[dileepr1]

Summary

TheRock side of test-filter standardization for rocSHMEM, paired with ROCm/rocm-systems#6334.

rocm-systems#6334 adds native CTest integration for rocSHMEM: projects/rocshmem/tests/test_categories.yaml plus apply_ctest_category_labels() wiring that installs relocatable bin/rocshmem/**/CTestTestfile.cmake files (a root dispatcher that subdirs() into tests/functional and tests/unit) carrying the quick/standard/comprehensive/full tier labels. This PR wires TheRock so the generic test_runner.py can drive those tests from the installed artifact:

• comm-libs/CMakeLists.txt — pass -DBUILD_CTESTS=${THEROCK_BUILD_TESTING} to the rocshmem subproject. The CTest registration and the install-tree CTestTestfile.cmake generation in (feat): [AILIKFD-40] add kfdtest to TheRock CI #6334 are gated behind BUILD_CTESTS (default OFF), so without this the install tree would contain no bin/rocshmem/** and test_runner.py would fail with Test directory does not exist.
• comm-libs/artifact-rocshmem.toml — populate the previously-empty test component: bin/rocshmem/** (the install-tree CTest files), bin/rocshmem_unit_tests, the share/rocshmem/* CTest wrapper scripts, and the device_bitcode HSACOs. The functional binary and rocshmem_info already live in the run component (fetched via rocshmem_run).
• build_tools/install_rocm_from_artifacts.py — add a --rocshmem artifact selector that pulls rocshmem_lib/_run/_test plus OpenMPI (mpirun + libs) needed to launch the tests.
• build_tools/github_actions/fetch_test_configurations.py — add a rocshmem test job using test_runner.py (fetch_artifact_args = "--rocshmem --tests"), on Linux multi-GPU runners (the functional/unit tests launch via mpirun with RANKS 2..64, same setup as rccl).
• fetch_test_configurations_test.py — relax the assertion that assumed exactly one multi-GPU job repo-wide, now that rocshmem is a second multi-GPU job alongside rccl.

test_runner.py needs no COMPONENT_DIR_MAPPING entry (job name rocshmem == install dir bin/rocshmem).

Test plan

• python3 -m pytest build_tools/github_actions/tests/fetch_test_configurations_test.py (22 passed)
• black --check on the edited Python files
• --rocshmem registered in install_rocm_from_artifacts.py --help
• artifact-rocshmem.toml parses; test include resolves
• CI: build with THEROCK_BUILD_TESTING=ON, confirm bin/rocshmem/CTestTestfile.cmake + tests/{functional,unit}/CTestTestfile.cmake are installed and bundled in the test artifact
• CI: TEST_COMPONENT=rocshmem TEST_TYPE=quick python test_runner.py discovers labels and runs ctest -L quick from THEROCK_BIN_DIR/rocshmem

Test Results

Test Run: https://github.com/ROCm/TheRock/actions/runs/30112055659/job/89560360151?pr=5691

Made with Cursor

[geomin12]

tests fail + merge conflict

but it looks good

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

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

[therock-pr-bot]

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

[geomin12]

lgtm

[ScottTodd]

We don't build openmpi by default and we don't distribute a vendored version.

You may be interested in how rocprofiler-sdk has recently started running mpi tests by using a test dockerfile with a system install:

• ci: add no_rocm_image_ubuntu24_04_openmpi test image #6610
• fix(rocprofiler-sdk): enable rocprofv3 mpi-ranks tests in CI #5627

[ScottTodd]

fyi @amd-shiraz new 8 GPU runner like linux-gfx942-8gpu-ossci-rocm-8txs9-runner-wknrq in use here, e.g. https://github.com/ROCm/TheRock/actions/runs/30112055659/job/89560360151

[amd-shiraz]

this will cause huge queues with current capacity for gfx942-8gpu. do these actually need multi gpu? cc @suchitravemuri1

[ScottTodd]

What tests do you specifically expect to run with this script? Looking at the logs on https://github.com/ROCm/TheRock/actions/runs/30112055659/job/89560360151, I see

28/28 Test #191: pingpong_n2_w32_z1_uuid ..................................   Passed    1.53 sec

100% tests passed, 0 tests failed out of 28

Label Time Summary:
AMO                       =  21.74 sec*proc (4 tests)
OTHER                     =  47.31 sec*proc (7 tests)
RMA                       = 128.30 sec*proc (17 tests)
backend_all               = 129.94 sec*proc (17 tests)
backend_gda               =  55.17 sec*proc (7 tests)
backend_ipc               =  67.41 sec*proc (11 tests)
comprehensive             = 197.34 sec*proc (28 tests)
full                      = 197.34 sec*proc (28 tests)
functional                = 197.34 sec*proc (28 tests)
gpu_all                   = 197.34 sec*proc (28 tests)
launcher_slr              = 197.34 sec*proc (28 tests)
nightly                   =  96.18 sec*proc (12 tests)
pr                        =   7.85 sec*proc (2 tests)
pre-commit                = 197.34 sec*proc (28 tests)
quick                     = 197.34 sec*proc (28 tests)
smoke                     = 197.34 sec*proc (28 tests)
standard                  = 197.34 sec*proc (28 tests)
variant_default_stream    =  15.98 sec*proc (2 tests)
variant_uuid              = 189.14 sec*proc (27 tests)

Total Test time (real) =  88.85 sec
# TEST_COMPONENT: rocshmem -> Test Directory: build/bin/rocshmem
# TEST_TYPE: quick -> Category: quick
# AMDGPU_FAMILIES: gfx94X-dcgpu -> GPU Architecture: gfx94X

# Discovering available test labels...
# Warning: No GPU specific test suites available

# No GPU suite found for gfx94X, excluding all ex_gpu tests
# Running: ctest -L ^quick$ -LE ex_gpu --output-on-failure --parallel 1 --timeout 7200 --test-dir build/bin/rocshmem -V --tests-information 1,,1

That seems pretty light on tests, even for the "quick" filter? The "No GPU suite found" logs are also concerning given that this claims it should be running multi-GPU tests?