This isn't necessarily true, it just depends on the implementation. I can say that because I've published research which embeds steganographic text into the out

This isn't necessarily true, it just depends on the implementation. I can say that because I've published research which embeds steganographic text into the output of GPT-2 and we had to deal with this. Running everything locally was usually fine--the model was deterministic as long as you had the same initial conditions. The problems occurred when trying to run the model on different hardware.