New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
BLD: build Windows wheel for py39 against numpy 1.22.3 #18036
Conversation
Does it mean then that our compatibility for Windows is in fact NumPy >= 1.22.3? |
This is a follow-up to scipygh-17959, which bumped numpy from 1.19.5 to 1.21.6, which is wrong because that version was built with vc142 (see comment above the line added in this PR). [wheel build] [skip circle] [skip cirrus] [skip azp]
437d88c
to
f1369e2
Compare
For builds from source, no. For our wheels, yes. |
I am not sure I understand this right. Are you saying a user could build from source and use 1.21.6, but if they use our wheels there have to be on >=1.22.3? |
Yes, it's perfectly possible to build both NumPy and SciPy with the same set of compilers. The problem with numpy 1.21.6-1.22.2 is that a compiler was used that clashes with the one we use for SciPy. This is generically true by the way - constraints for wheels and source code does not have to be the same for any project. |
Ok but isn't it an issue if we say that we support 1.21.6 and then some users would not be able to get a wheel working with that? Since building from source is not an easy/transparent exercice (and from my understanding even less on Windows), shouldn't we only advertise for wheel compatibility or the easy/transparent path? |
Could we transition to using the same compilers as numpy then (or both projects decide on a common set of compilers)? Historically wasn't the reason why we avoided using MSVC for scipy redistributing the CRT? |
NumPy isn't moving away from MSVC. We could conceivably use MSVC + Intel Fortran, now that Intel has made its Fortran compiler more freely available. It's still nontrivial to install though - gh-16957 is working in that direction, but I put that on hold because I didn't have the time to keep messing with CI (and I'm terrible at Windows command-line stuff).
To some extent yes - but we've always had cases like this. We could lower our minimum numpy version (one of 1.21.3/4/5 should work I expect). I wasn't the one who pushed for a higher numpy minimum version:) |
This is looking good. Thanks Ralf. |
This is a follow-up to gh-17959, which bumped numpy from 1.19.5 to 1.21.6, which is wrong because that version was built with vc142 (see comment above the line added in this PR).
[wheel build]
[skip circle] [skip cirrus] [skip azp]