Except LLMs are not stochastic in nature. Correct me if I am wrong, but it's just the final layer which outputs a distribution across the output tokens. In reality, the rest of the model is deterministic.

This is not like a bayes model or something were it's distributions all the way down.