diff --git a/REQUIRE b/REQUIRE
index 25bf9de45..fa0ae4dc6 100644
--- a/REQUIRE
+++ b/REQUIRE
@@ -1,6 +1,7 @@
 julia 0.4
 AutoGrad 0.0.7
 Compat 0.16.0         # for @views
+Requires
 
 # We need these for some examples. They get automatically installed when needed.
 # ArgParse
diff --git a/src/Knet.jl b/src/Knet.jl
index 3b0fa575c..ebc5dabce 100644
--- a/src/Knet.jl
+++ b/src/Knet.jl
@@ -39,7 +39,7 @@ dir(path...) = joinpath(dirname(dirname(@__FILE__)),path...)
 
 
 # See if we have a gpu at initialization:
-function __init__()
+@init begin
     try
         r = gpu(true)
         # info(r >= 0 ? "Knet using GPU $r" : "No GPU found, Knet using the CPU")
diff --git a/src/karray.jl b/src/karray.jl
index 56884991e..4fbee7b39 100644
--- a/src/karray.jl
+++ b/src/karray.jl
@@ -42,7 +42,7 @@ operations.
 
 * Reduction operators: countnz, maximum, minimum, prod, sum, sumabs,
   sumabs2, vecnorm.
-    
+
 * Linear algebra: (*), axpy!, permutedims (up to 5D), transpose
 
 * Knet extras: relu, sigm, invx, logp, logsumexp, conv4, pool,
@@ -345,12 +345,12 @@ summary(a::KnetDisplay) = summary(a.a)
 summary(a::KnetArray) = string(Base.dims2string(size(a)), " ", typeof(a))
 display(a::KnetArray) = display(KnetDisplay(a))
 
-# Hack for JLD file load/save of KnetArrays:
-if Pkg.installed("JLD") != nothing
-    import JLD: writeas, readas
+using Requires
+
+@require JLD begin
     type KnetJLD; a::Array; end
-    writeas(c::KnetArray) = KnetJLD(Array(c))
-    readas(d::KnetJLD) = (gpu() >= 0 ? KnetArray(d.a) : d.a)
+    JLD.writeas(c::KnetArray) = KnetJLD(Array(c))
+    JLD.readas(d::KnetJLD) = (gpu() >= 0 ? KnetArray(d.a) : d.a)
 end
 
 # Array/KnetArray Transfer