CompatHelper: bump compat for Bijectors to 0.12, (keep existing compat)

[github-actions]

This pull request changes the compat entry for the Bijectors package from 0.8, 0.9, 0.10 to 0.8, 0.9, 0.10, 0.12.
This keeps the compat entries for earlier versions.

Note: I have not tested your package with this new compat entry.
It is your responsibility to make sure that your package tests pass before you merge this pull request.

[coveralls]

Pull Request Test Coverage Report for Build 4088997384

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at 0.0%

---

Totals
Change from base Build 4047530053:	0.0%
Covered Lines:	0
Relevant Lines:	1422

---

💛 - Coveralls

[codecov]

Codecov Report

Patch and project coverage have no change.

Comparison is base (35a1280) 0.00% compared to head (762fda9) 0.00%.

❗ Current head 762fda9 differs from pull request most recent head 8fc2f56. Consider uploading reports for the commit 8fc2f56 to get more accurate results

Additional details and impacted files

@@          Coverage Diff           @@
##           master   #1946   +/-   ##
======================================
  Coverage    0.00%   0.00%           
======================================
  Files          21      21           
  Lines        1422    1422           
======================================
  Misses       1422    1422           

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[ven-k]

New patch of Distributions breaks Bijector@0.10 as GLOBAL_RNG is now default_rng (diff)

So, Turing fails to precompile as Bijector's compat entry is "0.8, 0.9, 0.10".

Merging this PR should fix it.

[devmotion]

This PR can't be merged currently, Bijectors 0.12 (and also 0.11) is untested and had quite a few breaking changes.

[devmotion]

New patch of Distributions breaks Bijector@0.10 as GLOBAL_RNG is now default_rng (diff)

Why does that break Bijectors?

[ven-k]

Oh I see. I'll pin Distributions to 0.25.80 for now then.

Why does that break Bijectors?

This definition:

https://github.com/TuringLang/Bijectors.jl/blob/6f05a87647a05ab71c1602dffbb6d28d89dcd57e/src/transformed_distribution.jl#L218

Repro and error stack

(bijector) pkg> add Bijectors@0.10
(bijector) pkg> precompile
Precompiling environment...
  ✗ Bijectors
  0 dependencies successfully precompiled in 5 seconds. 52 already precompiled.

ERROR: The following 1 direct dependency failed to precompile:

Bijectors [76274a88-744f-5084-9051-94815aaf08c4]

Failed to precompile Bijectors [76274a88-744f-5084-9051-94815aaf08c4] to "/home/ub/.julia/compiled/v1.9/Bijectors/jl_63meTE".
ERROR: LoadError: UndefVarError: `GLOBAL_RNG` not defined
Stacktrace:
  [1] getproperty(x::Module, f::Symbol)
    @ Base ./Base.jl:31
  [2] top-level scope
    @ ~/.julia/packages/Bijectors/vUc4m/src/transformed_distribution.jl:218
  [3] include(mod::Module, _path::String)
    @ Base ./Base.jl:457
  [4] include(x::String)
    @ Bijectors ~/.julia/packages/Bijectors/vUc4m/src/Bijectors.jl:1
  [5] top-level scope
    @ ~/.julia/packages/Bijectors/vUc4m/src/interface.jl:172
  [6] include(mod::Module, _path::String)
    @ Base ./Base.jl:457
  [7] include(x::String)
    @ Bijectors ~/.julia/packages/Bijectors/vUc4m/src/Bijectors.jl:1
  [8] top-level scope
    @ ~/.julia/packages/Bijectors/vUc4m/src/Bijectors.jl:253
  [9] include
    @ ./Base.jl:457 [inlined]
 [10] include_package_for_output(pkg::Base.PkgId, input::String, depot_path::Vector{String}, dl_load_path::Vector{String}, load_path::Vector{String}, concrete_deps::Vector{Pair{Base.PkgId, UInt128}}, source::Nothing)
    @ Base ./loading.jl:1995
 [11] top-level scope
    @ stdin:2
in expression starting at /home/ub/.julia/packages/Bijectors/vUc4m/src/transformed_distribution.jl:218
in expression starting at /home/ub/.julia/packages/Bijectors/vUc4m/src/interface.jl:172
in expression starting at /home/ub/.julia/packages/Bijectors/vUc4m/src/Bijectors.jl:1
in expression starting at stdin:2

[devmotion]

Well, this code is.... bad. TuringLang/Bijectors.jl#248 should fix it.

[torfjelde]

Well, this code is.... bad. TuringLang/Bijectors.jl#248 should fix it.

Oh gee wow. Yeah that's from when I first started out with Julia 😅

[sourish-cmi]

So we are using Turing extensively in CRRao.jl - looks like it breaks our package. All tests are failing now. Our development is stuck.

[devmotion]

Can you explain what exactly is broken? The problem mentioned above was fixed right away, within a few hours after it had been brought to our attention. It should not occur with the latest release of Bijectors 0.10.x.

[sourish-cmi]

Can you explain what exactly is broken? The problem mentioned above was fixed right away, within a few hours after it had been brought to our attention. It should not occur with the latest release of Bijectors 0.10.x.

Thanks for quick reply. We were trying to update the ReadMe file and check of PRs failing as the log shows the Bayesian tests are failing due to Turing (https://github.com/xKDR/CRRao.jl/pull/105/checks)

[devmotion]

Based on these logs I'm very certain that there is no problem with Turing, and in particular not with Bijectors (the problem discussed above is indeed fixed and does not occur anymore). I'm also very certain that the test failures are caused by the recent switch in Distributions from Random.GLOBAL_RNG to Random.default_rng() (which should be used instead nowadays) as the default RNG. IMO this switch was correctly included in a non-breaking release since not even Julia itself does not promise reproducibility of the default RNG across different (non-breaking) versions and architectures. So I would suggest changing your tests, e.g., by using StableRNG.jl if you want guaranteed reproducibility or by relaxing your tests, increasing the number of samples to reduce variance, or, in case you know the true solution or moments of it, by measuring the distance of the samples and possibly performing hypothesis testing.

[sourish-cmi]

Based on these logs I'm very certain that there is no problem with Turing, and in particular not with Bijectors (the problem discussed above is indeed fixed and does not occur anymore). I'm also very certain that the test failures are caused by the recent switch in Distributions from Random.GLOBAL_RNG to Random.default_rng() (which should be used instead nowadays) as the default RNG. IMO this switch was correctly included in a non-breaking release since not even Julia itself does not promise reproducibility of the default RNG across different (non-breaking) versions and architectures. So I would suggest changing your tests, e.g., by using StableRNG.jl if you want guaranteed reproducibility or by relaxing your tests, increasing the number of samples to reduce variance, or, in case you know the true solution or moments of it, by measuring the distance of the samples and possibly performing hypothesis testing.

okay thanks for your reply. We will do the same and we will keep you posted.

[yebai]

This PR likely requires some further changes to support Bijectors@v0.12. See, e.g. #1945

cc @torfjelde let's try to push this through as a priority.

[yebai]

Closed by #1979