PyTorch, the Toolbox Behind Most AI, Just Fixed a Glitchy Test

A PyTorch pull request, "Update handler clearance logic" (#196102), fixes a logging test that failed when tests were run in shards with pytest instead of python. The original test checked all active handlers with an assertion that each logger had exactly two handlers, described in the test as stream and file handlers. Under Python only the PyTorch handlers were active, but under pytest four additional handlers were added, and the assertion failed reporting that scalars were not equal — expected 2 but got 6 — for torch._export.converter. The observed handlers included pytest-specific NOTSET handlers, which the article says are only present when test_logging.py is run with pytest. The change made was to check that the number of handlers managed by PyTorch is two, instead of asserting that the logger has exactly two handlers in total. Reviewers were bobrenjc93, xuhancn, and ezyang, who approved it. It was tracked by #195996, NVIDIA Windows nightly failures and upstream fixes, and an old PR #195974 was closed due to lint failures and fixed here.