Is Enzyme.jl compatible at all?

まず、issue にリンクされている GNNLux チュートリアルと issue #389 を確認し、次に、文書化されている微分バックエンドと、報告されている Enzyme および Zygote の挙動を比較します。サポートされていない互換性と関連する制限事項を文書化し、ユーザーが Enzyme の生の例外をそのまま受け取ることがないようにします。プロアクティブなエラーハンドリングを追加するかどうかは、引き続き決定が必要です。

Trying to reproduce bug 2472 on Enzyme, I kept reducing my examples until I found that the most trivial use of GCNConv results in Enzyme throwing an error about "Constant memory is stored (or returned) to a differentiable variable", and sometimes even in segfaults. Here's the example:

using Lux, GNNLux, Random, Enzyme, Optimisers

Lux.@concrete struct GraphPolicyLux <: GNNContainerLayer{(:conv1, :conv2, :dense)}
    conv1
    conv2
    dense
end

function GraphPolicyLux(nin::Int, d::Int, n_nodes::Int)
    conv1 = GCNConv(nin => d)
    conv2 = GCNConv(d => d)
    dense = Dense(d => n_nodes, sigmoid)
    return GraphPolicyLux(conv1, conv2, dense)
end

function (model::GraphPolicyLux)(g::GNNGraph, x, ps, st)
    dense = StatefulLuxLayer{true}(model.dense, ps.dense, GNNLux._getstate(st, :dense))
    x, st_c1 = model.conv1(g, x, ps.conv1, st.conv1)
    x = tanh.(x)
    x, st_c2 = model.conv2(g, x, ps.conv2, st.conv2)
    x = relu.(x)
    new_weights = dense(x)
    return new_weights, (conv1 = st_c1, conv2 = st_c2)
end


# Dummy graph construction
n = 100
D = 2
A = sprand(Float32, n, n, 0.2)
X = rand(Float32, D, n)
g = GNNGraph(A; ndata=(features = X))
# Setup model and parameters
model = GraphPolicyLux(2, 50, g.num_nodes)
ps, st = Lux.setup(Random.default_rng(), model)
train_state = Training.TrainState(model, ps, st, Adam(0.01f0))

function compute_loss(model, ps, st, (g, X))
    prop_A, st = model(g, X, ps, st)
    Reward = 0.0f0
    # This is where a simulation with A prop_A would go
    for t in 1:100
        Reward -= t * 0.01
        # Reward -= rand(Float32)
    end
    return Reward, st, 0
end

function train_trigger_bug!(model, ps, st, g, nepochs::Int=100)
    tstate = Training.TrainState(model, ps, st, Adam(0.01f0))
    data = (g, g.ndata.x)
    for epoch in 1:nepochs
        _, loss, _, tstate = Training.single_train_step!(AutoEnzyme(), compute_loss, (g, g.ndata.x), tstate)
        @show epoch
    end
end

To make sure I wasn't doing something wrong, I copy-pasted the (only) example with Lux.jl. This problem is not limited to Lux.jl, though, and the same issue occurs when using Flux.jl. By changing a single line from

 _, loss, _, train_state = Lux.Training.single_train_step!(AutoZygote(), custom_loss,(g, g.x, g.y), train_state)

to

 _, loss, _, train_state = Lux.Training.single_train_step!(AutoEnzyme(), custom_loss,(g, g.x, g.y), train_state)

I started seeing the cryptic errors again:

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)* %157, {} addrspace(10)** %sret_return.repack.repack, align 8, !dbg !234, !noalias !244 const val:   %157 = call fastcc nonnull {} addrspace(10)* @julia_vcat_89903({} addrspace(10)* noundef nonnull align 8 dereferenceable(24) %15, {} addrspace(10)* noundef nonnull align 8 dereferenceable(24) %59) #77, !dbg !228
Type tree: {[-1]:Pointer, [-1,0]:Pointer, [-1,0,-1]:Integer, [-1,8]:Pointer, [-1,8,0]:Integer, [-1,8,1]:Integer, [-1,8,2]:Integer, [-1,8,3]:Integer, [-1,8,4]:Integer, [-1,8,5]:Integer, [-1,8,6]:Integer, [-1,8,7]:Integer, [-1,8,8]:Pointer, [-1,8,8,-1]:Integer, [-1,16]:Integer, [-1,17]:Integer, [-1,18]:Integer, [-1,19]:Integer, [-1,20]:Integer, [-1,21]:Integer, [-1,22]:Integer, [-1,23]:Integer}
 llvalue=  %157 = call fastcc nonnull {} addrspace(10)* @julia_vcat_89903({} addrspace(10)* noundef nonnull align 8 dereferenceable(24) %15, {} addrspace(10)* noundef nonnull align 8 dereferenceable(24) %59) #77, !dbg !228

Stacktrace:
 [1] add_self_loops
   @ /scratch/htc/amartine/julia/packages/GNNGraphs/Ldvz4/src/transform.jl:24

Stacktrace:
  [1] add_self_loops
    @ /scratch/htc/amartine/julia/packages/GNNGraphs/Ldvz4/src/transform.jl:24
  [2] #_#10
    @ /scratch/htc/amartine/julia/packages/GNNLux/AHHiN/src/layers/conv.jl:137 [inlined]
  [3] GCNConv
    @ /scratch/htc/amartine/julia/packages/GNNLux/AHHiN/src/layers/conv.jl:131 [inlined]
  [4] GCNConv
    @ /scratch/htc/amartine/julia/packages/GNNLux/AHHiN/src/layers/conv.jl:128 [inlined]
  [5] GCN
    @ ./REPL[21]:3
  [6] custom_loss
    @ ./REPL[29]:4
  [7] #4
    @ /scratch/htc/amartine/julia/packages/Lux/ptjU6/src/helpers/training.jl:257 [inlined]
  [8] augmented_julia__4_89363_inner_33wrap
    @ /scratch/htc/amartine/julia/packages/Lux/ptjU6/src/helpers/training.jl:0
  [9] macro expansion
    @ /scratch/htc/amartine/julia/packages/Enzyme/rwbr4/src/compiler.jl:5610 [inlined]
 [10] enzyme_call
    @ /scratch/htc/amartine/julia/packages/Enzyme/rwbr4/src/compiler.jl:5144 [inlined]
 [11] AugmentedForwardThunk
    @ /scratch/htc/amartine/julia/packages/Enzyme/rwbr4/src/compiler.jl:5083 [inlined]
 [12] autodiff
    @ /scratch/htc/amartine/julia/packages/Enzyme/rwbr4/src/Enzyme.jl:408 [inlined]
 [13] compute_gradients_impl
    @ /scratch/htc/amartine/julia/packages/Lux/ptjU6/ext/LuxEnzymeExt/training.jl:10 [inlined]
 [14] compute_gradients
    @ /scratch/htc/amartine/julia/packages/Lux/ptjU6/src/helpers/training.jl:200 [inlined]
 [15] single_train_step_impl!
    @ /scratch/htc/amartine/julia/packages/Lux/ptjU6/src/helpers/training.jl:327 [inlined]
 [16] #single_train_step!#6
    @ /scratch/htc/amartine/julia/packages/Lux/ptjU6/src/helpers/training.jl:292 [inlined]
 [17] single_train_step!(backend::AutoEnzyme{…}, obj_fn::typeof(custom_loss), data::Tuple{…}, ts::Lux.Training.TrainState{…})
    @ Lux.Training /scratch/htc/amartine/julia/packages/Lux/ptjU6/src/helpers/training.jl:285
 [18] train_model!(gcn::GCN{…}, ps::@NamedTuple{…}, st::@NamedTuple{…}, g::GNNGraph{…})
    @ Main ./REPL[35]:4
 [19] top-level scope
    @ REPL[36]:1
Some type information was truncated. Use `show(err)` to see complete types.

Now that I have come across #389 I realize why. I think it would be a good idea to include a warning on the documentation that Zygote.jl is not supported. Perhaps, better yet, catch it and let the user know before they get the scary-looking exception with bits of IR and so that they are less likely to spend a bunch of time wondering why it's just not working.