Wheels and platforms
What a release builds, which platforms are proven by CI and which aren’t, and why. The steps to build and release are in CONTRIBUTING.md.
What a release builds
A release builds and validates exactly 30 wheels
(expected-wheel-count in .github/workflows/publish-release-packages.yml
is the number that counts; if this page and that value disagree, the workflow
is right):
CPython 3.11 to 3.15, plus the free-threaded 3.15t, on manylinux_2_28 x86_64, manylinux_2_28 aarch64, Windows AMD64 and macosx_arm64: 24.
CPython 3.13 to 3.14, Android API 21 arm64_v8a and x86_64: 4.
CPython 3.13 to 3.14, Pyodide wasm32: 2.
macOS: CI-proven
macOS is enabled at the CI-proven tier: ship everything for macOS that CI
can prove. test-cpython.yml builds, installs and runs the unit, contract
and parity suite and tools/validate_api.py on macos-latest, and its
macos-wheels job runs the same cibuildwheel a release uses, compiling the
native extension for CPython 3.11 to 3.15 and 3.15t and running an import
smoke test in each wheel, on every push. It publishes nothing. Hardware audio
output on a real Mac is community-verified: a field report is what would
promote that claim.
The wheel is macosx_arm64 only. The runner is Apple Silicon, and an
x86_64 or universal2 wheel cross-compiled there couldn’t run its test
command. An untested wheel is a claim, not a capability, so an Intel Mac
field report is what would add x86_64.
Linux aarch64: CI-proven and shipped
[tool.cibuildwheel.linux] uses archs = ["auto64"], the runner’s own
64-bit architecture, with manylinux_2_28 images for both x86_64 and
aarch64. Naming ["x86_64", "aarch64"] explicitly would make the x86_64
release runner try to build aarch64 under QEMU, which no workflow configures.
auto64 resolves to x86_64 on an x86_64 runner and aarch64 on an ARM one, so
one config is correct everywhere.
test-cpython.yml proves aarch64 on native ARM runners (ubuntu-24.04-arm),
not under QEMU, because the bugs worth catching are the ones that differ by
architecture, and an emulator reproduces those only by accident:
the full matrix (build, install, unit and contract tests,
validate_apiand the parity verifiers) runs on aarch64 for CPython 3.11 to 3.14;a
linux-arm-wheelsjob runs the same cibuildwheel a release uses, compiling the extension and running each wheel’s import smoke test. Nothing is published.
Releases ship the aarch64 wheels too, built by the release workflow on the
same kind of native ARM runner, so pip install omnipython-audiodsp works
on 64-bit Raspberry Pi OS and other 64-bit ARM Linux machines. They have run
on a Raspberry Pi 4 and a Pi 3 B+. A Pi 5 runs the same OS and should take the
same wheels, but nobody has tried it yet. If you do,
open an issue and say what
worked.
The same bytes within one architecture
Bit-identical audio is required within one CPU architecture, not across them. No tolerance is introduced anywhere: the gate stays exact, and an architecture that genuinely differs is held to its own recorded baseline, accepted deliberately with its evidence written beside it. That is a stricter regression detector than a threshold.
aarch64 is the example. At first, verify_effects differed on multitap and
pitchshift, the two effects that interpolate along a delay line: a
compiler’s fused multiply-add changes the rounding, and AArch64 fuses by
baseline where x86-64 doesn’t without -mfma. Forbidding contraction in
every src/shared/ file that computes in float closed it (issue
#79), so aarch64’s own
baseline was removed from tests/parity/golden/effects_component.json: it is
held to the oracle hash itself, the same one x86_64 is held to.
What isn’t built
musllinux stays skipped: it doubles the Linux matrix for the Alpine family. aarch64 is built because its users can be named (every Pi and 64-bit SBC); Alpine’s can’t. It remains in
pyproject.toml’s cibuildwheelskiplist.win32 (32-bit Windows): every Windows target here is 64-bit, and nothing in this repository has been built or tested 32-bit. Also in the
skiplist.MicroPython
.mpy: the native modules are firmware. MicroPython and CircuitPython build them fromsrc/as a user module, so there is nothing for mip to carry. The pure-Python tier publishes to mip as source, because compiling it to.mpywould pin it to one bytecode version per MicroPython release.