Right, the difference is that I am using the developmental version of ArviZ. Do you have an example that you could share, even if made up. I have not tested BART with categorical likelihoods, it may need some internal tweak, like what we currently do for binomial data.
Note: Also take into account that pmx.bart.predict is making predictions for the values of μ, so if you want to obtain observations you need first need to apply the softmax function and then the categorical distribution, something like:
pvals = pm.math.softmax(pmx.bart.predict(idata, rng, X=X_new)[0], axis=1).eval()
new_y = [np.random.multinomial(n=1, pvals=pvals[i]).argmax() for i in range(120)]
plt.hist(new_y, alpha=0.5)
plt.hist(Y, alpha=0.5)
