how to version step functions for ML?

A Step Functions state machine usually doesn't come alone and typically relies on other resources such as Lambda, EC2, DynamoDB, etc. You might want to package these dependent artifacts/resources altogether within a version otherwise you might have a state machine that doesn't fully work (eg, state machine version doesn't match Lambda version). I guess the simplest way to achieve this is to provision these resources together as code (eg, CDK or CloudFormation) and store them in a Git repo. You could then use Git tags for versioning.