docs: Add FAQ on backend support on Apple silicon Macs

[matthewfeickert]

Description

Superseded by PR #2119

---

Resolves #2111

At the moment, TensorFlow does not have Apple silicon compatible wheels on PyPI under the tensorflow project namespace, though it does under tensorflow-macos. As an alternative, note that conda-forge has linux-64, osx-64, and osx-arm64 compatible releases for all of pyhf's backends.

Checklist Before Requesting Reviewer

• Tests are passing
• "WIP" removed from the title of the pull request
• Selected an Assignee for the PR to be responsible for the log summary

Before Merging

For the PR Assignees:

• Summarize commit messages into a comprehensive review of the PR

* At the moment, TensorFlow does not have Apple silicon compatible
  wheels on PyPI. As an alternative, note that conda-forge has linux-64,
  osx-64, and osx-arm64 compatible releases for all of pyhf's backends.
   - c.f. https://github.com/tensorflow/tensorflow/issues/57185

[codecov]

Codecov Report

Base: 98.05% // Head: 98.30% // Increases project coverage by +0.24% 🎉

Coverage data is based on head (a02c668) compared to base (46ce316).
Patch has no changes to coverable lines.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2118      +/-   ##
==========================================
+ Coverage   98.05%   98.30%   +0.24%     
==========================================
  Files          69       69              
  Lines        4531     4531              
  Branches      645      645              
==========================================
+ Hits         4443     4454      +11     
+ Misses         54       45       -9     
+ Partials       34       32       -2     

Flag	Coverage Δ
contrib	97.88% <ø> (ø)
doctest	61.15% <ø> (?)
unittests-3.10	96.31% <ø> (ø)
unittests-3.8	96.33% <ø> (ø)
unittests-3.9	96.35% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
src/pyhf/tensor/jax_backend.py	98.61% <0.00%> (+0.69%)	⬆️
src/pyhf/tensor/tensorflow_backend.py	97.33% <0.00%> (+2.00%)	⬆️
src/pyhf/infer/test_statistics.py	100.00% <0.00%> (+3.33%)	⬆️
src/pyhf/tensor/pytorch_backend.py	98.48% <0.00%> (+3.78%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[matthewfeickert]

Relevant RTD build: https://pyhf--2118.org.readthedocs.build/en/2118/faq.html#why-can-t-pip-find-compatible-backends-on-pypi-for-apple-silicon-macs

[alexander-held]

You can also install tensorflow-macos (and tensorflow-metal) via pip, but I guess that's discussed in the issue you linked anyway.

[matthewfeickert]

You can also install tensorflow-macos (and tensorflow-metal) via pip, but I guess that's discussed in the issue you linked anyway.

Ah yeah I had forgotten about the Metal project to get TensorFlow to work on Mac GPUs. Thanks for the reminder, I'll mention something about them. Though I guess related to that, @alexander-held on your Apple silicon Mac are you able to do the following in your virtual environment that you tried to use for developing in Issue #2111?

# After cloning things locally
$ python -m pip install --upgrade tensorflow-macos
$ python -m pip install --upgrade '.[develop]'
$ pytest

Or given this comment tensorflow/tensorflow#57185 (comment) about tensorflow-macos needing Rosetta are things super slow (I assume Rosetta is being used for emulation which will slow things down) and to get normal speeds do you need to follow Apple's instructions on setting up tensorflow-deps from conda (not conda-forge!)?

[matthewfeickert]

on your Apple silicon Mac are you able to do the following in your virtual environment that you tried to use for developing in Issue #2111?

Ah, no I realize that this will fail regardless. Even though tensorflow-macos should install fine, we'll still run into trouble here as

python -m pip install --upgrade '.[develop]'

invokes the develop extra which includes the backends extra which includes the tensorflow extra

pyhf/pyproject.toml

Lines 64 to 67 in 29fb68e

	tensorflow = [
	"tensorflow>=2.7.0", # c.f. PR #1962
	"tensorflow-probability>=0.11.0", # c.f. PR #1657
	]

and so it will still try to do a pip install of tensorflow as that's going to be in the requested list, which will fail.

Normally this is where we would have some code to dynamically change determine the package name to use here based off of os.uname().machine or something, but we can't do that with TOML. I think this is again where using hatchling forces us into using hatch to setup environment workflows rather than using extras. We have Issue #2096 for that but we haven't done further investigation of how to orchestrate this.

[matthewfeickert]

@alexander-held I think this should work to actually get things installed. Though I think it is going to be annoying to get a full developer environment setup at the moment given that the test extra also pulls in the backends extra and so will break here.

[matthewfeickert]

This PR should no longer be needed as PR #2119 can work.