I did not know there was such a transformation. Looking at the code (https://github.com/pymc-devs/pymc3/blob/05e95dc01eee194a8e67e4e7a987906cf21cc8b2/pymc3/distributions/transforms.py#L401-L425) it seems the transformed values have some constraints, .e.g., beeing positive, and sum to something below 1. Using this in ADVI and the usual normal distributions to model the parameter distribution would violate the support constraint. That is the support of the fitted parameter distribution must coincide with the support of the posterior.
Using this transformation, SumTo1, with ADVI immediately yields the error:
FloatingPointError: NaN occurred in optimization.