Update EnzymeCoreExt.jl (#2565)

The inactive_noinl failed in broadcast assignment when using Enzyme=0.13.16 and CUDA=5.5.2. Switching to the version that prevents inlining seem to have fixed it.

Co-authored-by: William Moses <gh@wsmoses.com>

Author: simenhu

ext/EnzymeCoreExt.jl

@@ -25,7 +25,7 @@ end
 function EnzymeCore.EnzymeRules.inactive_noinl(::typeof(CUDA.is_pinned), args...)
     return nothing
 end
-function EnzymeCore.EnzymeRules.inactive_noinl(::typeof(CUDA.launch_configuration), args...; kwargs...)
+function EnzymeCore.EnzymeRules.inactive(::typeof(CUDA.launch_configuration), args...; kwargs...)
     return nothing
 end
 

test/extensions/enzyme.jl

@@ -145,6 +145,16 @@ end
     @test all(dx .≈ 1)
     @test all(dy .≈ 1)
 end
+
+sumabs2(x) = sum(abs2.(x))
+
+@testset "Reverse sum abs2" begin
+    x = CuArray([1.0, 2.0, 3.0, 4.0])
+    dx = CuArray([0., 0.0, 0.0, 0.0])
+    Enzyme.autodiff(Reverse, sumabs2, Active, Duplicated(x, dx))
+    @test all(dx .≈ 2 .* x)
+end
+
 # TODO once reverse kernels are in
 # function togpu(x)
 #     x = CuArray(x)