Make BLAS flags check lazy and more actionable

[ricardoV94]

It replaces the old eager warning that does not actually apply by a lazy yet more informative and actionable warning.

This however means users won't notice something is wrong with their installation until PyTensor actually accesses the config.blas__ldflags to decide on some behavior.

These flags are accessed to either not trigger the introduction of more efficient COps (CGEMv, sparse Usmm, ...), or to use the Python or C implementation for Ops that have both (Dot22, BatchedDot, ...).

The new warning will happen only when a function is compiled that triggers those code-paths.

---

Technically the old warning was wrong:
The specific warning message was about Ops that might use the alternative blas_headers, which rely on the Numpy C-API.

However, regular PyTensor user has not used this for a while. The only Op that would use C-code with this alternative headers is the GEMM Op which is not included in current rewrites. Instead Dot22 or Dot22Scalar are introduced, which refuse to generate C-code altogether if the BLAS flags are missing.

---

This complements #1161 and makes PyTensor a breeze to import compared to before.
Closes #1160 (although we could still try to speed the check up)

[jessegrabowski]

Love it. I suggest also mentioning that you can set the mode argument in .pytensorrc, so that's it becomes the global default. This would benefit users who are adamantly stanning pip (and should thus always use a non-C backend)

[ricardoV94]

Love it. I suggest also mentioning that you can set the mode argument in .pytensorrc, so that's it becomes the global default. This would benefit users who are adamantly stanning pip (and should thus always use a non-C backend)

Maybe update and link to https://pytensor.readthedocs.io/en/latest/troubleshooting.html#test-blas ?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.27%. Comparing base (6bdfbae) to head (b9cc961).
Report is 14 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1165      +/-   ##
==========================================
- Coverage   82.27%   82.27%   -0.01%     
==========================================
  Files         186      186              
  Lines       48023    48004      -19     
  Branches     8630     8623       -7     
==========================================
- Hits        39513    39495      -18     
- Misses       6347     6351       +4     
+ Partials     2163     2158       -5     

Files with missing lines	Coverage Δ
pytensor/link/c/cmodule.py	60.87% <100.00%> (+0.47%)	⬆️
pytensor/tensor/blas_headers.py	70.78% <100.00%> (+1.89%)	⬆️

... and 11 files with indirect coverage changes

[jessegrabowski]

Do we want points (2) and (3) in here anymore?

[ricardoV94]

Sure, it's the most hardcore version to get it up and running, why not mention it?

[jessegrabowski]

fair enough!

[maresb]

This is awesome!!!

Note to self: I'll need to rework the conda-forge tests since that assumes eager BLAS

[ricardoV94]

This is awesome!!!

Note to self: I'll need to rework the conda-forge tests since that assumes eager BLAS

Doesn't it explicitly check the pytensor.config.blas__ldflags? that will trigger the logic

[maresb]

I run python -c "import pytensor.configdefaults; import pytensor.tensor". Thanks for the suggestion, I'll just change it to that.

[Ch0ronomato]

Yup, this makes a lot of sense as other backends get developed.