Implementation of NPO/SimNPO is different for MUSE v/s TOFU

[molereddy]

The MUSE implementation uses logits and the TOFU implementation uses loss to find the log-ratios. Both are not equivalent and the latter is the only one faithful to the formula from the paper. The former ignores the softmax denominator term which is part of the probability expression.

NPO for MUSE:

NPO for TOFU:

SimNPO for MUSE:

SimNPO for TOFU:

Is there a reason different implementations have been used? Does the latter version lead to the same results on MUSE?