diff --git a/examples/eg1.jl b/examples/eg1.jl
index 0aa7b5c..ef6c897 100644
--- a/examples/eg1.jl
+++ b/examples/eg1.jl
@@ -9,7 +9,7 @@ let
     !(tstpath in LOAD_PATH) && push!(LOAD_PATH, tstpath)
 end
 
-using Lux
+using Lux, ComponentArrays
 using LuxDeviceUtils, CUDA, LuxCUDA
 using BenchmarkTools
 
@@ -20,14 +20,28 @@ device = Lux.gpu_device()
 function main()
     x = rand32(rng, 1, 1000) |> device
 
-    mlp = Chain(Dense(1, 32), Dense(32, 32), Dense(32, 1),)
-    kan = Chain(KDense(1, 8, 15), KDense(8, 8, 15), KDense(8, 1, 15))
+    mlp = Chain(
+        Dense(1, 32, tanh),
+        Dense(32, 32, tanh),
+        Dense(32, 1),
+    )
 
-    pM, stM = Lux.setup(rng, mlp) |> device
-    pK, stK = Lux.setup(rng, kan) |> device
+    kan = Chain(
+        KDense(1, 8, 15; use_base_act = false),
+        KDense(8, 8, 15; use_base_act = false),
+        KDense(8, 1, 15; use_base_act = false),
+    )
 
-    display(mlp)
-    display(kan)
+    # display(mlp)
+    # display(kan)
+
+    pM, stM = Lux.setup(rng, mlp)
+    pK, stK = Lux.setup(rng, kan)
+
+    pM = ComponentArray(pM) |> device
+    pK = ComponentArray(pK) |> device
+
+    stM, stK = device(stM), device(stK)
 
     @btime CUDA.@sync $mlp($x, $pM, $stM)
     @btime CUDA.@sync $kan($x, $pK, $stK)
@@ -37,5 +51,4 @@ end
 
 main()
 
-
 nothing
diff --git a/src/type.jl b/src/type.jl
index 4692aae..c33ddb5 100644
--- a/src/type.jl
+++ b/src/type.jl
@@ -2,7 +2,7 @@
 #======================================================#
 # Kolmogorov-Arnold Layer
 #======================================================#
-@concrete struct KDense <: LuxCore.AbstractExplicitLayer
+@concrete struct KDense{use_base_act} <: LuxCore.AbstractExplicitLayer
     in_dims::Int
     out_dims::Int
     grid_len::Int
@@ -11,7 +11,6 @@
     base_act
     init_W1
     init_W2
-    use_base_activation
 end
 
 function KDense(
@@ -22,26 +21,31 @@ function KDense(
     base_act = silu,
     init_W1 = glorot_uniform,
     init_W2 = glorot_uniform,
-    use_base_activation = true,
-    use_fast_activation::Bool = true,
+    use_base_act = true,
+    use_fast_act::Bool = true,
 )
-    normalizer = use_fast_activation ? tanh_fast : tanh
+    normalizer = use_fast_act ? tanh_fast : tanh
+    base_act = use_fast_act ? NNlib.fast_act(base_act) : base_act
 
-    KDense(
+    KDense{use_base_act}(
         in_dims, out_dims, grid_len,
         denominator, normalizer, base_act,
-        init_W1, init_W2, use_base_activation,
+        init_W1, init_W2,
     )
 end
 
-function LuxCore.initialparameters(rng::AbstractRNG, l::KDense)
+function LuxCore.initialparameters(
+    rng::AbstractRNG,
+    l::KDense{use_base_act}
+) where{use_base_act}
     p = (;
         W1 = l.init_W1(rng, l.out_dims, l.grid_len * l.in_dims),
     )
 
-    if l.use_base_activation
+    if use_base_act
         p = (;
-            p..., W2 = l.init_W2(rng, l.out_dims, l.in_dims),
+            p...,
+            W2 = l.init_W2(rng, l.out_dims, l.in_dims),
         )
     end
 
@@ -54,30 +58,43 @@ function LuxCore.initialstates(::AbstractRNG, l::KDense,)
     (; grid,)
 end
 
-LuxCore.statelength(l::KDense) = l.grid_len
-function LuxCore.parameterlength(l::KDense)
+function LuxCore.statelength(l::KDense)
+    l.grid_len
+end
+
+function LuxCore.parameterlength(
+    l::KDense{use_base_act},
+) where{use_base_act}
     len = l.in_dims * l.grid_len * l.out_dims
-    if l.use_base_activation
+    if use_base_act
         len += l.in_dims * l.out_dims
     end
 
     len
 end
 
-function (l::KDense)(x::AbstractVecOrMat, p, st)
+function (l::KDense{false})(x::AbstractVecOrMat, p, st)
+    K = size(x, 2)                # [I, K]
+    x_resh = reshape(x, 1, :)     # [1, I * K]
+    x_norm = l.normalizer(x_resh) # ∈ [-1, 1]
+
+    basis = rbf.(x_norm, st.grid, l.denominator)      # [G, I * K]
+    basis = reshape(basis, l.grid_len * l.in_dims, K) # [G * I, K]
+    y = p.W1 * basis                                  # [O, K]
+
+    y, st
+end 
+
+function (l::KDense{true})(x::AbstractVecOrMat, p, st)
     K = size(x, 2)                # [I, K]
     x_resh = reshape(x, 1, :)     # [1, I * K]
     x_norm = l.normalizer(x_resh) # ∈ [-1, 1]
 
-    basis = rbf.(x_norm, st.grid, l.denominator)       # [G, I * K]
+    basis = rbf.(x_norm, st.grid, l.denominator)      # [G, I * K]
     basis = reshape(basis, l.grid_len * l.in_dims, K) # [G * I, K]
     spline = p.W1 * basis                             # [O, K]
 
-    y = if l.use_base_activation
-        spline + p.W2 * l.base_act.(x)
-    else
-        spline
-    end
+    y = spline + p.W2 * l.base_act.(x)
 
     y, st
 end