Sampling log-likelihood from prior predictive distribution fails with `KeyError: 'posterior'`

I’m trying to sample from \log p(y| \theta, d) with \theta \stackrel{\text { i.i.d. }}{\sim} p(\theta) and d being a given design. As far as I understand, this means sampling from the prior predictive distribution and then calculating the log probabilities under p(y| \theta, d). I was hoping to achieve this with:

pymc.sample_prior_predictive(10, model=model, idata_kwargs={'log_likelihood': True})

but ’ getting a KeyError: 'posterior'. Am I doing something wrong?

Here is the (very simple) model:

k = Normal('k', mu=k_mu, sigma=k_sigma)
x0 = Normal('x0', mu=x0_mu, sigma=x0_sigma)

likelihood = Bernoulli('y', logit_p=k * (x - x0), observed=y)

Hi! Could you share a reproducible example I could run? Thanks!