[leifdenby]

Follow-up to #6. Resuming intrigued-sow-455 in place showed the limits of continuing inside the original run: MLflow rejected the changed trainer.max_time param (params are immutable), and the run's status/console log get overwritten by each continuation.

Change

• --config mlflow:<run> now logs the continuation as a new run named <original>-from-step<N>, tagged:
  • mlflow.parentRunId = original (the UI nests it under the original)
  • mlcast.resumed_from_run = original run id
  • mlcast.resumed_from_checkpoint = checkpoint filename
• Step numbers carry on from the checkpoint, so the runs' curves line up in the compare view.
• --config mlflow-inplace:<run> keeps logging into the original run (previous behaviour).
• resume_training gains inplace and run_name. For a linked run it clears any run_id, run_name and links inherited from a config saved by an earlier resume (otherwise it would log into the old run again).
• When a run has several config files (old temp-named one + config.yaml from an in-place resume), config/config.yaml is used.

Testing

• End to end against a local MLflow store: continuing a 40-step run created upset-calf-592-from-step40 with the three tags pointing at the original, its own config/ and logs/, checkpoints at step 60; the original run was untouched.
• New/updated tests for the linked-run tags and name, dropping inherited ids, inplace, the mlflow-inplace: prefix, and config.yaml preference. Full suite: 177 passed, 5 skipped.

🤖 Generated with Claude Code