[michaelzhang-ai]

Motivation

Eight test files carry an register_amd_ci call that never dispatches, so the AMD coverage their owners registered does not run. Two independent causes:

1. #37436 dropped AMD registrations. That PR collapsed each file's two registrations into one, kept only the CUDA line, and folded the AMD est_time into it (e.g. test_engine_child_pids.py CUDA est_time went 38 → 77, the old AMD value). The AMD coverage was not retired on purpose. #38581 restored five such files; these seven are the remainder:

File	Registered by
core/test_engine_child_pids.py	#25208 (AMD)
openai_server/basic/test_http2_server.py	#24572 (AMD)
core/test_request_queue_validation.py	#16426
observability/test_priority_metrics.py	#21740
openai_server/validation/test_large_max_new_tokens.py	#16326
openai_server/validation/test_matched_stop.py	#16326
openai_server/validation/test_request_length_validation.py	#16326

2. The 3FS hicache test has never run on AMD. Its AMD registration named a non-AMD suite from the day it was added in #17116 — the sibling file-backend test got stage-b-test-large-2-gpu-amd, this one was written without the -amd suffix. #25420's mechanical stage- → base- rename then carried the AMD line along with the CUDA one, leaving suite="base-b-test-2-gpu-large": a CUDA suite name that no AMD workflow invokes.

Nothing caught either regression. run_suite.py selects on effective_suite, and AMD sits outside its _SUITE_CHECKED_BACKENDS, so a suite name belonging to another backend matches zero tests and the job prints No tests found ... Skipping instead of failing.

Modifications

Three commits, one per logical change:

1. Restore the seven dropped registrations, each at the suite and est_time it carried before [Test] Consolidate test cleanup and CI taxonomy (net -11.4K lines) #37436. All seven target stage-b-test-1-gpu-small-amd. CUDA/CPU registrations are untouched — the est_time values [Test] Consolidate test cleanup and CI taxonomy (net -11.4K lines) #37436 folded into the CUDA lines are refreshed weekly from sglang-ci-stats, so they self-correct.

2. Retarget the 3FS hicache AMD registration at stage-b-test-2-gpu-large-amd, matching test_hicache_storage_file_backend.py, whose mixin this test derives from. The 3FS client is mocked (use_mock_hf3fs_client), so no 3FS hardware is required.

3. Add the AMD analogue of the existing non-dispatchable-CUDA-suite check to scripts/lint/check_registered_tests.py: an AMD suite name must carry amd as a hyphen-delimited token, or be a nightly-* suite (nightly-test-amd.yml names those after the model rather than the vendor, e.g. nightly-perf-8-gpu-grok2). All 124 AMD suite names currently in test/registered/ satisfy this rule; the only violation was the 3FS typo, so the guard is a no-op on main apart from catching recurrences.

Accuracy Tests

No model, kernel, or forward-path code is touched; this is CI registration metadata plus a lint rule.

Speed Tests and Profiling

Both affected jobs keep comfortable headroom:

Suite	Partitions	Est/partition	timeout-minutes
stage-b-test-1-gpu-small-amd	14	1087s → 1122s (+35s)	75
stage-b-test-2-gpu-large-amd	2	2666s → 2816s (+150s)	120

Verified locally that strict collection still succeeds over all 2491 registrations, that each restored registration resolves to stage-b-test-1-gpu-small-amd with an executable __main__ entry, and that the new lint rule fails on the pre-fix 3FS suite name and passes on the whole tree after it.

Because these tests have not run on AMD hardware recently (and the 3FS one never has), please trigger the AMD suites on this PR so the restored files are validated before merge.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Notes for reviewers

Per write-sglang-test, register_amd_ci should not be added unless a test exercises ROCm paths. This PR adds no new coverage decisions — every registration here restores one an AMD CI owner already made, at its original suite and est_time. If any of the seven is no longer wanted on AMD, the right outcome is an explicit disabled="reason" rather than the current silent drop.

Two adjacent gaps found while auditing, left out of scope: nightly-amd-8-gpu and nightly-amd-vlm are in NIGHTLY_SUITES with zero registered tests and no workflow dispatch, and 94 of the 124 registered AMD suite names are absent from run_suite.py's catalog (harmless today only because AMD is exempt from validate_all_suites). Adding AMD to _SUITE_CHECKED_BACKENDS would make the catalog authoritative, but it requires cataloguing all 94 names and adds friction for AMD contributors, so it belongs in a separate PR with AMD CI owners' input.

  

---

CI States

Latest PR Test (Base): ❌ Run #35623020492
Latest PR Test (Extra): ❌ Run #35623020053
Latest PR Test (AMD ROCm 10): ❌ Run #35623020472