Skip to content

Keep tuple structure of broadcast cotangents #858

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lkdvos wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Keep tuple structure of broadcast cotangents #858

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2652fa4
Keep tuple structure of broadcast cotangents
lkdvos Oct 29, 2026
6b8be24
Merge branch 'main' into tuple-tangent-broadcast
devmotion Oct 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 9 additions & 4 deletions src/rulesets/Base/broadcast.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,11 @@ using Base.Broadcast: Broadcast, broadcasted, Broadcasted, BroadcastStyle
const RCR = RuleConfig{>:HasReverseMode}
const TRI_NO = (NoTangent(), NoTangent(), NoTangent())

# The cotangent of a tuple-valued broadcast is a `Tangent` (`Tangent{Any}` from Zygote). Broadcasting
# over that would `collect` it into a `Vector`, losing the element types, so use its backing tuple.
Comment on lines +5 to +6

@devmotion devmotion Oct 30, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

With JuliaDiff/ChainRulesCore.jl#724, broadcasting over a Tangent throws instead of collecting it. ChainRules supports both versions, so maybe:

Suggested change
# The cotangent of a tuple-valued broadcast is a `Tangent` (`Tangent{Any}` from Zygote). Broadcasting
# over that would `collect` it into a `Vector`, losing the element types, so use its backing tuple.
# The cotangent of a tuple-valued broadcast can be a `Tangent`

_bc_cotangent(dys::Tangent{<:Any, <:Tuple}) = ChainRulesCore.backing(dys)
_bc_cotangent(dys) = dys

function rrule(::typeof(copy), bc::Broadcasted)
uncopy(Δ) = (NoTangent(), Δ)
return copy(bc), uncopy
Expand Down Expand Up @@ -59,7 +64,7 @@ function split_bc_derivatives(f::F, arg) where {F}
@debug("split broadcasting derivative", f)
ys = f.(arg)
function bc_one_back(dys) # For f.(x) we do not need StructArrays / unzip at all
delta = broadcast(unthunk(dys), ys, arg) do dy, y, a
delta = broadcast(_bc_cotangent(unthunk(dys)), ys, arg) do dy, y, a
das = only(derivatives_given_output(y, f, a))
dy * conj(only(das)) # possibly this * should be made nan-safe.
end
Expand All @@ -72,7 +77,7 @@ function split_bc_derivatives(f::F, args::Vararg{Any,N}) where {F,N}
@debug("split broadcasting derivatives", f, N)
ys = f.(args...)
function bc_many_back(dys)
deltas = unzip_broadcast(unthunk(dys), ys, args...) do dy, y, as...
deltas = unzip_broadcast(_bc_cotangent(unthunk(dys)), ys, args...) do dy, y, as...
das = only(derivatives_given_output(y, f, as...))
map(da -> dy * conj(da), das) # possibly this * should be made nan-safe.
end
Expand Down Expand Up @@ -110,7 +115,7 @@ function split_bc_inner(frule_fun::R, cfg::RuleConfig, f::F, arg) where {R,F}
frule_fun(cfg, (NoTangent(), one(a)), f, a)
end
function back_forwards(dys)
delta = broadcast(ydots, unthunk(dys), arg) do ydot, dy, a
delta = broadcast(ydots, _bc_cotangent(unthunk(dys)), arg) do ydot, dy, a
ProjectTo(a)(conj(ydot) * dy) # possibly this * should be made nan-safe.
end
return (TRI_NO..., ProjectTo(arg)(delta))
Expand All @@ -129,7 +134,7 @@ function split_bc_pullbacks(cfg::RCR, f::F, args::Vararg{Any,N}) where {F,N}
rrule_via_ad(cfg, f, a...)
end
function back_generic(dys)
deltas = unzip_broadcast(backs, dys) do back, dy # (could be map, sizes match)
deltas = unzip_broadcast(backs, _bc_cotangent(dys)) do back, dy # (could be map, sizes match)
map(unthunk, back(dy))
end
dargs = map(unbroadcast, args, Base.tail(deltas))
Expand Down
14 changes: 14 additions & 0 deletions test/rulesets/Base/broadcast.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,6 +197,20 @@ BT1 = Broadcast.BroadcastStyle(Tuple)
@test y3 == ()
@test bk3(Tangent{Tuple{}}()) == (NoTangent(), NoTangent(), NoTangent(), ZeroTangent(), ZeroTangent())
end
@testset "tuple broadcast with Tangent{Any} cotangent" begin
# Zygote passes the cotangent of a tuple as `Tangent{Any}`, which broadcasting used to
# `collect` into a `Vector`, losing the element types.
Comment on lines +201 to +202

@devmotion devmotion Oct 30, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same as above:

Suggested change
# Zygote passes the cotangent of a tuple as `Tangent{Any}`, which broadcasting used to
# `collect` into a `Vector`, losing the element types.

x = (1.0, 2.0f0)
y2, bk2 = rrule(CFG, copy∘broadcasted, BT1, log, x) # split 2
dx2 = @inferred bk2(Tangent{Any}(1.0, 1.0f0))
@test dx2[4] isa Tangent{typeof(x), typeof(x)}
@test all(ChainRulesCore.backing(dx2[4]) .≈ (1.0, 0.5))

y4, bk4 = rrule(CFG, copy∘broadcasted, BT1, sin∘cos, x) # split 4
dx4 = bk4(Tangent{Any}(1.0, 1.0f0))
@test dx4[4] isa Tangent{typeof(x), typeof(x)}
@test all(ChainRulesCore.backing(dx4[4]) .≈ (-cos(cos(1.0)) * sin(1.0), -cos(cos(2.0f0)) * sin(2.0f0)))

@devmotion devmotion Oct 30, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe also cover thunked cotangents and the multi-argument path (bc_many_back), which is also inferable:

Suggested change
@test all(ChainRulesCore.backing(dx4[4]) .≈ (-cos(cos(1.0)) * sin(1.0), -cos(cos(2.0f0)) * sin(2.0f0)))
@test all(ChainRulesCore.backing(dx4[4]) .≈ (-cos(cos(1.0)) * sin(1.0), -cos(cos(2.0f0)) * sin(2.0f0)))
@test bk2(@thunk(Tangent{Any}(1.0, 1.0f0))) == dx2
@test bk4(@thunk(Tangent{Any}(1.0, 1.0f0))) == dx4
_, bkm = rrule(CFG, copy∘broadcasted, BT1, atan, x, x) # split 2, several arguments
dxm = @inferred bkm(Tangent{Any}(1.0, 1.0f0))
@test dxm[4] isa Tangent{typeof(x), typeof(x)}
@test dxm[5] isa Tangent{typeof(x), typeof(x)}

end
@testset "unbroadcast with NTuple" begin # https://github.com/JuliaDiff/ChainRules.jl/pull/661
@test ChainRules.unbroadcast((1, 2, [3]), [4, 5, [6]]) isa Tangent # earlier, NTuple demanded same type
@test ChainRules.unbroadcast(broadcasted(-, (1, 2), 3), (4, 5)) == (4, 5) # earlier, called ndims(::Tuple)
Expand Down
Loading