Strange Behaviour of sample_prior_predictive()

Questions
1

Hello everyone,

I am currently working on a model using NormalMixture() as Likelihood. I encounter some (in my opinion) strange behaviour of sample_prior_predictive(). I expect to see a bunch of normal mixture distributions, as in my second plot (green lines). However, neither does plot_ppc() from arviz produce those (first plot), nor are the prior predictive samples of the observed variable distributed that way (red lines in second plot). Is this a bug, or did I misunderstand anything?

I installed pymc3 according to the explanation (for MacOS) on GitHub.
I’m running:
pymc: 3.11.4 (pip)
theano-pymc: 1.1.2 (pip)
arviz: 0.11.2 (pip)
python: 3.9.7
OS: macOS Big Sur (M1 chip)

A short example of my problem:

import pymc3 as pm
import arviz as az
import numpy as np
import matplotlib.pyplot as plt

random_seed = np.random.seed(123)

# create some multimodal data
data = pm.NormalMixture.dist(w=[0.75,0.25],mu=[212,213],sigma=0.05).random(size=1000)

with pm.Model() as model:
    # priors
    # positions of normal distributions
    pos = pm.Normal("pos",mu=data.mean(),sigma=1,shape=2)
    # weights sum to 1
    w1 = pm.Uniform("w1",lower=0,upper=1)
    w2 = 1-w1
    weights = [w1,w2]
    # standard deviations of normal distributions
    sigma = pm.Uniform("sigma",lower=0,upper=1,shape=2)

    # Likelihood
    obs = pm.NormalMixture("obs",w=weights,mu=pos,sigma=sigma,observed=data)

    # strange behaviour?
    priorpc = pm.sample_prior_predictive(random_seed=random_seed)
    az.plot_ppc(az.from_pymc3(prior=priorpc),group="prior")
    plt.show()

# what I would expect

rand_indices = np.random.choice(np.arange(500),size=50,replace=False)
for r in rand_indices:
    w1 = priorpc["w1"][r]
    pos = priorpc["pos"][r]
    sigma = priorpc["sigma"][r]
    mixture_sample = pm.NormalMixture.dist(w=[w1,1-w1],mu = pos,sigma = sigma).random(size=1000)
    # in green: sampled mixture distributions (that look like mixture distributions)
    az.plot_kde(mixture_sample,plot_kwargs={"color":"green"})
    # in red: corresponding arrays of observed variable in prior predicitve dict
    az.plot_kde(priorpc["obs"][r],plot_kwargs={"color":"red"})

az.plot_kde(data,plot_kwargs={"color":"black"})
plt.show()

my output:
plot1
plot2

Thank you very much in advance for any help!