Skip to content

docs: Track wheel targets in file entries - #4731

Merged
luhenry merged 2 commits into
mainfrom
python-versions
Nov 6, 2026
Merged

luhenry merged 2 commits into
mainfrom
python-versions

Conversation

@luhenry

@luhenry luhenry commented Nov 5, 2026

Copy link
Copy Markdown
Member

Move Python ABI and Linux platform targets into files so each target can carry its published wheel metadata or remain empty. Empty targets keep the version pending and force complete release rebuilds when coverage changes.

Move Python ABI and Linux platform targets into files so each target can carry its published wheel metadata or remain empty. Empty targets keep the version pending and force complete release rebuilds when coverage changes.
Use matrix.platform in dynamic build selectors, job names, and artifact names so musllinux rows no longer duplicate manylinux builds. Also pretty-print the setup matrix and document the requirement in the porting guide.
@luhenry
luhenry merged commit 5857b2c into main Nov 6, 2026
2272 of 12880 checks passed
@luhenry
luhenry deleted the python-versions branch November 6, 2026 01:57
luhenry added a commit that referenced this pull request Nov 7, 2026
Convert the 1.27.3 entry from python-versions to the files schema
introduced in #4731, which pending_versions.py now requires.
luhenry added a commit that referenced this pull request Nov 15, 2026
* Add openturns riscv64 wheel build

OpenTURNS has no riscv64 wheel on PyPI (only manylinux_2_28 x86_64/aarch64,
macOS arm64, Windows) and is missing from pypi.riseproject.dev. Upstream
ships no sdist/pyproject.toml at all: the published wheels are produced by
a separate repo (openturns/build-wheels) that drives a prebuilt manylinux
image directly with CMake + SWIG and hand-assembles the wheel. This
workflow adapts that recipe to the generic manylinux_riscv64 image, with
USE_NLOPT/USE_TBB/USE_HIGHS/USE_CMINPACK on (TBB and HiGHS already have
riscv64 wheels in this repo at the exact pinned versions upstream's image
uses) and USE_IPOPT/USE_BONMIN/USE_CERES/USE_PAGMO/USE_CUBA/USE_DLIB off:
their native deps (the COIN-OR suite with its Fortran Mumps solver,
Ceres+glog, Pagmo, dlib) have no riscv64 precedent anywhere, and
OpenTURNS gracefully falls back to its own bundled implementations when
they're absent, the same way upstream's own "mindeps" CI leg
(.ci_support/run_docker_mindeps.sh) is built and tested.

* Build oneTBB from source instead of dnf tbb-devel

Rocky10's tbb-devel is the legacy pre-oneAPI TBB, which does not provide the
oneTBB CMake package config OpenTURNS' FindTBB expects. Build the same
pinned oneTBB release this repo's own tbb.yaml port already validates on
riscv64 instead.

* openturns: Record target Python versions

* docs: openturns: Track wheel targets in file entries

Convert the 1.27.3 entry from python-versions to the files schema
introduced in #4731, which pending_versions.py now requires.

* ci: openturns: Upgrade pip in test venv before installing wheel

The cp310 venv is seeded with ensurepip's bundled pip 23.0.1 (packaging
21.3), which has no manylinux_*_riscv64 tags and rejects the wheel as
not supported on this platform. cp312+ bundle a newer pip.

* openturns: fix declared target to match the abi3 wheel tag

The build produces a stable-ABI wheel tagged cp310-abi3-manylinux_2_39_riscv64,
but the yaml declared the slot as plain 3.10:manylinux, so the publish job
fails with:
  built wheels have undeclared python targets: 3.10-abi3:manylinux

* openturns: pass matrix input to _publish-wheel.yml

_publish-wheel.yml now requires a matrix input (main's "publish:
Validate effective wheel targets"), but this PR branch forked before
that landed and openturns has never been merged to main, so
build-openturns.yml never picked up the matching one-line update every
already-published package's workflow got. The merge ref combines the
new required-input _publish-wheel.yml with this branch's old
build-openturns.yml, so any run has failed startup with "Input matrix is
required, but not provided".

Pass it the same way the mass-update commit did for every other
package's publish job.

* openturns: declare only the real cp310-abi3 target; test the rest via a hardcoded matrix

docs/packages/openturns.yaml declared 4 separate files: entries (3.10-abi3,
3.12, 3.13, 3.14, all :manylinux), but build_wheels only ever builds ONE
wheel (hardcoded WHEEL_TAG: cp310-abi3-manylinux_2_39_riscv64) - the
cp312/3.13/3.14 entries existed purely to drive test_wheels' per-interpreter
fan-out via needs.setup.outputs.matrix, which this port's author repurposed
for test coverage instead of its documented build-target role.

ci_scripts/pending_versions.py's python_matrix() has no concept of a
"test-only" declared target: a non-abi3-suffixed "3.12:manylinux" entry is
treated exactly like a real cp312 build target, becoming its own entry in
BUILD_MATRIX with target "3.12:manylinux". ci_scripts/update_doc.py's
merge_publication_files() then requires a wheel matching that exact target
string before it'll consider the release complete, and the single real
wheel (tagged cp310-abi3, parsed to target "3.10-abi3:manylinux") can never
satisfy it - "release is incomplete; no wheel was built for: 3.12:manylinux,
3.13:manylinux, 3.14:manylinux", even though the build itself succeeds and
the wheel is correctly import-tested on all four interpreters.

Confirmed via a survey of every other abi3 package's docs/packages/*.yaml
in this repo (aardwolf, advent-of-code, aioquic, akquant, angr, anki,
argon2-cffi-bindings, astropy, ...) that a single abi3 entry per platform,
covering every interpreter at or above its floor, is the universal
convention - no other package declares extra interpreter-specific entries
alongside an abi3 one. openturns.yaml's 4-entry declaration was a one-off
mistake from how this port was originally written, not a shared-script bug:
pending_versions.py/update_doc.py behave correctly for every other package.

Fix: trim openturns.yaml back to its one real target, and give test_wheels
its own hardcoded python list (cp310, cp312, cp313, cp314) crossed with
needs.setup.outputs.versions, independent of needs.setup.outputs.matrix -
the same way build_wheels already hardcodes its own build interpreter via
PYDIR/WHEEL_TAG rather than consuming the matrix. This keeps the intended
4-interpreter test coverage without any shared ci_scripts/_setup.yml change.
luhenry added a commit that referenced this pull request Nov 15, 2026
build-aioquic.yml has only ever built a single cp310-abi3 manylinux
wheel (CIBW_BUILD lists manylinux_riscv64 targets only; no musllinux
leg was ever added). PR #4731's mechanical files: expansion added a
musllinux placeholder to every version's target list regardless, which
1.3.0 absorbed harmlessly since it was already tagged, but which
permanently blocks 1.4.0-1.7.0's publish validation ("release is
incomplete; no wheel was built for: 3.11-abi3:musllinux") since they
have no tag yet. Drop the target this workflow was never built to
satisfy.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant