Add Logging module to the code. #524
Conversation
…s/logging"" This reverts commit 22bdac3.
| except ImportError: | ||
| warnings.warn("Unable to determine the package version. " "This is likely a broken installation.") | ||
|
|
||
| # This is needed for compatibility with some older compilers: erfinv needs to be | ||
| # imported before other packages though I'm not sure why. | ||
| from scipy.special import erfinv |
There was a problem hiding this comment.
When this line is removed, the tests no longer run on Baldur for me. I get the error:
File "/astro/users/jkubica/kbmod_logs/kbmod/src/kbmod/__init__.py", line 14, in <module>
from kbmod.search import Logging
ImportError: /astro/users/jkubica/kbmod_logs/kbmod/src/kbmod/search.cpython-312-x86_64-linux-gnu.so: undefined symbol: _ZNSt15__exception_ptr13exception_ptr10_M_releaseEv
But as soon as I put it back in all the tests pass fine. I have no idea why or what is happening . But the PR works fine with the include.
There was a problem hiding this comment.
Now that the PR is merged, I'm seeing the same error on baldur. Though no need for a rollback since
from scipy.special import erfinv in __init__.py does seem to fix it for me as well.
Wondering however if a better solution for the environment issue was found?
Fix typo Co-authored-by: Jeremy Kubica <[email protected]>
|
This is not a blocker to anyone, nor add a critical functionality. So I would not be for the idea that blocks one dev from continuing to contribute. I'm still available for a debugging session if you want. |
|
Ok, we sorted this out so I'm merging. |
Following some discussion the issue that was the cause for reverting the original logging module merge seems to be environment inconsistency. Cuda can build packages with a different standard library implementation than the system default. Merging of logging seems to have updated the env packages, or included something not previously required, which resulted in an ABI incompatibility.
There are many instances of discussions around the internet as how to properly handle this issue (see for example pybind/pybind11#3623), but it seems the issue is that ultimately C/C++ have no way of propagating, validating or enforcing compiler requirements across packages resulting in the fact that one can build two packages in the same environment, that link against 2 different standard library implementations. The culprit seems to have been somewhere in SciPy.
Consequently, I've removed the reverted logging PR and removed the SciPy patch from #506. I'm unable to reproduce the error, so merging of this PR should wait until the environment is fixed.