diff --git a/test/alloc_tests.jl b/test/alloc_tests.jl
index 50d09b54..d67a1691 100644
--- a/test/alloc_tests.jl
+++ b/test/alloc_tests.jl
@@ -7,6 +7,8 @@ using DiffEqGPU
 
 # Note: AllocCheck.@check_allocs is not compatible with GPU kernels and complex
 # dispatch, so we use @allocated instead for testing allocation counts.
+# @allocated must be evaluated inside a function: measured at top-level scope,
+# older Julia versions also count boxing the measured call's result.
 
 # Test Lorenz system for allocation tests
 function lorenz(u, p, t)
@@ -26,11 +28,11 @@ end
         t = 0.0f0
 
         # Warmup
-        lorenz(u, p, t)
+        measure() = @allocated lorenz(u, p, t)
+        measure()
 
         # Test allocations
-        allocs = @allocated lorenz(u, p, t)
-        @test allocs == 0
+        @test measure() == 0
     end
 
     @testset "make_prob_compatible should be low allocation" begin
@@ -42,12 +44,12 @@ end
         prob = ODEProblem{false}(lorenz, u0, tspan, p)
 
         # Warmup
-        DiffEqGPU.make_prob_compatible(prob)
+        measure() = @allocated DiffEqGPU.make_prob_compatible(prob)
+        measure()
 
         # Test - some allocations are expected for problem conversion
-        allocs = @allocated DiffEqGPU.make_prob_compatible(prob)
         # Should be reasonably low (less than 1KB)
-        @test allocs < 1024
+        @test measure() < 1024
     end
 
     @testset "diffeqgpunorm should not allocate for SVector" begin
@@ -55,11 +57,11 @@ end
         t = 0.0f0
 
         # Warmup
-        DiffEqGPU.diffeqgpunorm(u, t)
+        measure() = @allocated DiffEqGPU.diffeqgpunorm(u, t)
+        measure()
 
         # Test allocations
-        allocs = @allocated DiffEqGPU.diffeqgpunorm(u, t)
-        @test allocs == 0
+        @test measure() == 0
     end
 
     @testset "diffeqgpunorm should not allocate for scalars" begin
@@ -67,10 +69,10 @@ end
         t = 0.0f0
 
         # Warmup
-        DiffEqGPU.diffeqgpunorm(u, t)
+        measure() = @allocated DiffEqGPU.diffeqgpunorm(u, t)
+        measure()
 
         # Test allocations
-        allocs = @allocated DiffEqGPU.diffeqgpunorm(u, t)
-        @test allocs == 0
+        @test measure() == 0
     end
 end
diff --git a/test/runtests.jl b/test/runtests.jl
index e26fe6b0..ac7e777f 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -136,6 +136,9 @@ if GROUP == "CPU"
     @time @safetestset "Adaptive endpoint termination" begin
         include("gpu_kernel_de/adaptive_endpoint.jl")
     end
+    @time @safetestset "Allocation Tests" begin
+        include("alloc_tests.jl")
+    end
 end
 
 if GROUP == "JLArrays"
@@ -168,10 +171,3 @@ if GROUP == "CUDA"
         include("enzyme_environment.jl")
     end
 end
-
-# Allocation tests run separately to avoid precompilation interference
-if GROUP == "all" || GROUP == "nopre"
-    @time @safetestset "Allocation Tests" begin
-        include("alloc_tests.jl")
-    end
-end