DynamicPPL 0.36

[penelopeysm]

This PR adds compatibility with DynamicPPL 0.36.

Given that there are some reasonably breaking changes in the way that submodels behave, I've decided to increment the minor version number for Turing.

Because submodels now also prefix VarNames in a 'proper' way, submodel variables now have non-identity lenses. In order to allow Gibbs to sample models with submodels, this PR also fixes #2403 by letting GibbsContext store a tuple of VarNames rather than just Symbols. Performance is compared below.

There's some nasty lack of reproducibility in HMCDA sampling, which is causing the failures in test/mcmc/Inference. I haven't yet been able to repro this locally. However, I would like to deal with this in a separate PR, as the failure isn't introduced by this PR (the last CI run on main was already failing).

This PR supersedes the CompatHelper ones at #2501 #2532 #2533

Benchmarks

I'm not 100% convinced of these, because my benchmarks have returned wildly variable results upon being run multiple times / when I restart my Julia session. Sometimes when I run sample(rng, model, spl, 10000) it runs in < 1 second and sometimes it takes 4 seconds. (This is after running it multiple times in the same session, so it's not an issue of JIT compilation.) I think I would probably feel a little bit more reassured if somebody else could test them out.

However, broadly speaking, I think it's safe to say that I don't have any evidence of performance getting worse. Another noteworthy point is that CI for test/mcmc/gibbs (on ubuntu-latest, 1 thread) for this PR ran in 1 hr 25 min, which is the same time it took for the same job on main. So it seems unlikely that there are any serious performance implications (good or bad).

using AbstractMCMC: step
using DynamicPPL, Turing
using StableRNGs: StableRNG
using Chairmarks

@model function f()
    x ~ Normal()
    y ~ Normal()
end
rng = StableRNG(468)
model = f()
spl = Gibbs(@varname(x) => MH(), @varname(y) => MH())
dspl = DynamicPPL.Sampler(spl)

# This benchmark compares the initial step of MCMC
@be step($rng, $model, $dspl)

# Before: 91.250 μs
# After: 91.208 µs

# This compares the 2nd step onwards
_, st = step(rng, model, dspl)
@be step($rng, $model, $dspl, $st)

# Before: 66.500 μs
# After: 66.417 μs

[codecov]

Codecov Report

Attention: Patch coverage is 88.88889% with 2 lines in your changes missing coverage. Please review.

Project coverage is 84.27%. Comparing base (3901096) to head (daf8088).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/mcmc/Inference.jl	0.00%	1 Missing ⚠️
src/mcmc/gibbs.jl	94.11%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2535      +/-   ##
==========================================
+ Coverage   84.05%   84.27%   +0.21%     
==========================================
  Files          21       21              
  Lines        1455     1456       +1     
==========================================
+ Hits         1223     1227       +4     
+ Misses        232      229       -3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[coveralls]

Pull Request Test Coverage Report for Build 14910188111

Details

• 16 of 18 (88.89%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.2%) to 84.272%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/mcmc/Inference.jl	0	1	0.0%
src/mcmc/gibbs.jl	16	17	94.12%

Totals
Change from base Build 14887164727:	0.2%
Covered Lines:	1227
Relevant Lines:	1456

---

💛 - Coveralls

[mhauru]

This is great! Leaving it unapproved just because I want to confirm that nothing bad has happened to type stability and hence runtimes.

Might it also make sense to throw in some non-trivially lensed varnames in the test models of the "type stability" and "Sampler call order" testsets in test/mcmc/gibbs.jl?

[github-actions]

Turing.jl documentation for PR #2535 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2535/