Compare three model trainings with exact same seeding but different parallelization. (Losses must be 100% equivalent)

Ensure that runs are deterministic even across different parallelization strategies. 
Loss / Parameters / Outputs must be the same. 