JuliaNLSolvers / Optim.jl

Did you test on CUDA devices too?
My quick tests resulted in errors

Scalar indexing is disallowed.

Invocation of getindex resulted in scalar indexing of a GPU array.

This is typically caused by calling an iterating implementation of a method.

Such implementations *do not* execute on the GPU, but very slowly on the CPU,

and therefore are only permitted from the REPL for prototyping purposes.

If you did intend to index this array, annotate the caller with @allowscalar.

    error(::String)@error.jl:35
    assertscalar(::String)@GPUArraysCore.jl:103
    getindex(::CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, ::Int64)@indexing.jl:48
    initial_state(::Optim.Adam{Float64, Optim.Flat}, ::Optim.Options{Float64, Nothing}, ::NLSolversBase.OnceDifferentiable{Float32, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}}, ::CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer})@adam.jl:48
    optimize(::NLSolversBase.OnceDifferentiable{Float32, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}}, ::CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, ::Optim.Adam{Float64, Optim.Flat}, ::Optim.Options{Float64, Nothing}, ::Optim.AdamState{CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, Float32, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, Vector{Float32}, Int64})@optimize.jl:36[inlined]
    var"#optimize#91"(::Bool, ::Symbol, ::typeof(Optim.optimize), ::NLSolversBase.InplaceObjective{Nothing, SwissVAMyKnife.var"#fg!#19"{SwissVAMyKnife.var"#L_VAM#18"{SwissVAMyKnife.var"#fwd2#12"{SwissVAMyKnife.var"#AS_abs2#11"{WaveOpticsPropagation.AngularSpectrum3{CUDA.CuArray{ComplexF32, 3, CUDA.Mem.DeviceBuffer}, Float32, CUDA.CUFFT.cCuFFTPlan{ComplexF32, -1, true, 3}}, Int64}, StepRangeLen{Float32, Float64, Float64, Int64}}, CUDA.CuArray{Bool, 3, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Bool, 3, CUDA.Mem.DeviceBuffer}, Tuple{Float32, Float32}, SwissVAMyKnife.var"#loss_f3#16"{typeof(abs2), CUDA.CuArray{Bool, 2, CUDA.Mem.DeviceBuffer}}}}, Nothing, Nothing, Nothing}, ::CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, ::Optim.Adam{Float64, Optim.Flat}, ::Optim.Options{Float64, Nothing})@interface.jl:143
    optimize(::NLSolversBase.InplaceObjective{Nothing, SwissVAMyKnife.var"#fg!#19"{SwissVAMyKnife.var"#L_VAM#18"{SwissVAMyKnife.var"#fwd2#12"{SwissVAMyKnife.var"#AS_abs2#11"{WaveOpticsPropagation.AngularSpectrum3{CUDA.CuArray{ComplexF32, 3, CUDA.Mem.DeviceBuffer}, Float32, CUDA.CUFFT.cCuFFTPlan{ComplexF32, -1, true, 3}}, Int64}, StepRangeLen{Float32, Float64, Float64, Int64}}, CUDA.CuArray{Bool, 3, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Bool, 3, CUDA.Mem.DeviceBuffer}, Tuple{Float32, Float32}, SwissVAMyKnife.var"#loss_f3#16"{typeof(abs2), CUDA.CuArray{Bool, 2, CUDA.Mem.DeviceBuffer}}}}, Nothing, Nothing, Nothing}, ::CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, ::Optim.Adam{Float64, Optim.Flat}, ::Optim.Options{Float64, Nothing})@interface.jl:139
    optimize_patterns(::CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}, ::SwissVAMyKnife.WaveOptics{CUDA.CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}, Float32, StepRangeLen{Float32, Float64, Float64, Int64}, Nothing}, ::SwissVAMyKnife.GradientBased{Optim.Adam{Float64, Optim.Flat}, Int64, typeof(abs2), Symbol, Float32})@optimization.jl:110
    macro expansion@[Local: 35](http://localhost:1234/edit?id=6ff68db6-be92-11ee-0059-63ce1cf34809#)[inlined]
    macro expansion@[Local: 621](http://localhost:1234/edit?id=6ff68db6-be92-11ee-0059-63ce1cf34809#)[inlined]
    top-level scope@[Local: 1](http://localhost:1234/edit?id=6ff68db6-be92-11ee-0059-63ce1cf34809#)[inlined]

No I did not. I suppose it's this

Optim.jl/src/multivariate/solvers/first_order/adam.jl

Line 48 in b0ba898

u = fill(zero(m[1]^2), length(m))

and I suppose maybe zeros(m) should suffice?

Maybe try this? #1075

Fixed in #1075

CUDA and Adam errors