trunk/8376f7cabcc19198d1a827b2eb0c4a9e57137381: [inductor] Hold GPU lock across multi-kernel comparison (#192876)
- PyTorch: 182 events in the last 90 days
- PyTorch: 180th Release in the last 90 days
- Previous: earlier the same day · trunk/76588348eb3e6d543e06f04751268a19f57bca8d: Fix post-local SGD checkpoint reload race (#194614)
What happened
Summary Benchmark all candidates in a runtime multi-kernel comparison under one registered GPU benchmark lock. This prevents concurrent compile workers from interleaving candidate measurements on the same GPU. Existing per-candidate benchmark locking remains nested through the reentrant lock contract. Candidate ordering, forced and cached choices, and benchmark failure behavior are unchanged. Test plan python -m pyt…
Summary assembled by rule from the sources below