Switch to using ADTypes for AD choice

[jClugstor]

Checklist

• Appropriate tests were added
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

This will switch over to using ADTypes for the autodiff keyword in the construction of the solver types.

• Changes the default autodiff keyword to be AutoForwardDiff() instead of true.
• Adds a deprecation path so that using true or false for autodiff uses AutoForwardDiff() and AutoFiniteDiff() respectively.
• Also ensures that if the chunk_size and diff_type kwargs are used with a Bool autodiff, they are respected.

[jClugstor]

@oscardssmith It's able to precompile in this state, so at least anything in the precompile workloads will work. I'm pretty sure some tests won't pass though, and I still need to add some tests as well.

[jClugstor]

@oscardssmith All CI tests are passing now. Some integration tests are broken.

[oscardssmith]

The StochasticDiffEq error looks very odd. @gdalle any idea what would cause this?

 In-place methods: Test Failed at /home/runner/work/OrdinaryDiffEq.jl/OrdinaryDiffEq.jl/downstream/test/complex_tests.jl:42
  Expression: solve(prob, alg)
    Expected: ArgumentError
      Thrown: DivideError
      DivideError: integer division error
      Stacktrace:
        [1] div
          @ ./int.jl:295 [inlined]
        [2] div
          @ ./div.jl:310 [inlined]
        [3] div
          @ ./div.jl:365 [inlined]
        [4] cld
          @ ./div.jl:322 [inlined]
        [5] generate_chunked_partials(x::Matrix{ComplexF64}, colorvec::UnitRange{Int64}, cs::Val{0})
          @ SparseDiffTools ~/.julia/packages/SparseDiffTools/m4eae/src/differentiation/compute_jacobian_ad.jl:80
        [6] SparseDiffTools.ForwardColorJacCache(f::SciMLBase.UJacobianWrapper{true, SciMLBase.SDEFunction{true, SciMLBase.FullSpecialize, Main.var"##Complex Number Tests#233".var"#f#2", Main.var"##Complex Number Tests#233".var"#f#2", LinearAlgebra.UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, typeof(SciMLBase.DEFAULT_OBSERVED), Nothing, Nothing, Nothing}, Float64, SciMLBase.NullParameters}, x::Matrix{ComplexF64}, _chunksize::Val{0}; dx::Nothing, tag::Type, colorvec::UnitRange{Int64}, sparsity::Nothing)
          @ SparseDiffTools ~/.julia/packages/SparseDiffTools/m4eae/src/differentiation/compute_jacobian_ad.jl:37
        [7] ForwardColorJacCache
          @ ~/.julia/packages/SparseDiffTools/m4eae/src/differentiation/compute_jacobian_ad.jl:22 [inlined]

[gdalle]

It appears you have a chunksize of 0 somewhere. Maybe an ill-advised choice in AutoForwardDiff(; chunksize=...)? The default is nothing and not 0.

[jClugstor]

Yeah, my guess is that it has something to do with the chunksize,

[jClugstor]

I think it's because the fallback _get_fwd_chunksize(AD) = Val(0). Since the Stochastic algs aren't using ADTypes yet, it falls back to that inside of get_chunksize.

[gdalle]

If anything, that fallback should be Val(1). Unless we're explicitly trying to throw an error by using zero.

[oscardssmith]

What's happening here is that we were using Val(0) to mean automatic chunksize construction. Does DI have a good way to ask for that?

[gdalle]

That's what AutoForwardDiff(; chunksize=nothing) does. Do you need to manipulate chunk sizes for other backends? if so, why?

[jClugstor]

Actually, I think I'm wrong, for a Stochastic algorithm, it should be getting the chunksize from one of these.

OrdinaryDiffEq.get_chunksize(alg::StochasticDiffEqNewtonAlgorithm{CS,AD,FDT,ST,CJ,Controller}) where {CS,AD,FDT,ST,CJ,Controller} = Val(CS)
OrdinaryDiffEq.get_chunksize(alg::StochasticDiffEqNewtonAdaptiveAlgorithm{CS,AD,FDT,ST,CJ,Controller}) where {CS,AD,FDT,ST,CJ,Controller} = Val(CS)
OrdinaryDiffEq.get_chunksize(alg::StochasticDiffEqJumpNewtonAdaptiveAlgorithm{CS,AD,FDT,ST,CJ,Controller}) where {CS,AD,FDT,ST,CJ,Controller} = Val(CS)
OrdinaryDiffEq.get_chunksize(alg::StochasticDiffEqJumpNewtonDiffusionAdaptiveAlgorithm{CS,AD,FDT,ST,CJ,Controller}) where {CS,AD,FDT,ST,CJ,Controller} = Val(CS)

[gdalle]

Note that depending on the semantics of chunk size in these libraries (which I don't know), setting nothing as the default may not work either. All I can comment on is ADTypes and DifferentiationInterface:

• In ADTypes, Auto(Polyester)ForwardDiff is the only backend with a configurable chunk size, and it can be either nothing or an integer > 0
• In DI, I leverage the choice of chunk size in AutoForwardDiff and set it to 1 for every other backend (except Enzyme which has a custom rule). Chunk size handling is internal right now because I'm not yet sure I have nailed the right API.

[gdalle]

To get a clearer idea of which tests are actually failing and which tests are just Codecov complaining about your uploads, you can temporarily set these settings to false:

OrdinaryDiffEq.jl/.github/workflows/CI.yml

Line 97 in ce43d57

fail_ci_if_error: true

OrdinaryDiffEq.jl/.github/workflows/Downstream.yml

Line 80 in ce43d57

fail_ci_if_error: true

[jClugstor]

Oh thanks, that's helpful.

[jClugstor]

@oscardssmith all CI tests pass, and all of the failing integration tests are failing on master as well.

[oscardssmith]

looks good! Thanks so much

[gdalle]

This Codecov throttling is really frustrating

[ChrisRackauckas]

Yeah and we need to delete it repo by repo for ones that don't use the high level CI description 😱 . At least there aren't too many repos left that need to do it.

[jClugstor]

I believe all the tests that are failing here are also failing on master.

[jClugstor]

@ChrisRackauckas @oscardssmith I think this is ready

[ChrisRackauckas]

Can you make sure to do the downstream in StochasticDiffEq and DelayDiffEq?

[jClugstor]

Yes, will do.

[gdalle]

hot damn, this is a great christmas gift! thanks for all your hard work