David Widmann

I would guess (without having checked the details) that the problem are some of the `tilde` dispatches in src/modes or DynamicPPL. Both `x'` and `collect(x')` are matrices, and hence `x...

`SMC` is a bit special, these options don't make sense. The returned samples are just a set of particles that were propagated together through the model. So there's not any...

IMO an `ArgumentError` would be more appropriate, usually `DomainError` is used for actual domain errors such as `sqrt(-1)`. However, I think it would be more consistent with other samplers and...

I think this issue is relevant here: https://github.com/TuringLang/Turing.jl/issues/1588 Most importantly, it seems you use the keyword argument `init_theta`. However, as also discussed in the linked issue, initial parameter values have...

It's already supported by AbstractMCMC, I don't think there are any general limitations that disallow this use case. As far as I can tell the main issue might be a...

> `x` is not stored in`state.vi` after condition, I assume due to how condition works. That's what I meant, I assume you would have to update `state.vi` manually currently. >...

IIRC these messages are generated by AdvancedVI. https://github.com/TuringLang/Turing.jl/issues/1398#issuecomment-687259280 contains two examples for how to suppress them.

As explained on Slack, `Turing.setadbackend(:reverse_diff)` is incorrect and should not be used - it does not work even if you run it in a script. On the other hand `Turing.setadbackend(:reversediff)`...

Ah nice, so it seems my hypothesis was correct and the precompilation problem is caused by Requires. The error messages that you get now with your Turing fork without Requires...

> Or is it possible to have ReverseDiffRules.jl pkg similar to ZygoteRules? ReverseDiff will support ChainRules (but it is not clear yet when) (which BTW should also be used to...