Effect of not calling disable_logging/Why calling it causes error code -11

[ilan-gold]

Hello,

I have a very bizarre thing I'm seeing. We are trying to run something with asv for benchmarking. We have a PR for reference if you want to look. We pulled the memory logger from the asv-rapids benchmark hosted here but when we call disable_logging here it causes exactly one of our benchmarks to fail with error code -11 and no thrown error or anything. Deleting the line makes the benchmark pass.

So a workaround for us being deleting that line, what is the effect of not calling it? Is there anything I should be worried about? enable_logging will be called for each benchmark with a different name each time.

Also any insight into why this is happening would also be rad!

[harrism]

error 11 generally means a seg fault. Can you step into the rmm.disable_logging() call with a debugger to see where it fails internally?

rmm.disable_logging is not something I've worked on myself, but looking at the code it basically unwraps any per-device memory resource that is wrapped in a LoggingResourceAdapter. It could seg fault, I guess, if the wrapped memory resource was deleted somehow before that point. Does your code set the current device resource at all (from either Python or C++)?

[harrism]

If you are able to provide a simple, standalone reproducer then feel free to open an issue.