That prior is way too broad. Specially since you are taking the exponent for v. One standard devation is already enormous: np.exp(105) is 3.989519570547216e+45.
The mu / sigma parameterization is very sensitive, as sigma is only valid if sigma < np.sqrt(mu * (1 - mu)). You have to at least set a valid testval for each Beta, but this hard constraint will make sampling very inefficient (no reason to use find_MAP, if you can use MCMC sampling). The lack of testval is probably what is giving you the invalid initial evaluation.
You can parameterize your Betas with alpha = w*(k-2) + 1 and beta = (1-w)*(k-2) + 1 where w = r[ti-1] and k = 1 / v_[ti-1]: Beta distribution - Wikipedia