Wheel support for Mac Arm (M1)

[PhilipVinc]

There's no issue on this, so I'll open one to track this.

Currently LLVMLite does not provide Wheels on PyPi for Mac Arm (M1) processors, but they are only provided on Conda.
This is annoying because users cannot install using pip packages depending on Numba when running native. Releasing PyPi wheels for LLVMLite would be the first step towards addressing this.

[esc]

@PhilipVinc thank you for opening this issue. We hope to tackle this in Q1/Q2 2022 so that M1 wheels will be available for the 0.39.0 release.

[esc]

A quick update on this one. We are getting closer.

M1 support for llvmlite is now (probably) complete: #813
M1 support for Numba is still in progress: numba/numba#7822

We are still hoping to publish M1 wheels with the next Numba and llvmlite releases.

[pikeas]

Until #813 is merged and released, what's the best way to install llvmlite in a Python 3.8+ virtualenv on an M1 Mac?

[PhilipVinc]

In my experience, annoying as it may be, conda-forge

[esc]

Good news, both llvmlite and numba have been extensively tested on an M1 now:

• Fix M1 Unit Test Failures. numba#7822
• *Add m1 support to conda build scripts #813)

So, we are now in a position to build wheels too. If everything goes according to plan (which it rarely does) I'll be uploading some initial builds somewhere for testing, come next week.

[esc]

An initial set of wheels that I built are now available via anaconda.org:

https://anaconda.org/numba/llvmlite/files?channel=wheels_experimental_m1

While I have not tried this yet, the installation of the native wheels should simply be:

pip install -i https://pypi.anaconda.org/numba/label/wheels_experimental_m1/simple llvmlite

Anyone waiting on these wheels is invited to "give them a spin". Please report back any issues you encounter with m1 wheels on this issue, thank you! 🙏

[esc]

Unfortunately, there seems to be an issue here, I just tried the instructions listed on:

https://anaconda.org/numba/llvmlite

However, I am unable to install the wheels.

(wheeltest) ci@colo-osxarm-02 ~ % pip install -i https://pypi.anaconda.org/numba/label/wheels_experimental_m1/simple llvmlite

Looking in indexes: https://pypi.anaconda.org/numba/label/wheels_experimental_m1/simple
ERROR: Could not find a version that satisfies the requirement llvmlite (from versions: none)
ERROR: No matching distribution found for llvmlite

I will continue to look into this and keep you posted about progress.

[esc]

It seems like something went wrong during the wheel build process:

(wheeltest) ci@colo-osxarm-02 ~ % pip install llvmlite-0.39.0.dev0+61.gf27ac6f-cp310-cp310-macosx_10_9_x86_64.whl
ERROR: llvmlite-0.39.0.dev0+61.gf27ac6f-cp310-cp310-macosx_10_9_x86_64.whl is not a supported wheel on this platform.
(wheeltest) ci@colo-osxarm-02 ~ % file llvmlite-0.39.0.dev0+61.gf27ac6f-cp310-cp310-macosx_10_9_x86_64.whl
llvmlite-0.39.0.dev0+61.gf27ac6f-cp310-cp310-macosx_10_9_x86_64.whl: Zip archive data, at least v2.0 to extract, compression method=deflate

[esc]

OK, so for some reason the wheels end up having the wrong tag:

(wheeltest) ci@colo-osxarm-02 llvmlite-0.39.0.dev0+61.gf27ac6f.dist-info % cat WHEEL
Wheel-Version: 1.0
Generator: bdist_wheel (0.37.1)
Root-Is-Purelib: false
Tag: cp310-cp310-macosx_10_9_x86_64

My assumption is that the suffix of the Tag entry should be osx-arm64 and not x86_64.

[esc]

After some more fiddling with the CI system that builds the wheels, I managed to get some that work:

(wheeltest) ci@colo-osxarm-02 ~ % pip install -i https://pypi.anaconda.org/numba/label/wheels_experimental_m1/simple llvmlite
Looking in indexes: https://pypi.anaconda.org/numba/label/wheels_experimental_m1/simple
Collecting llvmlite
  Downloading https://pypi.anaconda.org/numba/label/wheels_experimental_m1/simple/llvmlite/0.39.0.dev0%2B61.gf27ac6f/llvmlite-0.39.0.dev0%2B61.gf27ac6f-cp310-cp310-macosx_11_0_arm64.whl (23.1 MB)
     |████████████████████████████████| 23.1 MB 1.3 MB/s
Installing collected packages: llvmlite
Successfully installed llvmlite-0.39.0.dev0+61.gf27ac6f

These are for Python 3.8, 3.9 and 3.10 only. The system we use depends on python being available from defaults on anaconda.org and Python 3.7 is not (was not at the time of writing) available.

[henryiii]

Python 3.7 was never officially adapted to work on M1. 3.8 is the oldest the M1 changes were backported to (and officially, only 3.9+ is fully supported, 3.8 is considered experimental by CPython, but Apple provides their patched version)

[esc]

Python 3.7 was never officially adapted to work on M1. 3.8 is the oldest the M1 changes were backported to (and officially, only 3.9+ is fully supported, 3.8 is considered experimental by CPython, but Apple provides their patched version)

Excellent, thank you for clarifying that @henryiii 🙏 looks like we may be "in the clear" regarding targeted Python versions for now.

[esc]

For reference, there are Numba wheels now too. See also: numba/numba#7951

[esc]

An announcement for the wheels and a request to test them is here: https://numba.discourse.group/t/wheels-for-apple-silicon-m1/1282

[dbuades]

An announcement for the wheels and a request to test them is here: https://numba.discourse.group/t/wheels-for-apple-silicon-m1/1282

I confirm that they work properly. Thank you!

[esc]

An announcement for the wheels and a request to test them is here: https://numba.discourse.group/t/wheels-for-apple-silicon-m1/1282

I confirm that they work properly. Thank you!

Thank you for checking and confirming!

[ghost]

An announcement for the wheels and a request to test them is here: https://numba.discourse.group/t/wheels-for-apple-silicon-m1/1282

I can also confirm that the wheels are working as expected.

Thank you for finally releasing them!

[esc]

With llvmlite 0.38.1 M1 wheels are now available:

https://numba.discourse.group/t/llvmlite-0-38-1/1340