I think Simon is right that the problem is the variance parameter. iirc BUGS/JAGS always parameterizes Normal distributions with the precision (tau), rather than the standard deviation (sigma). This is also available in PyMC. You can do this instead:
alphas=pm.Normal('alphas',mu=mu, tau=tau[1]**0.5,dims='school')
tau and sigma are linked by tau = 1/sigma, so if you pass tau to the sigma argument you’ll end up with ultra-informative priors.