Use more aggressive gc thresholds for a big speedup by JukkaL · Pull Request #18306 · python/mypy (original) (raw)
In some cases gc was consuming a significant fraction of CPU, so run gc less often.
This made incremental checking of torch 27% faster for me (based on 100 measurements), and also speeds up incremental self check by about 20% and non-incremental self check by about 10%. All measurements were on Python 3.13.
In some cases gc was consuming a significant fraction of CPU, so run gc less often. Also tune gc very early, to run gc less during import time -- not sure if this has any significant impact, but in an unrelated project this was somewhat helpful.
This makes incremental checking of torch 27% faster (based on 100 measurements), and also speeds up self check by about 10% (both incremental and non-incremental).
Don't merge yet -- I'm still performing some additional measurements. I'm not sure if tuning very early is actually useful.
This comment has been minimized.
I simplified the PR, since tuning GC early before most imports seemed to make this a little bit slower instead of faster.
Use more aggressive gc thresholds, and tune earlier Use more aggressive gc thresholds for a big speedup
According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅