varinfo: store evaluation outputs only in accumulators

[yebai]

This replaces the value-owning VarInfo with an accumulator collection and removes the separate OnlyAccsVarInfo name. VarInfo() collects log densities without recording parameters. For example, VarInfo(VectorValueAccumulator()) opts into recording parameter values during evaluation; retrieve them with get_vector_values(vi) instead of vi.values. The VarInfo(rng, model) convenience constructor continues to collect both parameters and densities. This changes the public storage API.

Closes #1376.

[codecov]

Codecov Report

❌ Patch coverage is 86.70886% with 21 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.34%. Comparing base (13dd27f) to head (5e60233).

Files with missing lines	Patch %	Lines
src/varinfo.jl	82.81%	11 Missing ⚠️
src/accumulators/priors.jl	25.00%	3 Missing ⚠️
src/test_utils/model_interface.jl	0.00%	3 Missing ⚠️
src/accumulators/vnt.jl	66.66%	2 Missing ⚠️
src/threadsafe.jl	91.30%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1500      +/-   ##
==========================================
+ Coverage   84.09%   84.34%   +0.25%     
==========================================
  Files          54       53       -1     
  Lines        4741     4638     -103     
==========================================
- Hits         3987     3912      -75     
+ Misses        754      726      -28     

☔ View full report in Codecov by Harness.
📢 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.

[github-actions]

DynamicPPL.jl documentation for PR #1500 is available at:
https://TuringLang.github.io/DynamicPPL.jl/previews/PR1500/

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[github-actions]

Benchmarks @ 5e60233

Performance Ratio: gradient time divided by log-density time.

For very small models these ratios are noisy across runs and machines; raw primal and gradient timings are more reliable. The benchmarks are aimed at DynamicPPL developers and mainly catch obvious allocation or type-stability regressions. See benchmark notes for details.

===================================================================================================
                                               eval                       gradient                 
                                            ----------  -------------------------------------------
Model                        dim    linked      primal     FwdDiff    RvsDiff    Mooncake    Enzyme
---------------------------------------------------------------------------------------------------
Simple assume observe*         1     false     4.37 ns       10.69    1602.21       13.04      2.75
Simple assume observe*         1      true     4.38 ns       10.81    1657.82       13.10      2.74
Smorgasbord                  201     false     5.05 μs       76.07     143.21        6.55      3.91
Smorgasbord                  201      true      6.1 μs       75.81     156.71        6.77      2.26
Loop univariate 1k          1000     false     12.2 μs     1127.49     377.91        4.63      1.30
Loop univariate 1k          1000      true     18.1 μs     1279.04     264.35        3.18      0.82
Multivariate 1k             1000     false     1.51 μs     1357.40     605.74        1.61      5.48
Multivariate 1k             1000      true     1.48 μs     1485.69     639.14        1.72      5.62
Loop univariate 10k        10000     false    112.0 μs    31951.38     468.34        4.78      1.10
Loop univariate 10k        10000      true    167.0 μs    26870.14     323.16        3.25      0.76
Multivariate 10k           10000     false     15.8 μs    25756.63     588.85        1.87      4.03
Multivariate 10k           10000      true     15.3 μs    25078.29     634.44        1.76      4.19
Dynamic                      err     false      1.4 μs         err      68.82       19.63      7.29
Dynamic                       10      true     1.78 μs        1.87      75.43       16.72      9.45
Submodel*                      1     false     4.38 ns       11.51    1879.12       13.30     10.86
Submodel*                      1      true     4.38 ns       11.83    1914.98       13.08     10.73
LDA                            6      true      3.9 μs        1.36      14.17       39.54     41.21
===================================================================================================

Main @ 13dd27f

===================================================================================================
                                               eval                       gradient                 
                                            ----------  -------------------------------------------
Model                        dim    linked      primal     FwdDiff    RvsDiff    Mooncake    Enzyme
---------------------------------------------------------------------------------------------------
Simple assume observe*         1     false     5.56 ns        9.11    1326.76       10.75      8.21
Simple assume observe*         1      true     5.56 ns        9.55    1286.77       10.19      7.56
Smorgasbord                  201     false     5.39 μs       68.76     146.27        5.86      2.88
Smorgasbord                  201      true     6.44 μs       72.01     169.36        6.10      2.09
Loop univariate 1k          1000     false     16.9 μs      818.43     325.64        2.98      0.76
Loop univariate 1k          1000      true     18.3 μs     1316.48     303.45        2.80      0.73
Multivariate 1k             1000     false    906.0 ns     1612.44    1102.63        1.95      6.99
Multivariate 1k             1000      true    953.0 ns     1569.78    1092.17        2.03      7.46
Loop univariate 10k        10000     false    163.0 μs    22189.92     360.27        2.99      0.73
Loop univariate 10k        10000      true    178.0 μs    23844.82     334.28        2.84      0.68
Multivariate 10k           10000     false     8.56 μs    28293.08    1172.16        1.70      5.99
Multivariate 10k           10000      true     8.97 μs    26282.83    1148.59        1.67      6.13
Dynamic                       15     false     1.29 μs         err      60.13       15.10      5.92
Dynamic                       10      true     1.84 μs        1.61      63.34       12.69      6.91
Submodel*                      1     false     5.57 ns       10.06    1572.16       10.09      8.06
Submodel*                      1      true     5.57 ns        9.26    1496.11       10.73      8.04
LDA                            6      true     4.19 μs        1.38      11.19       31.54     28.94
===================================================================================================

Environment

Julia Version 1.13.1
Commit 96ca370cf0e (2026-09-25 19:34 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 4 × INTEL(R) XEON(R) PLATINUM 8573C
  WORD_SIZE: 64
  LLVM: libLLVM-20.1.8 (ORCJIT, emeraldrapids)
  GC: Built with stock GC
Threads: 1 default, 1 interactive, 1 GC (on 4 virtual cores)

[sunxd3]

thanks for this. in general, the PR looks right to me. I think the only sampler in Turing still uses varinfo's value is EMCEE, which should be an easy migration.

[yebai]

Thanks, Xianda. Both addressed on traces. A missing accumulator now throws an ArgumentError naming it (c85a691). On LinkAll(): the root cause was get_transform_strategy(vi) reconstructing the strategy from outputs, which #1469 rules out. I removed it along with link!!(vi, vns, model) (e654fbc); pass the strategy to init!! instead, e.g. LinkSome(Set(vns), UnlinkAll()). Inference also returns UnlinkAll() for empty values (f00ab3d). Agreed on Emcee.

[sunxd3]

thanks!

On LinkAll(): the root cause was get_transform_strategy(vi) reconstructing the strategy from outputs, which #1469 rules out. I removed it along with link!!(vi, vns, model) (e654fbc); pass the strategy to init!! instead, e.g. LinkSome(Set(vns), UnlinkAll()).

makes sense, I notices the removal of partial links, but didn't connect to this.

[sunxd3]

Should we merge to a preparation branch? I don't mind merging to main, just a reminder.
I misunderstood the stacked PR feature, pls ignore this comment