Tullio and the nograd option

General Usage
3

I will probably have to create a MWE, which I am loathe to do because of the multiple interacting pieces. But here is an error occurring in Zygote, probably due to some kind of array content corruption. I ran with --inline=no and --bound-check=yes, and found that there was still inlining performed. Strange. I also tried with --optimization=0.

Here is the error, in case something jumps out:

ERROR: UndefVarError: 🐳 not defined
Stacktrace:
  [1] βˆ‡π’œπ’Έπ“‰!
    @ ~/.julia/packages/Tullio/NGyNM/src/macro.jl:1211 [inlined]
  [2] tile_halves(fun!::var"#βˆ‡π’œπ’Έπ“‰!#15", ::Type{Union{Nothing, Vector{Any}}}, As::Tuple{Vector{Float64}, Nothing, Nothing, Nothing, Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Tuple{Int64, Int64}}, Vector{Float64}}, Is::Tuple{}, Js::Tuple{UnitRange{Int64}, UnitRange{Int64}}, keep::Nothing, final::Nothing)
    @ Tullio ~/.julia/packages/Tullio/NGyNM/src/threads.jl:139
  [3] tile_halves(fun!::var"#βˆ‡π’œπ’Έπ“‰!#15", ::Type{Union{Nothing, Vector{Any}}}, As::Tuple{Vector{Float64}, Nothing, Nothing, Nothing, Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Tuple{Int64, Int64}}, Vector{Float64}}, Is::Tuple{}, Js::Tuple{UnitRange{Int64}, UnitRange{Int64}}, keep::Nothing, final::Nothing) (repeats 2 times)
    @ Tullio ~/.julia/packages/Tullio/NGyNM/src/threads.jl:146
  [4] tile_halves
    @ ~/.julia/packages/Tullio/NGyNM/src/threads.jl:136 [inlined]
  [5] βˆ‡threader(fun!::var"#βˆ‡π’œπ’Έπ“‰!#15", #unused#::Type{Union{Nothing, Vector{Any}}}, As::Tuple{Vector{Float64}, Nothing, Nothing, Nothing, Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Tuple{Int64, Int64}}, Vector{Float64}}, I0s::Tuple{}, J0s::Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, block::Int64)
    @ Tullio ~/.julia/packages/Tullio/NGyNM/src/threads.jl:103
  [6] (::var"#17#βˆ‡β„³π’Άπ“€β„―#16"{var"#βˆ‡π’œπ’Έπ“‰!#15"})(π›₯β„›::Vector{Float64}, β„›::Vector{Float64}, coef::Vector{Float64}, x::Vector{Float64}, I::Vector{Tuple{Int64, Int64}}, y::Vector{Float64})
    @ Main ~/.julia/packages/Tullio/NGyNM/src/macro.jl:1300
  [7] tullio_back
    @ ~/.julia/packages/Tullio/NGyNM/src/eval.jl:53 [inlined]
  [8] (::Zygote.ZBack{Tullio.var"#tullio_back#156"{Tullio.Eval{var"#ℳ𝒢𝓀ℯ#17"{var"#π’œπ’Έπ“‰!#14"}, var"#17#βˆ‡β„³π’Άπ“€β„―#16"{var"#βˆ‡π’œπ’Έπ“‰!#15"}}, Tuple{Vector{Float64}, Vector{Float64}, Vector{Tuple{Int64, Int64}}, Vector{Float64}}, Vector{Float64}}})(dy::Vector{Float64})
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/chainrules.jl:206
  [9] Pullback
    @ ~/src/2022/basic_UODE/custom_lux_layer/Polynomial.jl:25 [inlined]
 [10] (::typeof(βˆ‚(Ξ»)))(Ξ”::Vector{Float64})
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface2.jl:0
 [11] Pullback
    @ ~/src/2022/basic_UODE/custom_lux_layer/Polynomial_layer.jl:53 [inlined]
 [12] (::typeof(βˆ‚(Ξ»)))(Ξ”::Tuple{Vector{Float64}, Vector{Float64}, Nothing})
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface2.jl:0
 [13] Pullback
    @ ~/src/2022/basic_UODE/custom_lux_layer/testing_lux_layer_test.jl:78 [inlined]
 [14] Pullback
    @ ~/src/2022/basic_UODE/custom_lux_layer/testing_lux_layer_test.jl:132 [inlined]
 [15] (::typeof(βˆ‚(Ξ»)))(Ξ”::Float64)
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface2.jl:0
 [16] (::Zygote.var"#60#61"{typeof(βˆ‚(Ξ»))})(Ξ”::Float64)
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface.jl:45
 [17] gradient(f::Function, args::ComponentVector{Float32, Vector{Float32}, Tuple{Axis{(coeffs = ViewAxis(1:30, ShapedAxis((2, 15), NamedTuple())),)}}})
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface.jl:97
 [18] main(tstate::Lux.Training.TrainState{NamedTuple{(:coeffs,), Tuple{Matrix{Float32}}}, NamedTuple{(), Tuple{}}, NamedTuple{(:coeffs,), Tuple{Optimisers.Leaf{Adam{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}, Tuple{Float32, Float32}}}}}, Polylayer{typeof(Lux.rand32)}}, vjp::Lux.Training.ZygoteVJP, data::Tuple{Matrix{Float32}, Matrix{Float64}}, epochs::Int64)
    @ Main ~/src/2022/basic_UODE/custom_lux_layer/testing_lux_layer_test.jl:132
 [19] top-level scope
    @ ~/src/2022/basic_UODE/custom_lux_layer/testing_lux_layer_test.jl:156

I was able to resolve the mutation issue (which I still do not understand), via:

Ix = 0:degree
    Jx = 0:degree 
    #I = [(x=i,y=j) for i in Ix for j in Jx if i+j <= degree]
    I = [(i,j) for i in Ix for j in Jx if i+j <= degree]

	coef = randn(length(I))
     println("=====> Polynomial, length coef: ", length(coef))

    Poly(coef) = @tullio poly[i] := coef[j] * x[i]^I[j][1] * y[i]^I[j][2] grad=Dual nograd=x nograd=y nograd=I

I still do not understand why the nograd=I was not respected. I probably misunderstand something.

Thanks,