VizTracer --log_torch triggers an error and segfault on exit when used in conjunction with PyTorch profiler and the multiprocessing mode of DataLoader, where kineto_results is None.

[PEKEW]

When using VizTracer's --log_torch feature to sample PyTorch code (DataLoader with multiple workers and large HDF5 files (60GB+) ), the program attempts to export the torch profiler trace on exit, resulting in an AttributeError / RuntimeError. Ultimately, the process segfaults and exits. It appears that VizTracer fails to robustly handle the scenario where the profiler does not properly generate kineto results (with kineto_results being None) upon exit.

The program throws an AttributeError / RuntimeError on exit, and eventually crashes with a segmentation fault (core dumped).

OS: ubuntu 22.04
Python ：3.10.19
PyTorch : '2.9.0+cu130'

[gaogaotiantian]

Do you have a traceback? If you already saw exit routine, the segfault might happen in torch profiler, because viztracer does not have any native code after that point. If the torch profiler crashes when exporting data, viztracer can't really do anything.