diff --git a/src/lib/broadcast.jl b/src/lib/broadcast.jl
index 8c0d3c54c..086fc6d13 100644
--- a/src/lib/broadcast.jl
+++ b/src/lib/broadcast.jl
@@ -286,3 +286,10 @@ using GPUArraysCore  # replaces @require CUDA block, weird indenting to preserve
 
   pull_block_vert(sz, Δ::AbstractGPUArray, A::Number) = @allowscalar Δ[sz]
 
+  ∇getindex(x::T, inds::Tuple{AbstractArray{<:Integer}}) where {T <: AbstractGPUArray} = dy -> begin
+    inds1_cpu = Array(inds[1])
+    dx = zeros(eltype(dy), length(x))
+    dxv = view(dx, inds1_cpu)
+    dxv .= accum.(dxv, _droplike(Array(dy), dxv))
+    return _project(x, T(dx)), nothing
+  end
diff --git a/test/cuda.jl b/test/cuda.jl
index 5cb1c8cdc..403ff553a 100644
--- a/test/cuda.jl
+++ b/test/cuda.jl
@@ -140,3 +140,9 @@ end
   @test_skip gradient((x,y) -> sum(vcat(x,y)), 1f0, r, 2f0, r)[2] isa CUDA.CuArray{Float32}
 end
 
+@testset "repeated indexing" begin
+  f(a) = sum(view(a, [1, 1, 2]))
+  a = CUDA.CuArray([1.0f0, 1.0f0, 1.0f0])
+  @test f(a) == 3.0f0
+  @test Array(gradient(f, a)[1]) == [2.0f0, 1.0f0, 0.0f0]
+end