While this may not have much impact for numpy based calculations, any autodiff implementation would benefit greatly (in terms of numerical stability) from allowing users to feed log_weights directly instead of weights. This is particularly valid for a few reasons:

• Most packages out there prefer outputting logpdf for random variables
• Most techniques (bregman projections, smooth ot at least) can (should?) be written purely in the logspace without much work for numerical speed and stability