Valentin Churavy

Ok this improves the situation for #2395 but is not yet enough ``` warning: linking module flags 'Dwarf Version': IDs have conflicting values ('i32 4' from start with 'i32 2'...

Note for self: ``` %18 = call {}*** @julia.get_pgcstack(), !dbg !42 %19 = inttoptr i64 ptrtoint (void ({ i8 addrspace(1)*, i64, [1 x i64], i64 } addrspace(11)*)* @julia_f_20562 to i64)...

Well this is new... ``` ERROR: LoadError: Cannot compile MethodInstance for Enzyme.var"##266".f(::Float64) for world 33088; method is only valid from world 33386 onwards Stacktrace: [1] error(s::String) @ Base ./error.jl:35 [2]...

> Does that mean you should set compat for GPUCompiler to v1.6? I did that in #2451 to triple-check that the release is backwards compatible.

Yay buildkite is green!

@wsmoses did you see my comment in https://github.com/EnzymeAD/Enzyme.jl/pull/2263#discussion_r2051753861 I.e., my thinking is that we are looking at the wrong type and that we should be looking at the parent layout...

So yeah we are constructing a `primal_job` that is hella wrong ``` primal_job = GPUCompiler.CompilerJob{GPUCompiler.NativeCompilerTarget, Enzyme.Compiler.PrimalCompilerParams}(MethodInstance for f(::CuDeviceVector{Float32, 1}), CompilerConfig for GPUCompiler.NativeCompilerTarget, 0x00000000000068a4) ```

This is the annouced change from https://github.com/JuliaGPU/GPUCompiler.jl/pull/668#issuecomment-2665477244 We are matching the wrong function and should be looking at the `config` instead. I will try rewriting all that tomorrow morning.

*sigh* missed the extension itself. I was looking at ImplicitDifferentiation for AD through non linear problems and adjoint models for implicit time integration.

Set's up the integration tests such that https://github.com/EnzymeAD/Enzyme.jl/pull/2309#issuecomment-2868907605 is not an issue, but highlights an issue with Bijectors and 1.11