Comment by mh-

To my understanding, with batched inference and other "optimizations" you wouldn't get the exact same token predictions even with temp=0.0.