Distributions Wishart 1.12 error

[penelopeysm]

To be absolutely clear, I don't know if this is the only error with Wishart. This is at least the first.

using LinearAlgebra: LAPACK
using Enzyme

function div2(A::AbstractMatrix, B::AbstractVecOrMat)
    AA = copyto!(similar(A), A)
    lpt = LAPACK.getrf!(AA)
    BB = zeros(2, 2)
    copy!(BB, B)
    LAPACK.getrs!('N', lpt[1], lpt[2], BB)
    return sum(BB)
end

A = [1.0 0.2; 0.2 2.0]
B = [3.0 0.5; 0.5 4.0]
f(x) = div2(A, x)
f(B)

Enzyme.gradient(Enzyme.Reverse, Enzyme.Const(f), B)

#=
ERROR: EnzymeNoDerivativeError:
No augmented forward pass found for ejlstr$dgetrf_64_$@rpath/libblastrampoline.5.dylib
 at context:   call void @"ejlstr$dgetrf_64_$@rpath/libblastrampoline.5.dylib"(ptr nonnull %"new::RefValue52", ptr nonnull %"new::RefValue56", ptr %25, ptr nonnull %"new::RefValue61", ptr %29, ptr nonnull %"new::RefValue") #33 [ "jl_roots"(ptr addrspace(10) null, ptr addrspace(10) %31, ptr addrspace(10) null, ptr addrspace(10) %27, ptr addrspace(10) null, ptr addrspace(10) null) ], !dbg !122

Stacktrace:
 [1] #getrf!#1
   @ ~/.julia/juliaup/julia-1.12.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/src/lapack.jl:587

Failure within method:

var"#getrf!#1"(::Bool, ::typeof(LinearAlgebra.LAPACK.getrf!), ::Matrix{Float64}, ::Vector{Int64})
     @ LinearAlgebra.LAPACK ~/.julia/juliaup/julia-1.12.4+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/LinearAlgebra/src/lapack.jl:580
=#

julia> versioninfo()
Julia Version 1.12.4
Commit 01a2eadb047 (2026-01-06 16:56 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: macOS (arm64-apple-darwin24.0.0)
  CPU: 10 × Apple M1 Pro
  WORD_SIZE: 64
  LLVM: libLLVM-18.1.7 (ORCJIT, apple-m1)
  GC: Built with stock GC
Threads: 1 default, 1 interactive, 1 GC (on 8 virtual cores)

(test) pkg> st
Status `~/ppl/test/Project.toml`
  [7da242da] Enzyme v0.13.127

[wsmoses]

We shuoldn't ever hit this? We have custom rules for div!, / and friends that should make this unnecessary?

Can you pop up the MWE a bit (specifically identifying where it fails on 1.12, but passes on 1.11)?

[penelopeysm]

Interesting, then it might be some methods that changed in LinearAlgebra.jl. I'll try to retrace the steps a bit.

[penelopeysm]

using LinearAlgebra: tr, Cholesky, UpperTriangular, Symmetric, ldiv

A = Cholesky(UpperTriangular([1.0 0.2; 0.0 1.4]))
B = Symmetric([3.0 0.5; 0.5 4.0])
f(x) = tr(A \ x)
f(B)
Enzyme.gradient(Enzyme.Reverse, Enzyme.Const(f), B)

This passes on 1.11 but not on 1.12. If you change A \ x to ldiv(A, x) (that's the definition @ https://github.com/JuliaLang/LinearAlgebra.jl/blob/7249e662d66b53da6c9335ea748eac2750c2924f/src/LinearAlgebra.jl#L690 1.12, and https://github.com/JuliaLang/LinearAlgebra.jl/blob/72d1a8fdb683fda7d019b2c365e87c2691476f80/src/LinearAlgebra.jl#L631 1.11) then it works on both versions. I guess the rule for \ is not being triggered on 1.12 (or maybe the other way round?), though I didn't look into why.