[aalabsi-amd]

JIRA ID : ALMIOPEN-1926

Motivation

#14169 (merged 2026-10-26, cec66d06) fixed the data race in RNNDescriptor::RNNForwardMS that made bias-enabled
Full/GPU_LstmMSRnn_FP32.TestFloat cases fail intermittently. Every quarantine of this family still in the tree predates that fix:

Where	Pattern	Added by	Data it was based on
Jenkins MIOPEN_TEST_GFX94X negative filter	Full/GPU_LstmMSRnn_FP32*	#13107	Jenkins gfx942, 2026-08-24..10-23
comprehensive.exclude, full.exclude	Full/GPU_LstmMSRnn_FP32.TestFloat/1095	#13107	Jenkins per-index filter
exclude_gpu_gfx94X	Full/GPU_LstmMSRnn_FP32*	#13107	Jenkins gfx942 (above)
exclude_gpu_gfx110X	Full/GPU_LstmMSRnn_FP32*	#13107	Windows gfx110X runs of 10-15/10-16
exclude_gpu_gfx115X	Full/GPU_LstmMSRnn_FP32*	#13107	Windows gfx1151 run of 10-20

The #14169 author planned this cleanup as a follow-up once the fix was confirmed. While these quarantines stay, a regression in the
LSTM multi-stream forward pass (reachable from miopenRNNForward, miopenRNNForwardTraining and miopenRNNForwardInference) goes
untested on gfx942, gfx110X and gfx1151.

Technical Details

• projects/miopen/test/gtest/test_categories.yaml:
  • remove /1095 from the comprehensive and full excludes;
  • remove Full/GPU_LstmMSRnn_FP32* from exclude_gpu_gfx110X and exclude_gpu_gfx115X;
  • remove the exclude_gpu_gfx94X key. Its only pattern was this family, so gfx94X goes back to the base suites (-LE ex_gpu),
    as before test(miopen): add Jenkins-only gtest families to test_categories.yaml… #13107. That's 21 suites instead of 25.
• projects/miopen/test/gtest/CMakeLists.txt: remove the Jenkins MIOPEN_TEST_GFX94X block, which only held this filter.
• Kept: exclude_gpu_gfx115X_windows Full/GPU_GRU_Base_FP16…biasMode_1_*. GRU doesn't use the multi-stream LSTM forward that
  fix(miopen): order LSTM multi-stream fwd bias add before GEMM #14169 changed, so the fix doesn't cover that flake.
• Kept: standard.exclude Full/GPU_LstmMSRnn_*. That's tier placement (the Full/ cases run in comprehensive), not a quarantine.

Test Plan

1. parse_test_categories.py: rc 0, 21 suites.
2. yaml_select.py diff --old git:origin/develop, using the 69,069 case names from miopen_gtest --gtest_list_tests of the post-fix build:

GPU / OS	standard	comprehensive / full
gfx942 Linux	0	+1,536 Full/GPU_LstmMSRnn_FP32
gfx1100 Windows	0	+1,536
gfx1151 Windows / Linux	0	+1,536
gfx1201 Linux, gfx950 Linux	0	+1 (/1095; they already ran the rest)

3. Local A/B stress on gfx942 (MI300X). Same method as fix(miopen): order LSTM multi-stream fwd bias add before GEMM #14169. The pre-fix build is TheRock CI artifacts of rl run 39358709264
   (2026-10-20); the post-fix build is rl run 41582846076 (develop 9953cb5 + PR test(miopen)[ALMIOPEN-2966]: exclude TheRock bump failures on gfx120X and Windows gfx1151 #14882). Two modes: (A) one process with
   GPU_MAX_HW_QUEUES=16; (B) 4 gtest shards sharing one GPU, as Jenkins does. Rounds alternate pre/post, each with a fresh user db
   and kernel cache. Script: /scratch/users/aalabsi/tmp/lstm_ab/ab.sh.
4. CI on this PR, labels test:miopen + test_filter:comprehensive + ci:gfx1151. This runs
   comprehensive (including all 1,536 Full/GPU_LstmMSRnn_FP32 cases) on Linux gfx94X, Windows gfx110X and Windows/Linux gfx1151.
   These are the lanes that failed before. Read every attempt.
5. Jenkins PR: the gfx942 stages run the family again now that the filter is gone.

Test Result

Local A/B stress, gfx942 (MI300X, banff-cyxtera-s70-4), Slurm jobs 68262769 (3 rounds) and 68262798 (20 rounds),
23 rounds per build per mode, pre/post alternating. Each round runs all 1,536 cases: 1,152 execute and 384 skip by design at
rnn_seq_api.cpp:61 in both builds.

Build	Mode A: 1 process, GPU_MAX_HW_QUEUES=16	Mode B: 4 shards on 1 GPU (Jenkins-like)
pre-fix (rl run 39358709264, 2026-10-20)	0 failures / 26,496	150 failures / 26,496 (0.57%), in 23 of 23 rounds, 113 distinct indices, all at rnn_seq_api.hpp:272 (the ALMIOPEN-1926 signature)
post-fix (rl run 41582846076, develop 9953cb5)	0 / 26,496	0 / 26,496, in 0 of 23 rounds

At the pre-fix rate, 26,496 post-fix executions would give about 150 failures. Seeing 0 by chance is effectively impossible
(p ≈ e^-150). Mode A didn't reproduce on this machine for either build; #14169 also needed mode B or many passes. A round takes
77–81 s for both builds, so the fix costs no measurable test time. Logs:
/scratch/users/aalabsi/tmp/lstm_ab/results-68262769/, …/results-68262798/.

Jenkins develop, gfx942 (QE dashboard): 20 failures / 11,719 executions from 2026-09-01 to 10-26; 0 / 1,144 from 10-27 to 10-30,
after which #13107 filtered the whole family. gfx90a: 0 / 11,615 before and 0 / 1,872 after; it was never quarantined.

CI on this PR: pending. Windows gfx110X and gfx1151 are the open question: no Windows run of this family exists after the fix,
and the stress test above covers gfx942 only.

Submission Checklist

• Every removed entry cites ALMIOPEN-1926 and is covered by fix(miopen): order LSTM multi-stream fwd bias add before GEMM #14169
• YAML parses; selection diff shows only the intended cases
• gfx942 A/B: pre-fix fails under stress (150 / 26,496), post-fix doesn't (0 / 26,496)
• PR CI comprehensive: Windows gfx110X / gfx1151 and Linux gfx94X run Full/GPU_LstmMSRnn_FP32 with 0 failures in every attempt
• Jenkins gfx942 PR stages green with the filter removed
• Remove the test_filter:comprehensive label before merge
• Close ALMIOPEN-1926, or narrow it to the GRU FP16 flake

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 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.

[JonathanLichtnerAMD]

The yaml and CMake cleanup is correct and complete. MIOPEN_TEST_GFX94X is still used elsewhere, so removing only the filter block is right. The kept GRU FP16 and standard-tier entries are also right.

Needs a follow-up when merging with develop: #15021 added projects/miopen/AGENTS.md, whose example ctest ... -L '^ex_gpu_gfx94X$' uses the label this PR deletes. After merge that command fails with --no-tests=error. Please update it in this PR once you rebase or merge develop, or in a follow-up. Suggested change (untested; ex_gpu_gfx950 still exists at test_categories.yaml:753):

-# test_categories.yaml, often a family name (gfx942 -> ex_gpu_gfx94X).
-ctest ... -L '^quick$' -L '^ex_gpu_gfx94X$'
+# test_categories.yaml, often a family name (gfx950 -> ex_gpu_gfx950).
+ctest ... -L '^quick$' -L '^ex_gpu_gfx950$'

Request: read every Windows gfx110X and gfx1151 attempt of Full/GPU_LstmMSRnn_FP32, because the stress evidence covers gfx942 only.

No blockers. 0 suggestion blocks are included, since the AGENTS.md line is not in this PR's diff.

Co-authored by Claude (AI).