different target score when using logits from sample_autoregressive

[marton-avrios]

I am currently try to implement returning logits along with prediction from sample_autoregressive to calculate score from them. However the score calculated from these logits are slightly different from the ones I get with mode="scoring". Any explanation?