diff --git a/lib/MadNLPGPU/Project.toml b/lib/MadNLPGPU/Project.toml
index 9d99df38..212febaa 100644
--- a/lib/MadNLPGPU/Project.toml
+++ b/lib/MadNLPGPU/Project.toml
@@ -10,7 +10,7 @@ LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MadNLP = "2621e9c9-9eb4-46b1-8089-e8c72242dfb6"
 
 [compat]
-CUDA = "~4"
+CUDA = "4, 5"
 CUSOLVERRF = "0.2"
 KernelAbstractions = "0.9"
 MadNLP = "0.7"
diff --git a/lib/MadNLPGPU/test/densekkt_gpu.jl b/lib/MadNLPGPU/test/densekkt_gpu.jl
index 9c806360..7ea3ff55 100644
--- a/lib/MadNLPGPU/test/densekkt_gpu.jl
+++ b/lib/MadNLPGPU/test/densekkt_gpu.jl
@@ -3,7 +3,6 @@ using CUDA
 using MadNLPTests
 
 function _compare_gpu_with_cpu(KKTSystem, n, m, ind_fixed)
-
     for (T,tol,atol) in [
         (Float32,1e-3,1e-1),
         (Float64,1e-8,1e-6)
@@ -16,13 +15,9 @@ function _compare_gpu_with_cpu(KKTSystem, n, m, ind_fixed)
             :tol=>tol
         )
 
+        # Host evaluator
         nlph = MadNLPTests.DenseDummyQP(zeros(T,n); m=m, fixed_variables=ind_fixed)
-
-        # Some weird issue: there's some non-deterministic behavior in generating the model for the first call
-        # Not sure where this error is originating, but seems to be resolved in v1.10
-        # Here, we call this twice to avoid this error
-        nlpd = MadNLPTests.DenseDummyQP(CUDA.zeros(T,n); m=m, fixed_variables=CuArray(ind_fixed))
-        
+        # Device evaluator
         nlpd = MadNLPTests.DenseDummyQP(CUDA.zeros(T,n); m=m, fixed_variables=CuArray(ind_fixed))
 
         # Solve on CPU
diff --git a/lib/MadNLPTests/src/Instances/dummy_qp.jl b/lib/MadNLPTests/src/Instances/dummy_qp.jl
index 0c3db1cb..27894d70 100644
--- a/lib/MadNLPTests/src/Instances/dummy_qp.jl
+++ b/lib/MadNLPTests/src/Instances/dummy_qp.jl
@@ -25,7 +25,7 @@ function NLPModels.hess_structure!(qp::DenseDummyQP, I::AbstractVector{T}, J::Ab
     copyto!(J, qp.hcols)
 end
 
-function NLPModels.obj(qp::DenseDummyQP, x::AbstractVector)
+function NLPModels.obj(qp::DenseDummyQP{T}, x::AbstractVector{T}) where T
     mul!(qp.buffer, qp.P, x)
     return 0.5 * dot(x, qp.buffer) + dot(qp.q, x)
 end
@@ -84,9 +84,15 @@ function DenseDummyQP(
     end
 
     Random.seed!(1)
+    # Generate random values.
+    # N.B.: we need to allocate the matrix P_ right after the vector
+    # q_ if we want to obtain a deterministic behavior: the seed is not working
+    # if we call the function `randn` after allocating vectors on the device.
+    q_ = randn(n)
+    P_ = randn(n, n)
 
     y0 = fill!(similar(x0, m), zero(T))
-    q = copyto!(similar(x0, n), randn(n))
+    q = copyto!(similar(x0, n), q_)
     buffer = similar(x0, n)
 
     # Bound constraints
@@ -100,7 +106,7 @@ function DenseDummyQP(
     xl[fixed_variables] .= @view(xu[fixed_variables])
 
     # Build QP problem 0.5 * x' * P * x + q' * x
-    P = copyto!(similar(x0, n , n), randn(n,n))
+    P = copyto!(similar(x0, n , n), P_)
     P = P*P' # P is symmetric
     P += T(100.0) * I
 
@@ -109,10 +115,6 @@ function DenseDummyQP(
     A = fill!(similar(x0, m, n), zero(T))
     A[1:m+1:m^2] .= one(T)
     A[m+1:m+1:m^2+m] .=-one(T)
-    # for j in 1:m
-    #     A[j, j]  = one(T)
-    #     A[j, j+1]  = -one(T)
-    # end
 
     nnzh = div(n * (n + 1), 2)
     hrows = copyto!(similar(x0, Int, nnzh), [i for i in 1:n for j in 1:i])