lowering_stats

lowering_stats()

Return a snapshot of rewrite counters since the last reset.

Keys are "rewritten" plus one "skipped:<reason>" entry per rejection cause (e.g. "skipped:spatial-too-large"). A pass that silently does nothing is hard to debug; this is how you tell “did not match” from “matched but rejected”, and why.

Note

Counters only advance when the pass actually runs. Inductor caches compiled artifacts on disk, and a cache hit skips every pre-grad pass — so empty counters can mean “served from cache” rather than “did not fire”. Set torch._inductor.config.force_disable_caches = True when using this to verify behaviour.

Returns:

A copy of the counter dict.

Return type:

dict[str, int]