-
-
Notifications
You must be signed in to change notification settings - Fork 5.1k
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
Error with _fblas #11826
Comments
I have the same issue, couldn't find a fix either |
Could you please add some details on how you installed Python, how you installed SciPy, and provide relevant debugging info like |
i am on windows, installed python via windows 10 store. the ls you ask for:
the output of the config:
also, apart from |
There's a file
unless there's a recurrence of the DLL load library path issue here .... @zoeherri did you also install Python from the Windows Store? |
|
@rgommers didn't work, same error "cannot find module"
|
Yes, I think this is a duplicate of gh-10171, Windows Store Python 3.7 is problematic. Could you upgrade to 3.8? |
Oh this brings back bad memories. @tylerjereddy the
This is a new micro release. I suspect the DLL path resolution was changed in that release. @petacreepers23 could you download https://github.com/MacPython/scipy-wheels/blob/master/_distributor_init.py and put it in the main |
ok @rgommers i uninstalled windows store version, and installed python 3.7.7 from python.org, I cannot use python 3.8 because of other dependencies. It seems the problem is microsoft store python... too bad for that UWP api that might use... |
* we've been getting reports of new Python 3.7 patch releases from Windows store causing SciPy DLL load issues: scipy/scipy#11826 * so, extend the `_distributor_init.py` machinery usage to include Python 3.7 and up for our wheels * simplify the appveyor `matplotlib` install now that there are stable releases available for Python `3.8`
I've opened the cross-linked wheels PR above to include Python 3.7 for |
Thanks Tyler. Let's see if more reports come in and if we get confirmation that the "copy in |
I found some weird behaviour, not affecting my development of apps, but just as a curiosity,
Also the _distributor_init.py is in this normal version of python, but the file just have some commented lines. I'll try the win10store version and let you know. |
Yeah there's a default dummy file, which needs to be replaced depending on how binaries are built. |
ok, reinstalled win10 python, and still have problems while importing linalg.
and it works! , with that file i can import signal and linalg |
Great thanks @petacreepers23. That's at least an easy workaround for others who hit this issue. We can include this fix in the next release. |
I have a question here since I've been struggling a bit with the fix in our wheels repo--would it be "ok" to only apply the fix to 64-bit Windows wheels with Python 3.7? I noticed in the Windows store that Python 3.7 has "x64" as minimum arch requirement, so technically users should have the option to use 64-bit mode? Another perhaps-reasonable excuse for this shortcut is that Python 3.7 in windows store is described as experimental, so that 32-bit DLL resolution rabbit hole might best be revisited later:
|
So is adding the Leaving out 32-bit for now seems okay if needed. I would make sure it's unconditionally applied for 64-bit though, otherwise we'll run into it again for 3.6 or 3.9 |
I did not install from the windows store. I used https://www.python.org/downloads/ to get the latest version and to pull 3.7 in case it was a but with the latest. I used pip to uninstall and reinstall scipy but when reinstalling it calls from a cache. I am currently uninstalling python and all related libraries to start from scratch. |
I ended up clearing the cache, reinstalling and then telling it to update scipy once installed. I can now import scipy and associated libraries. |
I think I'm reasonably happy with the current version of the proposed fix in: MacPython/scipy-wheels#70 It adds a The main trick was to add a The only remaining failures in that PR are unrelated: #11095 |
* MAINT: init file and mpl updates * we've been getting reports of new Python 3.7 patch releases from Windows store causing SciPy DLL load issues: scipy/scipy#11826 * so, extend the `_distributor_init.py` machinery usage to include Python 3.7 and up for our wheels * simplify the appveyor `matplotlib` install now that there are stable releases available for Python `3.8` * MAINT: update .travis.yml * fix warnings in `.travis.yml` based on feedback from https://config.travis-ci.com/explore * bump distro to `bionic` since NumPy did this recently too * attempting to restore Travis CI runs to our wheels build matrix * MAINT: revise PR 70 * try using `[System.Version]` for more robust Python version comparisons in powershell * MAINT: PR 70 revisions * update minimum NumPy version to `1.14.5` * the distutils override for appveyor/Windows builds now only happens for Python < 3.7, since we have now added the `_distributor_init.py` for Python 3.7 (not just Python 3.8) * DEBUG: try newer NumPy * try using newer NumPy version for Windows Python 3.7 builds, to see if this helps `check_installed_package.py` * Add debug prints to _distributor_init.py * Debug print contents of .libs path for DLLs * Make sure libs_path exists before trying to debug print contents. * DEBUG: simplify CI matrix for iteration. * DEBUG: more debug prints near point of 32-bit Python 3.7 DLL resolution failure. * DEBUG: try moving into DLL dir * try to solve DLL resolution issues for older (32-bit) Python versions by temporarily moving into the path where they are stored prior to load attempts in `_distributor_init.py` * Revert "DEBUG: simplify CI matrix for iteration." This reverts commit 46e69e1. * MAINT: PR 70 cleanup * remove some debug prints * revert some debug changes * use `_distributor_init.py` with Python 3.6+ (all supported Python versions) * MAINT: PR 70 revisions * wrap the `_distributor_init.py` `os.chdir` code in a `try .. finally` block to ensure restoration of the working directory if something fails * substantially expand the comments describing the rationale for the working directory changes in `_distributor_init.py` * MAINT: PR 70 revisions * sync multibuild submodule to latest `devel` branch checkout in attempt to deal with Travis CI failures that just started appearing
The fix was merged, so let's close this. The 1.5.0 wheels will have the fix, if anyone runs into this with 1.4.1 please use this workaround: #11826 (comment) (a 1.4.2 may follow, but not guaranteed). Thanks everyone! |
I am trying to use scipy.spatial.distance.cdist
I can import scipy but cannot call or import scipy.spatial
I have searched the web and found other posts similar to this but none have worked for me. I tried different versions of scipy, numpy and python itself. I continue to get the error shown in the image below. I do not appear to have _fblas.py. I searched the C drive of my computer and all it returned was test_fblas.py, which is present in the test folder within the scipy folder. I appreciate any help leading to the resolution of this hours long quest.
The text was updated successfully, but these errors were encountered: