Only project sentinels in gradient

Project.toml

@@ -1,6 +1,6 @@
 name = "Zygote"
 uuid = "e88e6eb3-aa80-5325-afca-941959d7151f"
-version = "0.6.69"
+version = "0.7.0"
 
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"

src/compiler/interface.jl

@@ -114,8 +114,14 @@ sensitivity(y::AbstractArray) = error("Output is an array, so the gradient is no
 sensitivity(y) = error("Output should be scalar; gradients are not defined for output $(repr(y))")
 
 # Preserves output as tuple when gradients are collapsed
-_project_all(::NTuple{N}, ::Nothing) where {N} = ntuple(_ -> nothing, N)
-_project_all(x::Tuple, dx::Tuple) = map(_project, x, dx)
+_project_grad(::NTuple{N}, ::Nothing) where {N} = ntuple(_ -> nothing, N)
+_project_grad(x::Tuple, dx::Tuple) = map(_project_grad, x, dx)
+_project_grad(::Any, ::NoTangent) = nothing
+_project_grad(::Any, ::ZeroTangent) = nothing
+_project_grad(::Any, ::Nothing) = nothing
+_project_grad(::Any, dx::Any) = dx
+_project_grad(x::AbstractArray, dx::Tuple) = _project(x, dx)
+_project_grad(x::Any, dx::Base.RefValue) = _project(x, dx)
 
 """
     gradient(f, args...)
@@ -146,7 +152,7 @@ julia> gradient([7, 11], 0, 1) do x, y, d
 function gradient(f, args...)
   y, back = pullback(f, args...)
   grad = back(sensitivity(y))
-  return _project_all(args, grad)
+  return _project_grad(args, grad)
 end
 
 # Base.adjoint(f::Function) = x -> gradient(f, x)[1]  # piracy!
@@ -212,7 +218,7 @@ function withgradient(f, args...)
   else
     back(sensitivity(y))
   end
-  results = _project_all(args, grad)
+  results = _project_grad(args, grad)
   (val=y, grad=results)
 end
 
@@ -473,7 +479,7 @@ function pullback(f, ps::Params)
 end
 
 # No conversion required here
-_project_all(_, dx::Grads) = dx
+_project_grad(_, dx::Grads) = dx
 
 # Code Reflection
 

test/gradcheck.jl

@@ -2125,3 +2125,9 @@ end
   @test gradient(x -> @.(x * x * x), 2.0) == gradient(x -> x * (x * x), 2.0)
   @test gradient(x -> @.(3.0*x*2.0*x), 2.0) == gradient(x -> 6(x^2), 2.0)
 end
+
+@testset "Sparse input" begin
+  g1 = Zygote.gradient(sum, zeros(1,1))[1]
+  g2 = Zygote.gradient(sum, spzeros(1,1))[1]
+  @test g1 == g2
+end