Some broadcasting fixes

Project.toml

@@ -1,6 +1,6 @@
 name = "ChainRules"
 uuid = "082447d4-558c-5d27-93f4-14fc19e9eca2"
-version = "1.48.0"
+version = "1.48.1"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"

src/rulesets/Base/broadcast.jl

@@ -92,12 +92,13 @@ end
 # and split broadcasting may anyway change N^2 executions into N, e.g. `g.(v ./ f.(v'))`.
 # We don't know `f` is cheap, but `split_bc_pullbacks` tends to be very slow.
 
-function may_bc_forwards(cfg::C, f::F, args::Vararg{Any,N}) where {C,F,N}
+may_bc_forwards(cfg, f, args...) = false
+function may_bc_forwards(cfg::C, f::F, arg) where {C,F}
     Base.issingletontype(F) || return false
-    N==1 || return false  # Could weaken this to 1 differentiable
+    TA = _eltype(arg)
+    TA <: Real || return false
     cfg isa RuleConfig{>:HasForwardsMode} && return true  # allows frule_via_ad
-    TA = map(_eltype, args)
-    TF = Core.Compiler._return_type(frule, Tuple{C, Tuple{NoTangent, TA...}, F, TA...})
+    TF = Core.Compiler._return_type(frule, Tuple{C, Tuple{NoTangent, TA}, F, TA})
     return isconcretetype(TF) && TF <: Tuple
 end
 
@@ -344,6 +345,16 @@ function unbroadcast(x::T, dx_raw) where {T<:Tuple{Vararg{Any,N}}} where {N}
 end
 unbroadcast(x::Tuple, dx::AbstractZero) = dx
 
+# zero(::Tangent) is some Zero, which means sum(dx; dims) fails unless you do this:
+function Base.reducedim_init(
+    f::typeof(identity),
+    op::Union{typeof(+), typeof(Base.add_sum)},
+    A::AbstractArray{<:ChainRulesCore.AbstractTangent},
+    dims,
+)
+    return Base.reducedim_initarray(A, dims, ZeroTangent(), Union{ZeroTangent, eltype(A)})
+end
+
 # Scalar types
 
 unbroadcast(x::Number, dx) = ProjectTo(x)(sum(dx))

test/rulesets/Base/broadcast.jl

@@ -40,7 +40,7 @@ BT1 = Broadcast.BroadcastStyle(Tuple)
     @testset "split 3: forwards" begin
         # In test_helpers.jl, `flog` and `fstar` have only `frule`s defined, nothing else.
         test_rrule(copy∘broadcasted, BS1, flog, rand(3))
-        test_rrule(copy∘broadcasted, BS1, flog, rand(3) .+ im)
+        @test_skip test_rrule(copy∘broadcasted, BS1, flog, rand(3) .+ im)  # not OK, assumed analyticity, fixed in PR710
         # Also, `sin∘cos` may use this path as CFG uses frule_via_ad
         # TODO use different CFGs, https://github.com/JuliaDiff/ChainRulesTestUtils.jl/issues/255
     end
@@ -177,7 +177,21 @@ BT1 = Broadcast.BroadcastStyle(Tuple)
     end
 
     @testset "bugs" begin
-        @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)
+        @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)
+        end
+        @testset "unbroadcast with Matrix{Tangent}" begin  # https://github.com/JuliaDiff/ChainRules.jl/issues/708
+            x = Base.Fix1.(*, 1:3.0)
+            dx1 = [Tangent{Base.Fix1}(; x = i/2) for i in 1:3, _ in 1:1]
+            @test size(ChainRules.unbroadcast(x, dx1)) == size(x)
+            dx2 = [Tangent{Base.Fix1}(; x = i/j) for i in 1:3, j in 1:4]
+            @test size(ChainRules.unbroadcast(x, dx2)) == size(x)  # was an error, convert(::ZeroTangent, ::Tangent)
+            # sum(dx2; dims=2) isa Matrix{Union{ZeroTangent, Tangent{Base.Fix1...}}, ProjectTo copies this so that
+            # unbroadcast(x, dx2) isa Vector{Tangent{...}}, that's probably not ideal.
+
+            @test sum(dx2; dims=2)[end] == Tangent{Base.Fix1}(x = 6.25,)
+            @test sum(dx1) isa Tangent{Base.Fix1}  # no special code required
+        end
     end
 end