PyTorch tests are often generated at import time, so CI failures may show device/dtype-specific names that differ from the source template. For local debugging, pytest -k and test/run_test.py are usually the fastest ways to reproduce generated test failures.