Add Accelerate framework blas__ldflags tests

[lucianopaz]

Description

This just adds a test of the Accelerate framework flag to test in blas__ldflags. I haven't written a test because it proved to be a PITA. I've tested it locally and it seems to work. The only thing that I can say is that we should run the test suite on Mac and ensure that the blas__ldflags should be -framework Accelerate -rpath some_path. I written a test to mock this, but I've also added a macos-latest runner that tries to test blas, and also asserts that the blas__ldflags get set to accelerate.

Related Issue

• Closes pytensor and blas problems on on MacOS 15 Sequoia with Apple Silicon  #1005
• Related to #

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

---

📚 Documentation preview 📚: https://pytensor--1056.org.readthedocs.build/en/1056/

[ricardoV94]

The only thing that I can say is that we should run the test suite on Mac and ensure that the blas__ldflags should be -framework Accelerate -rpath some_path.

Running the whole test suite is probably overkill. Can we just pick some smoke tests besides the flag check?

[lucianopaz]

Running the whole test suite is probably overkill. Can we just pick some smoke tests besides the flag check?

I suppose that the blas tests

[codecov]

Codecov Report

Attention: Patch coverage is 97.50000% with 1 line in your changes missing coverage. Please review.

Project coverage is 82.09%. Comparing base (6132203) to head (8307af1).
Report is 11 commits behind head on main.

Files with missing lines	Patch %	Lines
pytensor/tensor/blas.py	94.44%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1056      +/-   ##
==========================================
+ Coverage   81.95%   82.09%   +0.14%     
==========================================
  Files         182      182              
  Lines       47872    47901      +29     
  Branches     8618     8631      +13     
==========================================
+ Hits        39233    39325      +92     
+ Misses       6472     6410      -62     
+ Partials     2167     2166       -1     

Files with missing lines	Coverage Δ
pytensor/link/c/cmodule.py	60.50% <100.00%> (+3.44%)	⬆️
pytensor/tensor/blas.py	66.18% <94.44%> (+0.63%)	⬆️

... and 7 files with indirect coverage changes

[lucianopaz]

I'll have to refactor some tests because the extra condition I added doesn't rely on the same file existence logic that the other conditions did, and the mocks I had in place fail there

[lucianopaz]

I hope to have fixed my mocks now. @maresb, I'm having a dumb problem with the macos-latest runner. I don't know why, the runner doesn't have md5sum and fails to generate the matrix id. Do you have any idea of an alternative that could work on that runner? Locally, I do have both md5 and md5sum, so I have no idea why the runner shouldn't have md5sum.

[lucianopaz]

@ricardoV94, I have no idea what went wrong with the single blas test that failed on the mac run. Could you have a look? Is it worth fixing or should we skip that test on Mac?

[ricardoV94]

@ricardoV94, I have no idea what went wrong with the single blas test that failed on the mac run. Could you have a look? Is it worth fixing or should we skip that test on Mac?

Can you reproduce locally? Seems meaningful

[lucianopaz]

@ricardoV94, I have no idea what went wrong with the single blas test that failed on the mac run. Could you have a look? Is it worth fixing or should we skip that test on Mac?

Can you reproduce locally? Seems meaningful

You were totally right, @ricardoV94! This small test turned out to indicate that the link flags were wrong and that pytensor was not compiling against blas at all! There were a bunch of problems, the main one was that pytensor was programed with GNU C compiler, gcc, in mind. This only has -l, -L and -I flags for linking libraries, adding library search dirs and hearder include dirs. clang adds the notion of -framework, that is the same as implicitly adding a bunch of -l, -L and -I flags to the compilation command. The problem was that pytensor actually tried to enforce all blas flags to follow gcc convention, and failed horribly when faced with -framework Accelerate. I had to add a hackish solution to work around the limitations. A proper solution would be to add a c_frameworks to the Ops and handle that in the C linkers, but that would involve a massive refactor that is completely not worth the effort. At the moment, we only really want to handle Accelerate. If the situation changes in the future, we will have to revise this.

[lucianopaz]

@ricardoV94, the codecov patch difference isn't significant. The changes I made removed coverage from a couple of lines, and I don't think that it's worth it to prevent the merge just because of that.

[maresb]

The silent failures should be resolved by mamba-org/setup-micromamba#241