Here is a little program that illustrates the problem:

using Turing, ReverseDiff, Memoization
using Statistics
using Random
Turing.setadbackend(:reversediff)
turing.setrdcache(true)
Turing.turnprogress(false)
using Distributions
for i in 1:10
    N = 1000
    d = Normal(0.0, i)
    data=rand(d,N)
    @model g(rn,T) = begin
        sigma ~ Uniform(0.0,20.0)
        rn ~ MvNormal(zeros(T),sigma)
    end
    chn = sample(g(data,N),NUTS(0.65),1000)
    println("Std of chain: ",mean(Array(chn[:sigma])),"Std of data: ",std(data))
end

I am creating a dataset that samples from a Normal distributions and our model is estimating the sigma of the distribution. When you run the code you can see that the sampler samples only from the first dataset and therefore the estimated mean for sigma is always around 1. I tried to run Turing.emptyrdcache() after sampling but that did not solve the problem.

Here is my Project.toml:
Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
Memoization = "6fafb56a-5788-4b4e-91ca-c0cea6611c73"
Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
StatsPlots = "f3b207a7-027a-5e70-b257-86293d7955fd"
Turing = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"