diff --git a/crates/specialized/rtx-cfd/src/solvers/incompressible/mod.rs b/crates/specialized/rtx-cfd/src/solvers/incompressible/mod.rs
index ada95cf..5819242 100644
--- a/crates/specialized/rtx-cfd/src/solvers/incompressible/mod.rs
+++ b/crates/specialized/rtx-cfd/src/solvers/incompressible/mod.rs
@@ -4,9 +4,6 @@
 //! - SIMPLE (Semi-Implicit Method for Pressure Linked Equations)
 //! - PISO (Pressure-Implicit with Splitting of Operators)
 //! - SIMPLER (SIMPLE Revised)
-// PERF-2 (2026-09-16): the legacy GPU modules have not compiled since the
-// initial commit (cudarc API drift); kept out of the `cuda` build until
-// someone needs them — the batched V-cycle benchmark uses cudarc directly.
 
 use crate::{CfdConfig, CfdError, CfdResult};
 // use nalgebra::{DMatrix, DVector};
@@ -29,16 +26,16 @@ pub mod overset;
 /// PISO algorithm implementation
 pub mod piso;
 /// GPU-accelerated PISO algorithm implementation
-// #[cfg(feature = "cuda")]
-// pub mod piso_gpu;
+#[cfg(feature = "cuda")]
+pub mod piso_gpu;
 /// Five-point Poisson problems and the multigrid-preconditioned CG solver
 pub mod poisson;
 pub mod polygon_sdf;
 /// SIMPLE algorithm implementation
 pub mod simple;
 /// GPU-accelerated SIMPLE algorithm implementation
-// #[cfg(feature = "cuda")]
-// pub mod simple_gpu;
+#[cfg(feature = "cuda")]
+pub mod simple_gpu;
 /// CSR matrix + Jacobi-BiCGSTAB for the curvilinear pressure equation
 pub mod sparse_bicgstab;
 
@@ -65,8 +62,8 @@ pub use overset::{
     OversetResult, OversetSolverState, ResidualBucket, StepTimers,
 };
 pub use piso::{PisoParameters, PisoResult, PisoSolver};
-// #[cfg(feature = "cuda")]
-// pub use piso_gpu::PisoGpuSolver;
+#[cfg(feature = "cuda")]
+pub use piso_gpu::PisoGpuSolver;
 pub use poisson::{
     LevelExport, MgPrecision, MgSmoother, MultigridParameters, PcgCache, PoissonProblem,
     PoissonSolution, export_hierarchy, vcycle_f32_reference, vcycle_f32_work,
@@ -74,8 +71,8 @@ pub use poisson::{
 };
 pub use polygon_sdf::PolygonSdf;
 pub use simple::{ConvectionScheme, SimpleParameters, SimpleResult, SimpleSolver};
-// #[cfg(feature = "cuda")]
-// pub use simple_gpu::SimpleGpuSolver;
+#[cfg(feature = "cuda")]
+pub use simple_gpu::SimpleGpuSolver;
 
 /// Common solver parameters
 #[derive(Debug, Clone)]
diff --git a/crates/specialized/rtx-cfd/src/solvers/lbm/mod.rs b/crates/specialized/rtx-cfd/src/solvers/lbm/mod.rs
index 2d4c0db..4a74f0b 100644
--- a/crates/specialized/rtx-cfd/src/solvers/lbm/mod.rs
+++ b/crates/specialized/rtx-cfd/src/solvers/lbm/mod.rs
@@ -5,30 +5,27 @@
 //! - D3Q19: 3D model with 19 discrete velocities
 //! - Boundary conditions: bounce-back, Zou-He
 //! - Collision operators: BGK, MRT
-// PERF-2 (2026-09-16): the legacy GPU modules have not compiled since the
-// initial commit (cudarc API drift); kept out of the `cuda` build until
-// someone needs them — the batched V-cycle benchmark uses cudarc directly.
 
 pub mod boundary;
 pub mod d2q9;
 /// GPU-accelerated D2Q9 LBM solver
-// #[cfg(feature = "cuda")]
-// pub mod d2q9_gpu;
+#[cfg(feature = "cuda")]
+pub mod d2q9_gpu;
 pub mod d3q19;
 /// GPU-accelerated D3Q19 LBM solver
-// #[cfg(feature = "cuda")]
-// pub mod d3q19_gpu;
+#[cfg(feature = "cuda")]
+pub mod d3q19_gpu;
 
 pub use boundary::{
     BounceBackBc, BoundaryConditionSet, BoundaryOrientation, LbmBoundaryCondition, ZouHeBc,
     ZouHeBoundaryType,
 };
 pub use d2q9::{D2Q9Parameters, D2Q9Solver};
-// #[cfg(feature = "cuda")]
-// pub use d2q9_gpu::D2Q9GpuSolver;
+#[cfg(feature = "cuda")]
+pub use d2q9_gpu::D2Q9GpuSolver;
 pub use d3q19::{D3Q19Parameters, D3Q19Solver, MacroscopicVariables3D};
-// #[cfg(feature = "cuda")]
-// pub use d3q19_gpu::D3Q19GpuSolver;
+#[cfg(feature = "cuda")]
+pub use d3q19_gpu::D3Q19GpuSolver;
 
 /// Common LBM traits and utilities
 pub mod common {
diff --git a/crates/specialized/rtx-cfd/src/turbulence/mod.rs b/crates/specialized/rtx-cfd/src/turbulence/mod.rs
index 8e35cb4..bfac356 100644
--- a/crates/specialized/rtx-cfd/src/turbulence/mod.rs
+++ b/crates/specialized/rtx-cfd/src/turbulence/mod.rs
@@ -7,23 +7,18 @@
 //! - **Wall Functions**: Log-law, enhanced wall treatment
 //! - **Transition Models**: γ-Reθ, k-kL-ω
 
-// PERF-2 (2026-09-16): the legacy GPU modules have not compiled since the
-// initial commit (cudarc API drift); kept out of the `cuda` build until
-// someone needs them — the batched V-cycle benchmark uses cudarc directly.
 pub mod k_epsilon;
-// PERF-2 (2026-09-16): the legacy k-epsilon GPU model no longer matches
-// `KEpsilonConstants` and has not compiled since the initial commit; kept out
-// of the `cuda` build until someone needs it.
-// pub mod k_epsilon_gpu;
+#[cfg(feature = "cuda")]
+pub mod k_epsilon_gpu;
 pub mod smagorinsky;
-// #[cfg(feature = "cuda")]
-// pub mod smagorinsky_gpu;
+#[cfg(feature = "cuda")]
+pub mod smagorinsky_gpu;
 pub mod transition;
 pub mod wall_functions;
 
 pub use k_epsilon::{KEpsilonConstants, KEpsilonModel, KEpsilonVariant};
-// #[cfg(feature = "cuda")]
-// pub use k_epsilon_gpu::KEpsilonGpuModel;
+#[cfg(feature = "cuda")]
+pub use k_epsilon_gpu::KEpsilonGpuModel;
 pub use smagorinsky::{SmagorinskyConstants, SmagorinskyModel};
 pub use wall_functions::{EnhancedWallTreatment, LogLawWallFunction, WallFunction};