trunk/b9f83ac8558c010d8a221c819dac4e2fac7bd980: Fix remaining checkpoint filesystem races flagged in #194614 (#196591)
- PyTorch: 1077 events in the last 90 days
- PyTorch: 1064th Release in the last 90 days
- Previous: earlier the same day · ciflow/trunk/197366
What happened
Follow-up to #194614 . That PR fixed a checkpoint reload race in post-local SGD tests but noted in review ( comment ) that the same bug shapes exist in several other places. This PR fixes those remaining sites. Root cause Three related races, all stemming from sharing a checkpoint through the filesystem across ranks instead of through a collective: Fixed-name collision: a checkpoint path is built from a hardcoded na…
Summary assembled by rule from the sources below