I am trying to test a function that I am writing that uses torch.arange, and am running into some floating-point issues. Namely, I am trying to figure out the (reasonable) minimum and maximum step size for which I can guarantee that torch.arange will work. I believe the minimum value is torch.finfo().eps - i.e. the machine epsilon, and the maximum value is torch.finfo().max (ignoring, for the moment, that the step size might be greater than the total length of the sequenc). Is this sound reasoning?