Publish Linux ARM64 Wheels to PyPI

[johnthagen]

🚀 The feature, motivation and pitch

We would like to deploy executorch into a container-based application to perform inference on a pre-trained model. We value the smaller filesize footprint of executorch compared to torch as minimizing container image size is important to us.

Our developers works on both Linux and macOS. Apple has transitioned new hardware to ARM64, which means that containers created on current Apple hardware must run on Linux/ARM64.

executorch is currently missing Linux/ARM64 wheels, which makes installing it as a Python dependency prohibitive.

• https://pypi.org/project/executorch/0.4.0/#files

To make Python development smooth across both Linux and macOS, we'd request that Linux/ARM64 wheels be uploaded to PyPI.

Looking at other similar packages, tflite-runtime from the Tensorflow team publishes Linux/ARM64 wheels

• https://pypi.org/project/tflite-runtime/2.14.0/#files
• tflite_runtime-2.14.0-cp311-cp311-manylinux_2_34_aarch64.whl

Alternatives

executorch could be built from source within the container, but that is slow an inefficient compared to installing a pre-built wheel.

Additional context

No response

RFC (Optional)

No response

[Jack-Khuu]

Hmmm interesting that we don't have this spin up already

@cccclai were you working on the initial pypi stuff with Dave? Do you know how to enable new wheels?

[johnthagen]

In case it helps, cibuildwheel is used for this kind of thing by other Python projects

• https://github.com/pypa/cibuildwheel