David Widmann

According to the linked benchmarks, yes. I checked where exactly Distributions is used in MCMCDiagnosticTools. It turns out mainly in the discrete diagnostics (for sampling and for evaluating CDFs, the...

I think it is nevertheless questionable if the FFT-based approach should be considered rhe best one. It is difficult to comment on accuracy and efficiency without inspecting the actual implementation...

> Which split-rhat paper? If you mean Vehtari, 2021 (cited in the docstrings), there are several differences, and as I noted, the key innovation in that paper is computing ESS...

Isn't it possible to leverage some of the existing CV implementations, e.g. in MLJ? (Edit: Just to be clear, this question is orthogonal to whether/how the package should be split)

The - currently disabled - tests still fail: https://github.com/TuringLang/NormalizingFlows.jl/actions/runs/10279602667/job/28445284206

Julia 1.6 crashes if runtime activity is enabled: https://github.com/TuringLang/NormalizingFlows.jl/actions/runs/10279812092/job/28445874772?pr=38#step:6:280

I'm not sure, it's not clear to me that this functionality should be included in https://github.com/JuliaMath/ChangesOfVariables.jl/pull/2. I would prefer if such wrapping/reshaping/flattening etc. is handled by the user.

An alternative with zero overhead would be to a function instead of a `const` `Ref`: ```julia julia> using Arblib, BenchmarkTools julia> enable_debug() = false enable_debug (generic function with 1 method)...

> If I understand it correctly it would mean that you pay the cost of recompilation when switching the flag but otherwise it should be zero overheard since the check...

The same with `const debug`: ```julia julia> const debug = Ref(true) Base.RefValue{Bool}(true) julia> g() = debug[] ? 1 : 1.0 g (generic function with 1 method) julia> @code_llvm g() ;...