[cgmillette]

Motivation

tests/test_rocke_ci_static.py::TestIrParityCoverage::test_ir_cases_match_golden_sha256 lowers and hashes every representative-IR case at every LLVM flavor (llvm20, llvm22, llvm23, llvm24) inside a single test. In the installed rocke_pytest CTest lane, which runs pytest with --timeout=60, that one test runs close to the per-test limit and times out intermittently on develop and on unrelated PRs.

ISSUE ID : AICK-2895

Tracking issue: #14733

Technical Details

• The golden check is now one test per LLVM flavor, test_ir_cases_match_golden_sha256[<flavor>], parametrized over GOLDEN_FLAVORS (which is LLVM_FLAVORS, the list of qualified compiler lanes), so a newly added flavor gets its own test with no edit here. Each test lowers a quarter of the work the old test did. A drift message names its flavor.
• The checks the old all-flavor call made on the golden itself (a document for each emission profile, the expected aliases, nothing for a flavor no longer qualified) move to one fast test, test_golden_holds_every_llvm_flavor, which lowers nothing.
• rocke.core.ir_golden gains check_golden_flavors(golden_path), the flavor-set part of an all-flavor check_golden call, so a caller that checks flavors one at a time keeps the same coverage. check_golden returns the same results as before; the library golden tests that call it are unchanged.
• The file's golden tests are now pytest functions (unittest has no parametrization); the other tests in the file stay unittest classes. Running the file directly now runs pytest.
• The installed-tree golden gate (rocke_installed_golden_test.py, CTest rocke_golden_static) is the same check but runs as its own CTest process with no 60 s per-test limit, so it is left as one all-flavor call.
• BUILDING.md drops its note that the golden test reports conv drift without the C++ engine: the Python lowerer matches the goldens at every flavor.

No emitter, IR or golden changes.

Test Plan

• Coverage equality: with lowering replaced by a recorder, count the (case, flavor) pairs the old single test and the new per-flavor tests lower, and compare them with the golden's entries.
• New unit tests in tests/core/test_ir_golden.py: check_golden_flavors accepts a correct golden and reports a stale profile document; checking each flavor separately plus check_golden_flavors gives exactly the errors of one all-flavor check_golden call.
• The new tests and the old test pass on the unchanged golden; rocke_installed_golden_test.py passes.
• Byte identity at llvm20, llvm22, llvm23, llvm24; ctest; full platform suite against develop; pre-commit.

Test Result

• Coverage: old single test lowers 1440 (case, flavor) pairs (360 cases x 4 flavors); the new tests lower the same 1440 pairs (identical multiset), equal to the golden's 1440 entries.
• tests/core/test_ir_golden.py and tests/test_rocke_ci_static.py pass (both engines, strict).
• rocke_installed_golden_test.py passes at all four flavors.
• Byte identity GREEN at llvm20, llvm22, llvm23 and llvm24; ctest 35/35; pre-commit clean.
• Full platform suite (tests/core, python/rocke/tests, tests/test_rocke.py, tests/test_rocke_ci_static.py): no failures. The remaining platform test files and the library tests that reach the touched code show no new failures against develop.

🤖 Generated with Claude Code

[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.

[cgmillette]

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

PR #14820 — check the representative-IR golden one LLVM flavor per test (cgmillette, AICK-2895) — 🟡 First-round: Coverage equal; CI pending

Prior item	Status
Reviewers	🟢 None yet • code-owner teams requested: rocke-platform, rocke-lib
Develop issue 14733 (golden test timeout)	🟡 OPEN: this PR is the fix candidate • same 1440 case-flavor pairs; check_golden same on valid alias sets (ran) • slowest split test 26% of the old one (ran) • needs gfx94X margin over several runs: pending
Bots / CI	🟡 running • 35 ✓ · 3 running (math-libs builds) · tests not started • fails: Multi-Arch CI Summary — first run, cancelled by the label re-run (read) • required: pre-commit, Component, Math ✓; Multi-Arch pending
Our checks	🟢 pass • merge vs d1de2108a7 clean; ctest 34 ✓ / 1 skip (ran) • goldens, byte identity, dg, 31 tests, installed layout ✓ (ran) • open 12167 conflicts in test_rocke_ci_static.py (ran)

#	Sev	file:line	Finding
N1	🟡	rocke/BUILDING.md:124	Edited note still promises conv drift without the C++ engine • impact: no drift at any flavor without the engine (ran) • fix: drop the residual paragraph

Utility: 🟢 Positive — splits the near-limit golden test four ways; same checks, no IR change

---

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

First-round review: first pass by a local qwen3.6 model, verified by the orchestrator (Claude Opus 5.5). Checks: merge with develop, build + ctest, golden gate (4 flavors), byte identity (llvm22), differential golden, the PR's tests, a coverage recorder (old vs new tests), an old-vs-new check_golden differential, an installed-layout run, the library golden tests, a same-time timing A/B, merges with open PRs on the same files. Flags whether a closer look is needed; not an approval or rejection.

[cgmillette]

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

PR #14820 — check the representative-IR golden one LLVM flavor per test (cgmillette, AICK-2895) — 🟡 Follow-up: N1 fixed; CI running

Prior item	Status
Develop issue 14733 (golden test timeout)	🟢 Done in CI 6bf24a89 • gfx94X shards 1–2: 4 per-flavor tests ✓, each ≤ 24% of the per-test limit (read) • test code unchanged at head; its gfx94X run pending
Our 10-31 N1: BUILDING.md promises conv drift without the engine	🟢 FIXED a2fea5b4 • paragraph removed; datalayout note and closing line kept (ran) • Python-lowerer goldens ✓ ×4: gate (ran), CI gfx94X (read)
shumway (11-02) approval	🟢 APPROVED 6bf24a89 • older head; still counts on develop • since then: only the BUILDING.md paragraph removal (ran)
shumway (11-02) BUILDING.md:122: remove the stale paragraph	🟢 FIXED a2fea5b4 • paragraph removed (ran) • thread resolved by cgmillette
Bots / CI	🟡 running • running: Math Libs builds ×4, HipDNN Superbuild ×2; no test jobs yet • 19 ✓ · 0 ✗; duplicate therock-pr-bot run cancelled; its later run ✓ • merge: blocked; required Multi-Arch CI Summary pending
Our checks	🟢 pass • merge vs 650a69da clean; ctest 34 ✓ / 1 skip; 31 tests ✓ (ran) • byte identity, golden gate, differential golden ✓ ×4; rules clean (ran) • open 12167 still conflicts in test_rocke_ci_static.py (ran)

Net-new: none.

Utility: 🟢 Positive — splits the near-limit golden test four ways; CI shows ample margin

---

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: range-diff vs 6bf24a89, merge with develop, build + ctest, golden gate and byte identity (4 flavors), differential golden, the PR's tests, gfx94X CI logs at 6bf24a89, merge with open 12167. Flags whether a closer look is needed; not an approval or rejection.

[aosewski]

Small comment to improve doc - not blocking.

[assistant-librarian]

TheRock Submodule Bump Activity

Newest first

• 2026-11-03 — Included in TheRock via ROCm/TheRock#9698.