Find a training recipe that fits your model and GPUs.
Recent revisions are kept. Restoring one creates a new revision.