Hierarchical regression of large dataset with uncertainties on two observed variables

Questions v5
1

I have large data samples with two measured observables, R and V. One sample is smaller, around 1700 data points, the others are x500 bigger.
I would like to fit these data with a model, taking into account the uncertainties of both observables and the scatter.
I have been trying real models without success, so I decided to start with a simple linear one.
The linear model is as follow:

    with pm.Model() as pymc_model:
        # DEFINING PRIORS FOR THE MODEL PARAMETERS
        V0 = pm.Uniform(name='V0', lower=0., upper=1.e3)
        slope = pm.Uniform(name='slope', lower=-100, upper=100)

        obs_r = velocity_obs[0]
        obs_r_err = velocity_obs[1]
        obs_v = velocity_obs[2]
        obs_v_err = velocity_obs[3]

        true_r = pm.Uniform(name='true_r', lower=0., upper=50., shape=len(velocity_obs[0]))
 
        # EXPECTED VALUES FOR VELOCITY
        mu_vrot = V0 + slope * true_r

        truev_ = pm.Normal(name='truev_', mu=0, sigma=1, shape=len(velocity_obs[0]))  # reparametrization
        intr_scatter = pm.TruncatedNormal(name='intr_scatter', mu=20, sigma=1.e4, lower=1.e-4)
        true_v = truev_ * intr_scatter + mu_vrot. # reparametrization

        # LIKELIHOOD
        pm.Normal(name='Radius', mu=true_r, sigma=obs_r_err, observed=obs_r)
        pm.Normal(name='Velocity', mu=true_v, sigma=obs_v_err, observed=obs_v)

        # STARTING THE CALCULATIONS
        idata = pm.sample_prior_predictive(samples=n_samp)
        idata.extend(pm.sampling_jax.sample_numpyro_nuts(draws=n_samp, **kwargs))
        idata.extend(pm.sample_posterior_predictive(idata))

This is what I get:

pymcdisc
pymcdisc773×429 93.7 KB

And this is the model with the scatter:

pymcdisc1
pymcdisc11048×735 59.9 KB

However, using real models I don’t get convergence at all.
The real models have a flat behaviour in the region of the data, like the linear model, they have only a few more parameters. So they should fit the data as well, as they do if I bin the data over R.

Do you have any suggestions on how to proceed?
Thank you very much.