Fix a Gibbs bug with models where linking changes variable dimension (#2472)

* Add a test that catches a bug in Gibbs

* Bump patch version to 0.36.1

* Bump DPPL compat to 0.34.1

* Expand comment

Author: mhauru

Project.toml

@@ -1,6 +1,6 @@
 name = "Turing"
 uuid = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
-version = "0.36.1"
+version = "0.36.2"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
@@ -63,7 +63,7 @@ Distributions = "0.23.3, 0.24, 0.25"
 DistributionsAD = "0.6"
 DocStringExtensions = "0.8, 0.9"
 DynamicHMC = "3.4"
-DynamicPPL = "0.33, 0.34"
+DynamicPPL = "0.34.1"
 EllipticalSliceSampling = "0.5, 1, 2"
 ForwardDiff = "0.10.3"
 Libtask = "0.8.8"

test/mcmc/gibbs.jl

@@ -759,6 +759,26 @@ end
             check_numerical(chain, [:m1, :m2], [-0.2, 0.6]; atol=0.1)
         end
     end
+
+    # Test a model that where the sampler needs to link a variable, which consequently
+    # changes dimension. This used to error because the initial value `VarInfo`,
+    # obtained from just `VarInfo(model)`, had a value of dimension 2 for `w`, and the one
+    # coming out of the initial step of the component sampler had a dimension of 1, since
+    # the latter was linked. `merge` of the varinfos couldn't handle that before DPPL
+    # 0.34.1.
+    @testset "linking changes dimension" begin
+        @model function dirichlet_model()
+            K = 2
+            w ~ Dirichlet(K, 1.0)
+            for i in 1:K
+                0.1 ~ Normal(w[i], 1.0)
+            end
+        end
+
+        model = dirichlet_model()
+        sampler = Gibbs(:w => HMC(0.05, 10))
+        @test (sample(model, sampler, 10); true)
+    end
 end
 
 end