You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Four multi-processing tests fail in environments with CUDA available when the test suite is run (or when at least one other test initializes CUDA)
#14979
Closed
speediedan opened this issue
Oct 3, 2022
· 4 comments
· Fixed by #14982
The four multi-processing tests below (two versions for each PL and Lite) are failing in environments with CUDA available when the test suite is run (or at least one other test initializes CUDA). I'll be submitting a PR shortly to fix:
==================================================================================== short test summary info ====================================================================================
FAILED tests/tests_lite/strategies/launchers/test_multiprocessing.py::test_multiprocessing_launcher_start_method[fork] - RuntimeError: Lightning can't create new processes if CUDA is already...
FAILED tests/tests_lite/strategies/launchers/test_multiprocessing.py::test_multiprocessing_launcher_restore_globals[fork] - RuntimeError: Lightning can't create new processes if CUDA is alre...
FAILED tests/tests_pytorch/strategies/launchers/test_multiprocessing.py::test_multiprocessing_launcher_start_method[fork] - RuntimeError: Lightning can't create new processes if CUDA is alre...
FAILED tests/tests_pytorch/strategies/launchers/test_multiprocessing.py::test_multiprocessing_launcher_restore_globals[fork] - RuntimeError: Lightning can't create new processes if CUDA is a...
=========================================================================== 4 failed, 1 passed
@speediedan Thanks for noticing and proposing the fix. I noticed it too and wanted to address it for a while.
@carmocca You closed #14982 but I think we should still land it, as it fixes the two remaining tests (on lite side) that were failing and not covered in #14550
@speediedan Thanks for noticing and proposing the fix. I noticed it too and wanted to address it for a while.
@carmocca You closed #14982 but I think we should still land it, as it fixes the two remaining tests (on lite side) that were failing and not covered in #14550
Bug description
The four multi-processing tests below (two versions for each PL and Lite) are failing in environments with CUDA available when the test suite is run (or at least one other test initializes CUDA). I'll be submitting a PR shortly to fix:
How to reproduce the bug
Error messages and logs
(error due to fork after CUDA already initialized)
...
Environment
More info
No response
The text was updated successfully, but these errors were encountered: