Just to clarify the relationship with #703: Even with safe mode 'off', the exception that would be thrown for sample rejection would not cause the sampling to fail; only to reject the sample. Anyway, perhaps the dev team will decide on a different implementation than an exception.

During sampling, at least for HMC, we reject samples for which the energy is Inf or NaN. So, if you simply implement you model such that it returns Inf when you got errors, the inference would work correctly.

Why not exception? Simply because it is slow. But you can always catch error and return Inf in your model definition (or logpdf for customised distribution).

One can now do @logpdf() = -Inf; return to reject a sample in #965.

One can now do @logpdf() = -Inf; return to reject a sample in #965.

Might be worth to add this to the docs.

Sure!

Thanks for implementing this!

What's the current way to do it? @logpdf have been removed in fa44e8b#diff-107480759ab6311e0ad3c775750a1047

The current way is

Turing.acclogp!(_varinfo, -Inf)

which uses the internal variable _varinfo. In upcoming Turing versions you should use

@addlogprob!(-Inf)

instead (see TuringLang/DynamicPPL.jl#143 and the issue linked therein).

Closed by #1391.