0.5.1

This release keeps long runs from failing on their saved checkpoints, and lets you serve Gemma 4 and branch a supervised run from its full training state.

Changes

  • A reinforcement run no longer saves the frozen base model in its checkpoints, only the adapter, so its pauses, resumes and branches restore less.

  • A run removes each checkpoint from where it trains once the snapshot is stored. Previously, a long run could fail for lack of space.

  • A run retries recording a snapshot for about five minutes when the service does not answer, such as while the service restarts. Previously, the save failed.

  • An endpoint over Gemma 4 starts and answers. Previously, it failed as it started.

  • A trainingState branch of a supervised run trains. Previously, it failed after it restored its parent’s state, before its first step.