I'd revised the initial ballpark fitting procedure so that the scaling and offset (beta and baseline or slope and intercept, however you like to describe it) would be calculated rather than estimated. This speeds up the fit since we can simply calculate these values.
I'd assumed that we always want non-negative scaling. This turns out not to be the case, as there are scenarios where we might want to model negative BOLD, for instance. Then I added some clunky functionality for letting users overload the ballpark estimate so that these can be negative. This ended up being cumbersome in python3 due to RecursionError.
Long story short—I removed the non-negative requirement on the beta/slope/scaling. This has simplified things so much that I'm tagging a new release.