In the documentation of the total_count parameter of the tdf.NegativeBinomial we can read:
"Its components should be equal to integer values."
However, the distribution works fine with fractional values e.g the log_prob is differentiable wrt total_count.
Is there a particular reason for such a constrain since, mathematically, the fractional value is perfectly valid when NegativeBinomial is interpreted as GammaPoisson distribution.
In fact, the sampling method is based on this interpretation.