TuringLang/Turing.jl: v0.38.0

Kai Xu; Hong Ge; Cameron Pfiffer; David Widmann; Martin Trapp; Tor Erlend Fjelde; Mohamed Tarek; Emile Mathieu; Markus Hauru; Will Tebbutt; Penelope Yong; Adam Scibior; Killian Q. Zhuo; Hessam Mehr; Xianda Sun; Emma Smith; Philipp Gabler; Jaime RZ; Harry Wilde; Arthur Lui; FredericWantiez; Pietro Monticone; Tom Röschinger; azev77; Kyurae Kim; Peifan Wu; Phillip Alday; Ramon Diaz-Uriarte; Saranjeet Kaur

doi:10.5281/zenodo.15371698

Published May 9, 2025 | Version v0.38.0

Software Open

TuringLang/Turing.jl: v0.38.0

1. University of Cambridge
2. @dottxt-ai
3. @PumasAI
4. Aalto University
5. @TuringLang
6. University of British Columbia
7. None
8. University of Glasgow
9. UCL
10. @nestauk @dssg @turinglang @knowlab @BHFDSC
11. University of Trento
12. Fowler College of Business, SDSU
13. University of Pennsylvania
14. Google
15. Beacon Biosignals
16. Universidad Autonoma de Madrid

Turing v0.38.0

Diff since v0.37.0

:

Breaking changes

DynamicPPL compatibility has been bumped to 0.36. This brings with it a number of changes: the ones most likely to affect you are submodel prefixing and conditioning. Variables in submodels are now represented correctly with field accessors. For example:

using Turing
@model inner() = x ~ Normal()
@model outer() = a ~ to_submodel(inner())

keys(VarInfo(outer())) now returns [@varname(a.x)] instead of [@varname(var"a.x")]

Furthermore, you can now either condition on the outer model like outer() | (@varname(a.x) => 1.0), or the inner model like inner() | (@varname(x) => 1.0). If you use the conditioned inner model as a submodel, the conditioning will still apply correctly.

Please see the DynamicPPL release notes for fuller details.

Gibbs sampler

Turing's Gibbs sampler now allows for more complex VarNames, such as x[1] or x.a, to be used. For example, you can now do this:

@model function f()
    x = Vector{Float64}(undef, 2)
    x[1] ~ Normal()
    return x[2] ~ Normal()
end
sample(f(), Gibbs(@varname(x[1]) => MH(), @varname(x[2]) => MH()), 100)

Performance for the cases which used to previously work (i.e. VarNames like x which only consist of a single symbol) is unaffected, and VarNames with only field accessors (e.g. x.a) should be equally fast. It is possible that VarNames with indexing (e.g. x[1]) may be slower (although this is still an improvement over not working at all!). If you find any cases where you think the performance is worse than it should be, please do file an issue.

Merged pull requests:

Add check_model argument to optimisation (#2518) (@penelopeysm)
Add PR auto assign workflow (#2528) (@penelopeysm)
DynamicPPL 0.36 (#2535) (@penelopeysm)
DocsGHA: Grant write permission to pull-requests event (#2544) (@shravanngoswamii)
Update docstrings to account for non-identity varnames (#2546) (@penelopeysm)

Closed issues:

New Turing domain: please visit turinglang.org (#2041)
Allow Gibbs sampler to have non-identity lenses for target variables (#2403)
Figure out a consistent and robust way of defining and testing interfaces (#2434)
Clean up exports (#2468)
Clean up LogDensityFunctions interface code + setADtype (#2473)
prefix should be exported (#2481)
Enzyme illegal type analysis on many Turing models (#2510)
NaN error during sampling with NUTS in Hidden Markov Model (#2513)
Bug: Incorrect Variance Test in check_dist_numerical (#2514)
NaN in dataset leads to NaN parameters and difficult-to-interpret errors (#2515)
Can't find initial parameters in HMM with deterministic emissions (#2526)
Run formatter, start using always_use_return = true (#2530)
Remove extract_priors (#2537)

Files