Use AbstractPPL AD interface in tests. by yebai · Pull Request #460 · TuringLang/Bijectors.jl

yebai · 2026-04-23T23:31:09Z

Replacing #458

TODO

uncomment min / LTS tests.
remove AbstractPPL pinned sources.

github-actions · 2026-04-24T09:24:20Z

Bijectors.jl documentation for PR #460 is available at:
https://TuringLang.github.io/Bijectors.jl/previews/PR460/

yebai · 2026-05-18T14:19:14Z

 @testset "Mooncake bijector AD" begin
    for c in generate_ad_testcases(), adtype in adtypes
-        run_ad_case(c, adtype)
+        # AbstractPPL's Mooncake extension routes `AutoMooncakeForward` through


This is a Mooncake bug -- hopefully it will get fixed as part of chalk-lab/Mooncake.jl#1151

`src/vector/test_utils.jl` and `test/test_resources.jl` route all AD calls through `AbstractPPL.prepare` + `value_and_jacobian!!` / `value_and_gradient!!` in place of `DifferentiationInterface.jacobian` / `DI.gradient`. The package no longer depends on DifferentiationInterface directly; it remains a transitive test dep so AbstractPPL can dispatch to the DI extension for the `AutoForwardDiff` reference and for any other adtype that lacks a direct AbstractPPL extension. Other changes that follow from the routing switch: - `Project.toml` and the per-integration-suite `Project.toml`s bump `AbstractPPL = "0.15"` and add `ADTypes = "1"`; DI moves to a test-only dep. - `src/vector/test_utils.jl`: `test_optics` and `test_logjac` skip the AD-based check for discrete distributions — `AbstractPPL.prepare` rejects the integer vectors that `to_vec(d)` returns, and those checks are tautological for identity link transforms anyway. - `test/integration_tests/mooncake/main.jl`: PlanarLayer-inverse cases under `AutoMooncakeForward` hit Mooncake's batched NDual code path, which has no rule for `find_alpha`, so mark them `broken=true`. - `test/integration_tests/enzyme/main.jl`: the triple-nested product cases throw an Enzyme runtime `TypeError` that `@test_broken` doesn't reliably catch; switch to `skip=true` so non-AD checks still run. - `docs/src/defining_examples.md`: the StereographicProj example now uses `ForwardDiff.jacobian` directly instead of `DI.jacobian`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

yebai · 2026-05-20T16:18:20Z

@sunxd3, this should be ready for a final look.

sunxd3

sorry for nagging,

Bijectors.jl/docs/src/vector.md

Lines 110 to 123 in ff2818a

    
           ## Testing 
        
           Because the scope of a vector bijector is very well-defined, there is a well-established testing framework to verify correctness of an implementation (`Bijectors.VectorBijectors.test_all()`), which you can use in the test suite. 
        
           This function contains additional keyword arguments to control the exact testing procedure. 
        
           For example, you can test that the transformations do not cause extra allocations, should you know this to be the case for your bijector (note that this is not always possible). 
        
           For more information about generally testing bijectors (and in particular how to test Jacobians for transformations that modify the number of dimensions), see [the documentation on examples of defining bijectors](@ref bijectors-defining-examples). 
        
           One of the most tricky parts of testing Bijectors is ensuring that the transforms are compatible with automatic differentiation. 
        
           This is important for DynamicPPL: we need to be able to compute the gradient of the log-density with respect to (possibly transformed) parameters, which may include the log-abs-det-Jacobian of the transformation. 
        
           `test_all` accepts an `adtypes` keyword argument and defaults to `[AutoForwardDiff()]`, so the main test suite exercises ForwardDiff as the reference backend. 
        
           ReverseDiff, Mooncake, and Enzyme are covered by standalone integration suites under `test/integration_tests/<backend>/`, each of which passes its own `adtypes` list through to `test_all`. 
        
           It is acceptable to skip tests for a particular backend if there are genuine upstream bugs, especially with ReverseDiff, which is not actively maintained. 
        
           Where possible it is best to ensure that all backends are supported, and to use `@test_broken` to mark any known issues with specific backends.

need to be updated again on because default_adtypes is now empty

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

yebai · 2026-05-21T09:49:20Z

All addressed @sunxd3!

github-actions Bot assigned yebai Apr 23, 2026

yebai mentioned this pull request May 18, 2026

Move AD backends to standalone integration test suites #470

Merged

yebai force-pushed the replace-di-with-abstractppl branch 3 times, most recently from 876b60e to a0d363b Compare May 18, 2026 14:13

yebai commented May 18, 2026

View reviewed changes

yebai force-pushed the replace-di-with-abstractppl branch 5 times, most recently from a7a9eb1 to 4ccf99a Compare May 19, 2026 16:56

yebai force-pushed the replace-di-with-abstractppl branch from 4ccf99a to 795cb59 Compare May 19, 2026 20:37

yebai requested a review from sunxd3 May 19, 2026 21:39

Merge branch 'main' into replace-di-with-abstractppl

ff2818a

sunxd3 reviewed May 21, 2026

View reviewed changes

Comment thread docs/Project.toml

Fix vector.md docs: default_adtypes is empty

2c49e3f

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

yebai merged commit d758493 into main May 21, 2026
18 of 21 checks passed

yebai deleted the replace-di-with-abstractppl branch May 21, 2026 09:49

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use AbstractPPL AD interface in tests.#460

Use AbstractPPL AD interface in tests.#460
yebai merged 3 commits into
mainfrom
replace-di-with-abstractppl

yebai commented Apr 23, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Apr 24, 2026

Uh oh!

yebai May 18, 2026

Uh oh!

yebai commented May 20, 2026

Uh oh!

sunxd3 left a comment

Uh oh!

Uh oh!

yebai commented May 21, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	## Testing

	Because the scope of a vector bijector is very well-defined, there is a well-established testing framework to verify correctness of an implementation (`Bijectors.VectorBijectors.test_all()`), which you can use in the test suite.
	This function contains additional keyword arguments to control the exact testing procedure.
	For example, you can test that the transformations do not cause extra allocations, should you know this to be the case for your bijector (note that this is not always possible).

	For more information about generally testing bijectors (and in particular how to test Jacobians for transformations that modify the number of dimensions), see [the documentation on examples of defining bijectors](@ref bijectors-defining-examples).

	One of the most tricky parts of testing Bijectors is ensuring that the transforms are compatible with automatic differentiation.
	This is important for DynamicPPL: we need to be able to compute the gradient of the log-density with respect to (possibly transformed) parameters, which may include the log-abs-det-Jacobian of the transformation.
	`test_all` accepts an `adtypes` keyword argument and defaults to `[AutoForwardDiff()]`, so the main test suite exercises ForwardDiff as the reference backend.
	ReverseDiff, Mooncake, and Enzyme are covered by standalone integration suites under `test/integration_tests/<backend>/`, each of which passes its own `adtypes` list through to `test_all`.
	It is acceptable to skip tests for a particular backend if there are genuine upstream bugs, especially with ReverseDiff, which is not actively maintained.
	Where possible it is best to ensure that all backends are supported, and to use `@test_broken` to mark any known issues with specific backends.

Conversation

yebai commented Apr 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Apr 24, 2026

Uh oh!

yebai May 18, 2026

Choose a reason for hiding this comment

Uh oh!

yebai commented May 20, 2026

Uh oh!

sunxd3 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

yebai commented May 21, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

yebai commented Apr 23, 2026 •

edited

Loading