CompatHelper: bump compat for Bijectors to 0.13, (keep existing compat)

[github-actions]

This pull request changes the compat entry for the Bijectors package from 0.12 to 0.12, 0.13.
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.

[github-actions]

Pull Request Test Coverage Report for Build 5317209593

• 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 5313671228:	0.0%
Covered Lines:	0
Relevant Lines:	1427

---

💛 - Coveralls

[codecov]

Codecov Report

Patch and project coverage have no change.

Comparison is base (1b67694) 0.00% compared to head (e42eec9) 0.00%.

❗ Current head e42eec9 differs from pull request most recent head bc815c0. Consider uploading reports for the commit bc815c0 to get more accurate results

Additional details and impacted files

@@          Coverage Diff           @@
##           master   #2018   +/-   ##
======================================
  Coverage    0.00%   0.00%           
======================================
  Files          22      22           
  Lines        1473    1473           
======================================
  Misses       1473    1473           

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

[yebai]

@harisorgn It seems some fixes are probably required here as well to support new LKJ priors.

[harisorgn]

@harisorgn It seems some fixes are probably required here as well to support new LKJ priors.

Sorry, promised to have a look but couldn't make it before starting my holidays. If someone else wants to take this over, please feel free. Otherwise I'll try to fix tests by the end of next week : )

[yebai]

A gentle reminder @harisorgn

[yebai]

#2040 fixed some formatting issues for CI results. No functionality change was added.

[torfjelde]

TuringLang/DynamicPPL.jl#499 should address the failing tests

[harisorgn]

TuringLang/DynamicPPL.jl#499 should address the failing tests

thanks @torfjelde . Found something that probably still needs to be addressed. It's the change to fill! in triu_mask in Bijectors.jl . Basically fill!(x::Union{UpperTriangular, LowerTriangular}, some_value) complains that it can not access matrix indices that do not belong to the relevant triangular. Not sure if this is expected behaviour or a LinearAlgebra bug, seems like the latter to me, will open an issue there.

We can revert to the old way without fill! (you recently changed it IIRC).

[torfjelde]

We can revert to the old way without fill! (you recently changed it IIRC).

Maybe easier to just call parent on the input, no?

[harisorgn]

Maybe easier to just call parent on the input, no?

Yes, see, that's why I mentioned the problem before trying to fix it. Your way is easier (edit: and better overall), thanks 😅

[yebai]

TuringLang/DynamicPPL.jl#503 and #2049 should fix all remaining tests.

EDIT: there is still some issue with Bijectors.

[yebai]

After #2049, we have only two AD-related errors remaining. They are likely related to recent Bijectors changes, see e.g., #2049 (comment)

[torfjelde]

I've made a fix for ReverseDiff in TuringLang/Bijectors.jl#280, but this won't work for Tracker since it does not support ChainRulesCore.

[torfjelde]

I believe after TuringLang/Bijectors.jl#280 and TuringLang/DynamicPPL.jl#513, CI here should be fairly likely to pass

[torfjelde]

Btw, are we bothered about Tracker.jl tests failing, given that we're now warning the user that Tracker isn't officially supported anymore?

[devmotion]

I guess in principle it's fine to have worse support for Tracker - but then I think we should change the default reverse-mode backend to e.g. ReverseDiff. However, in this case here it seems the fix in Bijectors can be extended to Tracker without much effort by using the same pullbacks?

[torfjelde]

Once Bijectors has been released, I'll bump the compat version of the tests and give this another go:)

[yebai]

Thanks, @torfjelde @harisorgn -- please feel free to merge if CI passes.

[torfjelde]

A bit confused as to why we're getting complaints still; we shouldn't even be hitting cholesky with TrackedArray anymore.

[yebai]

@torfjelde It looks like a Julia 1.7 issue. The following tests run successfully on Julia 1.9, but fails on Julia 1.7:

julia> using Turing, LinearAlgebra, ReverseDiff

julia> @model function wishart()
           theta ~ Wishart(4, Matrix{Float64}(I, 4, 4))
       end
wishart (generic function with 2 methods)

julia> Turing.setadbackend(:reversediff)
:reversediff

julia> sample(wishart(), HMC(0.01, 1), 1000)

[torfjelde]

Hmm, need to check if there's a discrepancy between 1.6 and 1.7 then, since the tests were passing on 1.6 with Bijectors

[yebai]

It might be caused by other parts of Turing, e.g. DynamicPPL. CI tests for Bijectors pass for all Julia releases betwee 1.6-1.9. See TuringLang/Bijectors.jl#284

[yebai]

@torfjelde All tests for ReverseDiff pass, but this revealed one failure case for Zygote. Can you take a look?

julia> using Turing, LinearAlgebra, Zygote

julia> @model function wishart()
           theta ~ Wishart(4, Matrix{Float64}(I, 4, 4))
       end
wishart (generic function with 2 methods)

julia> Turing.setadbackend(:zygote)
:zygote

julia> sample(wishart(), HMC(0.01, 1), 1000)

[coveralls]

Pull Request Test Coverage Report for Build 5882947478

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

• For more information on this, see Tracking coverage changes with pull request builds.
• To avoid this issue with future PRs, see these Recommended CI Configurations.
• For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

• 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 5860942594:	0.0%
Covered Lines:	0
Relevant Lines:	1473

---

💛 - Coveralls

[torfjelde]

Thanks for finishing this off Hong!

[harisorgn]

Thanks for finishing this off Hong!

Seconded 🙏