Normalisation by in strength may provoke division by zero.
The way it is currently handled leads to numerical overflows.
The previous solution didn't make a lot of sense.
The normalisation by the in-strength is done, so that if we sum all the weights of a row the result is 1.
If the in-strength is zero, then it does not make sense to normalise that row. This only means that that node will be disconnected.
To avoid division by zero a possible solution would be to divide by 1; or normalise only where the in strength is > 0.
It's very likely that the overflows are a result of very small in strength values in combination with the LB model.