MvNormal much slower in pymc5 than pymc3?

Questions v5
2

The model sampled fine for me in 5.3.0, took about 3 minutes using the following data:


nt = 100
n = 30
innovations = np.random.normal(scale=10, size=(n, nt))
true_rho = np.random.uniform()
ensemble = np.zeros((n, nt))
for t in range(1, nt):
    ensemble[:, t] = true_rho * ensemble[:, t-1] + innovations[:, t]

Are you running it in a clean environment? How is your data generated? In my experience, when a model takes an eternity to run it’s a sign that some unexpected broadcasting is happening. In this case, I also wonder if there’s an environment problem from having multiple backends installed - it’s not recommended to have more than one of theano, aesara, or pytensor in a single environment.