-
Notifications
You must be signed in to change notification settings - Fork 154
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
[BUG]: Tests fail with "Core dumped" when using latest jax(lib) 0.4.28 #578
Comments
Logs:
|
More precisely, the issue is fixed since/by eed8201. |
Would you be able to provide a backtrace of a crashing build in debug mode? It's really difficult to say what the problem might be based on this information. |
I am not sure on how to extract such a backtrace, knowing that it is the python process itself that is crashing. |
I'm concerned that there may be another issue. The commit you listed doesn't really explain why one version crashes and the other one works. Could you run |
ping @GaetanLepage |
Sorry for the delay :/
Am I doing something wrong ? |
Hi @GaetanLepage , it's expected that Python itself would not have interesting debug symbols, it's the plugin that will provide them. To get CMake to build the nanobind test suite with debug symbols, I don't think that |
Thank you for those precise instructions. Here is the backtrace:
|
We are preparing the upgrade to nanobind 2.0 and there, this issue does not occur. |
Awesome. One last question: which version of nanobind is this? Can you tell me what's on |
This is on the Here are lines 311 - 313 of #if defined(NB_ABORT_ON_LEAK)
abort(); // Extra-strict behavior for the CI server
#endif |
Ok. So this is intentional. There is a reference leak, and the test suite crashes at the end to point everyone's attention to this. (Reference leaks are detected all the way at the end when the interpreter shuts down, and at that point this is the only way to make sure the issue doesn't go unnoticed). I will close this then. |
Thanks for helping to localize the issue! |
The issue here very likely lies with one of the other tensor frameworks. They sometimes hold on to the last ndarray converted and don't release a reference to a nanobind object by the time this shutdown routine is called. It's a benign issue. |
Ok great ! Thanks for your patience. |
Problem description
When running the test suite while the latest jax/jaxlib (v0.4.28) is installed,
pytest
will suddenly crash withAborted (core dumped)
after the tests have (supposedly) all succeeded.This weird behavior doesn't happen if I uninstall the jax library (the tests are then skipped and
pytest
quits without error).More interestingly,
pytest
runs fine when I use the latest commit (c545446 as of today). It only occurs on tag v0.9.2Context: Updating jax in the nixpkgs repo.
Reproducible example code
No response
The text was updated successfully, but these errors were encountered: