Skip to content

ci: migrate 14 workflows to needs.setup.outputs.matrix (batch 4/8) - #4924

Open
luhenry wants to merge 14 commits into
mainfrom
fix/matrix-batch-4
Open

luhenry wants to merge 14 commits into
mainfrom
fix/matrix-batch-4

Conversation

@luhenry

@luhenry luhenry commented Nov 18, 2026 •

Copy link
Copy Markdown
Member

Part of the repo-wide migration off hardcoded strategy.matrix python/libc arrays
onto needs.setup.outputs.matrix (see _setup.yml, workflow-anatomy.md). This
batch covers 14 packages:

  • kasa-crypt
  • lcm
  • lhafile
  • libscrc
  • libtorrent
  • line-profiler
  • littlefs-python
  • memory-allocator
  • miniupnpc
  • mlx
  • moderngl
  • mutf8
  • mwparserfromhell
  • nanoarrow

For each build_wheels job (and, for mlx, build_frontend/test_wheels), the
strategy.matrix block's separate version/python/libc arrays are replaced
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, and every remaining
${{ matrix.libc }} reference (job name:, cibuildwheel only:/CIBW_BUILD,
ternary CIBW_* knobs, and upload-artifact name:) is renamed to
${{ matrix.platform }}. ${{ matrix.python }}/${{ matrix.version }} are
unchanged. No publish job in this batch references matrix.libc, so none
needed updating. mlx's build_backend job has no python dimension (one
backend build per version) and is left untouched.

Pre-existing mismatches this surfaces (verified via uv run ci_scripts/pending_versions.py --python-matrix <pkg>)

The structural fix makes docs/packages/<pkg>.yaml the sole source of truth for
which python/platform combos get built, which changes what gets built for most
of these packages (same root cause as #4914/#4915 — a hardcoded list had drifted
out of sync with the YAML). Not fixing the YAMLs themselves here (out of scope
for a .github/workflows/*-only branch) — just flagging:

  • kasa-crypt, mutf8, nanoarrow: clean adds only — gain cp310/cp311
    (declared in the YAML but never built under the old hardcoded cp312+ list).
  • lcm: gains cp310/cp311, but loses cp314t (not declared in
    docs/packages/lcm.yaml for 1.5.2) even though the workflow has a cp314t-specific
    gotcha-127 fix (PYTHON_GIL=1) that becomes dead weight once cp314t drops out
    of the matrix.
  • lhafile: gains cp310/cp311, loses cp314 (not declared for 0.3.1).
  • libtorrent: gains cp310/cp311, loses cp314 (not declared for 2.1.1).
  • libscrc: biggest mismatch — gains cp310/cp311 (musllinux only where
    declared), but loses cp313/cp314/cp314t entirely; the YAML only declares up to
    3.12 for 1.8.1, far short of the old hardcoded cp312-cp314t list.
  • miniupnpc: gains cp310/cp311 (manylinux only, matching the YAML's partial
    musllinux coverage), loses cp314/cp314t (not declared).
  • memory-allocator: the old exclude: stanza dropping cp312 for 0.2.1 is now
    redundant and removed — docs/packages/memory-allocator.yaml's 0.2.1 entry
    already omits cp312 from files, so the generated matrix excludes it
    naturally.
  • line-profiler: for pending version 5.1.0, docs/packages/line-profiler.yaml
    declares manylinux only (no musllinux at all), even though the workflow has an
    explicit comment ("Upstream 5.1.0 added *-musllinux_* skip list; we build
    musllinux") recording a deliberate choice to keep building musllinux anyway.
    After this change, musllinux wheels stop being built for 5.1.0 until the YAML
    gains musllinux entries for it. Also fixes a real cross-product bug: 5.0.2 and
    5.1.0 were sharing one hardcoded python list regardless of what each version
    actually supports.
  • moderngl: same cross-product bug — the old hardcoded cp312-cp314t list was
    applied to both pending versions uniformly, even though 5.12.0's YAML entry
    doesn't declare cp314/cp314t at all. Also gains cp310/cp311 and, for 5.13.0,
    cp315/cp315t.
  • mwparserfromhell: the workflow's own comment says "PyPI ships no cp314...
    wheel either", yet the old hardcoded list built cp314 anyway, contradicting its
    own comment; docs/packages/mwparserfromhell.yaml has no 3.14 entries, so this
    fixes that self-contradiction. Also gains cp310/cp311.
  • mlx: build_frontend/test_wheels gain cp310/cp311 and, for the pending
    0.32.4, cp315/cp315t; also fixes the same version/python cross-product issue
    (one hardcoded list was shared across all three pending versions).
  • littlefs-python: gains cp310/cp311 only; no change to the existing cp314t
    exclusion (YAML doesn't declare it either).

None of these required a python-exclude addition on the setup: job: every
case where the old hardcoded list omitted something is either (a) not declared
in the YAML at all (so the generated matrix already omits it, no action needed
per the structural-fix rule), or (b) the memory-allocator per-version files
gap, which the generated matrix already encodes correctly.

…ython/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/kasa-crypt.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311 (previously unbuilt despite docs/packages/kasa-crypt.yaml declaring them)
Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/lcm.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311; drops cp314t (not declared in docs/packages/lcm.yaml for 1.5.2)
…on/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/lhafile.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311; drops cp314 (not declared in docs/packages/lhafile.yaml for 0.3.1)
…on/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/libscrc.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311 (musllinux only where declared); drops cp313/cp314/cp314t (not declared in docs/packages/libscrc.yaml for 1.8.1)
…ython/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/libtorrent.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311; drops cp314 (not declared in docs/packages/libtorrent.yaml for 2.1.1)
…d python/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/line-profiler.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311/cp315 for 5.1.0; drops musllinux entirely for 5.1.0 (not declared in docs/packages/line-profiler.yaml, though a comment there says musllinux is intentionally still built); also fixes a version/python cross-product so 5.0.2 and 5.1.0 no longer share one hardcoded list
…ded python/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/littlefs-python.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311 (no cp314t change, matches existing exclusion)
…oded python/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/memory-allocator.yaml, and renames matrix.libc to matrix.platform to match.

per-version files already encode the 0.2.1 cp312 gap, so the old 'exclude:' stanza is now redundant and dropped
…thon/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/miniupnpc.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311 (manylinux only, matching docs/packages/miniupnpc.yaml); drops cp314/cp314t (not declared)
Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/mlx.yaml, and renames matrix.libc to matrix.platform to match.

build_frontend/test_wheels gain cp310/cp311 and (for 0.32.4) cp315/cp315t, and per-version correctness replaces one hardcoded list shared across all pending versions; build_backend is untouched (version-only, no python dimension)
…hon/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/moderngl.yaml, and renames matrix.libc to matrix.platform to match.

fixes a real cross-product bug: the old hardcoded list applied cp314/cp314t to ALL pending versions including 5.12.0, which docs/packages/moderngl.yaml does not declare them for; also adds cp310/cp311 and (for 5.13.0) cp315/cp315t
…/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/mutf8.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311, otherwise unchanged
…oded python/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/mwparserfromhell.yaml, and renames matrix.libc to matrix.platform to match.

drops cp314 (the existing comment already says 'PyPI ships no cp314 wheel either', so the hardcoded list building cp314 anyway was itself a bug this fixes); adds cp310/cp311
…thon/libc

Replaces the hardcoded strategy.matrix python/libc arrays in build_wheels
with include: ${{ fromJSON(needs.setup.outputs.matrix) }}, derived from
docs/packages/nanoarrow.yaml, and renames matrix.libc to matrix.platform to match.

adds cp310/cp311, otherwise unchanged

This branch has not been deployed

No deployments
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