Without an example data set I can try running your model on, I can’t offer any specific suggestions, but one thing that that jumps out to me is that your data is non-negative, but your likelihood pm.StudentT allows negative numbers. You might want to transform your linear predictor μ_f using pm.math.exp or log1pexp (softplus). The same is true for your scale parameter, σ_f. That one must be transformed to be positive.
I would also remove the two pm.gp.cov.WhiteNoise components in favor of using the jitter argument of pm.gp.Latent.
Also, why sample_smc instead of the default NUTS? NUTS should work fine here