conversion to pointer not defined for ConcretePJRTArray{Float32, 2, 1}

[grero]

I have a custom loss function for a GLM type model where I'm trying to fit an integer variable as Poisson distributed where the mean $$\lambda$$ is computed as $$\lambda = \exp(X'\beta)$$ where X is some input variable and $$\beta$$ are the coefficients I'm trying to estimate. Because $$X$$ lives in a non-euclidean space, I wanted to include a laplacian regularisation term $$X'\beta L \beta' X$$ where L is the Laplacian. However, when I try to compile my cost function in Reactant, I get the error in the title of this issue. The following MWE illustrates the issue

using Lux
using Reactant
using Distributions
using SpecialFunctions
using Random
using LinearAlgebra

function poisson_loss(η,n)
    lg = loggamma(n+1)
    -n*η + lg + exp(η)
end


function lossfunc(model, ps, st, (x,y,L))
    η,st_new = model(x, ps, st)
    ll = sum(poisson_loss.(η,y))
    ll2 = 0.01f0*η*L*η'
    ll = ll + ll2
    return ll, st_new, (;y_pred=η)
end 

function test()
    β = randn(Float32, 3)
    X = randn(Float32, 3, 1000)
    λ = exp.(X'*β)
    y = reshape(rand.(Poisson.(λ)),1,1000)
    @show size(y)

    # Simple model
    model = Lux.Dense(3,1)

    dev = reactant_device()

    d,n = size(X)
    rng = Random.default_rng()
    _ps,_st = Lux.setup(rng, model)
    ps,st = dev((_ps, _st))

    L = diagm(ones(Float32, 1000)) # dummy; this would be a graph laplacian
    (xe,ye,Le) = dev.((X,y,L))
    model_compiled = @compile model(xe, ps, Lux.testmode(st))

    ll,st_new,_ = lossfunc(model_compiled, ps, st, (xe,ye,Le))
end

julia> test()
size(y) = (1, 1000)
ERROR: conversion to pointer not defined for ConcretePJRTArray{Float32, 2, 1}
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] unsafe_convert(::Type{Ptr{Float32}}, a::ConcretePJRTArray{Float32, 2, 1})
    @ Base ./pointer.jl:68
  [3] gemm!(transA::Char, transB::Char, alpha::Float32, A::ConcretePJRTArray{…}, B::ConcretePJRTArray{…}, beta::Float32, C::ConcretePJRTArray{…})
    @ LinearAlgebra.BLAS ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/blas.jl:1644
  [4] gemm_wrapper!(C::ConcretePJRTArray{…}, tA::Char, tB::Char, A::ConcretePJRTArray{…}, B::ConcretePJRTArray{…}, _add::LinearAlgebra.MulAddMul{…})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:657
  [5] generic_matmatmul!
    @ ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:381 [inlined]
  [6] _mul!
    @ ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:287 [inlined]
  [7] mul!
    @ ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:285 [inlined]
  [8] mul!
    @ ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:253 [inlined]
  [9] *
    @ ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:124 [inlined]
 [10] _tri_matmul(A::ConcretePJRTArray{…}, B::ConcretePJRTArray{…}, C::Adjoint{…}, δ::Float32)
    @ LinearAlgebra ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:1144
 [11] *
    @ ~/.julia/juliaup/julia-1.11.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/matmul.jl:1199 [inlined]
 [12] lossfunc(model::Reactant.Compiler.Thunk{…}, ps::@NamedTuple{…}, st::@NamedTuple{}, ::Tuple{…})
    @ Main ~/.julia/dev/RecurrentNetworkModels/src/glmtest.jl:17
 [13] test()
    @ Main ~/.julia/dev/RecurrentNetworkModels/src/glmtest.jl:43
 [14] top-level scope
    @ REPL[3]:1
Some type information was truncated. Use `show(err)` to see complete types.

julia> 

[grero]

Sorry for got add version info.

julia> versioninfo()
Julia Version 1.11.4
Commit 8561cc3d68d (2025-03-10 11:36 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: macOS (arm64-apple-darwin24.0.0)
  CPU: 20 × Apple M1 Ultra
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, apple-m1)
Threads: 1 default, 0 interactive, 1 GC (on 16 virtual cores)
Environment:
  JULIA_EDITOR = code
  JULIA_VSCODE_REPL = 1

(RecurrentNetworkModels) pkg> st
Project RecurrentNetworkModels v0.10.0
Status `~/.julia/dev/RecurrentNetworkModels/Project.toml`
  [cdddcdb0] ChainRulesTestUtils v1.13.1
  [7da242da] Enzyme v0.13.108
⌅ [033835bb] JLD2 v0.5.15
  [b2108857] Lux v1.27.0
  [3bd65402] Optimisers v0.4.6
  [92933f4c] ProgressMeter v1.11.0
  [3c362404] Reactant v0.2.180
  [276daf66] SpecialFunctions v2.6.1
  [860ef19b] StableRNGs v1.0.4
  [aedffcd0] Static v1.3.1
⌅ [2913bbd2] StatsBase v0.34.5
  [e88e6eb3] Zygote v0.7.10
  [8bf52ea8] CRC32c v1.11.0
  [37e2e46d] LinearAlgebra v1.11.0
  [de0858da] Printf v1.11.0
  [9a3f8284] Random v1.11.0
  [8dfed614] Test v1.11.0
Info Packages marked with ⌅ have new versions available but compatibility constraints restrict them from upgrading. To see why use `status --outdated`

[avik-pal]

ll,st_new,_ = lossfunc(model_compiled, ps, st, (xe,ye,Le))

Try compiling the entire lossfunc instead of calling a partially compiled model from there

[wsmoses]

Generally conversion to pointer intentionally does not exist [as it does not exist on accelerators/shareded devices], so yeah the resolution here is compiling the code which computes on the data