Part 1 of autograd backend: DifferentiationInterface extention

[yebai]

No description provided.

[github-actions]

AbstractPPL.jl documentation for PR #158 is available at:
https://TuringLang.github.io/AbstractPPL.jl/previews/PR158/

[codecov]

Codecov Report

❌ Patch coverage is 5.74713% with 82 lines in your changes missing coverage. Please review.
✅ Project coverage is 77.53%. Comparing base (ceec39b) to head (17ebe0e).

Files with missing lines	Patch %	Lines
ext/AbstractPPLTestExt.jl	0.00%	46 Missing ⚠️
ext/AbstractPPLDifferentiationInterfaceExt.jl	0.00%	36 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (ceec39b) and HEAD (17ebe0e). Click for more details.

HEAD has 1 upload less than BASE

Flag	BASE (ceec39b)	HEAD (17ebe0e)
	6	5

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #158      +/-   ##
==========================================
- Coverage   85.10%   77.53%   -7.58%     
==========================================
  Files          13       14       +1     
  Lines         705      779      +74     
==========================================
+ Hits          600      604       +4     
- Misses        105      175      +70     

☔ 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.

[yebai]

@shravanngoswamii, can you help take a look at this PR as you reviewed #157?

[yebai]

Thanks, @shravanngoswamii, for the review—I'll try to address them later.

[yebai]

Addressed all review feedback in 5797c83:

• value_and_gradient!! / value_and_jacobian!! now run a _check_vector_length before the Val(0) short-circuit and the DI call, so a wrong-length x (including a non-empty x against a length-0 Prepared) raises DimensionMismatch matching prepared(x).
• AutoReverseDiff(; compile=true) Jacobian path: unified gradient/jacobian DI prep into _prepare_di(prep, ...); the no-context carve-out now applies to both, restoring tape recording for vector outputs.
• Docs wording tweak applied.

Cleanup

• Tightened the gradient of vector-valued output edge case from Exception → ArgumentError.
• Documented the prep-time evaluator(x) call and reserved Val{:vector}/Val{:edge} testcase keys.

EDIT: the DynamicPPL failures are unrelated.

[shravanngoswamii]

I think value_and_gradient!! / value_and_jacobian!! still bypass the other VectorEvaluator validation for integer vectors.

julia> using AbstractPPL

julia> using ADTypes: AutoForwardDiff

julia> using DifferentiationInterface

julia> using ForwardDiff

julia> p = AbstractPPL.prepare(AutoForwardDiff(), x -> sum(abs2, x), zeros(3));

julia> p([1, 2, 3])
ERROR: ArgumentError: VectorEvaluator requires a vector of floating-point values, but received an `Vector{Int64}`. Convert to a floating-point vector (e.g. `Float64.(x)`) before calling.
Stacktrace:
 [1] _reject_integer_input(x::Vector{Int64})
   @ AbstractPPL.Evaluators ~/Work/vectorly-ai/AbstractPPL.jl/src/evaluators/Evaluators.jl:167
 [2] (::AbstractPPL.Evaluators.VectorEvaluator{true, var"#2#3"})(x::Vector{Int64})
   @ AbstractPPL.Evaluators ~/Work/vectorly-ai/AbstractPPL.jl/src/evaluators/Evaluators.jl:182
 [3] (::AbstractPPL.Evaluators.Prepared{…})(x::Vector{…})
   @ AbstractPPL.Evaluators ~/Work/vectorly-ai/AbstractPPL.jl/src/evaluators/Evaluators.jl:41
 [4] top-level scope
   @ REPL[6]:1
Some type information was truncated. Use `show(err)` to see complete types.

julia> 

But the derivative call goes into DI and gives a backend preparation error:

julia> AbstractPPL.value_and_gradient!!(p, [1, 2, 3])
ERROR: PreparationMismatchError (inconsistent types between preparation and execution):
  - f: ✅
  - backend: ✅
  - x: ❌
    - prep: Vector{Float64}
    - exec: Vector{Int64}
  - contexts: ✅
If you are confident that this check is superfluous, you can disable it by running preparation with the keyword argument `strict=Val(false)` inside DifferentiationInterface.

Stacktrace:
 [1] check_prep(f::Function, ::DifferentiationInterfaceForwardDiffExt.ForwardDiffGradientPrep{…}, backend::AutoForwardDiff{…}, x::Vector{…}, contexts::Constant{…})
   @ DifferentiationInterface ~/.julia/packages/DifferentiationInterface/IS0Dg/src/utils/prep.jl:177
 [2] value_and_gradient(f::typeof(AbstractPPLDifferentiationInterfaceExt._call_evaluator), prep::DifferentiationInterfaceForwardDiffExt.ForwardDiffGradientPrep{…}, backend::AutoForwardDiff{…}, x::Vector{…}, contexts::Constant{…})
   @ DifferentiationInterfaceForwardDiffExt ~/.julia/packages/DifferentiationInterface/IS0Dg/ext/DifferentiationInterfaceForwardDiffExt/onearg.jl:411
 [3] value_and_gradient!!(p::AbstractPPL.Evaluators.Prepared{…}, x::Vector{…})
   @ AbstractPPLDifferentiationInterfaceExt ~/Work/vectorly-ai/AbstractPPL.jl/ext/AbstractPPLDifferentiationInterfaceExt.jl:74
 [4] top-level scope
   @ REPL[7]:1
Some type information was truncated. Use `show(err)` to see complete types.

julia> 

Same thing happens for value_and_jacobian!! on a vector-valued function.

Can we reuse the same input validation as VectorEvaluator{true} before calling DI, not only the length check?

[yebai]

Good catch.

[shravanngoswamii]

Just a personal opinion: I don't think those emojis as error looks good.

[yebai]

These are from DifferentiationInterface, not AbstractPPL.