Enable opting out of TSVI

[penelopeysm]

Performance gains in #1113, #1125, etc. are only available when not using TSVI.

TSVI gets used more than it should (#1086).

This PR provides a mechanism to opt out of TSVI, in a way that is fully backwards-compatible.

[github-actions]

Benchmark Report for Commit 9a74692

Computer Information

Julia Version 1.11.7
Commit f2b3dbda30a (2025-09-08 12:10 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 4 × AMD EPYC 7763 64-Core Processor
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, znver3)
Threads: 1 default, 0 interactive, 1 GC (on 4 virtual cores)

Benchmark Results

┌───────────────────────┬───────┬─────────────┬───────────────────┬────────┬────────────────┬─────────────────┐
│                 Model │   Dim │  AD Backend │           VarInfo │ Linked │ t(eval)/t(ref) │ t(grad)/t(eval) │
├───────────────────────┼───────┼─────────────┼───────────────────┼────────┼────────────────┼─────────────────┤
│ Simple assume observe │     1 │ forwarddiff │             typed │  false │            6.9 │             1.7 │
│           Smorgasbord │   201 │ forwarddiff │             typed │  false │          740.1 │            41.4 │
│           Smorgasbord │   201 │ forwarddiff │ simple_namedtuple │   true │          419.7 │            53.5 │
│           Smorgasbord │   201 │ forwarddiff │           untyped │   true │          780.7 │            35.5 │
│           Smorgasbord │   201 │ forwarddiff │       simple_dict │   true │         6686.5 │            27.8 │
│           Smorgasbord │   201 │ forwarddiff │      typed_vector │   true │          743.3 │            40.2 │
│           Smorgasbord │   201 │ forwarddiff │    untyped_vector │   true │          782.7 │            35.8 │
│           Smorgasbord │   201 │ reversediff │             typed │   true │          930.3 │            43.3 │
│           Smorgasbord │   201 │    mooncake │             typed │   true │          749.0 │             5.7 │
│           Smorgasbord │   201 │      enzyme │             typed │   true │          923.4 │             3.7 │
│    Loop univariate 1k │  1000 │    mooncake │             typed │   true │         3844.5 │             5.8 │
│       Multivariate 1k │  1000 │    mooncake │             typed │   true │          985.9 │             9.0 │
│   Loop univariate 10k │ 10000 │    mooncake │             typed │   true │        42510.0 │             5.3 │
│      Multivariate 10k │ 10000 │    mooncake │             typed │   true │         8341.9 │            10.2 │
│               Dynamic │    10 │    mooncake │             typed │   true │          120.1 │            13.1 │
│              Submodel │     1 │    mooncake │             typed │   true │            9.3 │            11.2 │
│                   LDA │    12 │ reversediff │             typed │   true │          990.7 │             2.1 │
└───────────────────────┴───────┴─────────────┴───────────────────┴────────┴────────────────┴─────────────────┘

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.73%. Comparing base (7deaaab) to head (9a74692).
⚠️ Report is 1 commits behind head on py/fastinit.

Additional details and impacted files

@@               Coverage Diff               @@
##           py/fastinit    #1128      +/-   ##
===============================================
+ Coverage        81.72%   81.73%   +0.01%     
===============================================
  Files               41       41              
  Lines             3956     3959       +3     
===============================================
+ Hits              3233     3236       +3     
  Misses             723      723              

☔ 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.

[penelopeysm]

Technically, Turing overloads this function, so releasing this in a patch would cause upstream breakage. I don't feel bad about this though because this function is not exported.

On top of that, the method that Turing defines for this is bogus. It silently disables TSVI for PG/SMC, because Libtask can't handle it, and it generally doesn't work for SMC anyway because observations for all particles need to be in step. The problem is that for models that do need threadsafe eval, this will lead to silent wrong results (TuringLang/Turing.jl#2658).

The correct solution for this should be to emit an error saying:

PG/SMC can't run with models that require threadsafe eval. If you know your model doesn't need it, then you should call DynamicPPL.use_threadsafe_eval!(false).

[penelopeysm]

Given that FastLDF is going into breaking instead, we should just rebase this on breaking.

[github-actions]

DynamicPPL.jl documentation for PR #1128 is available at:
https://TuringLang.github.io/DynamicPPL.jl/previews/PR1128/