FredericWantiez

Can I give this one a try ? Made some progress [here](https://github.com/FredericWantiez/aesara/pull/1) (it doesn't quite work yet). It's mostly a few tweaks on: https://doi.org/10.1145/42372.42381 and the Numpyro implementation. Seems like...

https://github.com/TuringLang/AdvancedPS.jl/pull/80#discussion_r1328727017

1) We can drop this one, that really only applies when AdvancedPS is used with Libtask outside of Turing. We will probably sunset that (or target people who supposedly know...

@willtebbutt I think 2) might also be a problem for Turing, when looking at this part: https://github.com/TuringLang/Turing.jl/blob/afb5c44d6dc1736831f45620328c9d5681748111/src/mcmc/particle_mcmc.jl#L140-L142

Two small issues I found cleaning up the tests. Libtask returns a value after the last produce statement: ```julia function f() Libtask.produce(1) Libtask.produce(2) end t1 = TapedTask(nothing, f) consume(t1) #...

That should work, I have a branch against Turing that tries to do this but seems like one copy is not quite correct. The other solution is to use one...

@willtebbutt running models against this PR I see a large performance drop: ```julia using Libtask using AdvancedPS using Distributions using Random mutable struct NormalModel

@willtebbutt if you're testing against the released version of Libtask/AdvancedPS you need to explicitly pass the RNG in the model definition, something like that: ```julia function (model::Model)(rng::Random.AbstractRNG) # Add the...

Yes, that fixed the issue with the return value ! Thanks