diff --git a/ext/RecursiveArrayToolsCUDAExt.jl b/ext/RecursiveArrayToolsCUDAExt.jl
index 950bf709..cc52a1de 100644
--- a/ext/RecursiveArrayToolsCUDAExt.jl
+++ b/ext/RecursiveArrayToolsCUDAExt.jl
@@ -1,6 +1,6 @@
 module RecursiveArrayToolsCUDAExt
 
-using RecursiveArrayTools: AbstractVectorOfArray
+using RecursiveArrayTools: AbstractVectorOfArray, ArrayPartition
 import CUDA: CuArray
 
 # Disambiguate CuArray(::AbstractVectorOfArray) vs CuArray(::AbstractArray{T,N}) from CUDA.jl.
@@ -10,4 +10,12 @@ function CuArray(VA::AbstractVectorOfArray{T, N}) where {T, N}
     return CuArray{T, N}(stack(VA.u))
 end
 
+# Assemble each state from its GPU partitions before stacking the states.
+# Stacking ArrayPartitions directly reads their elements on the CPU.
+function CuArray(
+        VA::AbstractVectorOfArray{T, N, <:AbstractVector{<:ArrayPartition}}
+    ) where {T, N}
+    return CuArray{T, N}(stack(map(ap -> vcat(vec.(ap.x)...), VA.u)))
+end
+
 end
diff --git a/test/GPU/vectorofarray_gpu.jl b/test/GPU/vectorofarray_gpu.jl
index b5bf351f..a14a68b2 100644
--- a/test/GPU/vectorofarray_gpu.jl
+++ b/test/GPU/vectorofarray_gpu.jl
@@ -52,3 +52,20 @@ _b = Adapt.adapt(CuArray, b)
 @test _b isa DiffEqArray
 @test _b.u isa Vector{<:CuArray}
 @test _b.t isa CuArray
+
+# Conversion of partitioned solution states must stay on the GPU.
+states = [
+    ArrayPartition(CUDA.fill(1.0f0, 2), CUDA.fill(2.0f0, 3)),
+    ArrayPartition(CUDA.fill(3.0f0, 2), CUDA.fill(4.0f0, 3)),
+]
+partitioned_sol = DiffEqArray(states, Float32[0, 1])
+partitioned_cu = CuArray(partitioned_sol)
+@test partitioned_cu isa CuArray
+@test size(partitioned_cu) == (5, 2)
+@test Array(partitioned_cu) == Float32[1 3; 1 3; 2 4; 2 4; 2 4]
+
+matrix_states = [
+    ArrayPartition(CUDA.fill(1.0f0, 2, 2), CUDA.fill(2.0f0, 1)),
+    ArrayPartition(CUDA.fill(3.0f0, 2, 2), CUDA.fill(4.0f0, 1)),
+]
+@test Array(CuArray(VectorOfArray(matrix_states))) == Float32[1 3; 1 3; 1 3; 1 3; 2 4]