Merge pull request #814 from pxl-th/pxl-th/eachslice

Unthunk each element in `∇eachslice`

Project.toml

@@ -1,6 +1,6 @@
 name = "ChainRules"
 uuid = "082447d4-558c-5d27-93f4-14fc19e9eca2"
-version = "1.72.1"
+version = "1.72.2"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"

src/rulesets/Base/indexing.jl

@@ -262,20 +262,22 @@ end
 # Using Val(dim) here is worth a factor of 2 in this, on Julia 1.8-
 # @btime rrule(eachcol, $([1 2; 3 4]))[2]($([[10, 20], [30, 40]]))
 function ∇eachslice(dys_raw, x::AbstractArray, vd::Val{dim}) where {dim}
-    dys = unthunk(dys_raw)
+    dys = unthunk.(unthunk(dys_raw))
     i1 = findfirst(dy -> dy isa AbstractArray, dys)
     if i1 === nothing  # all slices are Zero!
         return _zero_fill!(similar(x, float(eltype(x)), axes(x)))
     end
+
     T = Base.promote_eltype(dys...)
     # The whole point of this gradient is that we can allocate one `dx` array:
     dx = similar(x, T, axes(x))
     for i in axes(x, dim)
         slice = selectdim(dx, dim, i)
-        if dys[i] isa AbstractZero
+        dy = dys[i]
+        if dy isa AbstractZero
             _zero_fill!(slice)  # Avoids this: copyto!([1,2,3], ZeroTangent()) == [0,2,3]
         else
-            copyto!(slice, dys[i])
+            copyto!(slice, dy)
         end
     end
     return ProjectTo(x)(dx)

test/rulesets/Base/indexing.jl

@@ -261,4 +261,17 @@ end
         Val(3);
         check_inferred=(VERSION >= v"1.7"),
     )
+
+    # eachslice: Make sure pulling back an array of thunks unthunks them and does not return all zeros.
+    x = ones(Float32, 3)
+    Δ = ones(Float32, 1)
+    _, norm_back = ChainRules.rrule(norm, x)
+    dx = norm_back(Δ)[2]
+    @test dx isa AbstractThunk
+
+    x = ones(Float32, 3, 1)
+    _, eachcol_back = ChainRules.rrule(eachcol, x)
+    Δ2 = [dx]
+    dx2 = eachcol_back(Δ2)[2]
+    @test all(dx2 .≉ 0f0)
 end