Invlink if linked=false

Author: penelopeysm

src/test_utils/ad.jl

@@ -4,7 +4,7 @@ using ADTypes: AbstractADType, AutoForwardDiff
 using Chairmarks: @be
 import DifferentiationInterface as DI
 using DocStringExtensions
-using DynamicPPL: Model, LogDensityFunction, VarInfo, AbstractVarInfo, link
+using DynamicPPL: Model, LogDensityFunction, VarInfo, AbstractVarInfo, link, invlink
 using LogDensityProblems: logdensity, logdensity_and_gradient
 using Random: Random, Xoshiro
 using Statistics: median
@@ -184,6 +184,8 @@ function run_ad(
 )::ADResult
     if linked
         varinfo = link(varinfo, model)
+    else
+        varinfo = invlink(varinfo, model)
     end
     if isnothing(params)
         params = varinfo[:]

test/ad.jl

@@ -23,7 +23,6 @@ using DynamicPPL: LogDensityFunction
             varinfos = DynamicPPL.TestUtils.setup_varinfos(m, rand_param_values, vns)
 
             @testset "$(short_varinfo_name(varinfo))" for varinfo in varinfos
-                # TODO: This runs unlinked. Should we test linked as well?
                 f = LogDensityFunction(m, varinfo)
                 x = DynamicPPL.getparams(f)
                 # Calculate reference logp + gradient of logp using ForwardDiff
@@ -57,14 +56,14 @@ using DynamicPPL: LogDensityFunction
                             ref_ldf, adtype
                         )
                     else
-                        # TODO: Should we test linked as well?
-                        DynamicPPL.TestUtils.AD.run_ad(
+                        # TODO: test linked as well
+                        @test DynamicPPL.TestUtils.AD.run_ad(
                             m,
                             adtype;
                             linked=false,
                             varinfo=varinfo,
                             expected_value_and_grad=(ref_logp, ref_grad),
-                        )
+                        ) isa Any
                     end
                 end
             end