diff --git a/src/Flux.jl b/src/Flux.jl
index 0d78024b..dac36973 100644
--- a/src/Flux.jl
+++ b/src/Flux.jl
@@ -37,6 +37,6 @@ include("layers/normalise.jl")
 
 include("data/Data.jl")
 
-@require CuArrays include("cuda/cuda.jl")
+@init @require CuArrays="3a865a2d-5b23-5a0f-bc46-62713ec82fae" include("cuda/cuda.jl")
 
 end # module
diff --git a/src/onehot.jl b/src/onehot.jl
index 07206dfe..f19fa6f8 100644
--- a/src/onehot.jl
+++ b/src/onehot.jl
@@ -32,7 +32,7 @@ import Adapt.adapt
 
 adapt(T, xs::OneHotMatrix) = OneHotMatrix(xs.height, adapt(T, xs.data))
 
-@require CuArrays begin
+@init @require CuArrays="3a865a2d-5b23-5a0f-bc46-62713ec82fae" begin
   import CuArrays: CuArray, cudaconvert
   Base.Broadcast._containertype(::Type{<:OneHotMatrix{<:CuArray}}) = CuArray
   cudaconvert(x::OneHotMatrix{<:CuArray}) = OneHotMatrix(x.height, cudaconvert(x.data))
diff --git a/src/treelike.jl b/src/treelike.jl
index 13e562e6..e65ac41a 100644
--- a/src/treelike.jl
+++ b/src/treelike.jl
@@ -53,7 +53,7 @@ cpu(m) = mapleaves(x -> adapt(Array, x), m)
 
 gpu_adaptor = identity
 
-@require CuArrays begin
+@init @require CuArrays="3a865a2d-5b23-5a0f-bc46-62713ec82fae" begin
   global gpu_adaptor = CuArrays.cu
 end