feat: overlay Zygote.gradient and use Enzyme instead

Project.toml

@@ -46,6 +46,7 @@ Random123 = "74087812-796a-5b5d-8853-05524746bad3"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 YaoBlocks = "418bc28f-b43b-5e0b-a6e7-61bbc1a2c1df"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [sources]
 ReactantCore = {path = "lib/ReactantCore"}
@@ -67,6 +68,7 @@ ReactantRandom123Ext = "Random123"
 ReactantSpecialFunctionsExt = "SpecialFunctions"
 ReactantStatisticsExt = "Statistics"
 ReactantYaoBlocksExt = "YaoBlocks"
+ReactantZygoteExt = "Zygote"
 
 [compat]
 AbstractFFTs = "1.5"
@@ -107,6 +109,7 @@ Sockets = "1.10"
 SpecialFunctions = "2.4"
 Statistics = "1.10"
 YaoBlocks = "0.13, 0.14"
+Zygote = "0.7"
 julia = "1.10"
 unzip_jll = "6"
 

docs/src/api/config.md

@@ -33,6 +33,11 @@ Reactant.PrecisionConfig
 Reactant.DotGeneralAlgorithm
 ```
 
+### Zygote Overlay
+
+- `OVERLAY_ZYGOTE_CALLS`: Whether to overlay `Zygote.gradient` calls with `Enzyme.autodiff`
+  calls.
+
 ## Environment Variables
 
 The following environment variables can be used to configure Reactant.

ext/ReactantZygoteExt.jl

@@ -0,0 +1,29 @@
+module ReactantZygoteExt
+
+using Reactant:
+    Reactant, CallWithReactant, @reactant_overlay, use_overlayed_version, call_with_reactant
+using Zygote: Zygote
+using Enzyme: Enzyme, Reverse, Active, Const, Duplicated
+
+# TODO: overload the following as well
+#       - Zygote.pullback
+#       - Zygote.jacobian
+#       - Zygote.hessian
+
+@reactant_overlay function Zygote.gradient(f::F, args...) where {F}
+    # TODO: check `f` as well once #1642 is merged
+    if use_overlayed_version(args) && Reactant.OVERLAY_ZYGOTE_CALLS[]
+        @warn "Reactant doesn't support using Zygote for computing gradients. Replacing \
+               `Zygote.gradient` with `Enzyme.autodiff` call. Please update your code to \
+               not use `Zygote.gradient` and instead use `Enzyme.gradient` inside \
+               `Reactant.@compile`."
+        dargs = map(Enzyme.make_zero, args)
+        duplicated = map(Duplicated, args, dargs)
+        Reactant.overload_autodiff(Reverse, Const(f), Active, duplicated...)
+        return dargs
+    else
+        return Base.inferencebarrier(Zygote.gradient)(CallWithReactant(f), args...)
+    end
+end
+
+end

src/Configuration.jl

@@ -30,6 +30,11 @@ scope will use the provided values.
     or [`PrecisionConfig`](@ref). Defaults to `PrecisionConfig.DEFAULT`.
   - `convolution_precision`: Precision for `stablehlo.convolution`. Can be `nothing`,
     or [`PrecisionConfig`](@ref). Defaults to `PrecisionConfig.DEFAULT`.
+
+### Zygote Overlay
+
+  - `overlay_zygote_calls`: Whether to overlay `Zygote.gradient` calls with
+    `Enzyme.autodiff` calls. Defaults to `true`.
 """
 function with_config(
     f;
@@ -38,6 +43,7 @@ function with_config(
     convolution_precision=missing,
     lower_partialsort_to_approx_top_k=missing,
     fallback_approx_top_k_lowering=missing,
+    overlay_zygote_calls=missing,
 )
     config_vars = ()
     dot_general_algorithm !== missing &&
@@ -58,6 +64,9 @@ function with_config(
             FALLBACK_APPROX_TOP_K_LOWERING => fallback_approx_top_k_lowering,
         )
     )
+    overlay_zygote_calls !== missing && (
+        config_vars = (config_vars..., OVERLAY_ZYGOTE_CALLS => overlay_zygote_calls)
+    )
 
     return ScopedValues.with(f, config_vars...)
 end
@@ -379,3 +388,6 @@ function DotGeneralAlgorithm(
 
     return nothing
 end
+
+# Overlay Zygote.jl
+const OVERLAY_ZYGOTE_CALLS = ScopedValue(true)

test/Project.toml

@@ -31,6 +31,7 @@ StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
 Adapt = "4.1"
@@ -65,6 +66,7 @@ StableRNGs = "1"
 Statistics = "1.10"
 StatsBase = "0.34"
 Test = "1.10"
+Zygote = "0.7"
 
 [extras]
 Reactant_jll = "0192cb87-2b54-54ad-80e0-3be72ad8a3c0"

test/integration/zygote.jl

@@ -0,0 +1,14 @@
+using Zygote, Reactant, Enzyme, Test
+
+sumabs2(x) = sum(abs2, x)
+
+@testset "Zygote" begin
+    @testset "Zygote.gradient" begin
+        x = Reactant.to_rarray(rand(Float32, 32, 10))
+
+        zyg_grad = @jit Zygote.gradient(sumabs2, x)
+        enz_grad = @jit Enzyme.gradient(Reverse, Const(sumabs2), x)
+        @test zyg_grad[1] isa Reactant.ConcreteRArray
+        @test enz_grad[1] ≈ zyg_grad[1]
+    end
+end

test/runtests.jl

@@ -51,6 +51,7 @@ const REACTANT_TEST_GROUP = lowercase(get(ENV, "REACTANT_TEST_GROUP", "all"))
         @safetestset "Python" include("integration/python.jl")
         @safetestset "Optimisers" include("integration/optimisers.jl")
         @safetestset "FillArrays" include("integration/fillarrays.jl")
+        @safetestset "Zygote" include("integration/zygote.jl")
     end
 
     if REACTANT_TEST_GROUP == "all" || REACTANT_TEST_GROUP == "neural_networks"