Every diffusion checkpoint records its optimizer, scheduler, AMP scaler, Python/ NumPy/Torch RNG state, resolved configuration, and a run manifest. The manifest contains a configuration hash, code revision when Git is available, platform and Torch versions, CUDA availability, and the SHA-256 of recording metadata.
To resume, retain latest_checkpoint.pt and the exact recording directory. A
resumed job restores model, optimizer, scheduler, scaler, and RNG state. It does
not yet restore worker scheduling or an external game emulator state; therefore
bit-identical continuation is not claimed for multi-worker or environment runs.
Only report a benchmark with its immutable checkpoint, config, recording metadata, evaluator version, and generated outputs. Paper values are reference values unless a local artifact set proves otherwise.
Use save_evaluation_report to write metric results atomically with their run
manifest; do not rely on terminal output as an evaluation record.