Pin libgcc for windows installs#1399
Conversation
575494c to
9bfb689
Compare
| "scipy>=1,<2", | ||
| "numpy>=1.17.0", | ||
| "filelock>=3.15", | ||
| "libgcc<15 ; platform_system == 'Windows'", |
There was a problem hiding this comment.
The c backend is optional for PyTensor, this is something you want to do at the conda level not here
There was a problem hiding this comment.
Ya, libgcc isn't even on PyPI. Would anyone be available for a quick call to explain to me what's going on?
There was a problem hiding this comment.
Yep I quickly discovered this PR is extremely stupid :)
I can close it once we have another one addressing the actual issue.
We only do a smoke test on mac but why is it redundant. They are very different oses these days, so much mac breaks every other day and I don't remember the last time we had issues on linux |
|
I did a bit more digging and I have a vague hypothesis. I don't think it's a matter of pinning libgcc, but rather that pytensor is finding the wrong compiler. Perhaps we are picking up the system compiler, or there is a mismatch within the conda environment itself between the gxx version and libgcc version? (ping @lucianopaz) (Note that we had to change to g++ on Windows, but it has been apparently working well for several months.) conda-forge/pytensor-suite-feedstock#133 (comment) I reproduced a Windows environment via pixi, and I'm wondering why Environment list |
I checked on linux-64 and it was the same, but it still works on Linux. 🤷 |
|
Closing this PR since it's entirely wrong, we can move the discussion on how to proceed to #1398 |
Closes #1398 . Issue was reported and debugged by user DrEntropy here.
Maybe there's a better solution for this issue, but it's probably good to patch it quickly for now. We should also consider running the CI on windows -- testing linux and mac is a bit redundant?
📚 Documentation preview 📚: https://pytensor--1399.org.readthedocs.build/en/1399/