using Flux: Flux
using Enzyme: Enzyme

function reproducer()
    model = Flux.Chain(Flux.Dense(28^2 => 32, Flux.sigmoid), Flux.Dense(32 => 10), Flux.softmax)
    # this allocates space for the gradient
    model_shadow = Enzyme.Duplicated(model)
    input = randn(28^2, 1)
    label = [i == 3 for i in 1:10]
    grad_model = Flux.gradient(
        (m, x, y) -> sum(abs2, m(x) .- y),
        model_shadow,
        # input and label are const because we differentiate w.r.t. to the model parameters, not input or output.
        Enzyme.Const(input),
        Enzyme.Const(label),
    )
end

reproducer()

ERROR: Constant memory is stored (or returned) to a differentiable variable.
As a result, Enzyme cannot provably ensure correctness and throws this error.
This might be due to the use of a constant variable as temporary storage for active memory (https://enzyme.mit.edu/julia/stable/faq/#Runtime-Activity).
If Enzyme should be able to prove this use non-differentable, open an issue!
To work around this issue, either:
 a) rewrite this variable to not be conditionally active (fastest, but requires a code change), or
 b) set the Enzyme mode to turn on runtime activity (e.g. autodiff(set_runtime_activity(Reverse), ...) ). This will maintain correctness, but may slightly reduce performance.
Mismatched activity for:   store {} addrspace(10)* %68, {} addrspace(10)* addrspace(10)* %.repack65, align 8, !dbg !331, !tbaa !333, !alias.scope !268, !noalias !335 const val:   %68 = call fastcc nonnull {} addrspace(10)* @julia_summary_3629({} addrspace(10)* nocapture nofree noundef nonnull readonly align 16 dereferenceable(40) %1) #308, !dbg !323
Type tree: {[-1]:Pointer}
 llvalue=  %68 = call fastcc nonnull {} addrspace(10)* @julia_summary_3629({} addrspace(10)* nocapture nofree noundef nonnull readonly align 16 dereferenceable(40) %1) #308, !dbg !323

Stacktrace:
 [1] #invokelatest#2
   @ ./essentials.jl:894
 [2] invokelatest
   @ ./essentials.jl:889
 [3] macro expansion
   @ ./logging.jl:377
 [4] _match_eltype
   @ ~/.julia/packages/Flux/fteq2/src/layers/stateless.jl:60

function sum_converted(x)
    sum(convert(AbstractArray{Float32}, x))
end

function reproducer_reduced()
    input = randn(10)
    Enzyme.gradient(Enzyme.Reverse, sum_converted, input)
end
# returns `([1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],)`