-
Notifications
You must be signed in to change notification settings - Fork 139
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
Update SciPy version #949
Update SciPy version #949
Conversation
Trying to go further, 1.6.3 version seems reachable! |
For versions 1.7.x and 1.8.x I still have issues with numpy |
I put quotation marks around all the version numbers to prevent things like 0.10 being parsed as 0.1, so I've merged that into this PR. This also picks up all the changes mentioned in #648 (comment), so please let me know if that causes any trouble. |
Still working for me, I just had to reinstall requirements to get |
Sorry, tests are not passing with 1.6.3 because of requirements: |
Ok now tests OK |
Or maybe better to downgrade numpy to a <1.23.0,>=1.16.5 version, but which one ? We also need to provide it |
Yes, it's probably best to respect the limits declared by the package, even if they added them for questionable reasons.
1.19.5 is the obvious choice:
|
…fore lib release: scipy/scipy#12862)" This reverts commit c27a650.
Seemed obvious at first but can't make it build with 1.19.5, not sure why... :( And as this version is not built yet for Python 3.9 (and still need some patch-fixing because of 52cef01 ) we are free to use another. I propose 1.20.3 for Python 3.9 And tests are passing |
By the way, to build numpy 1.20.3, just need to remove last update from the patch. |
Was this the error you received when building NumPy 1.19.5 for Python 3.9? I got this on x86 and armeabi-v7a, but not on x86_64 and arm64-v8a, which built successfully. So it's probably a 32 vs 64-bit issue. And unfortunately it looks like 1.20.3 has the same problem.
|
Nope, numpy 1.19.5 is building, but Scipy is not when using this version as a dependency |
Numpy 1.23.3 doesn't build anymore either on armeabi-v7a |
Ok found, my mistake : we should not remove last patch, but alter it temporarily to build 1.20.3, the error is due to setuptools version
So Numpy 1.23.3 is still building on armeabi-v7a with the existing patch. |
Thanks, I'll try that now.
Yes, but I'd prefer not to invest time into replicating yet another conda-build feature, because our long-term plan is to migrate to using conda-build directly. |
I've built and released NumPy 1.20.3 for Python 3.9, and saved the recipe in #957. But now SciPy 1.6.3 is also failing on the 32-bit ABIs, with the error:
I think I know how to fix this, so I'll look into it tomorrow. |
I think both of the 32-bit failures were caused by the fact that the monkey patches in sitecustomize.py, which were supposed to filter out include and library paths from outside of the build directory, didn't work on the older versions of setuptools which were requested by the pyproject.toml files of these NumPy and SciPy versions. #935, which involved a newer version of setuptools, was probably caused by the same thing. So I've moved these directory filters into a compiler wrapper script instead. This means it'll take effect on all build systems, both Python and non-Python. However, for Python packages, it would still be a good idea to look into using crossenv, so the invalid paths don't get put on the command line in the first place. This should also allow one of the remaining monkey patches in sitecustomize.py to be removed. |
I've removed a couple of things from the patch which weren't needed anymore, and reordered the files so they're still in alphabetical order. SciPy 1.6.3 is now in the public repository for Python 3.9. Thanks again! |
A small step forward for #648 , I could update the existing patch and reach SciPy 1.6.3
Tested with Python 3.9 and arm64-v8a ABI