request normalization for tuning.
8	N/A	This PR does not change normalize_knobs or knob value coercion.
9	N/A	This PR modifies GPU test reference bundle replay, not dispatch tuning_id/spec_hash benchmark replay.
10	N/A	This PR does not touch KnobSpace.find or tuning_id resolution in dispatch.
11	N/A	This PR does not touch sweep entry points, pin_to_spec, or sweep harnesses.
12	N/A	This PR does not change production stacks, accept_default, or base specs for auto tuning selection.
13	N/A	This PR does not touch benchmark or sweep harnesses or make claims about auto selection.

§L graded 10/10

#	Verdict	Evidence
1	N/A	This PR reorganizes pinned GPU test reference bundles by variant and does not add or modify any gfx hardware target or ISA backend.
2	N/A	This PR does not add or modify any MMA op_id or layout map.
3	N/A	This PR does not touch MMA selection, MfmaAtom catalogs, or kernel lowering.
4	N/A	This PR only relocates existing test adapter code and does not alter ArchTarget.wave_size or lane math.
5	N/A	This PR does not add or modify arch-specific kernel recipes or tile loaders.
6	N/A	This PR does not modify pipeline policies, ArchTarget, or rocke/ imports.
7	N/A	This PR does not add or modify logical dtypes, bit packing, instruction fragments, or FP8 dialects.
8	N/A	This PR does not touch scaled-MMA contracts or WMMA/MFMA scaled emission.
9	N/A	This PR does not introduce a new architecture target.
10	N/A	This PR does not modify intrinsic arch-domain artifacts or LLVM flavor membership lists.

§M graded 10/10

#	Verdict	Evidence
1	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff only modifies test reference files under rocke/library/tests/ and rocke/platform/cmake/, with no changes to core/ IR, builders, or kernel…
2	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not touch legalize_for_target, architecture data, or lowering files.
3	N/A	The PR diff does not touch ck.dsl.ir serialization, print_ir, wire versions, attribute tagging, SSA ids, or ABI parameter ordering.
4	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not touch portable IR schemas, importers, recipe expansion/VM, or recording builders.
5	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not touch AsyncTileLoader, LDS layouts, or kernel memory lowering.
6	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not emit or modify kernel control flow, barriers, memory loads, or cross-lane shuffles.
7	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not touch ISA backends, buffer resource word 3, or TemporalHint handling.
8	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not add intrinsic declarations, modify LLVM profile tables, or change backend selection precedence.
9	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not touch public C entry points, headers under cpp/include/rocke/, or C++ host launch behavior.
10	N/A	Resolved since the previous review by rebasing onto develop: the current PR diff does not emit or consume LLVM debug info (!dbg) or DWARF metadata.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The PR diff does not add or modify helpers under platform/python/rocke/helpers/.
2	N/A	The PR diff does not author or modify kernels or their use of shared helper mechanisms.
3	N/A	The PR diff does not touch SoftwarePipeline or any kernel pipeline code.
4	N/A	The PR diff does not add or modify fused kernels, epilogue fusion logic, fusion validation, or fusion legalizers.
5	N/A	The PR diff does not introduce or change any public tiling API under helpers/tiling/.
6	N/A	The PR diff does not add or modify front-door or toolbox tiling helpers or any LayoutStyle implementation.
7	N/A	The PR diff does not touch tiling layout overrides, accumulator descriptors, or MMA validation.
8	N/A	The PR diff does not touch tiling memory operations, LDS allocation, global load/store emission, clipping, padding, or wide-load selection.
9	N/A	The PR diff does not use or modify N-D tensor axis roles, batch reductions, flow-stage relabeling, or AB-swap routing in tiling.
10	N/A	The PR diff does not modify tiling emit, LDS staging, MMA issue logic, or tiling fail-fast raises.
11	N/A	The PR diff does not touch MMA traits, TileMma, lds_swizzle, scheduling group helpers, or related tiling policies.
12	N/A	The PR diff makes no LDS bank-conflict or coalescing claims and does not modify related simulator or coalescing analysis tools.
13	N/A	The PR diff does not construct or modify WarpDistributionEncoding, make_tile_desc, RegisterMapper, or tiling store casts.
14	N/A	The PR diff does not use tiling verbs, modify synchronization or barrier placement, or add a tiling pipeline variant.
15	N/A	The PR diff does not touch tiling allocation or bounds-checking logic for transformed address ranges.
16	N/A	The PR diff does not use or modify optimize_layout, recommend, or evaluate_transform.
17	N/A	The PR diff does not touch the tiling recorder, round-trip analysis, witness flags, or visualization modules.
18	N/A	The PR diff does not touch interleaved layouts, tiling demo kernels, dispatch registration of tiling demos, or tiling numeric gates.

§O graded 16/16

#	Verdict	Evidence
1	N/A	This PR does not touch core/verify.py, cpp/core/verify.cpp, or IR address spaces/attributes.
2	N/A	This PR does not modify scf.for, region structure, or IR block emission.
3	N/A	This PR does not modify IRBuilder, LLVM intrinsic declaration tables, or metadata emission.
4	N/A	This PR does not modify IR types, SmemType, or IR parsers.
5	N/A	This PR does not modify core/passes.py, cpp/core/passes.cpp, or op purity tables.
6	N/A	This PR does not touch constant folding in Python or C++ passes.
7	N/A	This PR does not modify kernel attributes, COMGR options, or cluster dimension checks.
8	N/A	This PR does not touch CodegenPolicy, SchedulerStrategy, or kernel compilation options.
9	N/A	This PR does not modify waitcnt encoding or barrier lowering.
10	N/A	This PR does not touch LLVM lowerers or parameter attribute/type lowering.
11	N/A	This PR does not touch ArchTarget or target catalog resolution.
12	N/A	This PR does not touch vector/LDS width caps or codegen arch defaults.
13	N/A	This PR does not touch public C headers under cpp/include/rocke/ or ABI versions.
14	N/A	This PR does not touch C++ engine sources, headers, bindings, or C++ compile flags in platform/CMakeLists.txt.
15	N/A	This PR does not touch C++ engine fallback handling or alternative lowerers.
16	N/A	This PR does not touch known_gaps.md or C++ library/builder capabilities.

§P graded 14/14

#	Verdict	Evidence
1	N/A	The current PR diff (Files list and Diff) does not touch any kernel builders, specs, or REQUIRED_FIELDS; the earlier NEEDS-INFO came from un-rebased base changes.
2	N/A	The current PR diff does not touch any kernel signatures, conv_abi, instance_conv_abi.h, or runtime argument packing.
3	N/A	The current PR diff does not change any GEMM, conv, or block-scaled GEMM launch geometry or grid calculation logic (conv_reference/architectures/gfx942/implicit_gemm_fwd/init.py:131 is a pure file move).
4	N/A	The current PR diff does not touch any LDS capacity gates, admission checks, or C++ validators.
5	N/A	The current PR diff does not touch WMMA/RDNA GEMM, GemmMultiD/GemmMultiAbd, or wave-size spec validation.
6	N/A	The current PR diff does not touch Stream-K, MoE sorting, fused MoE, or split-K workspace sizing or zeroing.
7	N/A	The current PR diff does not touch Stream-K, persistent GEMM, GroupedGemmLauncher, or instances/SUPPORT_MATRIX.md.
8	N/A	The current PR diff does not touch quantization helpers, MX GEMM, or MX scale decode.
9	N/A	The current PR diff does not touch numeric emulation or decode paths in either engine.
10	N/A	The current PR diff changes pinned GPU test reference bundle metadata and CMake staging, not rocKE runtime example manifests under helpers/manifest.py.
11	N/A	The current PR diff does not modify runtime launch helpers, stream synchronization, WorkspacePool, or benchmark timing code.
12	N/A	The current PR diff does not touch COMGR or HIP library loading in runtime/comgr.py or runtime/hip_module.py.
13	N/A	The current PR diff only reorganizes test reference bundles by variant and does not refactor host-side grid, tile-count, workspace math, kernel-cache keys, or admission guards.
14	N/A	The current PR diff does not touch examples, expected.json files, or add new kernel/helper modules.

§Q graded 7/7

#	Verdict	Evidence
1	NEEDS-INFO	The diffs for dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt (+7 -2 lines) and dnn-providers/hip-kernel-provider/rocke/platform/tests/test_gpu_references.py (+485 -80 lines) remain omitted to fit the model budget in the current diff, so CTest registration and installed test changes cannot be verified.
2	N/A	This PR refactors the GPU reference test bundle layout and does not add a new kernel family or variant to the platform IR or builders.
3	PASS	Tests and documentation grade the reference harness accurately, enforcing failure controls for perturbed outputs and missing launches in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:168-276 and noting subset verification caveats in dnn-providers/hip-kernel-provider/rocke/TESTING.md:154-156.
4	N/A	This PR does not re-bless IR goldens or widen the valid spec space in a lowerer or builder.
5	PASS	Tests remain portable across CPU/GPU and OS environments, skipping cleanly when hardware or bundles are absent (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:76-160) and querying the active device processor dynamically (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pytest_support.py:80-115).
6	PASS	The harness refactor preserves per-configuration semantics, scaling timeouts per installed variant (dnn-providers/hip-kernel-provider/rocke/TESTING.md:1009-1011), keeping at most one slot's workers alive at a time (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:82-163), and preserving error classes (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:227-232).
7	N/A	This PR does not modify engine backend rejection contracts or ROCKE_BACKEND / ROCKE_CPP_STRICT fallback behavior.

§R graded 11/11

#	Verdict	Evidence
1	N/A	This PR only restructures pinned GPU test reference harnesses and bundles and changes no kernel body, algorithm, tiling strategy, pipeline variant, or performance lever.
2	N/A	No performance lever or architecture-specific optimization recipe is modified in this test-reference refactoring PR.
3	N/A	No optimization lever covered by the utilities skills is touched by this test-infrastructure PR.
4	N/A	The diff does not alter kernel codegen, lowering, or performance primitives, so static ISA and occupancy probe evidence does not apply.
5	N/A	No global memory accesses in GPU kernels are added or modified by this diff.
6	N/A	No LDS or MMA layouts are added or changed in this pull request.
7	N/A	The PR makes no kernel bottleneck claims and implements no performance fixes.
8	N/A	No waits, barriers, or scheduling hints in GPU kernels are changed.
9	N/A	This PR is not a performance regression fix and changes no shape-dependent performance lever.
10	N/A	No compiler codegen changes, assembly parity claims, or performance benchmarks are included in this diff.
11	N/A	This PR introduces no new GPU optimization lessons, probes, or analysis tools.

§S graded 8/8

#	Verdict	Evidence
1	N/A	The PR does not add a public rocke package symbol, <Op>Spec, build_<op>, grid helper, compile_kernel/KernelLauncher/time_launches parameter, run_manifest option, or rocke_* C entry point.
2	NEEDS-INFO	The previous NEEDS-INFO is resolved: existing pinned-reference cohort and contract tests are updated in the same PR to resolve variant adapters without dropping coverage at dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py:409 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:67. — no diff line supports this verdict
3	N/A	No public rocke package or C engine API is removed or renamed; the moved reference adapters are internal to library/tests/.
4	N/A	No __all__ list in rocke, rocke.core, rocke.runtime, rocke.analysis, or rocke.helpers.tiling is modified by this PR.
5	PASS	The test suites assert concrete contracts and negative failure cases, including wrong variant rejection at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:402 and cohort digest mismatch rejection at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:588.
6	PASS	CPU-checkable reference contracts and CLI commands are tested on host without a GPU at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312, while GPU execution and numeric checks are marked @​pytest.mark.gpu at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:190.
7	N/A	The PR does not add or modify a platform API with a C++ twin or a C ABI binding.
8	NEEDS-INFO	The PR body lists reference suite runs, but does not state that PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=python:../library python tests/test_rocke.py and python tests/run_all.py from rocke/platform/ were run.

[brockhargreaves-amd]

Thanks @yraparti. On the re-reviews at 01e4921ab16 and e49fb56d9f8:

• §B.4 (HIGH): fixed in af517e7. Details are on its thread.
• §D.1 (HIGH): the same harness-only environment as before, answered on each thread.
• NEEDS-INFO:

  • Most rows grade hunks from the develop merges rather than this PR's changes:

    • 10-31: the §A, §M and §P rows, plus §B.1–§B.3, §B.5, §B.6, §G.2, §G.9, §G.11, §G.12 and §G.17;
    • 11-04: §G.20–§G.34.

    They concern kernels, lowering, dispatch, the runtime, benchmarks and goldens. Your 11-04 §A note says that the PR diff touches none of them.

  • The truncation rows (§C.1, §E.13, §E.15, §G.4–§G.6, §G.16, §J.4, §Q.1, §S.2) concern the same hunks as before, and my 10-30 answers still hold.

  • §F.6: pre-commit passes over the PR's files at af517e79a8b, which also merges develop.

  • §F.7: every review thread has a reply and is resolved.

  • §G.3, §S.8: run_checks.py's byte-identity, IR-validity and parity stages and test_rocke.py cover none of the changed files. The pytest stage's reference suites do, and at af517e79a8b they pass:

    • the reference host suites and the Torch audit: 745 passed, 18 skipped;
    • the CMake and install checks in platform/tests/test_gpu_references.py: 85 passed, 2 skipped.

[brockhargreaves-amd]

Kicked off another multi-arch workflow for architectures not covered per PR: https://github.com/ROCm/rocm-libraries/actions/runs/41945289255

[yraparti]

rocKE checklist re-review of the fixes since e49fb56d9f82

ROCm/rocm-libraries#14164 · af517e79a8bd · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 1, MEDIUM 0, LOW 0, NEEDS-INFO 63

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (e49fb56d9f82)

Resolved 1 · still open 1 · new 0

• ✅ §B.4 was FAIL, now PASS — had asked: In dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:46 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:65, replace request.get("variant") with request["variant"] and require isinstance(request["variant"], str) in…
• ⏳ §D.1 still FAIL (HIGH): Put the knob on the KernelSpec and delete the environment read.

Findings

• HIGH §D · dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 · Put the knob on the KernelSpec and delete the environment read.

Needs info

• §C.1 Re-review resolves the prior NEEDS-INFO: repeated architecture/variant lookup is centralized in AdapterRegistry at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and operation packages delegate to it at dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:10 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:10. — no diff line supports this verdict
• §C.2 New kernel-family code is exposed through spec-driven family builders rather than shape scripts, e.g. DwFwdWmmaSpec is the family spec in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:79 and the builder takes that spec in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:260. — no diff line supports this verdict
• §C.5 The current re-review diff has inflated blast radius beyond the pinned-reference variant migration, adding unrelated benchmark fill tooling at dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/conv_fill.py:1 and new kernel implementation code at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:1. — no diff line supports this verdict
• §C.8 New and revised helpers reuse shared current utilities instead of local reimplementations: dgrad WMMA code imports shared align_up/ceil_div at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_dgrad_patch_wmma.py:53 and reuses shared WMMA gate symbols at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_dgrad_smallc_wmma.py:53. — no diff line supports this verdict
• §D.20 The diff is truncated around the platform/cpp instance hunks needed to verify C++ mirrors for the kernel-level Python selector/spec-builder changes; provide the untruncated C++ mirror hunks or parity-test lines for those changes.
• §E.13 The rebase onto develop since the previous review did not resolve this: cli.py (truncated at line 1445), test_reference_common.py (truncated at line 1676), test_gpu_references.py, and reference_common/docs/*.md are still truncated or omitted in the prompt, although dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61 and dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107 enforce that no pre-variant locks or DVC pointers remain.
• §E.15 The rebase onto develop since the previous review did not resolve this: all visible new variant parameters in artifact.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), registry.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65), and the visible portion of cli.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349) are read and used, but the tail of cli.py (after line 1445) and test_gpu_references.py are truncated/omitted.
• §F.6 The new commits since e49fb56 do not include local pre-commit execution output, which cannot be verified from the diff alone.

Section notes

• §A: Re-review at head af517e7: the PR was rebased onto develop, which accounts for upstream commits in the e49fb56..af517e7 range; the PR's own diff against base (Files list) remains strictly limited to pinned GPU reference test harnesses, documentation, and CMake reference staging.
• §C: Current head includes additional rocKE kernel/benchmark/platform changes beyond the earlier reference-variant migration; item 5 is graded against that expanded blast radius.
• §D: This PR only reorganizes pinned GPU reference test bundles and harnesses by (operation, architecture, variant); it does not touch dispatcher registration, routing, or KernelSpec definitions.
• §D: The current diff is truncated; the C++ mirror hunks needed for item 20 are not fully visible.
• §G: Re-reviewed items 18-34 against head af517e7; the branch update since e49fb56 is a base-branch sync while the PR's own file set remains scoped to pinned GPU reference variant splitting.
• §H: Trace from visible hunks only; many kernel hunks are omitted/truncated, so this trace is incomplete. Visible conv_direct_dw_fwd_wmma: buffer SRD extents come from launch/runtime params A_bytes/B_bytes/D_bytes, while the visible compile-time visit space is grid=(ceil(Wo/tile_w), ceil(groups/block_ch), ceil(N/w_fold)ceil(Ho/block_h)); the spec only bounds sentinel-addressable X/Y bytes, so agreement with actual launch tensors cannot be fully checked from the truncated diff. Visible masks/bounds: ch_ok source is runtime p_groups vs compile-time channel chunk block_ch; s range is compile-time KW; output drain appears to combine runtime p_N/p_Wo/p_G with compile-time FTW*PER_PX but the hunk is truncated. Visible attention_dense_spec: seqlen_q/seqlen_kv are spec fields checked against block_m/block_n for ragged/sliding-window, not runtime kv_lens/cu_seqlens. No visible block_tables/paged indirection values are introduced, but omitted hunks prevent a complete trace.

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	N/A	The PR diff against base (dnn-providers/hip-kernel-provider/rocke/library/tests/... and platform/cmake/InstallGpuReferences.cmake) touches only pinned GPU reference test infrastructure and no kernel-body emitters.
2	N/A	No core/lower_* or core/ir files are modified in this PR's file list.
3	N/A	This PR does not modify any kernel selectors, spec builders, lowering, or codegen knobs.
4	N/A	This PR modifies no Python or C++ IR emitters.
5	N/A	No representative-IR or library IR golden files are touched in this PR (only test reference baseline_lock.json and .dvc files are moved/updated).
6	N/A	This PR makes no IR or engine changes requiring parity or golden proof.
7	N/A	No kernel or lowering logic is added or modified on any IR emission path.
8	N/A	No C++ rocKE engine, lowering, helper, or instance code is touched in this PR.
9	N/A	No shared op-handler fix or lowerer arch divergence is introduced in this PR.
10	N/A	No new kernel or IR feature flag is introduced in this PR.
11	N/A	No new IR op is added in this PR.
12	N/A	core/backend.py and ROCKE_BACKEND handling are not touched by this PR.
13	N/A	.ll emission and IR gate environments are not modified by this PR.
14	N/A	No classic helper files under platform/python/rocke/helpers/ or platform/cpp/helpers/ are modified by this PR.

§B graded 6/6

#	Verdict	Evidence
1	N/A	The PR diff against base (develop) only modifies test reference infrastructure, CMake staging, and docs, and does not add any kernel or dispatcher rejection gates.
2	N/A	This PR does not touch any kernel spec constructors or preflight gates.
3	N/A	This PR only reorganizes pinned GPU reference test bundles by variant and does not change kernel routing or fallback behavior.
4	PASS	The previous FAIL is resolved: source_variant in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:55 now reads request["variant"] and requires isinstance(variant, str), and dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:52 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:71 call get_adapter(architecture, source_variant(request)) in…
5	PASS	The new variant-selection, cohort-digest, and CMake reference-wiring gates (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:77) are covered by host tests (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:402…
6	N/A	This PR does not add or modify any is_valid_spec or supports_* predicates.

§C graded 10/10

#	Verdict	Evidence
1	NEEDS-INFO	Re-review resolves the prior NEEDS-INFO: repeated architecture/variant lookup is centralized in AdapterRegistry at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and operation packages delegate to it at dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/__init__.py:10 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/__init__.py:10. — no diff line supports this verdict
2	NEEDS-INFO	New kernel-family code is exposed through spec-driven family builders rather than shape scripts, e.g. DwFwdWmmaSpec is the family spec in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:79 and the builder takes that spec in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:260. — no diff line supports this verdict
3	PASS	Names are operation/algorithm/knob based: the new depthwise forward WMMA spec defaults to rocke_direct_dw_fwd_wmma in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:87, while reference variants use family names such as implicit_gemm_fwd at dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4 and dense at…
4	PASS	Arch-independent reference lookup lives in common AdapterRegistry at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, processor packages now only name the processor at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/__init__.py:9, and new shared conv kernel logic is placed under library/kernels/common at…
5	NEEDS-INFO	The current re-review diff has inflated blast radius beyond the pinned-reference variant migration, adding unrelated benchmark fill tooling at dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/conv_fill.py:1 and new kernel implementation code at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:1. — no diff line supports this verdict
6	PASS	The reference migration is refactor-first through the shared AdapterRegistry abstraction at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, and the benchmark fill behavior is factored into a shared helper instead of copied per benchmark at dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/conv_fill.py:40.
7	PASS	The reference bundle migration preserves the DVC object identity while changing only the archive path, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:2 and dnn-providers/hip-kernel-provider/rocke/library/tests/reference_bundles/conv/gfx942/implicit_gemm_fwd.tar.gz.dvc:5, and locks now pin the cohort digest at…
8	NEEDS-INFO	New and revised helpers reuse shared current utilities instead of local reimplementations: dgrad WMMA code imports shared align_up/ceil_div at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_dgrad_patch_wmma.py:53 and reuses shared WMMA gate symbols at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_dgrad_smallc_wmma.py:53. — no diff line supports this verdict
9	PASS	The variant source/replay bug class is handled across sibling operation workers: convolution dispatches source mode through the selected adapter at dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/worker.py:45, SDPA applies the same source/replay split at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/worker.py:64, and CMake rejects stale pre-variant locks at…
10	PASS	The widened shared key (operation, architecture, variant) is propagated through option parsing at dnn-providers/hip-kernel-provider/rocke/library/tests/conftest.py:55, bundle path construction at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/paths.py:23, pytest parametrization at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:178, and installation staging at…

§D graded 21/21

#	Verdict	Evidence
1	FAIL	dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:892 reads the environment
2	N/A	This PR does not add or modify any dispatcher KernelCandidate or Capability declarations.
3	N/A	This PR does not add, modify, or call any dispatcher candidate _supports or admits predicates.
4	N/A	This PR does not touch dispatcher selectors or spec builders on the selection path.
5	N/A	This PR does not modify any KernelSpec identity, kernel_name, or cache-key logic.
6	N/A	This PR does not modify cohort routing, spec builders, launch metadata, or cache keys in the dispatcher.
7	N/A	This PR does not add or modify dispatcher heuristics, hand-pinned overrides, or resolvers.
8	N/A	This PR does not define or modify any dispatcher routing cohorts or spec override predicates.
9	N/A	This PR does not register any new kernel or architecture in the dispatcher registries.
10	N/A	This PR does not introduce any new KernelSpec fields or features.
11	N/A	This PR does not register, promote, or change precedence or opt-in status for any dispatcher candidates.
12	N/A	This PR does not add, rename, or reprioritize production dispatcher candidates in the registry priority table.
13	N/A	This PR does not modify CandidateRegistry.select or add a production dispatcher wrapper or entry point that could catch PinRefused and retry with auto.
14	N/A	This PR does not change dispatcher compiled-binary caches, routing logs, tuning records, KernelId identity, or candidate kernarg ABI versions.
15	N/A	This PR does not add or edit production dispatcher candidate registration files or arch modules as dispatcher candidates.
16	N/A	This PR does not introduce a new dispatcher operator family requiring registered combos, sweep space, or dispatch-all helpers.
17	N/A	This PR does not add a production dispatcher candidate whose CPU routing invariants or family fixture coverage must be extended.
18	N/A	This PR does not modify CandidateRegistry ratchets, binding requirements, route-only registries, or binding exempt sets.
19	N/A	This PR does not over-claim unified attention auto routing through run_unified_attention_torch or change that production auto path.
20	NEEDS-INFO	The diff is truncated around the platform/cpp instance hunks needed to verify C++ mirrors for the kernel-level Python selector/spec-builder changes; provide the untruncated C++ mirror hunks or parity-test lines for those changes.
21	N/A	This PR does not present a dispatcher chooser bug fix where a selected spec should be changed instead of editing emitted artifacts.

§E graded 17/17

#	Verdict	Evidence
1	PASS	library/tests imports rocke (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:70), while platform/cmake/InstallGpuReferences.cmake invokes artifact.py via an isolated subprocess (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:208) without importing library into platform.
2	PASS	Device queries in the reference test infrastructure use from rocke.runtime.hip_module import get_device_target_id (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/worker.py:72) rather than torch.
3	PASS	Torch remains optional: dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/check_torch_reference.py:13 gates with pytest.importorskip("torch") and pinned reference tests verify assert not report["torch_imported"] (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:210).
4	N/A	This PR only modifies files under library/tests/, platform/, and TESTING.md, not production library/{kernels,dispatch,builders,benchmarks} layering or KNOWN_VIOLATIONS/KNOWN_BAD lists.
5	PASS	Platform CMake rules use relative directory variables (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:20) and Python helpers use tempfile and pathlib (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:18).
6	N/A	This PR does not touch emitter trees (core/, helpers/, instances/, library/kernels/).
7	PASS	Testing and workflow documentation are updated alongside the reference layout changes (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/docs/conv-test-reference.md:13, dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md:386).
8	PASS	No Jira/tracking IDs, reviewer handles, or PR numbers appear in code or comments in the diff (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:1).
9	PASS	All PR-modified files are scoped to splitting pinned GPU reference bundles by operation, architecture, and variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	The PR was rebased onto current develop and explicitly guards against concurrent branches restoring pre-variant locks or DVC pointers at legacy paths (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61, dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107).
11	PASS	Only deliverable test harness code, locks, DVC pointers, CMake rules, and required reference documentation are committed (dnn-providers/hip-kernel-provider/rocke/TESTING.md:123, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22).
12	PASS	All enrolled architectures (gfx942, gfx950, gfx1151, gfx1200, gfx1201) are migrated to the variant layout (dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:483, dnn-providers/hip-kernel-provider/rocke/platform/cmake/PublishedGpuReferences.cmake:11).
13	NEEDS-INFO	The rebase onto develop since the previous review did not resolve this: cli.py (truncated at line 1445), test_reference_common.py (truncated at line 1676), test_gpu_references.py, and reference_common/docs/*.md are still truncated or omitted in the prompt, although dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:61 and dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:107 enforce that no pre-variant locks or DVC pointers remain.
14	PASS	After rebasing onto develop, moved adapter files preserve upstream comments, historical API compatibility checks, and docstrings (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/__init__.py:4, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/__init__.py:43).
15	NEEDS-INFO	The rebase onto develop since the previous review did not resolve this: all visible new variant parameters in artifact.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:109, read at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143), registry.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65), and the visible portion of cli.py (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:349) are read and used, but the tail of cli.py (after line 1445) and test_gpu_references.py are truncated/omitted.
16	N/A	This item is a meta-instruction for review grading rather than a property of the pull request diff.
17	PASS	No performance numbers or superiority claims appear in the PR description, docs, or code (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/brockhargreaves-amd/pinned-reference-variants matches users/<user>/… (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).
2	PASS	title test(rocke): split pinned GPU references by operation, architecture, and variant matches type(scope): summary and is 80 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py is in the diff
5	PASS	no key, certificate, or secret filename in the diff
6	NEEDS-INFO	The new commits since e49fb56 do not include local pre-commit execution output, which cannot be verified from the diff alone.
7	NEEDS-INFO	The changes since e49fb56 do not include GitHub review thread metadata, so bot/Copilot thread resolution and deferred ticket IDs cannot be verified.
8	PASS	Base branch is develop (body notes Stacked on #13651), and the branch prefix, Conventional Commit title, description with issue ref, unit tests, and absence of forbidden files satisfy policy.yml (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312).
9	PASS	The full variant-keyed reference migration and test suite remain together in one PR without artificial splitting for a size cap (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:32).
10	PASS	Tests in dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:353 exercise real variant discovery, cohort digesting, and CMake staging invariants rather than throwaway checks.
11	PASS	The fetched PR metadata shows the populated Title and complete Motivation/Technical Details/Test Plan/Test Result Body are intact (dnn-providers/hip-kernel-provider/rocke/TESTING.md:120).

§G graded 34/34

#	Verdict	Evidence
1	N/A	No visible change uses byte-identity or parity as a correctness proof rather than a parity/golden check.
2	NEEDS-INFO	New kernel work is present in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:4, but the diff provides no on-silicon fp32-reference numeric completion evidence for that kernel work. — no diff line supports this verdict
3	NEEDS-INFO	Previous NEEDS-INFO is not resolved: current changes add kernel/platform emission work, but the diff does not include output proving tools/run_checks.py from the rocKE root, the llvm23 byte-identity check, installed-golden IR, IR-validity, and relevant pytest/numeric gates were run or shown unchanged versus develop.
4	NEEDS-INFO	Previous NEEDS-INFO is not resolved: many added behavior/test hunks are omitted from the diff, including direct-kernel tests, platform tests, and the tail of reference_common/cli.py, so fail-before coverage cannot be verified for all new behavior.
5	NEEDS-INFO	Previous NEEDS-INFO is not resolved: visible reference tests exercise shipped paths, but omitted test hunks for new kernels, platform changes, and CLI/status paths prevent verifying falsifiability and reachability for all added tests.
6	NEEDS-INFO	New problem in current commits: the comment in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:66 says the block-wave cap is lower than the actual cap set at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:69, so load-bearing control/limit prose contradicts the code. — no diff line supports this verdict
7	N/A	This diff does not introduce a new design-doc policy for an fp32 parity-reference oracle.
8	N/A	No visible rocKE harness change adds or modifies an fp8-to-fp16 silent dtype fallback.
9	NEEDS-INFO	The new kernel adds load-bearing fragment/VGPR budget prose and constants at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:66 without a tool-sourced provenance note in the visible diff. — no diff line supports this verdict
10	N/A	The visible rocKE diff changes benchmark plumbing but makes no specific kernel performance claim requiring device, ratio, shipped-script, and parity provenance.
11	NEEDS-INFO	The current diff adds optimized kernel builders, but selector/caller and measurement hunks are omitted, so reachability through Python routing and live C++ selectors on a real measured path cannot be verified.
12	NEEDS-INFO	The current diff includes cross-engine Python/C++ kernel changes, but many C++ instance, binding, selector, and caller hunks are omitted, so the who-calls-it-now cross-layer gap cannot be graded.
13	N/A	No visible rocKE change recommends a mangled intrinsic symbol requiring .td overload-class verification.
14	N/A	The visible rocKE diff does not classify a specific performance bottleneck that would require dynamic hardware-counter evidence.
15	N/A	No wrong-branch risk is apparent because the review is against the provided current-head diff.
16	NEEDS-INFO	Previous NEEDS-INFO is resolved: AdapterRegistry is defined in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:22 and both operation packages delegate to it in dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/init.py:12 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/init.py:12. — no diff line supports this verdict
17	NEEDS-INFO	New shipped scripts and script-like benchmark/CLI paths were added or modified in the current head, but several relevant test hunks are omitted, so end-to-end execution coverage beyond import/parse cannot be verified for all shipped scripts.
18	PASS	Variant and cohort-digest guards are wired into bundle validation and CLI manifest authentication (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:143, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/artifact.py:163, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:604), and unit tests construct mismatched variant and cohort-digest cases…
19	PASS	Test fixtures construct multi-variant, sole-variant, published, unpublished, missing-bundle, and candidate-bundle states and assert each gate's inline handling (dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1273, dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1468, dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:1596).
20	N/A	The PR's file list touches only pinned GPU reference test organization and documentation and does not add or modify a claim-ban detector.
21	PASS	In the reference qualification and verification harness touched by this PR, case error budgets, tolerances, and cohort digests are explicitly validated and propagated (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:298, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:815), and the forward conv benchmark fill helper synced in the branch update documents its floor semantics…
22	PASS	Explicitly selected pinned-reference runs fail rather than skip when a published bundle is missing or when an unpublished variant is selected without a candidate bundle and lock (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:85, dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:145), and CMake fails configuration when published locks or archives are missing…
23	N/A	This PR reorganizes pinned GPU reference test bundles by variant and makes no performance or kernel-routing speedup claims.
24	PASS	The pinned GPU reference locks are migrated to variant subdirectories with cohort_sha256 added while preserving the existing baseline_revision and manifest_sha256 digests unchanged (dnn-providers/hip-kernel-provider/rocke/library/tests/conv_reference/architectures/gfx942/implicit_gemm_fwd/baseline_lock.json:1, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa_reference/architectures/gfx942/dense/baseline_lock.json:1).
25	N/A	This PR does not ship a codegen micro-optimization.
26	N/A	This PR does not report kernel performance benchmarks.
27	N/A	This PR does not perform or report a performance A/B experiment.
28	N/A	This PR does not report performance deltas or counterfactual benchmark tables.
29	N/A	This PR does not swap guarded math primitives for fast approximations.
30	N/A	This PR makes no register-spill or scratch-memory claims.
31	N/A	This PR does not claim or benchmark a fix for a kernel performance defect.
32	N/A	This PR reports test pass/skip counts rather than measured performance or hardware resource figures.
33	PASS	The new variant resolution and selection logic in AdapterRegistry.resolve_variant (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/registry.py:65) and select_bundle (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:85) is exercised across unenrolled, ambiguous, explicit-override, sole-variant, published, and unpublished combinations…
34	N/A	This PR does not add or modify an opt-in kernel output flag.

§H graded 5/5

#	Verdict	Evidence
1	NEEDS-INFO	The visible new kernel creates buffer resources from runtime A_bytes/B_bytes/D_bytes, but multiple kernel hunks that may create SRDs or launch validation are omitted/truncated; provide the full diff for all buffer_rsrc/resource-bound construction plus the host-side launch validation and negative tests.
2	NEEDS-INFO	No block-table or paged-indirection value lookup is visible in the supplied hunks, but the kernel/attention hunks are omitted/truncated, so it is not possible to confirm that no loaded value is used as a memory index or that such values are range-checked.
3	NEEDS-INFO	The visible kernel uses div/mod by compile-time constants, but the allowed constant sets and additional div/mod sites are in omitted/truncated hunks; provide the full spec validation showing those divisors are constrained to powers of two or the benchmark/justification for arbitrary divisors.
4	N/A	The visible diff does not introduce runtime sequence-length inputs such as kv_lens or cu_seqlens that build a device-side softmax mask over a compile-time attention tile range.
5	NEEDS-INFO	Visible runtime launch inputs feed device-side bounds/masks, including tensor byte bounds and geometry values, but the host-validation and negative-test hunks for all such inputs are omitted/truncated, so symmetry across sibling inputs cannot be checked.

§I graded 6/6

#	Verdict	Evidence
1	NEEDS-INFO	The benchmark path measures the candidate space rather than only a preselected winner: _measure_plans is documented to time every plan of one case, and cache mode enumerates compatible candidates before optional narrowing by a kernels file (dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/benchmark_direct_conv.py:2234-2250; dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/benchmark_implicit_gemm_conv.py:4740-4778). — no diff line supports this verdict
2	PASS	Invalid or half-wired reference states are rejected at configure time with explicit causes, including unenrolled published architectures, pre-variant locks, invalid variant names, missing adapters, missing publication entries, and missing DVC pointers (dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:47-87; dnn-providers/hip-kernel-provider/rocke/platform/cmake/InstallGpuReferences.cmake:94-118).
3	N/A	This change reorganizes pinned reference variants and related tooling; it does not alter a curated/generated coverage policy or add Hypothesis-style generated coverage.
4	NEEDS-INFO	The relevant benchmark result-emission/schema hunks are omitted or truncated, so it is not possible to verify whether shipped and achievable columns are both retained.
5	PASS	Qualified bundles are explicit snapshots: qualification records the selected variant in the manifest and writes a lock with baseline_revision, cohort_sha256, and manifest_sha256, while artifact.py published provides a regeneration/tooling path for locating published variant archives and locks (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:470-489…
6	NEEDS-INFO	The re-review changes mention platform primitive/lowering and hardware-pattern instance files, but their hunks are omitted; the full diffs for the public API/intrinsics, lowering, and equivalence tests are needed to verify this ★ interface rule.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The modified Markdown visible in the diff (TESTING.md, conv-test-reference.md, platform/tests/README.md) contains no mathematical equations or matrix products to type-check.
2	N/A	The PR does not introduce mathematical formula notation tables or multi-domain quantities such as decay versus log-decay.
3	N/A	The visible Markdown documents GPU reference packaging and test workflows rather than restating a kernel algorithm derivation owned by another design document.
4	NEEDS-INFO	Not resolved from the previous NEEDS-INFO: the rebase commit (e49fb56..af517e7) did not make the omitted PR Markdown hunks visible (dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/docs/README.md, gpu-ci-pinned-rocke-test-reference-plan.md, gpu-reference-workflow.md, quickstart.md, and sdpa_reference/docs/sdpa-test-reference.md remain omitted to fit the model budget), so novelty claims in those files still cannot be verified.
5	PASS	dnn-providers/hip-kernel-provider/rocke/TESTING.md:1010-1011 and dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md:389-390 state the CTest timeout as a per-variant aggregate (five minutes per installed variant / five minutes per variant).
6	N/A	This PR reorganizes pinned test reference bundles and does not add model config.json tables, tuned dispatch band edges, or kernel constants requiring deployment provenance.
7	N/A	No new Markdown design or algorithm document file is added in the PR file list; existing test documentation files are updated in place.

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR only reorganizes pinned GPU reference test bundles and adapters under library/tests/ and does not add or modify any tuned dispatch candidates or KnobSpace subclasses.
2	N/A	This PR does not add or modify any tunable kernel spec fields or KnobAxis declarations.
3	N/A	This PR does not change tuning_id, config_key, knob canonicalization rules, or tuned spec defaults.
4	N/A	This PR does not touch KnobSpace.defaults() or KnobSpace.recorded().
5	N/A	This PR does not modify knob dropping, refusal, inertness, or base_value() rules.
6	N/A	This PR does not modify KnobSpace.validate or policy-free tuning spec construction.
7	N/A	This PR does not touch KnobSpace.base(req) or request normalization for tuning.
8	N/A	This PR does not change normalize_knobs or knob value coercion.
9	N/A	This PR modifies GPU test reference bundle replay under library/tests/, not dispatch tuning_id/spec_hash benchmark replay.
10	N/A	This PR does not touch KnobSpace.find or tuning_id resolution in dispatch.
11	N/A	This PR does not touch dispatch sweep entry points, pin_to_spec, or sweep harnesses.
12	N/A	This PR does not change production stacks, accept_default, or base specs for auto tuning selection.
13	N/A	This PR does not touch dispatch tuning sweep harnesses or make claims about auto selection.

§L graded 10/10

#	Verdict	Evidence
1	N/A	This diff does not show a new gfx hardware target being added to the rocKE architecture fact sources or ISA backend registries.
2	NEEDS-INFO	The diff is truncated for relevant MMA lowering changes such as dnn-providers/hip-kernel-provider/rocke/platform/cpp/core/lower_hip/lower_hip_mma.cpp and does not include enough hunks to determine whether any new MMA op_id was added with matching fragment/layout metadata.
3	NEEDS-INFO	The diff is truncated for relevant WMMA/MFMA helper and kernel changes such as dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_fwd_wmma_gate.py and multiple new WMMA kernels, so I cannot verify that all MMA selection and lowering goes through the catalog and does not hard-code intrinsic names or lane maps.
4	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:81 hard-codes DW_FWD_WMMA_WAVE = 32 instead of deriving the wave size from the exact ArchTarget.wave_size. — no diff line supports this verdict
5	NEEDS-INFO	The diff is truncated for several arch-specific recipe files and new WMMA kernels, so I cannot verify that gfx942/gfx950/gfx1250 recipes and loaders remain properly gated and do not leak across targets.
6	NEEDS-INFO	The diff is truncated for relevant policy/instance and platform Python changes such as dnn-providers/hip-kernel-provider/rocke/platform/python/rocke/instances/gfx1250/block_scaled_gemm.py and related files, so I cannot verify pipeline vocabulary placement or import boundaries.
7	NEEDS-INFO	The diff is truncated for relevant dtype and packing changes including NVFP4/scale-format examples, tests, and lowering files, so I cannot verify the logical dtype, bit-packing, scalar-lowering, and FP8 dialect contracts.
8	NEEDS-INFO	The diff is truncated for relevant scaled-MMA changes in lowering, examples, and golden files, so I cannot verify all-or-nothing scale contracts or that retired scaled WMMA ids were not reintroduced.
9	N/A	This diff does not introduce a new architecture target, so the new-arch completion gates do not apply.
10	NEEDS-INFO	The diff is truncated for dnn-providers/hip-kernel-provider/rocke/platform/tests/tools/test_arch_domain_artifact.py and related tooling changes, so I cannot verify that arch-domain artifacts and LLVM flavor membership remain generated and non-hand-rolled.

§M graded 10/10

#	Verdict	Evidence
1	N/A	The PR diff only modifies test reference harnesses, locks, and CMake scripts under rocke/library/tests/ and rocke/platform/, with no changes to core/ IR, builders, or kernel control flow.
2	N/A	The PR diff does not touch legalize_for_target, architecture data, or lowering files.
3	N/A	The PR diff does not touch ck.dsl.ir serialization, print_ir, wire versions, attribute tagging, SSA ids, or ABI parameter ordering.
4	N/A	The PR diff does not touch portable IR schemas, importers, recipe expansion/VM, or recording builders.
5	N/A	The PR diff does not touch AsyncTileLoader, LDS layouts, or kernel memory lowering.
6	N/A	The PR diff does not emit or modify kernel control flow, barriers, memory loads, or cross-lane shuffles.
7	N/A	The PR diff does not touch ISA backends, buffer resource word 3, or TemporalHint handling.
8	N/A	The PR diff does not add intrinsic declarations, modify LLVM profile tables, or change backend selection precedence.
9	N/A	The PR diff does not touch public C entry points, headers under cpp/include/rocke/, or C++ host launch behavior.
10	N/A	The PR diff does not emit or consume LLVM debug info (!dbg) or DWARF metadata.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The PR diff only touches GPU reference test harness files and CMake rules, without adding or modifying helpers under platform/python/rocke/helpers/.
2	N/A	The PR diff does not author or modify kernels or their use of shared helper mechanisms.
3	N/A	The PR diff does not touch SoftwarePipeline or any kernel pipeline code.
4	N/A	The PR diff does not add or modify fused kernels, epilogue fusion logic, fusion validation, or fusion legalizers.
5	N/A	The PR diff does not introduce or change any public tiling API under helpers/tiling/.
6	N/A	The PR diff does not add or modify front-door or toolbox tiling helpers or any LayoutStyle implementation.
7	N/A	The PR diff does not touch tiling layout overrides, accumulator descriptors, or MMA validation.
8	N/A	The PR diff does not touch tiling memory operations, LDS allocation, global load/store emission, clipping, padding, or wide-load selection.
9	N/A	The PR diff does not use or modify N-D tensor axis roles, batch reductions, flow-stage relabeling, or AB-swap routing in tiling.
10	N/A	The PR diff does not modify tiling emit, LDS staging, MMA issue logic, or tiling fail-fast raises.
11	N/A	The PR diff does not touch MMA traits, TileMma, lds_swizzle, scheduling group helpers, or related tiling policies.
12	N/A	The PR diff makes no LDS bank-conflict or coalescing claims and does not modify related simulator or coalescing analysis tools.
13	N/A	The PR diff does not construct or modify WarpDistributionEncoding, make_tile_desc, RegisterMapper, or tiling store casts.
14	N/A	The PR diff does not use tiling verbs, modify synchronization or barrier placement, or add a tiling pipeline variant.
15	N/A	The PR diff does not touch tiling allocation or bounds-checking logic for transformed address ranges.
16	N/A	The PR diff does not use or modify optimize_layout, recommend, or evaluate_transform.
17	N/A	The PR diff does not touch the tiling recorder, round-trip analysis, witness flags, or visualization modules.
18	N/A	The PR diff does not touch interleaved layouts, tiling demo kernels, dispatch registration of tiling demos, or tiling numeric gates.

§O graded 16/16

#	Verdict	Evidence
1	N/A	This PR does not touch core/verify.py, cpp/core/verify.cpp, or IR address spaces/attributes.
2	N/A	This PR does not modify scf.for, region structure, or IR block emission.
3	N/A	This PR does not modify IRBuilder, LLVM intrinsic declaration tables, or metadata emission.
4	N/A	This PR does not modify IR types, SmemType, or IR parsers.
5	N/A	This PR does not modify core/passes.py, cpp/core/passes.cpp, or op purity tables.
6	N/A	This PR does not touch constant folding in Python or C++ passes.
7	N/A	This PR does not modify kernel attributes, COMGR options, or cluster dimension checks.
8	N/A	This PR does not touch CodegenPolicy, SchedulerStrategy, or kernel compilation options.
9	N/A	This PR does not modify waitcnt encoding or barrier lowering.
10	N/A	This PR does not touch LLVM lowerers or parameter attribute/type lowering.
11	N/A	This PR does not touch ArchTarget or target catalog resolution.
12	N/A	This PR does not touch vector/LDS width caps or codegen arch defaults.
13	N/A	This PR does not touch public C headers under cpp/include/rocke/ or ABI versions.
14	N/A	This PR does not touch C++ engine sources, headers, bindings, or C++ compile flags in platform/CMakeLists.txt.
15	N/A	This PR does not touch C++ engine fallback handling or alternative lowerers.
16	N/A	This PR does not touch known_gaps.md or C++ library/builder capabilities.

§P graded 14/14

#	Verdict	Evidence
1	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:260 adds build_dw_fwd_wmma(spec, arch="gfx1250"), so arch is a positional/defaulted parameter rather than keyword-only (spec, *, arch). — no diff line supports this verdict
2	NEEDS-INFO	The changes-since list shows ABI-relevant files such as platform/cpp/include/rocke/abi.h and platform/python/rocke/portable_ir/src/abi.py, but their hunks are omitted, so kernarg order and packing compatibility cannot be verified.
3	NEEDS-INFO	The changes-since list includes platform instance changes such as platform/python/rocke/instances/gfx1250/block_scaled_gemm.py, but the relevant launch-geometry hunks are omitted, so grid-source compliance cannot be verified.
4	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:234 gates spec.lds_bytes() against the literal DW_DGRAD_MFMA_LDS_BUDGET instead of the architecture target LDS capacity/fits helper. — no diff line supports this verdict
5	NEEDS-INFO	The changes-since list includes gfx1250 GEMM instance files and WMMA-related platform changes, but their hunks are omitted, so narrow WMMA/RDNA GEMM admission rules and wave-size validation cannot be checked.
6	N/A	The shown rocKE diff does not touch Stream-K, MoE sorting, fused MoE, or split-K workspace sizing or zeroing.
7	NEEDS-INFO	The changes-since list includes platform/python/rocke/instances/SUPPORT_MATRIX.md, but its hunk is omitted, so preserved rejection limits for Stream-K, persistent GEMM, grouped GEMM, and the support matrix cannot be verified.
8	NEEDS-INFO	The changes-since list includes quantization/NVFP4 and MX-related instance, example, and test files, but the relevant hunks are omitted, so use of the shared quantization helpers and MX constraints cannot be verified.
9	NEEDS-INFO	The changes-since list includes numeric emulation and NVFP4 test files, but the relevant hunks are omitted, so edge-case decode/emulation coverage in both engines cannot be verified.
10	N/A	The shown diff changes pinned GPU reference bundle metadata and CMake staging, not rocKE runtime example manifests under helpers/manifest.py.
11	N/A	The shown diff does not modify rocKE runtime launch helpers, stream synchronization, WorkspacePool, or benchmark timing infrastructure covered by this item.
12	N/A	The shown diff does not touch COMGR or HIP library loading code in rocKE runtime.
13	NEEDS-INFO	The changes-since diff visibly refactors host grid/tile math with shared ceil_div/align_up helpers, but the relevant numeric-test hunks and per-consumer proof are omitted, so the required consumer test coverage cannot be verified.
14	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:4 adds new kernel-family logic as another flat library/kernels/common module instead of placing new kernel logic under the appropriate instances/family structure required by this item. — no diff line supports this verdict

§Q graded 7/7

#	Verdict	Evidence
1	NEEDS-INFO	The previous NEEDS-INFO is not resolved in the visible current diff: the relevant registration hunks for dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt and dnn-providers/hip-kernel-provider/ROCKE_ENGINE_test_categories_external.yaml are still omitted, so CI/CTest registration and lane selection cannot be verified.
2	NEEDS-INFO	Current head adds new kernel variants, but the hunk for dnn-providers/hip-kernel-provider/rocke/platform/tests/instances/rocke_ir_parity_harness.py is omitted, so I cannot verify that cases() covers one case per variant per arch for byte-identity lowering.
3	NEEDS-INFO	Current head changes platform gates and related tests, but relevant hunks such as platform/tools/check_byte_identity.py, platform/tests/tools/test_run_checks.py, and platform/tests/instances/rocke_ir_parity_harness.py are omitted, so I cannot verify that statuses and gate claims are graded only for what each gate proves.
4	NEEDS-INFO	The current diff adds representative/differential golden files and lowerer/builder changes, but the diff does not include evidence that goldens were re-recorded from a green tree for every LLVM flavor or that fuzz_diff.py was run, or a stated reason why not.
5	NEEDS-INFO	The current head adds multiple platform tests with omitted hunks, including GPU/LLVM-specific tests and platform/tests/test_gpu_references.py, so portability details such as dynamic target/flavor selection, clean skips without HIP/device/clang, and newline normalization cannot be verified from the visible diff.
6	PASS	The refactor keeps per-variant installed timeout semantics in dnn-providers/hip-kernel-provider/rocke/TESTING.md:1007-1011, reuses at most one selected slot's workers at a time in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:76-163, and preserves worker failure error classes in dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/cli.py:226-232.
7	NEEDS-INFO	Current head changes backend/rejection-related code and tests, but the relevant hunks for files such as platform/python/rocke/core/backend.py and backend parity/rejection tests are omitted, so I cannot verify the ROCKE_BACKEND × ROCKE_CPP_STRICT × binding-present matrix or exact exception assertions.

§R graded 11/11

#	Verdict	Evidence
1	NEEDS-INFO	A redesign/new kernel body is visible in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:1, with tiling/pipeline knobs such as block_waves, block_h, w_fold, and prefetch_rows at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:114 and dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:117; missing the runbook Step 0 lever sweep, correctness pruning against fp32, and named best existing baseline.
2	NEEDS-INFO	The new kernel is arch-gated to gfx1250 at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:88 and validated through is_valid_dw_fwd_wmma_spec at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:229, but the diff does not provide the runbook §6/§7/§8 mapping, §13 convolution checklist evidence, arch/gfx1250.md recipe citation, or anti-pattern/failure-mode justification for the changed levers.
3	NEEDS-INFO	The visible levers include LDS sizing at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:162, launch geometry at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:138, and prefetch_rows at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:117; missing evidence that the PR followed the matching GEMM/LDS/prefetch/ISA/launch utilities skills or justified any contradiction to them.
4	NEEDS-INFO	The intended primitive changes to WMMA via wmma_fwd_op in dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:345 and the review-response diff also changes lowering/codegen files, but the PR evidence shown does not include before/after probe_occupancy.py, probe_regs.py, probe_intrinsic_counts.py or probe_isa_inspect.py, nor probe_lowering_compare.py for backend-affecting lowering changes.
5	NEEDS-INFO	The new kernel imports and uses buffer load/store helpers for global accesses at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:62 and dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:64, but the diff does not show analyze_coalescing/vector_transactions/addr_fn_from_strides summaries or assert_asm_backed gating against emitted ISA width for the added or changed accesses.
6	NEEDS-INFO	The new kernel allocates LDS staging at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:384 and uses WMMA at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:345, but the provided diff does not show verify_lds_roundtrip/verify_mma_soundness evidence, optimize_layout recommendations, or an equivalent bit-exact layout proof for this non-tiling kernel.
7	N/A	The visible rocKE diff does not make a bottleneck claim such as memory-bound, LDS-bound, barrier-bound, or latency-bound that would require runbook §3 or §8.6 classification evidence.
8	NEEDS-INFO	Scheduling validation changes are visible, for example replacing normalize_scheduler_strategy with scheduler_strategy_reason at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_dgrad_patch_wmma.py:245, and the new kernel describes row-level LDS/barrier staging at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:35; missing the runbook §8.3/§8.4 and scheduler-policy.md rationale plus ISA evidence for any changed waits, barriers, or scheduling hints.
9	NEEDS-INFO	The new gfx1250-only variant gate is at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:88 and its shape-dependent knobs are at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:114 through dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:117; missing the family shape sweep and any dispatch/heuristic gate for shapes that lose in the tradeoff.
10	NEEDS-INFO	The diff changes benchmark methodology through the new fill-mode helper at dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/conv_fill.py:1 and dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/conv_fill.py:16, and the review-response diff lists codegen/lowering changes, but the provided evidence does not show runbook §10 compiler-hazard/alias-semantics checks or §2.3 benchmark-hygiene details for any performance conclusions.
11	NEEDS-INFO	New optimization content is introduced by the Toeplitz WMMA depthwise forward kernel at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:1 and its new knobs at dnn-providers/hip-kernel-provider/rocke/library/kernels/common/conv_direct_dw_fwd_wmma.py:114, but the diff does not add a corresponding optimization lesson to runbook §14/§16/§17 or the relevant utilities/skills documentation. — no diff line supports this verdict

§S graded 8/8

#	Verdict	Evidence
1	N/A	The PR does not add a public rocke package symbol, <Op>Spec, build_<op>, grid helper, compile_kernel/KernelLauncher/time_launches parameter, run_manifest option, or rocke_* C entry point.
2	NEEDS-INFO	The previous NEEDS-INFO is resolved: existing pinned-reference cohort and contract tests are updated in the same PR to resolve variant adapters without dropping coverage at dnn-providers/hip-kernel-provider/rocke/library/tests/test_conv_reference_contract.py:435 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_reference_contract.py:67. — no diff line supports this verdict
3	N/A	No public rocke package or C engine API is removed or renamed; the moved reference adapters are internal to library/tests/.
4	N/A	No __all__ list in rocke, rocke.core, rocke.runtime, rocke.analysis, or rocke.helpers.tiling is modified by this PR.
5	PASS	The test suites assert concrete contracts and negative failure cases, including wrong variant rejection at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:402 and cohort digest mismatch rejection at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:588.
6	PASS	CPU-checkable reference contracts and CLI commands are tested on host without a GPU at dnn-providers/hip-kernel-provider/rocke/library/tests/test_reference_common.py:312, while GPU execution and numeric checks are marked @​pytest.mark.gpu at dnn-providers/hip-kernel-provider/rocke/library/tests/reference_common/pinned_tests.py:190.
7	N/A	The PR does not add or modify a platform API with a C++ twin or a C ABI binding.
8	NEEDS-INFO	The PR body lists reference suite runs, but still does not state that PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=python:../library python tests/test_rocke.py and python tests/run_all.py from rocke/platform/ were run.

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #14164 — split pinned GPU references by operation, architecture, and variant (brockhargreaves-amd, AICK-2299) — 🟠 Follow-up: fix verified; two quality should-fixes

Prior item	Status
Copilot (10-23) ×2: archive directory before pack; CONFIGURE_DEPENDS on the globs	🟢 FIXED a6450761, a010e8cf • still holds: test_pack_creates_the_archive_directory 4 ✓, installed (ran)
yraparti (bot, 10-31, 11-04) HIGH §B.4: a source request without variant runs the default	🟢 FIXED af517e79 • missing → KeyError, non-string → ValueError; every request writer names it (read) • the worker test fails with the fix reverted: "DID NOT RAISE" (ran) • bot marks it PASS at 04:18Z; source-mode GPU lanes pass at head (read)
yraparti (bot, 10-24 → 11-05) HIGH §D.1: environment read at test_reference_common.py:892	🟢 ANSWERED: Rule 3 exempts harness environment • :892 only sets a child pytest's env; unchanged since 6b2f933f (read) • the bot's own §D note: no KernelSpec touched (read) • 04:18Z re-post: no reply yet, the only open thread; author: reply, resolve
yraparti (bot) NEEDS-INFO rows (63 at head)	🟢 ANSWERED 11-05 • they grade the truncated diff or develop-merge hunks (read)
CongMa13 (10-26): threads without an author reply	🟢 ANSWERED 10-27 (read)
CongMa13 (10-27) ×2: an unenrolled listed architecture installs; verify / diff skip cohort_sha256	🟢 FIXED af45a8c8, bd2d28a6 • both tests pass at head (ran)
CongMa13 (10-27 23:38Z) APPROVED	🟢 APPROVED 964f9e9b • older head; still counts on develop, for rocke-platform • not covered: the 10-29/30 review commits, af517e79
AnaghaRaoAMD (10-29) ×5 code items: input digests, pre-variant manifests, globs, Torch audits, one-variant tests	🟢 FIXED 0de1b0d6, 1c37173e, a010e8cf, 4a4dc178, 353e8d92 • replies posted 10-30; threads resolved (read) • their tests pass at head, installed (ran)
AnaghaRaoAMD (10-29) ×3 docs items: one CTest per architecture; a new Case field; operation packages	🟢 FIXED 0edf7060, 93ee5b18, eee59587 • replies posted 10-30; threads resolved (read) • scaffold declined: a new operation adds one set(...) line; reason holds (read)
AnaghaRaoAMD (10-29): split test_reference_common.py by concern ("A tracked follow-up is fine")	🟡 DEFERRED • author: "a move of roughly 5k lines would bury the review fixes"; the rocKE team owns it • sound for after merge; the 14328 blocker is gone (landed 11-03) • untracked, no issue or AICK ticket: rocKE team links one before merge; 5,417 lines, ~11 concerns
Ours (10-24) cross-PR notes	🟢 14328 landed 11-03 • its file-end conflict: both tests kept in e49fb56d (ran) • 14440 adds a test there too: keep both when it rebases (ran) • 14927, 14846 merge clean; 14901's conflicts are with develop (ran)
Bots / CI	🟢 Required 4/4 ✓; PR CI 45 ✓, 0 ✗ • pinned ✓: gfx942 conv 35 / SDPA 11; gfx950 conv 35 / SDPA 10; gfx1151 SDPA 11; gfx1201 SDPA 11 on 10-30 (read) • fails, the author's extra-arch run only: gfx120X — provider C++ suites hang at the first GPU call; they pass on 3 other GPUs (read) • merge: blocked — rocke-lib code-owner approval; 1 open thread
Our checks	🟢 Clean • merge vs 8b6ada32: clean (ran) • build; ctest 46 ✓ / 1 skip; byte identity GREEN ×4; goldens pass (ran) • tests 18,923 ✓ / 537 skip; installed 734 ✓ (3 of 5 files); compile clean; rules 1–3 clean (ran)
Code quality	🟠 Should-fix • the registry keeps a default every caller must avoid: N2 (ran) • two-variant CTest cap documented, never enforced: N3 (ran) • both workers copy one run(); "published" re-derived in 5 places (read)

#	Sev	file:line	Finding
N1	🟡	PR description	"Stacked on" still names 13651, merged 10-22 • impact: the squash commits an out-of-date line • fix: drop the line
N2	🟠	library/tests/reference_common/registry.py:65	The registry still falls back to a processor's only variant • impact: an omitted variant breaks only once a second lands; source_variant is the third call-site guard, and the AST test missed the worker's explicit None • fix: require variant in the registry and the 12 CLI methods; default only where --variant is parsed (under that rule only the default's own test fails, ran)
N3	🟠	library/tests/reference_common/docs/gpu-reference-workflow.md:596	"At most two variants" per GPU CTest is stated 4× and enforced nowhere • impact: a third variant configures and installs (ran); two PRs can cross the cap silently, as the guide's overlap table says • fix: refuse a third in the configure wiring check, pointing at per-variant CTests; state the cap once

Utility: 🟢 Positive — source requests now name their variant; pinned references pass on three architectures.

---

Legend: 🔴 blocking · 🟠 should-fix · 🟡 minor nit · 🟢 info/good · ⚪ retracted

Follow-up review: first pass by a local qwen3.6 model, verified by the orchestrator (Claude Opus 5.5). Checks: merge with develop, build + ctest, byte identity, goldens, the PR's tests (source and installed), a revert test of the worker fix, trial merges with open PRs, CI shard logs, code quality with a strict-registry probe and a variant-cap probe. Flags whether a closer look is needed; not an approval or rejection.

[cgmillette]

lgtm