ML run a GPU that uses matrix multiplies isn't deterministic unless you go through great pains to lock things down at the expense of performance.