scikit-cuda
scikit-cuda copied to clipboard
lebedov
scikit-cuda test failed
Compiled scikit-cuda from source, build no problem but one test failed
The output for
python setup.py test
Ran 336 tests in 56.388s
OK (skipped=61)
Error in atexit._run_exitfuncs:
Traceback (most recent call last):
File "<frozen importlib._bootstrap>", line 991, in _find_and_load
File "<frozen importlib._bootstrap>", line 975, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 671, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 783, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/bernard/opt/python38/lib/python3.8/site-packages/pycuda-2019.1.2-py3.8-linux-x86_64.egg/pycuda/tools.py", line 34, in <module>
import pycuda.driver as cuda
File "/home/bernard/opt/python38/lib/python3.8/site-packages/pycuda-2019.1.2-py3.8-linux-x86_64.egg/pycuda/driver.py", line 8, in <module>
import numpy as np
File "/home/bernard/opt/python38/lib/python3.8/site-packages/numpy/__init__.py", line 142, in <module>
from . import core
File "/home/bernard/opt/python38/lib/python3.8/site-packages/numpy/core/__init__.py", line 24, in <module>
from . import multiarray
File "/home/bernard/opt/python38/lib/python3.8/site-packages/numpy/core/multiarray.py", line 14, in <module>
from . import overrides
File "/home/bernard/opt/python38/lib/python3.8/site-packages/numpy/core/overrides.py", line 16, in <module>
add_docstring(
RuntimeError: implement_array_function method already has a docstring
- OS platform : Ubuntu linux
- Python version: 3.8.3
- CUDA version: 10.0 from Nvidia and linked to /usr/local/cuda
- scikit-cuda version 0.5.3
In addition pycuda 2019.1.2 numpy 1.18.4
The error is not coming from a failed skcuda test because "OK" was printed after the tests were all run and the traceback doesn't come directly from a test. I can't replicate the problem with those package versions and Python 3.8.2; can you please try an earlier version of numpy?
Tried numpy 1.17.5 on scikit-cuda 0.5.3, pycuda from git and python-3.8.5 still same test error (which probably is harmless)
With sciki-cuda master there is one additional test error
Failure: OSError (libmagma.so not found)
I think you need a pay subscription to get libmagma.so from MAGMA.
There might be a fix for this issue in numpy 1.19+ - try that and see what happens.
Magma is free software - you don't need to pay any subscription to obtain it. I actually built a conda package for it once that contains the shared library, although it probably ought to be updated to the latest version.
@lebedov
Magma is free software
Really??!! From SAGE's manifesto 'Mission: Creating a viable free open source alternative to Magma, Maple, Mathematica and Matlab' https://www.sagemath.org/
So I take that to mean Magma is not free, rather like Maple, Mathematica and Matlab.
Really??!! From SAGE's manifesto 'Mission: Creating a viable free open source alternative to Magma, Maple, Mathematica and Matlab' https://www.sagemath.org/
So I take that to mean Magma is not free, rather like Maple, Mathematica and Matlab.
The MAGMA used by scikit-cuda is this one. The software that SAGE manifesto is referring to is a different MAGMA for computer algebra.
The reported error is not present anymore in numpy 1.19 (see https://github.com/numpy/numpy/issues/14384), so this can probably be closed. (Note that I did not reproduce this in scikit-cuda, but in another package I had the same issue when exiting the test, and this was fixed upgrading to numpy 1.19)
