Warning with Flux.Bilinear

[CarloLucibello]

When executing the following code

using Enzyme, Flux
loss(model, x) = sum(model(x))
model = Flux.Bilinear(2 => 3)
x = randn(Float32, 2, 4)
d_model = Flux.fmap(model) do x
    x isa Array ? zero(x) : x
end
Enzyme.autodiff(Reverse, loss, Active, Duplicated(model, d_model), Const(x))

I get the Warning (only showing a tiny part)

│ 
│ Stacktrace:
│  [1] getproperty
│    @ ./Base.jl:37
│  [2] _wait2
│    @ ./condition.jl:87
└ @ Enzyme.Compiler ~/.julia/packages/GPUCompiler/U36Ed/src/utils.jl:59
┌ Warning: TypeAnalysisDepthLimit
│ LLVM.StoreInst(store atomic {} addrspace(10)* %getfield14, {} addrspace(10)* addrspace(11)* %18 release, align 8, !dbg !171, !tbaa !142, !alias.scope !146, !noalias !174)
│ {[]:Pointer, [0]:Pointer, [0,0]:Pointer, [0,0,0]:Pointer, [0,0,0,0]:Pointer, [0,0,0,0,0]:Pointer, [0,0,0,8]:Pointer, [0,0,0,8,0]:Pointer, [0,0,0,8,0,0]:Pointer,
...

I haven't had time to reduce it yet, but the culprit could be the NNlib.batched_mul operation here

[wsmoses]

So that trace doesnt actually show the error on the inside. Can you post the full error log?

[CarloLucibello]

Here is the full output
out.txt

[wsmoses]

I don't see any error messages in your trace either -- just warnings.

You can silcence them with Enzyme.API.typeWarning!(false) right after importing Enzyme.

[CarloLucibello]

Ok. Should Enzyme.API.typeWarning!(false) be the default then?

[wsmoses]

Yeah that's something I've been leaning towards recently as we've added more type info regardless.

The point of that warning is to let the user know that we've hit a cut off for how much type information Julia is going to provide, which could lead to Enzyme not being able to differentiate a function at all (or as performantly).