Optimizing Large Model Training through Overlapped Activation Recomputation

Connections are not prepared for this record yet, or its source metadata has changed. The original record remains available while background snapshots are rebuilt.