Update for newer AbstractMCMC/Turing interface #49

penelopeysm · 2025-12-14T16:16:32Z

Comments in PR!

penelopeysm · 2025-12-14T16:18:32Z

ext/SliceSamplingTuringExt.jl

-# Required for using the slice samplers as `externalsampler`s in Turing
-# begin
-function Turing.Inference.getparams(
-    ::Turing.DynamicPPL.Model, sample::SliceSampling.Transition
-)
-    return sample.params
-end
-# end


The changes to the external sampler interface are described in https://github.com/TuringLang/Turing.jl/releases/tag/v0.42.0 -- the general aim is that you should not need to overload Turing internal functions (getparams was actually not exported afaik) and shifting this to AbstractMCMC means that it's easier for other packages to make use of this info.

Turing.Inference.getparams is gone now, it's replaced with AbstractMCMC.getparams (but called on the state instead of the transition).

penelopeysm · 2025-12-14T16:18:52Z

ext/SliceSamplingTuringExt.jl

-Turing.Inference.isgibbscomponent(::SliceSampling.RandPermGibbs) = true
-Turing.Inference.isgibbscomponent(::SliceSampling.HitAndRun) = true
-Turing.Inference.isgibbscomponent(::SliceSampling.Slice) = true
-Turing.Inference.isgibbscomponent(::SliceSampling.SliceSteppingOut) = true
-Turing.Inference.isgibbscomponent(::SliceSampling.SliceDoublingOut) = true


isgibbscomponent now defaults to true so it no longer needs to be overridden

penelopeysm · 2025-12-14T16:22:49Z

src/SliceSampling.jl

+abstract type AbstractStateWithTransition end
+AbstractMCMC.getparams(state::AbstractStateWithTransition) = state.transition.params
+AbstractMCMC.getstats(state::AbstractStateWithTransition) = state.transition.info
+function AbstractMCMC.setparams!!(
+    model::AbstractMCMC.LogDensityModel, state::AbstractStateWithTransition, params
+)
+    new_lp = LogDensityProblems.logdensity(model.logdensity, params)
+    new_transition = Transition(params, new_lp, NamedTuple())
+    return Accessors.@set state.transition = new_transition
+end


Since the definitions of these functions are the same for all states in this package, I thought it would be cleaner to just define the behaviour on an abstract type. It does necessitate an extra dep on Accessors, but that's fairly lightweight.

penelopeysm · 2025-12-14T17:45:06Z

The failing tests are to do with initial_sample here

SliceSampling.jl/ext/SliceSamplingTuringExt.jl

Lines 38 to 41 in 2dd6716

    
           function SliceSampling.initial_sample(rng::Random.AbstractRNG, ℓ::Turing.LogDensityFunction) 
        
               n_max_attempts = 1000 
        
               model, vi = ℓ.model, ℓ.varinfo

The fact is that when SliceSampling is used via Turing's externalsampler mechanism, that code path will never be hit, because externalsampler makes sure to provide a vector of initial params:

https://github.com/TuringLang/Turing.jl/blob/4dc7ad096f92fd571de312e7751986484ac6cb50/src/mcmc/external_sampler.jl#L151-L177

So this code path will only be hit if somebody actually manually constructs a LogDensityFunction and passes it to SliceSampling without specifying initial params. One option is therefore to delete the function and the tests. Even though this seems like a bit of a niche use case, I also don't see any justifiable reason to not support it, and deleting it seems overkill.

If we don't go down that route, we need a patch to make it work again, since DynamicPPL 0.39 removed some functionality that was needed for this. That patch is here: TuringLang/DynamicPPL.jl#1178. Thus, we need to wait for that to be merged, and I can revisit this after that.

Once that is done, I think that the Turing ext can be changed into a DynamicPPL ext, which will make life easier for everyone (more accurate compat bounds, no need to update when there are unrelated changes in Turing, etc.)

Red-Portal · 2025-12-14T19:45:32Z

So this code path will only be hit if somebody actually manually constructs a LogDensityFunction and passes it to SliceSampling without specifying initial params. One option is therefore to delete the function and the tests. Even though this seems like a bit of a niche use case, I also don't see any justifiable reason to not support it, and deleting it seems overkill.

I think it is fine to remove the custom initialization. After all, the purpose of that ext is to make things work automatically with Turing, nothing more.

Once that is done, I think that the Turing ext can be changed into a DynamicPPL ext, which will make life easier for everyone (more accurate compat bounds, no need to update when there are unrelated changes in Turing, etc.)

Yes, sounds good to me. Let me handle that once this PR is merged.

Red-Portal · 2026-01-15T17:27:52Z

Hi @penelopeysm , what are your plans for this PR?

penelopeysm · 2026-01-15T20:49:46Z

Apologies, I took an actual Christmas break and this slipped off the radar. I'll tidy it up tomorrow or something.

penelopeysm · 2026-01-17T02:15:52Z

I think it is fine to remove the custom initialization.

In that case, the entire extension can be removed! Essentially, all the functions that need to be overloaded are now in AbstractMCMC, so they just live in src. I've updated the PR by, well, simply removing the extension.

I think in Project.toml the extras section for test deps is not needed too, since you have a test/Project.toml, so I removed those too to avoid redundant information. It all works locally for me.

Finally, I also removed the tests that used to check the behaviour of initial_sample.

Now unfortunately CI will still fail -- that is due to a DynamicPPL bug, which I'll fix now.

penelopeysm · 2026-01-17T02:23:06Z

TuringLang/DynamicPPL.jl#1202 should fix it, but I'll only be able to check tomorrow.

Red-Portal

All looks great to me. Seems like we don't even need a DynamicPLL extension now, is that correct?

penelopeysm · 2026-01-17T22:11:10Z

Yup, that's right :)

Red-Portal · 2026-01-17T22:18:32Z

@penelopeysm Could you add a one-liner docstring to AbstractStateWithTransition so that Documenter doesn't complain? I'll merge the PR immediately after that.

penelopeysm · 2026-01-17T22:19:33Z

Ah apologies I should have checked CI haha

github-actions · 2026-01-17T22:32:48Z

SliceSampling.jl documentation for PR #49 is available at:
https://TuringLang.github.io/SliceSampling.jl/previews/PR49/

Red-Portal · 2026-01-17T22:45:14Z

This is really nice. Thank you very much!

penelopeysm force-pushed the py/turing branch from 17934e1 to 03943e3 Compare December 14, 2025 16:19

Update for newer AbstractMCMC/Turing interface

18fb8ec

penelopeysm force-pushed the py/turing branch from 03943e3 to 18fb8ec Compare December 14, 2025 16:22

penelopeysm commented Dec 14, 2025

View reviewed changes

penelopeysm marked this pull request as draft December 14, 2025 17:21

penelopeysm mentioned this pull request Dec 14, 2025

Implement DynamicPPL.rand_with_logdensity on LogDensityFunction TuringLang/DynamicPPL.jl#1178

Open

Remove Turing ext

9b061aa

Enforce minimum DynamicPPL bound

75e3b15

penelopeysm requested a review from Red-Portal January 17, 2026 12:24

penelopeysm marked this pull request as ready for review January 17, 2026 12:24

Red-Portal approved these changes Jan 17, 2026

View reviewed changes

Fix docs build

ed9decb

penelopeysm requested a review from Red-Portal January 17, 2026 22:35

Red-Portal merged commit ab02eb5 into main Jan 17, 2026
8 checks passed

Red-Portal deleted the py/turing branch January 17, 2026 22:45

Update for newer AbstractMCMC/Turing interface #49

Update for newer AbstractMCMC/Turing interface #49

Uh oh!

Conversation

penelopeysm commented Dec 14, 2025

Uh oh!

penelopeysm Dec 14, 2025

Choose a reason for hiding this comment

Uh oh!

penelopeysm Dec 14, 2025

Choose a reason for hiding this comment

Uh oh!

penelopeysm Dec 14, 2025

Choose a reason for hiding this comment

Uh oh!

penelopeysm commented Dec 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Red-Portal commented Dec 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Red-Portal commented Jan 15, 2026

Uh oh!

penelopeysm commented Jan 15, 2026

Uh oh!

penelopeysm commented Jan 17, 2026

Uh oh!

penelopeysm commented Jan 17, 2026

Uh oh!

Red-Portal left a comment

Choose a reason for hiding this comment

Uh oh!

penelopeysm commented Jan 17, 2026

Uh oh!

Red-Portal commented Jan 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

penelopeysm commented Jan 17, 2026

Uh oh!

github-actions bot commented Jan 17, 2026

Uh oh!

Uh oh!

Red-Portal commented Jan 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

penelopeysm commented Dec 14, 2025 •

edited

Loading

Red-Portal commented Dec 14, 2025 •

edited

Loading

Red-Portal commented Jan 17, 2026 •

edited

Loading