docs: add 3rd order AD example using Reactant

[skip ci]

Author: avik-pal

docs/make.jl

@@ -48,6 +48,8 @@ pages = [
         "manual/distributed_utils.md",
         "manual/nested_autodiff.md",
         "manual/compiling_lux_models.md",
+        "manual/exporting_to_jax.md",
+        "manual/nested_autodiff_reactant.md"
     ],
     "API Reference" => [
         "Lux" => [

docs/src/.vitepress/config.mts

@@ -346,6 +346,10 @@ export default defineConfig({
                 text: "Exporting Lux Models to Jax",
                 link: "/manual/exporting_to_jax",
               },
+              {
+                text: "Nested AutoDiff",
+                link: "/manual/nested_autodiff_reactant",
+              }
             ],
           },
           {

docs/src/manual/nested_autodiff.md

@@ -1,16 +1,16 @@
 # [Nested Automatic Differentiation](@id nested_autodiff)
 
-!!! note
-
-    This is a relatively new feature in Lux, so there might be some rough edges. If you
-    encounter any issues, please let us know by opening an issue on the
-    [GitHub repository](https://github.com/LuxDL/Lux.jl).
-
 In this manual, we will explore how to use automatic differentiation (AD) inside your layers
 or loss functions and have Lux automatically switch the AD backend with a faster one when
 needed.
 
-!!! tip
+!!! tip "Reactant Support"
+
+    Reactant + Lux natively supports Nested AD (even higher dimensions). If you are using
+    Reactant, please see the [Nested AD with Reactant](@ref nested_autodiff_reactant)
+    manual.
+
+!!! tip "Disabling Nested AD Switching"
 
     Don't wan't Lux to do this switching for you? You can disable it by setting the
     `automatic_nested_ad_switching` Preference to `false`.

docs/src/manual/nested_autodiff_reactant.md

@@ -0,0 +1,69 @@
+# [Nested AutoDiff with Reactant](@id nested_autodiff_reactant)
+
+We will be using the example from [issue 614](https://github.com/LuxDL/Lux.jl/issues/614).
+
+```@example nested_ad_reactant
+using Reactant, Enzyme, Lux, Random, LinearAlgebra
+
+const xdev = reactant_device(; force=true)
+const cdev = cpu_device()
+
+function ∇potential(potential, x)
+    dxs = stack(onehot(x))
+    ∇p = similar(x)
+    colons = [Colon() for _ in 1:ndims(x)]
+    @trace for i in 1:length(x)
+        dxᵢ = dxs[colons..., i]
+        res = only(Enzyme.autodiff(
+            Enzyme.set_abi(Forward, Reactant.ReactantABI), potential, Duplicated(x, dxᵢ)
+        ))
+        @allowscalar ∇p[i] = res[i]
+    end
+    return ∇p
+end
+
+function ∇²potential(potential, x)
+    dxs = stack(onehot(x))
+    ∇²p = similar(x)
+    colons = [Colon() for _ in 1:ndims(x)]
+    @trace for i in 1:length(x)
+        dxᵢ = dxs[colons..., i]
+        res = only(Enzyme.autodiff(
+            Enzyme.set_abi(Forward, Reactant.ReactantABI),
+            ∇potential, Const(potential), Duplicated(x, dxᵢ)
+        ))
+        @allowscalar ∇²p[i] = res[i]
+    end
+    return ∇²p
+end
+
+struct PotentialNet{P} <: AbstractLuxWrapperLayer{:potential}
+    potential::P
+end
+
+function (potential::PotentialNet)(x, ps, st)
+    pnet = StatefulLuxLayer{true}(potential.potential, ps, st)
+    return ∇²potential(pnet, x), pnet.st
+end
+
+model = PotentialNet(Dense(5 => 5, gelu))
+ps, st = Lux.setup(Random.default_rng(), model) |> xdev
+
+x_ra = randn(Float32, 5, 3) |> xdev
+
+@code_hlo model(x_ra, ps, st)
+
+@jit model(x_ra, ps, st)
+```
+
+```@example nested_ad_reactant
+sumabs2first(model, x, ps, st) = sum(abs2, first(model(x, ps, st)))
+
+function enzyme_gradient(model, x, ps, st)
+    return Enzyme.gradient(
+        Enzyme.Reverse, Const(sumabs2first), Const(model), Const(x), ps, Const(st)
+    )
+end
+
+@jit enzyme_gradient(model, x_ra, ps, st)
+```