Ray Tune functions and Train V2 runs

Ray Tune functions and Train V2 runs An architecture diagram generated by Archify. Tuner · searcher / trial scheduler · Architecture component Tuner searcher / trial scheduler Trial function · construct Trainer + fit() · Architecture component Trial function construct Trainer + fit() Train workers · model / data / collectives · Architecture component Train workers model / data / collectives Persistent checkpoint store · user-authored state / shards · Architecture component Persistent checkpoint store user-authored state / shards TuneReportCallback · metrics + checkpoint path · Architecture component TuneReportCallback metrics + checkpoint path trial config ScalingConfig checkpoint upload report callback rank 0 metrics / path Legend Backend Database

Explicit integration

  • • Do not pass a V2 Trainer instance directly to Tuner
  • • Use a unique Train run name and resource budget per trial

Reporting and recovery

  • • 2.58.0 V2 report is a barrier reached by every worker
  • • Callback neither reuploads checkpoints nor averages worker metrics