Backend switching for Mooncake

DifferentiationInterface/docs/src/explanation/backends.md

@@ -95,7 +95,7 @@ In general, using a forward outer backend over a reverse inner backend will yiel
 The wrapper [`DifferentiateWith`](@ref) allows you to switch between backends.
 It takes a function `f` and specifies that `f` should be differentiated with the substitute backend of your choice, instead of whatever true backend the surrounding code is trying to use.
 In other words, when someone tries to differentiate `dw = DifferentiateWith(f, substitute_backend)` with `true_backend`, then `substitute_backend` steps in and `true_backend` does not dive into the function `f` itself.
-At the moment, `DifferentiateWith` only works when `true_backend` is either [ForwardDiff.jl](https://github.com/JuliaDiff/ForwardDiff.jl) or a [ChainRules.jl](https://github.com/JuliaDiff/ChainRules.jl)-compatible backend.
+At the moment, `DifferentiateWith` only works when `true_backend` is either [ForwardDiff.jl](https://github.com/JuliaDiff/ForwardDiff.jl), [Mooncake.jl](https://github.com/chalk-lab/Mooncake.jl), or a [ChainRules.jl](https://github.com/JuliaDiff/ChainRules.jl)-compatible backend (e.g., [Zygote.jl](https://github.com/FluxML/Zygote.jl)).
 
 ## Implementations
 

DifferentiationInterface/docs/src/faq/differentiability.md

@@ -111,4 +111,5 @@ There are, however, translation utilities:
 ### Backend switch
 
 Also note the existence of [`DifferentiationInterface.DifferentiateWith`](@ref), which allows the user to wrap a function that should be differentiated with a specific backend.
-Right now it only targets ForwardDiff.jl and ChainRulesCore.jl, but PRs are welcome to define Enzyme.jl and Mooncake.jl rules for this object.
+
+Right now, it only targets [ForwardDiff.jl](https://github.com/JuliaDiff/ForwardDiff.jl), [Mooncake.jl](), [ChainRules.jl](https://juliadiff.org/ChainRulesCore.jl/stable/)-compatible backends (e.g., [Zygote.jl](https://github.com/FluxML/Zygote.jl)),  but PRs are welcome to define Enzyme.jl rules for this object.

DifferentiationInterface/ext/DifferentiationInterfaceMooncakeExt/DifferentiationInterfaceMooncakeExt.jl

@@ -3,14 +3,22 @@ module DifferentiationInterfaceMooncakeExt
 using ADTypes: ADTypes, AutoMooncake
 import DifferentiationInterface as DI
 using Mooncake:
+    Mooncake,
     CoDual,
     Config,
     prepare_gradient_cache,
     prepare_pullback_cache,
     tangent_type,
     value_and_gradient!!,
     value_and_pullback!!,
-    zero_tangent
+    zero_tangent,
+    rdata_type,
+    @is_primitive,
+    zero_fcodual,
+    MinimalCtx,
+    NoRData,
+    fdata,
+    primal
 
 DI.check_available(::AutoMooncake) = true
 
@@ -26,5 +34,6 @@ mycopy(x) = deepcopy(x)
 
 include("onearg.jl")
 include("twoarg.jl")
+include("differentiate_with.jl")
 
 end

DifferentiationInterface/ext/DifferentiationInterfaceMooncakeExt/differentiate_with.jl

@@ -0,0 +1,50 @@
+@is_primitive MinimalCtx Tuple{DI.DifferentiateWith,<:Union{Number,AbstractArray}}
+
+function Mooncake.rrule!!(dw::CoDual{<:DI.DifferentiateWith}, x::CoDual{<:Number})
+    primal_func = primal(dw)
+    primal_x = primal(x)
+    (; f, backend) = primal_func
+    y = zero_fcodual(f(primal_x))
+
+    # output is a vector, so we need to use the vector pullback
+    function pullback_array!!(dy::NoRData)
+        tx = DI.pullback(f, backend, primal_x, (fdata(y.dx),))
+        @assert only(tx) isa rdata_type(typeof(primal_x))
+        return NoRData(), only(tx)
+    end
+
+    # output is a scalar, so we can use the scalar pullback
+    function pullback_scalar!!(dy::Number)
+        tx = DI.pullback(f, backend, primal_x, (dy,))
+        @assert only(tx) isa rdata_type(typeof(primal_x))
+        return NoRData(), only(tx)
+    end
+
+    return y, typeof(primal(y)) <: Number ? pullback_scalar!! : pullback_array!!
+end
+
+function Mooncake.rrule!!(dw::CoDual{<:DI.DifferentiateWith}, x::CoDual{<:AbstractArray})
+    primal_func = primal(dw)
+    primal_x = primal(x)
+    fdata_arg = fdata(x.dx)
+    (; f, backend) = primal_func
+    y = zero_fcodual(f(primal_x))
+
+    # output is a vector, so we need to use the vector pullback
+    function pullback_array!!(dy::NoRData)
+        tx = DI.pullback(f, backend, primal_x, (fdata(y.dx),))
+        @assert first(only(tx)) isa rdata_type(typeof(first(primal_x)))
+        fdata_arg .+= only(tx)
+        return NoRData(), dy
+    end
+
+    # output is a scalar, so we can use the scalar pullback
+    function pullback_scalar!!(dy::Number)
+        tx = DI.pullback(f, backend, primal_x, (dy,))
+        @assert first(only(tx)) isa rdata_type(typeof(first(primal_x)))
+        fdata_arg .+= only(tx)
+        return NoRData(), NoRData()
+    end
+
+    return y, typeof(primal(y)) <: Number ? pullback_scalar!! : pullback_array!!
+end

DifferentiationInterface/src/misc/differentiate_with.jl

@@ -13,9 +13,13 @@ Moreover, any larger algorithm `alg` that calls `f2` instead of `f` will also be
 
 !!! warning
     `DifferentiateWith` only supports out-of-place functions `y = f(x)` without additional context arguments.
-    It only makes these functions differentiable if the true backend is either [ForwardDiff](https://github.com/JuliaDiff/ForwardDiff.jl) or automatically importing rules from [ChainRules](https://github.com/JuliaDiff/ChainRules.jl) (e.g. [Zygote](https://github.com/FluxML/Zygote.jl)). Some backends are also able to [manually import rules](https://juliadiff.org/ChainRulesCore.jl/stable/#Packages-supporting-importing-rules-from-ChainRules.) from ChainRules.
+    It only makes these functions differentiable if the true backend is either [ForwardDiff](https://github.com/JuliaDiff/ForwardDiff.jl), [Mooncake](https://github.com/chalk-lab/Mooncake.jl) or automatically importing rules from [ChainRules](https://github.com/JuliaDiff/ChainRules.jl) (e.g. [Zygote](https://github.com/FluxML/Zygote.jl)). Some backends are also able to [manually import rules](https://juliadiff.org/ChainRulesCore.jl/stable/#Packages-supporting-importing-rules-from-ChainRules.) from ChainRules.
     For any other true backend, the differentiation behavior is not altered by `DifferentiateWith` (it becomes a transparent wrapper).
 
+!!! warning
+    When using Mooncake as a substitute backend via `DifferentiateWith(f, AutoMooncake())`. The function `f` must not close over any active data.
+    As of now, we cannot differentiate with respect to parameters stored inside `f`.
+
 # Fields
 
 - `f`: the function in question, with signature `f(x)`

DifferentiationInterface/test/Back/DifferentiateWith/test.jl

@@ -1,18 +1,21 @@
 using Pkg
-Pkg.add(["FiniteDiff", "ForwardDiff", "Zygote"])
+Pkg.add(["FiniteDiff", "ForwardDiff", "Zygote", "Mooncake"])
 
 using DifferentiationInterface, DifferentiationInterfaceTest
 import DifferentiationInterfaceTest as DIT
 using FiniteDiff: FiniteDiff
 using ForwardDiff: ForwardDiff
 using Zygote: Zygote
+using Mooncake: Mooncake
 using Test
 
 LOGGING = get(ENV, "CI", "false") == "false"
 
 function differentiatewith_scenarios()
     bad_scens =  # these closurified scenarios have mutation and type constraints
-        filter(default_scenarios(; include_normal=false, include_closurified=true)) do scen
+        filter(
+            DIT.default_scenarios(; include_normal=false, include_closurified=true)
+        ) do scen
             DIT.function_place(scen) == :out
         end
     good_scens = map(bad_scens) do scen
@@ -22,7 +25,7 @@ function differentiatewith_scenarios()
 end
 
 test_differentiation(
-    [AutoForwardDiff(), AutoZygote()],
+    [AutoForwardDiff(), AutoZygote(), AutoMooncake(; config=nothing)],
     differentiatewith_scenarios();
     excluded=SECOND_ORDER,
     logging=LOGGING,