This seems to be unrelated to pm.Bernoulli(), because I can get the same thing with other distributions. In messing around a bit, it seems like the problem arise when having a parameter (e.g., Y) depend on an observed parameter that uses pm.Data() (e.g., X). So if you can work around that, there may be a short-term solution.
If you would like submit an issue via github, there may be additional develop feedback. Or I can submit it. Just let me know.