Test with Tapir

src/mcmc/abstractmcmc.jl

@@ -40,11 +40,17 @@ getstats(transition::AdvancedHMC.Transition) = transition.stat
 getparams(::DynamicPPL.Model, transition::AdvancedMH.Transition) = transition.params
 
 getvarinfo(f::DynamicPPL.LogDensityFunction) = f.varinfo
-getvarinfo(f::LogDensityProblemsAD.ADGradientWrapper) = getvarinfo(parent(f))
+function getvarinfo(f::LogDensityProblemsAD.ADGradientWrapper)
+    return getvarinfo(LogDensityProblemsAD.parent(f))
+end
 
 setvarinfo(f::DynamicPPL.LogDensityFunction, varinfo) = Accessors.@set f.varinfo = varinfo
-function setvarinfo(f::LogDensityProblemsAD.ADGradientWrapper, varinfo)
-    return Accessors.@set f.ℓ = setvarinfo(f.ℓ, varinfo)
+function setvarinfo(
+    f::LogDensityProblemsAD.ADGradientWrapper, varinfo, adtype::ADTypes.AbstractADType
+)
+    return LogDensityProblemsAD.ADgradient(
+        adtype, setvarinfo(LogDensityProblemsAD.parent(f), varinfo)
+    )
 end
 
 """
@@ -120,7 +126,7 @@ function AbstractMCMC.step(
             varinfo = DynamicPPL.link(varinfo, model)
         end
     end
-    f = setvarinfo(f, varinfo)
+    f = setvarinfo(f, varinfo, alg.adtype)
 
     # Then just call `AdvancedHMC.step` with the right arguments.
     if initial_state === nothing

test/mcmc/Inference.jl

@@ -2,6 +2,7 @@ module InferenceTests
 
 using ..Models: gdemo_d, gdemo_default
 using ..NumericalTests: check_gdemo, check_numerical
+import ..ADUtils
 using Distributions: Bernoulli, Beta, InverseGamma, Normal
 using Distributions: sample
 import DynamicPPL
@@ -14,7 +15,9 @@ import ReverseDiff
 using Test: @test, @test_throws, @testset
 using Turing
 
-@testset "Testing inference.jl with $adbackend" for adbackend in (AutoForwardDiff(; chunksize=0), AutoReverseDiff(; compile=false))
+ADUtils.install_tapir && import Tapir
+
+@testset "Testing inference.jl with $adbackend" for adbackend in ADUtils.adbackends
     # Only test threading if 1.3+.
     if VERSION > v"1.2"
         @testset "threaded sampling" begin

test/mcmc/abstractmcmc.jl

@@ -1,5 +1,6 @@
 module AbstractMCMCTests
 
+import ..ADUtils
 using AdvancedMH: AdvancedMH
 using Distributions: sample
 using Distributions.FillArrays: Zeros
@@ -15,14 +16,18 @@ using Test: @test, @test_throws, @testset
 using Turing
 using Turing.Inference: AdvancedHMC
 
+ADUtils.install_tapir && import Tapir
+
 function initialize_nuts(model::Turing.Model)
     # Create a log-density function with an implementation of the
     # gradient so we ensure that we're using the same AD backend as in Turing.
     f = LogDensityProblemsAD.ADgradient(DynamicPPL.LogDensityFunction(model))
 
     # Link the varinfo.
     f = Turing.Inference.setvarinfo(
-        f, DynamicPPL.link!!(Turing.Inference.getvarinfo(f), model)
+        f,
+        DynamicPPL.link!!(Turing.Inference.getvarinfo(f), model),
+        Turing.Inference.getADType(DynamicPPL.getcontext(LogDensityProblemsAD.parent(f))),
     )
 
     # Choose parameter dimensionality and initial parameter value
@@ -112,7 +117,6 @@ end
 
 @testset "External samplers" begin
     @testset "AdvancedHMC.jl" begin
-        # Try a few different AD backends.
         @testset "adtype=$adtype" for adtype in [AutoForwardDiff(), AutoReverseDiff()]
             @testset "$(model.f)" for model in DynamicPPL.TestUtils.DEMO_MODELS
                 # Need some functionality to initialize the sampler.

test/mcmc/gibbs.jl

@@ -2,6 +2,7 @@ module GibbsTests
 
 using ..Models: MoGtest_default, gdemo, gdemo_default
 using ..NumericalTests: check_MoGtest_default, check_gdemo, check_numerical
+import ..ADUtils
 using Distributions: InverseGamma, Normal
 using Distributions: sample
 using ForwardDiff: ForwardDiff
@@ -12,9 +13,9 @@ using Turing
 using Turing: Inference
 using Turing.RandomMeasures: ChineseRestaurantProcess, DirichletProcess
 
-@testset "Testing gibbs.jl with $adbackend" for adbackend in (
-    AutoForwardDiff(; chunksize=0), AutoReverseDiff(; compile=false)
-)
+ADUtils.install_tapir && import Tapir
+
+@testset "Testing gibbs.jl with $adbackend" for adbackend in ADUtils.adbackends
     @testset "gibbs constructor" begin
         N = 500
         s1 = Gibbs(HMC(0.1, 5, :s, :m; adtype=adbackend))

test/mcmc/gibbs_conditional.jl

@@ -2,6 +2,7 @@ module GibbsConditionalTests
 
 using ..Models: gdemo, gdemo_default
 using ..NumericalTests: check_gdemo, check_numerical
+import ..ADUtils
 using Clustering: Clustering
 using Distributions: Categorical, InverseGamma, Normal, sample
 using ForwardDiff: ForwardDiff
@@ -14,9 +15,9 @@ using StatsFuns: StatsFuns
 using Test: @test, @testset
 using Turing
 
-@testset "Testing gibbs conditionals.jl with $adbackend" for adbackend in (
-    AutoForwardDiff(; chunksize=0), AutoReverseDiff(; compile=false)
-)
+ADUtils.install_tapir && import Tapir
+
+@testset "Testing gibbs conditionals.jl with $adbackend" for adbackend in ADUtils.adbackends
     Random.seed!(1000)
     rng = StableRNG(123)
 

test/mcmc/hmc.jl

@@ -3,6 +3,7 @@ module HMCTests
 using ..Models: gdemo_default
 #using ..Models: gdemo
 using ..NumericalTests: check_gdemo, check_numerical
+import ..ADUtils
 using Distributions: Bernoulli, Beta, Categorical, Dirichlet, Normal, Wishart, sample
 import DynamicPPL
 using DynamicPPL: Sampler
@@ -16,7 +17,9 @@ using StatsFuns: logistic
 using Test: @test, @test_logs, @testset
 using Turing
 
-@testset "Testing hmc.jl with $adbackend" for adbackend in (AutoForwardDiff(; chunksize=0), AutoReverseDiff(; compile=false))
+ADUtils.install_tapir && import Tapir
+
+@testset "Testing hmc.jl with $adbackend" for adbackend in ADUtils.adbackends
     # Set a seed
     rng = StableRNG(123)
     @testset "constrained bounded" begin

test/mcmc/sghmc.jl

@@ -2,6 +2,7 @@ module SGHMCTests
 
 using ..Models: gdemo_default
 using ..NumericalTests: check_gdemo
+import ..ADUtils
 using Distributions: sample
 import ForwardDiff
 using LinearAlgebra: dot
@@ -10,7 +11,9 @@ using StableRNGs: StableRNG
 using Test: @test, @testset
 using Turing
 
-@testset "Testing sghmc.jl with $adbackend" for adbackend in (AutoForwardDiff(; chunksize=0), AutoReverseDiff(; compile=false))
+ADUtils.install_tapir && import Tapir
+
+@testset "Testing sghmc.jl with $adbackend" for adbackend in ADUtils.adbackends
     @testset "sghmc constructor" begin
         alg = SGHMC(; learning_rate=0.01, momentum_decay=0.1, adtype=adbackend)
         @test alg isa SGHMC
@@ -36,7 +39,7 @@ using Turing
     end
 end
 
-@testset "Testing sgld.jl with $adbackend" for adbackend in (AutoForwardDiff(; chunksize=0), AutoReverseDiff(; compile=false))
+@testset "Testing sgld.jl with $adbackend" for adbackend in ADUtils.adbackends
     @testset "sgld constructor" begin
         alg = SGLD(; stepsize=PolynomialStepsize(0.25), adtype=adbackend)
         @test alg isa SGLD

test/runtests.jl

@@ -7,6 +7,7 @@ import Turing
 
 include(pkgdir(Turing) * "/test/test_utils/models.jl")
 include(pkgdir(Turing) * "/test/test_utils/numerical_tests.jl")
+include(pkgdir(Turing) * "/test/test_utils/ad_utils.jl")
 
 Turing.setprogress!(false)
 

test/test_utils/ad_utils.jl

@@ -0,0 +1,21 @@
+module ADUtils
+
+import Pkg
+import Turing
+
+"""
+All the ADTypes on which we want to run the tests.
+"""
+adbackends = [
+    Turing.AutoForwardDiff(; chunksize=0), Turing.AutoReverseDiff(; compile=false)
+]
+
+# Tapir isn't supported for older Julia versions, hence the check.
+install_tapir = isdefined(Turing, :AutoTapir)
+if install_tapir
+    # TODO(mhauru) Is there a better way to install optional dependencies like this?
+    Pkg.add("Tapir")
+    push!(adbackends, Turing.AutoTapir(false))
+end
+
+end