[CongMa13]

JIRA ID: AICK-2822

Draft until CI reports.

---

Summary — For human reviewers

The short version

This is story S06b of the epic that covers gfx942 and gfx950 in rocKE tests (S06a is #14713). It covers the dispatch tests and the small test directories under platform/tests. Each single-arch test is now parametrized over both archs or marked arch_specific with a reason.

What changes

Three commits, one for each group of directories. Only test files change.

• dispatch_tests/gemm. 28 tests run on both archs, with per-arch expected values where the archs differ (MFMA intrinsic names, CU counts, warp tiles). 24 tests are arch_specific: the split-K cost model is fitted for gfx950 only, and split_k_cost_model("gfx942") returns None.
• dispatch_tests/core, moe, norm, tuning.
  • Core, build, binding and norm tests run on both archs.
  • test_moe_matches_the_example_exactly is gfx950 only. The example spec is invalid on gfx942 (bf16 warp tile (16, 16, 32)).
  • The moe tests that name one arch are marked.
  • The eight classes in tuning/test_tuning_template.py are marked gfx950. The toy family names gfx950 in 33 fixed labels. The tuning mechanics do not depend on the arch. Parametrizing needs about 30 label changes. Reviewers: say if you prefer parametrization.
• runtime, tools, analysis, numeric.
  • The comgr, device_info, arch-domain artifact, LDS registry and b32 LDS profile tests run on both archs.
  • The LDS profile tests with different bank behavior, the tf32 test and the f64 MFMA test are marked. Each reason states the difference.
  • test_lds_profile_gfx942.py and test_lds_profile_gfx950.py test their own profile only. Every test there is marked. The cases that both profiles share are in the new test_lds_profile_cdna_common.py, parametrized over both archs.
  • Last commit: test_f64_gemm_rejects_unsupported runs on both archs with per-arch messages and LDS caps. The fp64 pad budget test, the moe fp8 K=128 catalog test, the gfx942 lds_swizzle numeric test and five dsl_probes tests are marked.

Behavior changes to check

• The gemm runtime tests gated on gfx950 before. They now also run on a gfx942 runner. They passed on one gfx942 device.
• test_dispatch_gfx950_request is renamed test_dispatch_cdna_request.

How this was tested

• 478 tests passed, 8153 subtests passed, on the changed files before the last two commits. The run excluded core/test_example_parity.py::TestEveryRegisteredKernelBuilds. That pair builds every kernel on every arch and is too slow for a local run.
• Docker, llvm20 (rocm7.1.1), llvm22 (rocm7.2) and llvm23 (rocm10.1_therock): 473 passed, 10 skipped, 0 failed on the 27 files at the merge commit. The last commit's 4 files: 56 passed, 26 skipped, 0 failed on each flavor. The skips need a GPU. The containers have none.
• The local GPU is a gfx942. The gfx950 cases ran without a gfx950 device.
• ruff check without --fix: no new finding on the changed files. Pre-commit passes.
• Scan, with a script that reads arch_specific, scope of AICK-2822: unmarked single-arch tests go from 175 on develop to 7. The 7 are scan false positives (an arch name in a label, comment or docstring). Marked tests go from 2 to 78. Tests that run on both archs go from 28 to 145.

Review follow-up

• One constant, COVERED_ARCHES in rocke.tests.arch_marks, replaces the per-file copies of the gfx942 and gfx950 pair.
• The f64 MFMA probe mark states the real limit: the probe builds for gfx942 only.
• The common LDS profile file drops its duplicate helpers.
• test_support.py derives the 32x32 warp tile from the MMA catalog.
• The device_info tests and three comgr tests that only changed a label run on one arch again.

Not done

• The scan does not prove that both archs run. It reads arch names in source.
• No gfx950 hardware run.

---

Summary — For AI reviewers

Change surface

30 test files, one of them new (tests/analysis/test_lds_profile_cdna_common.py). No emitter, core, helpers, instances or library/kernels file changes. The byte-identity gate is not affected.

Properties to verify

1. Each parametrized assertion sits inside the subTest loop, so a gfx942 failure does not hide a gfx950 case.
2. Each arch_specific call states a reason that matches the code under test.
3. Per-arch expected values come from the arch catalog or the arch target, not from a copy of the production result.
4. Out-of-scope archs (gfx1151, gfx1250, gfx90a, gfx1201) are unchanged.

Where to look first

1. gemm/test_fp16_rcr.py: per-arch MFMA intrinsic table.
2. gemm/test_split_k_wgrad.py: per-arch default group counts.
3. tuning/test_tuning_template.py: the _toy_arch marker.
4. tools/test_arch_domain_artifact.py: ARCH class attribute and *Gfx950Test subclasses.

🤖 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 #15004 — test(rocke): cover gfx942 and gfx950 in the dispatch and small platform tests (CongMa13, AICK-2822) — 🟠 First-round: arch list copies; body figure

Prior item	Status
Reviews and threads	🟢 None yet • 0 reviews, 0 threads; pr-bot policy ✓ (read) • rocke-platform review requested (read)
Bots / CI	🟡 Red, develop's • fails: 4 math-libs builds + summary — develop's 14976 hipCCL break (read) • merge ref predates revert 15089; no test job ran; needs a develop merge or push • previous head 456b6abede, same test diff: rocKE shards ran these files; gfx942 runtime tests ✓ (read) • its only ✗: develop's fp64 doc test, fixed by 15002 (read) • 27 ✓, 48 skipped; merge: blocked
Our checks	🟢 Pass • merge vs 96e7c35607: clean, 88 behind (ran) • build + ctest 51 ✓ / 1 skip; byte identity ×4 GREEN; goldens unchanged (ran) • changed tests 499 ✓ / 36 skip; installed 28 of 30 files 388 ✓ / 37 skip, 36 device-gated (ran) • no assertion dropped (read); 10 per-arch mutants: these tests catch 10, develop's versions 6 (ran) • 2 tools files in no CI lane (develop's install excludes them); gfx950 cases ✓ from source (ran) • rules 1–3 clean; body has one measured figure (N1) (ran)
Code quality	🟠 Should-fix • one home per fact: the covered-arch pair has 13 (N2) (ran) • LDS common file repeats the per-arch files' _access/_predict; positional LdsAccess(0, 0, 0, 4) needs neither (read) • test_support.py:21 hand-lists an input select_largest_k gives (ran) • label-only loops (core, comgr, device info) add runs, not coverage; the tuning marks fit (read)

#	Sev	file:line	Finding
N1	🟠	PR body, "How this was tested"	Body states a measured test run time • impact: a squash commits the body; Compliance allows only measured ratios there (read) • fix: say it qualitatively ("too slow for a local run"); the test counts can stay
N2	🟠	platform/tests/dispatch/dispatch_tests/core/test_core.py:48	The covered-arch pair gets 13 homes under four names • evidence: _ARCHES, ARCHES, _CDNA_ARCHES (7 gemm files), _CDNA_TARGETS, plus 6 inline lists; the arch table's cdna family has five arches (read) • impact: test_core.py's fixtures keep their own pair; one more arch in _ARCHES fails 2 widened tests (ran) • fix: one public constant, e.g. COVERED_ARCHES in rocke.tests.arch_marks, imported at each site; build both _registry() fixtures from it: then passes (ran); runtime gates via current_device_arch_in
N3	🟡	platform/tests/numeric/test_gfx942_f64_mma_numeric.py:73	The f64 MFMA mark's reason is false at head • impact: gfx950's catalog lists mfma_f64_16x16x4_f64 (develop's 14328, merged here 11-03); that atom gets no probe case (ran) • fix: state the real limit (the probe builds for gfx942), or add gfx950 m=16 cases: they lower and compile (ran)

Utility: 🟢 Positive — widened tests catch four per-arch breaks develop's versions missed

---

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, byte identity, goldens, changed tests (source and installed), per-arch mutation runs, CI logs at this and the previous head, code quality. Flags whether a closer look is needed; not an approval or rejection.

[CongMa13]

Addressed in ddc24d40c0a (merged with latest develop).

• N1: The PR body now says the excluded test pair is "too slow for a local run". Test counts stay.
• N2: COVERED_ARCHES now lives in rocke.tests.arch_marks. The 18 test files import it. Both test_core.py _registry() fixtures use it. test_lds_registry.py and test_f64_gemm_numeric.py keep their own lists because they name different arch sets.
• N3: The f64 mark states the real limit: the probe builds for gfx942 only. No gfx950 m=16 cases added.
• LDS common file: The duplicate _access and _predict helpers are gone.
• test_support.py: The 32x32 warp tile comes from select_largest_k.
• Label-only loops: The two device_info tests and three comgr tests run on one arch again. The tuning marks stay.

Tests: 499 passed, 36 skipped on the changed files after the merge. Pre-commit (black 25.12.0) passes.

[cgmillette]

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

PR #15004 — test(rocke): cover gfx942 and gfx950 in the dispatch and small platform tests (CongMa13, AICK-2822) — 🟢 Follow-up: all three items fixed

Prior item	Status
cgmillette (11-06) N1 body states a measured run time	🟢 FIXED (body edit) • body now says "too slow for a local run" (read) • no measured time left in the body (ran)
cgmillette (11-06) N2 covered-arch pair has 13 homes	🟢 FIXED eddc69b975 • COVERED_ARCHES in rocke.tests.arch_marks; 18 files import it (ran) • mutated constant: every site follows, test_core fixtures too (ran) • kept lists (LDS registry, f64 GEMM) name other arch sets (read)
cgmillette (11-06) N3 f64 MFMA mark's reason false	🟢 FIXED eddc69b975 • reason matches: the probe lowers and builds for gfx942 only (read) • saved probe reads FIXED (ran)
Reviews and threads	🟢 None yet • 0 reviews, 0 threads; merge blocked on review (read)
Bots / CI	🟢 Green • attempt 1: Windows rocKE shard timed out on the Azure runner pool; attempt 2 ✓ (read) • gfx94X shards ran the gemm runtime tests: passed (read) • 45 ✓, 52 skipped, 1 cancelled (superseded PR-bot run) (read)
Our checks	🟢 Pass • merge vs a9f06669a3: clean, 27 behind (ran) • build + ctest 57 ✓ / 1 skip; byte identity ×4 GREEN; goldens unchanged (ran) • changed tests 494 ✓ / 36 skip; installed 28 of 30 files 383 ✓ / 37 skip (ran) • 4 COVERED_ARCHES mutants over the 18 importing files (ran) • rules 1–3 clean; body has no figure (ran)
Code quality	🟢 Good • one home per fact: COVERED_ARCHES, read at every site (ran) • LDS helpers gone; test_support takes the tile from select_largest_k (read) • comgr and device-info label loops gone; core's stay by design (read) • nit: test_core.py:47-48 comment still describes the removed tuple (read)

Net-new: none.

Utility: 🟢 Positive — one arch constant now drives every widened test; items fixed

---

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

Follow-up review: first pass by qwen3.6 (local), verified by the orchestrator (Claude Opus 5.5). Checks: merge with develop, build + ctest, byte identity, goldens, changed tests (source and installed), arch-constant mutants, CI logs at this head, code quality. Flags whether a closer look is needed; not an approval or rejection.