fix(amdsmi): isolate the tests package scriptlets and require the distro install matrix by marifamd · Pull Request #9110 · ROCm/rocm-systems

marifamd · 2026-07-23T19:20:18Z

Description

Isolate the component maintainer scripts and tighten the ABI/wheel build gates. Stacks on #9035.

Scriptlet ownership

fix: stop the tests package from running the main package scriptlets — the amd-smi-lib-tests package inherited the main package's post/preun scriptlets, so removing the tests package could run the main package's ld.so.conf.d cleanup and corrupt the main install.
fix: give every component its own maintainer scripts — scoping only the tests package left the same hole open for asan. An unprefixed CPACK_RPM_<KIND>_SCRIPT_FILE is copied into every component RPM, and a Debian component without its own control-extra inherits the generic one. The scripts manage absolute paths shared by every variant (/etc/ld.so.conf.d, /var/log/amd_smi_lib, /etc/logrotate.d), so an inheriting component clobbers the main package even under a different prefix. Name the owning component group for every script instead; anything not named ships nothing, so a component added later cannot inherit by accident.
fix: resolve the tests cleanup path at install time — the tests pre-uninstall built its path from the configure-time prefix instead of $RPM_INSTALL_PREFIX0, so a relocated install cleaned the wrong tree.

Build gates

test: fail the ABI gate when the binding scan matches nothing — the ABI compatibility gate silently passed when the ctypes binding scan matched zero symbols (e.g. a generator rename); it now fails loudly.
build: only require Python3 for the wheel build — scope the find_package(Python3) requirement to BUILD_PYTHON_WHEEL=ON so the default (non-wheel) build does not fail on a missing Python3.
fix: don't fall back to an unset Python3_SITELIB — with find_package(Python3) no longer running for a plain C build, the module-destination fallback to Python3_SITELIB would install to the prefix root. Fall back only when it holds a value, and fail loudly otherwise.

JIRA ID

ROCM-3941

therock-pr-bot · 2026-07-24T14:06:05Z

✅ All Policy Checks Passed

Check	Status	Details
📝 PR 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.

🙋 Wish to Override Policy?

amd-smi-lib-tests depends on amd-smi-lib, so removing or upgrading the tests package must never disturb the main package's state. #9110 fixes that; this adds the static regression guard it did not ship with. The assertions target the invariant rather than any one remedy: the tests component may carry its own scriptlets, or simply be left without the main package's, and either satisfies them. Verified green against #9110's implementation and red against the pre-fix templates (15 failures). Records the CPack 3.28.3 finding behind the RPM half of the bug: an unprefixed CPACK_RPM_<KIND> scriptlet variable is copied into every component RPM, so a global assignment has to be paired with a CPACK_RPM_TESTS_<KIND> override to keep it out of the tests package. JIRA ID: ROCM-3941

Copilot

Pull request overview

This PR hardens the amdsmi packaging and ABI gate behavior by isolating the tests-package uninstall hooks from the main package, tightening the ABI compatibility test, and scoping Python discovery to wheel-specific builds.

Changes:

Prevent amd-smi-lib-tests install/uninstall from inheriting main-package RPM/DEB scriptlets that could break a still-installed main package.
Make the Python ABI gate fail if the binding scan returns zero matches (avoids a silent no-op gate).
Require find_package(Python3) only when BUILD_PYTHON_WHEEL=ON, and add safer fallbacks for determining Python3_SITELIB.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
projects/amdsmi/tests/python/test_abi_compat.py	Adds a non-empty assertion to ensure the ABI scan doesn’t silently match nothing.
projects/amdsmi/RPM/tests-preun.in	Introduces tests-only RPM pre-uninstall cleanup for test `__pycache__`.
projects/amdsmi/RPM/tests-noop.in	Adds a no-op RPM scriptlet to prevent tests component from running main scriptlets.
projects/amdsmi/py-interface/CMakeLists.txt	Scopes Python discovery to wheel builds and hardens sitelib detection/fallbacks.
projects/amdsmi/DEBIAN/amd-smi-lib-tests/prerm.in	Replaces inherited main-package cleanup with tests-only `__pycache__` removal on remove.
projects/amdsmi/CMakeLists.txt	Wires per-component RPM scriptlets for the tests component to avoid inheriting global hooks.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

amd-smi-lib-tests depends on amd-smi-lib, so removing or upgrading the tests package must never disturb the main package's state. #9110 fixes that; this adds the static regression guard it did not ship with. The assertions target the invariant rather than any one remedy: the tests component may carry its own scriptlets, or simply be left without the main package's, and either satisfies them. Verified green against #9110's implementation and red against the pre-fix templates (15 failures). Records the CPack 3.28.3 finding behind the RPM half of the bug: an unprefixed CPACK_RPM_<KIND> scriptlet variable is copied into every component RPM, so a global assignment has to be paired with a CPACK_RPM_TESTS_<KIND> override to keep it out of the tests package. JIRA ID: ROCM-3941

Packaging install-path and multi-Python-version test coverage for the AMD SMI Python packaging rework. - Loader across Python versions: verify loader behavior across supported interpreters; run the 3.6/3.7 loader tests in python container images. - Install-path assertion: assert the system package ships the module on an importable path (site-packages/dist-packages). - Upgrade/downgrade round trip: exercise the deb/rpm upgrade and downgrade path in CI; select upgrade-test packages by build order (not version string); fail the upgrade test when the ld.so.conf.d linker registration is missing; run the CLI check in the upgrade test. - Python 3.6 compatibility: make the version and package tests run on Python 3.6. - CI hygiene: wire the packaging tests into the harness and CI, sparse-checkout the python-version and upgrade jobs, pin actions/setup-python to a commit SHA. JIRA ID: ROCM-3941

The RPM declares a versioned interpreter dependency whose form differs by distro family: `python(abi) = X.Y` on RHEL, but `pythonXY` with no dot (e.g. python311) on SLES/openSUSE. The GPU build matrix never exercises the SLES form, so a broken dependency token could ship unnoticed. Build the RPM in an openSUSE Leap 15.6 container and assert the pythonXY requirement both appears and resolves to an installable provider, then check the module ships on an import path. No GPU required. JIRA ID: ROCM-3941

Without timeout-minutes the job inherits the six-hour default, so a hung zypper refresh or a stalled container ties up a runner for the rest of the day instead of failing fast. Forty-five minutes matches the other packaging jobs, which build the project in a container the same way. JIRA ID: ROCM-3941

…iptlets - the global CPACK_RPM_*_SCRIPT_FILE vars stamp the main scriptlets onto every component RPM, so removing amd-smi-lib-tests ran the main package's ldconfig and cleanup and corrupted the still-installed main package - override the tests RPM hooks per-component: a tests-only pre-uninstall that clears just its own __pycache__, and no-op post/post-uninstall - rewrite the DEB tests prerm to the same tests-only cleanup so it no longer removes the main package's linker entry, bytecode, doc or symlink - clean only on a full erase, never on upgrade Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> (cherry picked from commit 64f9625)

- the no-clang branch copies the committed wrapper with plain cmake and never uses the interpreter, but still ran find_package(Python3 REQUIRED), so a C-only build failed at configure before reaching the sitelib step - gate the find_package on BUILD_PYTHON_WHEEL so a C-only build with an explicit AMDSMI_SYSTEM_PYTHON_SITELIB no longer needs an interpreter Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> (cherry picked from commit 16cc3e3)

- test_unguarded_bindings_are_stable subtracts a set from STABLE_SYMBOLS, so an empty scan trivially passes and the gate silently checks nothing - assert the scan is non-empty first, so a wrapper layout change that breaks _scan_unguarded_bindings() fails loudly Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> (cherry picked from commit af45ac4)

Requiring Python3 only for the wheel build means find_package(Python3) no longer runs for a plain C build: BUILD_PYTHON_WHEEL and BUILD_WRAPPER both default to OFF, so the default distro package build takes the NOT GOOD_CLANG_FOUND branch and never calls it. Python3_SITELIB is then undefined. The module-destination detection still fell back to that variable when the site-packages query failed, which would set AMDSMI_SYSTEM_PYTHON_SITELIB to "" and install the module to the prefix root -- the silent misplacement this detection exists to catch, reported only by a warning that trails off after "falling back to Python3_SITELIB=". Fall back only when Python3_SITELIB actually holds a value, and fail loudly otherwise, matching the neighbouring branch that already refuses to guess a destination no interpreter has on sys.path. JIRA ID: ROCM-3941

Scoping only the tests package left the same hole open for asan. An unprefixed CPACK_RPM_<KIND>_SCRIPT_FILE is copied into every component RPM, and on the Debian side a component without its own control-extra inherits the generic one, so amd-smi-lib-asan still shipped the main package's postinst/prerm/post/preun/postun. That matters even though asan installs under its own prefix: the state those scripts manage -- /etc/ld.so.conf.d/x86_64-libamd_smi_lib.conf, /var/log/amd_smi_lib and /etc/logrotate.d/amd_smi.conf -- is at absolute paths shared by every variant. Installing the asan package repoints the main package's linker registration at the asan library, and removing it deletes that registration, the log directory and the logrotate config out from under a still-installed amd-smi-lib. Name the owner explicitly for every script instead, which is what rocm-core, clr and rdc do. The name is the component GROUP, not the component: generic_package_post() puts dev into the "runtime" group, so CPack emits one package per group and a CPACK_*_DEV_* variable is silently ignored. Anything not named ships nothing, so a component added later cannot inherit by accident. Dropping the generic Debian control-extra also makes the tests package deterministic: it previously depended on two prerm files colliding by filename in the control archive. RPM/tests-noop.in is no longer needed now that nothing leaks, so it goes. Verified with a three-component CPack project: dev carries three scriptlets, tests one, and a component named nowhere carries none. JIRA ID: ROCM-3941

The tests pre-uninstall built its path from @CPACK_PACKAGING_INSTALL_PREFIX@, the prefix baked in at configure time, while every other RPM scriptlet uses $RPM_INSTALL_PREFIX0, the prefix the package was actually installed under. For a relocated install the two differ, so the cleanup would look in the wrong tree and leave the generated __pycache__ behind. JIRA ID: ROCM-3941

ammallya · 2026-08-11T21:54:58Z

Flaky github fetch sources error, build and test for smi are good

…ation (#9695) Stacked on #9110 — this targets that branch, not `develop`. ``` develop └─ #9034 packaging install-path and multi-Python coverage └─ #9035 SLES RPM interpreter dependency check └─ #9110 component scriptlet isolation └─ #9695 (this PR) ``` ## What this is The regression tests #9110 didn't ship with, plus fixes for several unrelated defects in the packaging test suite that let real breakage through. Nothing here changes #9110's fix — this PR adds tests and CI only, apart from the two behavior changes called out at the bottom. ## Regression coverage for #9110 #9110 stops `amd-smi-lib-tests` from tearing down the still-installed `amd-smi-lib`. Two tests pin it: - **`test_packaging_scriptlets.py`** — static guard on the templates and the CPack wiring. Written against the *invariant* rather than one particular remedy, so it accepts either "give the tests component its own scriptlets" or "scope the globals to the owner". - **`run_amdsmi_component_removal_test.py`** + a workflow step — installs both packages in the distro container, removes only the tests package, and asserts the main package survives: still installed, linker registration intact, module still importable and still owned by the main package, log directory and logrotate config untouched. The static guard alone is not sufficient. It can only confirm which CPack variables were *written*, never that CPack *honors* them, so it passes against packaging that is genuinely broken. The container test is what closes that gap. ## Also fixed — the test oracles themselves - **CPack path check** accepted any path merely *containing* `site-packages`. Verified false positives on real synthetic packages: `/usr/lib/python3/not-site-packages/…`, `/opt/acme/site-packages/…`, and a wrong Python minor. It now matches against the target interpreter's `site.getsitepackages()`, and picks that interpreter the way packaging does (`/usr/libexec/platform-python`, then `/usr/bin/python3`) rather than `sys.executable`, which differs on the RHEL family. It also parses `dpkg -c` without mangling symlink entries or paths containing spaces. - **Upgrade round trip** verified a bare `import amdsmi`. A pip wheel under `/usr/local` shadows the packaged module, so all three phases could validate the *same unchanged wheel* while the system package really did change — reproduced on Ubuntu 22.04 and AlmaLinux 9. It now requires the imported module to be owned by the package under test, comparing the owning package name **exactly**: a substring test accepts `amd-smi-lib-tests`, whose name contains the main package's. OLD and NEW must also differ, so equal versions can't no-op past the test. - **Build harness** ran that round trip *after* installing the wheel; moved ahead of the install and wheel stages. - **Post-install** now warns, non-fatally and destroying nothing, when a pip-installed `amdsmi` shadows the packaged module. - **`test_run_amdsmi_build.py`** never registered the module in `sys.modules` before `exec_module`, which `@dataclass` requires on Python 3.12+. All 16 of its tests were uncollectable. ## Testing The new unit tests are hermetic — no root, Docker, network or GPU — and run on Python 3.6 through 3.14 via the version harness. Both new guards fail on the pre-fix code for the right reason, not merely because a helper is missing: the scriptlet guard fails on the pre-fix templates, and the CPack path guard drives the old CLI through a fake `dpkg` shim and reports the unimportable path it used to accept. Each commit is self-contained and was verified individually, so the stack bisects cleanly. ## Notes for review - Two behavior changes only: the post-install warning, which deletes nothing and always exits 0, and stage ordering inside a test harness. Everything else is test coverage or CI. - The upgrade-round-trip ownership check uses exact package-name comparison rather than the `"amd-smi-lib"` prefix suggested in review, because that string is also a prefix of `amd-smi-lib-tests`. The sibling case is covered by a test. ## JIRA ID ROCM-3941

marifamd requested review from a team as code owners July 23, 2026 19:20

github-actions Bot added project: amdsmi github actions Pull requests that update GitHub Actions code labels Jul 23, 2026

systems-assistant Bot added the organization: ROCm label Jul 23, 2026

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from 79dcf9b to 1c078d6 Compare July 24, 2026 14:05

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from eff9ea2 to 8dca306 Compare July 24, 2026 14:05

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from 1c078d6 to 139becc Compare July 24, 2026 14:31

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 8dca306 to 8f4526d Compare July 24, 2026 14:31

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from 139becc to a2997ad Compare July 24, 2026 15:48

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 8f4526d to cd56f28 Compare July 24, 2026 15:48

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from a2997ad to d52c7bb Compare July 27, 2026 07:57

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from b953eee to 624a03f Compare July 27, 2026 07:57

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch 2 times, most recently from d52c7bb to b9aad05 Compare July 27, 2026 15:16

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 624a03f to 12ca26e Compare July 27, 2026 15:16

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from b9aad05 to 31d09ea Compare July 27, 2026 15:37

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 12ca26e to 021ac87 Compare July 27, 2026 15:37

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from 31d09ea to bcde5a9 Compare July 27, 2026 15:54

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch 2 times, most recently from e13b797 to 8290c33 Compare July 27, 2026 19:58

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from bcde5a9 to b3db3eb Compare July 27, 2026 19:58

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from b3db3eb to 0148360 Compare July 27, 2026 20:48

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 8290c33 to 84fb513 Compare July 27, 2026 20:48

marifamd force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 84fb513 to 16cc3e3 Compare July 29, 2026 06:10

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from 0148360 to 44bc5f3 Compare July 29, 2026 06:10

gabrpham approved these changes Jul 29, 2026

View reviewed changes

marifamd force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from c7a5d9f to cdb896e Compare July 30, 2026 17:42

dmitrii-galantsev requested a lite review from Copilot August 5, 2026 16:20

Copilot started reviewing on behalf of dmitrii-galantsev August 5, 2026 16:21

View session

Copilot AI reviewed Aug 5, 2026

View reviewed changes

Comment thread projects/amdsmi/RPM/tests-preun.in Resolved

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 3821035 to 3b3e9ad Compare August 5, 2026 17:51

dmitrii-galantsev approved these changes Aug 5, 2026

View reviewed changes

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from ce25324 to 6fb98a0 Compare August 5, 2026 22:15

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 3b3e9ad to 6d51660 Compare August 5, 2026 22:15

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from 6fb98a0 to f5e6a5f Compare August 5, 2026 22:19

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 6d51660 to ede02ff Compare August 5, 2026 22:19

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from f5e6a5f to c211c63 Compare August 7, 2026 17:08

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from ede02ff to 1ae025d Compare August 7, 2026 17:08

dmitrii-galantsev and others added 9 commits August 10, 2026 11:50

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-followup-sles branch from c211c63 to 4dd278e Compare August 10, 2026 16:50

dmitrii-galantsev force-pushed the users/marifamd/amdsmi-packaging-scriptlet-isolation branch from 1ae025d to cc983b3 Compare August 10, 2026 16:50

Base automatically changed from users/marifamd/amdsmi-packaging-followup-sles to develop August 11, 2026 21:54

ammallya merged commit 03dd0fa into develop Aug 11, 2026
180 of 205 checks passed

ammallya deleted the users/marifamd/amdsmi-packaging-scriptlet-isolation branch August 11, 2026 21:55

Navigation Menu

Repository navigation

fix(amdsmi): isolate the tests package scriptlets and require the distro install matrix - #9110

fix(amdsmi): isolate the tests package scriptlets and require the distro install matrix#9110
ammallya merged 9 commits into
developfrom
users/marifamd/amdsmi-packaging-scriptlet-isolation

marifamd commented Jul 23, 2026 •

edited by dmitrii-galantsev

Loading

Uh oh!

therock-pr-bot Bot commented Jul 24, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

ammallya commented Aug 11, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

marifamd commented Jul 23, 2026 • edited by dmitrii-galantsev Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

JIRA ID

Uh oh!

therock-pr-bot Bot commented Jul 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ All Policy Checks Passed

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

ammallya commented Aug 11, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

marifamd commented Jul 23, 2026 •

edited by dmitrii-galantsev

Loading

therock-pr-bot Bot commented Jul 24, 2026 •

edited

Loading