Utilize ChainRulesCore thunks (#966)

* Don't force unthunking of ChainRulesCore thunks

Introduces @_adjoint_keepthunks to mark adjoints that should pass
chunks through.

* Use @_adjoint_keepthunks where appropriate

* Use wrap_chainrules_output in unthunk_tangent

* Fix unthunk_tangent for array of thunks

* Don't unthunk explicitly in unbroadcast

* Define unthunk_tangent for IdDict to support Params

* Make unthunk_tangent for IdDict non-differentiable

Co-authored-by: Brian Chen <ToucheSir@users.noreply.github.com>

* Revert "Don't unthunk explicitly in unbroadcast"

This reverts commit 34865ea.

* Fix problems related to unthunk_tangent for IdDict

Co-authored-by: Brian Chen <ToucheSir@users.noreply.github.com>

* Resolve duplicate rrule for unthunk_tangent with IdDict

* Make unthunk_tangent recurse into arrays

* Fix tests

* Unthunk in collect(::Generator)

* Update deps

* Disable thunks for 2nd order AD

* Temporary use fork of CRC

* Remove hook

* Fix

* Cleanup

* Up deps

* Cleanup

* Remove extra unthunk_tangent

---------

Co-authored-by: Brian Chen <ToucheSir@users.noreply.github.com>
Co-authored-by: Anton Smirnov <tonysmn97@gmail.com>

Author: 3 people

Project.toml

@@ -38,8 +38,8 @@ ZygoteTrackerExt = "Tracker"
 
 [compat]
 AbstractFFTs = "1.3.1"
-ChainRules = "1.44.1"
-ChainRulesCore = "1.9"
+ChainRules = "1.72.2"
+ChainRulesCore = "1.25.1"
 ChainRulesTestUtils = "1"
 Colors = "0.12, 0.13"
 DiffRules = "1.4"

src/Zygote.jl

@@ -3,11 +3,12 @@ module Zygote
 using LinearAlgebra, Statistics
 using LinearAlgebra: copytri!, AbstractTriangular
 
+import ZygoteRules
 import ZygoteRules: @adjoint, @adjoint!, AContext, adjoint, _pullback, pullback,
   literal_getproperty, literal_getfield, unthunk_tangent
 
 using ChainRulesCore
-using ChainRules: ChainRules, rrule, unthunk, canonicalize
+using ChainRules: ChainRules, AbstractThunk, rrule, unthunk, canonicalize
 using IRTools
 using MacroTools, Requires
 using MacroTools: @forward

src/compiler/chainrules.jl

@@ -1,3 +1,14 @@
+# ToDo: Move some of this to ZygoteRules, or move unthunk_tangent for Tuple and NamedTuple from
+# Zygote rules here?
+function unthunk_tangent end
+@inline unthunk_tangent(x::AbstractThunk) = wrap_chainrules_output(unthunk(x))
+@inline unthunk_tangent(x::NTuple{N,<:Number}) where N = x
+@inline unthunk_tangent(x::AbstractArray{<:Number,N}) where N = x
+@inline unthunk_tangent(x::AbstractArray) = map(unthunk_tangent, x)
+unthunk_tangent(d::IdDict) = IdDict([unthunk_tangent(k) => unthunk_tangent(v) for (k, v) in d])
+@non_differentiable unthunk_tangent(::IdDict)
+
+
 struct ZygoteRuleConfig{CTX<:AContext} <: RuleConfig{Union{HasReverseMode,NoForwardsMode}}
   context::CTX
 end
@@ -107,7 +118,6 @@ is_kwfunc(k, ::Type{<:NamedTuple}, f, args...) = k===Core.kwftype(f)
 Convert `x` from the differentials types ChainRules uses to the format Zygote uses internally.
 """
 @inline wrap_chainrules_output(x) = x
-@inline wrap_chainrules_output(x::AbstractThunk) = wrap_chainrules_output(unthunk(x))  # For now we are just not going to deal with thunks
 @inline wrap_chainrules_output(x::Tuple) = map(wrap_chainrules_output, x)
 # Zygote convention: even if many AbstractZero partials (i.e. multi-input function), make just 1 nothing.
 @inline wrap_chainrules_output(x::Tuple{Vararg{ChainRules.AbstractZero}}) = nothing
@@ -261,7 +271,9 @@ function ChainRulesCore.rrule_via_ad(config::ZygoteRuleConfig, f_args...; kwargs
         _pullback(config.context, f_args...)
     end
 
-    ad_pullback(Δ) = zygote2differential(pb(wrap_chainrules_output(Δ)), f_args)
+    ad_pullback(Δ) = zygote2differential(
+        pb(wrap_chainrules_output(unthunk_tangent(Δ))),
+        f_args)
     return y, ad_pullback
 end
 

src/compiler/interface.jl

@@ -37,7 +37,13 @@ end
 _pullback(f, args...) = _pullback(Context(), f, args...)
 
 tailmemaybe(::Nothing) = nothing
-tailmemaybe(x::Tuple) = Base.tail(x)
+tailmemaybe(x::Tuple) = unthunk_tangent(Base.tail(x))
+
+# unthunking is essentially an identity operation on a lazy value, but
+# `@adjoint unthunk_tangent(x) = unthunk_tangent(x), ȳ -> (ȳ,)` is not enough to make
+# nested AD work, so define
+@adjoint tailmemaybe(xs::Tuple) = tailmemaybe(xs), x̄s -> ((nothing, x̄s...),)
+
 
 """
     pullback(f, args...)
@@ -351,6 +357,9 @@ function copy!(x::AbstractVector, ps::Params)
   x
 end
 
+_maybe_unthunk(x::AbstractThunk) = unthunk(x)
+_maybe_unthunk(x) = x
+
 """
     Grads(...)
 
@@ -385,7 +394,7 @@ end
 
 function Base.getindex(gs::Grads, x)
   isbits(x) && error("Only reference types can be differentiated with `Params`.")
-  return gs.grads[x]
+  return _maybe_unthunk(gs.grads[x])
 end
 
 """
@@ -468,7 +477,7 @@ function pullback(f, ps::Params)
       cache(cx)[p] = nothing
     end
     back(Δ)
-    Grads(cx.cache, ps) # TODO make a copy
+    Grads(_maybe_unthunk(cx.cache), ps)
   end
 end
 

src/compiler/reverse.jl

@@ -3,6 +3,18 @@ using IRTools: IR, Variable, Pipe, xcall, var, prewalk, postwalk,
   insertafter!, finish, expand!, prune!, substitute!, substitute,
   block, block!, branch!, return!, stmt, meta
 
+
+# TODO: Temporary, to be removed when ChainRulesCore rrules are required to
+# support thunks as an input and all instances of _adjoint_keepthunks in
+# Zygote have been replaces by rrules:
+macro _adjoint_keepthunks(ex)
+  ZygoteRules.gradm(ex, false, true)
+end
+macro _adjoint_keepthunks!(ex)
+  ZygoteRules.gradm(ex, true, true)
+end
+
+
 @inline tuple_va(N, xs) = xs
 @inline tuple_va(N, x, xs...) = (x, tuple_va(N, xs...)...)
 @inline tuple_va(::Val{N}, ::Nothing) where N = ntuple(_ -> nothing, Val(N))

src/lib/array.jl

@@ -237,7 +237,7 @@ function _pullback(cx::AContext, ::typeof(collect), g::Base.Generator)
     x̄ = reconstruct_if_dict(x̄, _keys) # return a dictionary if needed
     (nothing, (f = f̄, iter = x̄),)
   end
-  y, collect_pullback
+  y, collect_pullback ∘ unthunk_tangent
 end
 
 collect_if_dict(x::Dict) = collect(x), collect(keys(x))

src/lib/broadcast.jl

@@ -53,7 +53,8 @@ function Base.reducedim_init(::typeof(identity), ::typeof(accum), A::AbstractArr
   Base.reducedim_initarray(A, region, nothing, Union{Nothing,eltype(A)})
 end
 
-function unbroadcast(x::AbstractArray, x̄)
+function unbroadcast(x::AbstractArray, maybethunked_x̄)
+  x̄ = unthunk_tangent(maybethunked_x̄)
   N = ndims(x̄)
   if length(x) == length(x̄)
     _project(x, x̄)  # ProjectTo handles reshape, offsets, structured matrices, row vectors

src/lib/lib.jl

@@ -37,16 +37,23 @@ function accum(x::RefValue, y::RefValue)
   return x
 end
 
+accum(x::NamedTuple, y::ChainRulesCore.Tangent) = accum(x, wrap_chainrules_output(y))
+accum(x::ChainRulesCore.Tangent, y::NamedTuple) = accum(wrap_chainrules_output(x), y)
+
+accum(x, y::AbstractThunk) = @thunk(accum(x, unthunk(y)))
+accum(x::AbstractThunk, y) = @thunk(accum(unthunk(x), y))
+accum(x::AbstractThunk, y::AbstractThunk) = @thunk(accum(unthunk(x), unthunk(y)))
+
 # Core functions
-@adjoint deepcopy(x) = deepcopy(x), ȳ -> (ȳ,)
+@_adjoint_keepthunks deepcopy(x) = deepcopy(x), ȳ -> (ȳ,)
 
-@adjoint (::Type{V})(x...) where V<:Val = V(x...), _ -> nothing
+@_adjoint_keepthunks (::Type{V})(x...) where V<:Val = V(x...), _ -> nothing
 
-@adjoint ifelse(cond::Bool, t, f) =
+@_adjoint_keepthunks ifelse(cond::Bool, t, f) =
   ifelse(cond, t, f),
   Δ -> cond ? (nothing, Δ, zero(Δ)) : (nothing, zero(Δ), Δ)
 
-@adjoint Base.typeassert(x, T) = Base.typeassert(x, T), Δ -> (Δ, nothing)
+@_adjoint_keepthunks Base.typeassert(x, T) = Base.typeassert(x, T), Δ -> (Δ, nothing)
 
 accum_param(::Context{false}, _, Δ) = Δ
 @generated function accum_param(cx::Context, x, Δ)
@@ -70,11 +77,11 @@ end
 
 unwrap(x) = x
 
-@adjoint unwrap(x) = unwrap(x), x̄ -> (accum_param(__context__, x, x̄),)
+@_adjoint_keepthunks unwrap(x) = unwrap(x), x̄ -> (accum_param(__context__, x, x̄),)
 
 unwrap(ref, x) = x
 
-@adjoint unwrap(ref, x) = unwrap(x), function (x̄)
+@_adjoint_keepthunks unwrap(ref, x) = unwrap(x), function (x̄)
   accum_global(__context__, ref, x̄)
   (accum_param(__context__, x, x̄),)
 end
@@ -88,7 +95,7 @@ function global_set(ref, val)
   end
 end
 
-@adjoint! function global_set(ref, x)
+@_adjoint_keepthunks! function global_set(ref, x)
   global_set(ref, x), function (x̄)
     gs = cache(__context__)
     x̄ = accum(get(gs, ref, nothing), x̄)
@@ -101,9 +108,9 @@ end
 
 using Base: tail
 
-@adjoint tuple(xs...) = xs, identity
+@_adjoint_keepthunks tuple(xs...) = xs, identity
 
-@adjoint function literal_getindex(xs::NTuple{N,Any}, ::Val{i}) where {N,i}
+@_adjoint_keepthunks function literal_getindex(xs::NTuple{N,Any}, ::Val{i}) where {N,i}
   val = xs[i]
   function back(Δ)
     accum_param(__context__, val, Δ) === nothing && return
@@ -112,7 +119,7 @@ using Base: tail
   val, back
 end
 
-@adjoint function getindex(xs::NTuple{N,Any}, i::Integer) where N
+@_adjoint_keepthunks function getindex(xs::NTuple{N,Any}, i::Integer) where N
   val = xs[i]
   function back(Δ)
     accum_param(__context__, val, Δ) === nothing && return
@@ -121,10 +128,10 @@ end
   return val, back
 end
 
-@adjoint getindex(xs::NTuple{N,Any}, r::AbstractUnitRange) where N =
+@_adjoint_keepthunks getindex(xs::NTuple{N,Any}, r::AbstractUnitRange) where N =
   (xs[r], Δ -> (ntuple(j -> j in r ? Δ[findfirst(isequal(j), r)] : nothing, Val(N)), nothing))
 
-@adjoint function getindex(xs::NTuple{N,Any}, r::AbstractVector) where N
+@_adjoint_keepthunks function getindex(xs::NTuple{N,Any}, r::AbstractVector) where N
   val = xs[r]
   function back(Δ)
     dxs = ntuple(Val(length(xs))) do x
@@ -155,18 +162,18 @@ function _pullback(cx::AContext, ::typeof(literal_indexed_iterate), xs::Tuple, :
 end
 
 # Needed for iteration lowering
-@adjoint Core.getfield(xs::NTuple{N,Any}, i::Int) where N =
+@_adjoint_keepthunks Core.getfield(xs::NTuple{N,Any}, i::Int) where N =
   (xs[i], Δ -> (ntuple(j -> i == j ? Δ : nothing, Val(N)), nothing))
 
-@adjoint Core.getfield(xs::NamedTuple{K,<:NTuple{N,Any}}, i::Int) where {K,N} =
+@_adjoint_keepthunks Core.getfield(xs::NamedTuple{K,<:NTuple{N,Any}}, i::Int) where {K,N} =
   (xs[i], Δ -> (NamedTuple{K}(ntuple(j -> i == j ? Δ : nothing, Val(N))), nothing))
 
-@adjoint function Base.first(xs::Tuple)
+@_adjoint_keepthunks function Base.first(xs::Tuple)
   drest = map(_->nothing, tail(xs))
   first(xs), Δ -> ((Δ, drest...),)
 end
 
-@adjoint Base.tail(xs::Tuple) = tail(xs), x̄s -> ((nothing, x̄s...),)
+@_adjoint_keepthunks Base.tail(xs::Tuple) = tail(xs), x̄s -> ((nothing, x̄s...),)
 
 _empty(x) = length(x)
 _empty(x::Union{Tuple,NamedTuple}) = map(_->nothing, x)
@@ -188,7 +195,7 @@ end
 
 unapply(t, xs) = _unapply(t, xs)[1]
 
-@adjoint! function Core._apply(f, args...)
+@_adjoint_keepthunks! function Core._apply(f, args...)
   y, back = Core._apply(_pullback, (__context__, f), args...)
   st = map(_empty, args)
   y, function (Δ)
@@ -198,7 +205,7 @@ unapply(t, xs) = _unapply(t, xs)[1]
   end
 end
 
-@adjoint! function Core._apply_iterate(::typeof(iterate), f, args...)
+@_adjoint_keepthunks! function Core._apply_iterate(::typeof(iterate), f, args...)
   y, back = Core._apply(_pullback, (__context__, f), args...)
   st = map(_empty, args)
   y, function (Δ)
@@ -223,7 +230,7 @@ end
 @generated pair(::Val{k}, v, _=nothing) where k = :($k = v,)
 @generated pair(::Val{k}, v, ::NamedTuple{keys}) where {k,keys} = k isa Int ? :($(getfield(keys, k)) = v,) : :($k = v,)
 
-@adjoint function literal_getfield(x, ::Val{f}) where f
+@_adjoint_keepthunks function literal_getfield(x, ::Val{f}) where f
   val = getfield(x, f)
   function back(Δ)
     accum_param(__context__, val, Δ) === nothing && return
@@ -273,8 +280,7 @@ function _get!(default::Base.Callable, ch, x)
   end
 end
 
-
-@adjoint! function setfield!(x, f, val)
+@_adjoint_keepthunks! function setfield!(x, f, val)
   y = setfield!(x, f, val)
   g = grad_mut(__context__, x)
   y, function (_)
@@ -290,13 +296,13 @@ end
 
 Jnew{T}(g) where T = Jnew{T,typeof(g)}(g)
 
-@adjoint! function __new__(T, args...)
+@_adjoint_keepthunks! function __new__(T, args...)
   x = __new__(T, args...)
   g = !ismutabletype(T) || fieldcount(T) == 0 ? nothing : grad_mut(__context__, x)
   x, Jnew{T,typeof(g),false}(g)
 end
 
-@adjoint! function __splatnew__(T, args)
+@_adjoint_keepthunks! function __splatnew__(T, args)
   x = __splatnew__(T, args)
   g = !ismutabletype(T) || fieldcount(T) == 0 ? nothing : grad_mut(__context__, x)
   x, Jnew{T,typeof(g),true}(g)

test/gradcheck.jl

@@ -20,14 +20,14 @@ function ngradient(f, xs::AbstractArray...)
   return grads
 end
 
-function gradcheck(f, xs...)
+function gradcheck(f, xs...; rtol = 1e-5, atol = 1e-5)
   grad_zygote = gradient(f, xs...)
   grad_finite_difference = ngradient(f, xs...)
-  return all(isapprox.(grad_zygote, grad_finite_difference; rtol = 1e-5, atol = 1e-5))
+  return all(isapprox.(grad_zygote, grad_finite_difference; rtol = rtol, atol = atol))
 end
 
-gradtest(f, xs::AbstractArray...) = gradcheck((xs...) -> sum(sin.(f(xs...))), xs...)
-gradtest(f, dims...) = gradtest(f, rand.(Float64, dims)...)
+gradtest(f, xs::AbstractArray...; kwargs...) = gradcheck((xs...) -> sum(sin.(f(xs...))), xs...; kwargs...)
+gradtest(f, dims...; kwargs...) = gradtest(f, rand.(Float64, dims)...; kwargs...)
 
 # utilities for using gradcheck with complex matrices
 _splitreim(A) = (real(A),)
@@ -160,8 +160,8 @@ end
   @test gradient(y, x, z) == ([1, 1, 2], nothing)
 
   # https://github.com/FluxML/Zygote.jl/issues/376
-  _, back = Zygote._pullback(x->x[1]*im, randn(2))
-  @test back(1.0)[2] == real([-im, 0]) == [0, 0]
+  _, back = Zygote.pullback(x -> x[1] * im, randn(2))
+  @test back(1.0)[1] == real([-im, 0]) == [0, 0]
 
   # _droplike
   @test gradient(x -> sum(inv, x[1, :]'), ones(2, 2)) == ([-1 -1; 0 0],)
@@ -949,8 +949,8 @@ end
 _hermsymtype(::Type{<:Symmetric}) = Symmetric
 _hermsymtype(::Type{<:Hermitian}) = Hermitian
 
-function _gradtest_hermsym(f, ST, A)
-  gradtest(_splitreim(collect(A))...) do (args...)
+function _gradtest_hermsym(f, ST, A; kwargs...)
+  gradtest(_splitreim(collect(A))...; kwargs...) do (args...)
     B = f(ST(_joinreim(_dropimaggrad.(args)...)))
     return sum(_splitreim(B))
   end

test/interface.jl

@@ -269,5 +269,4 @@ end
     @test sgs[d.b] ≈ fill(1.f0, size(d.b))
   end
 
-
 end