could not deduce element type of value when `vcat`-ing array with abstract eltype

[penelopeysm]

The following errors on both forward- and reverse-mode on 1.11. (1.10 is fine for both modes)

import Enzyme: Enzyme, Forward, Reverse

function f(x)
    r = vcat(Any[1])
    return 0.0
end
params = rand(3)
f(params)

Enzyme.gradient(Forward, f, params)
Enzyme.gradient(Reverse, f, params)

This causes errors on any Turing model where there are two VarNames with the same symbol but different types, e.g. x.a and x.b since they get put together in the same vector and vcatted here.

Stack trace:

julia> Enzyme.gradient(Forward, f, params)
ERROR: Enzyme execution failed.
Enzyme: could not deduce element type of value within generic_memory_copyto of   %10 = phi {} addrspace(10)* [ %9, %L15 ], [ addrspacecast ({}* inttoptr (i64 4782741872 to {}*) to {} addrspace(10)*), %top ] within   call void @jl_genericmemory_copyto({} addrspace(10)* noundef %10, i64 %19, {} addrspace(10)* noundef %29, i64 %38, i64 %8) #13, !dbg !77

Stacktrace:
  [1] unsafe_copyto!
    @ ./genericmemory.jl:125 [inlined]
  [2] unsafe_copyto!
    @ ./array.jl:284 [inlined]
  [3] vcat
    @ ./array.jl:2234
  [4] f
    @ ./REPL[2]:2 [inlined]
  [5] fwddiffe3julia_f_8313wrap
    @ ./REPL[2]:0
  [6] macro expansion
    @ ~/.julia/packages/Enzyme/KdRPO/src/compiler.jl:5713 [inlined]
  [7] enzyme_call
    @ ~/.julia/packages/Enzyme/KdRPO/src/compiler.jl:5247 [inlined]
  [8] ForwardModeThunk
    @ ~/.julia/packages/Enzyme/KdRPO/src/compiler.jl:5138 [inlined]
  [9] autodiff
    @ ~/.julia/packages/Enzyme/KdRPO/src/Enzyme.jl:669 [inlined]
 [10] autodiff
    @ ~/.julia/packages/Enzyme/KdRPO/src/Enzyme.jl:538 [inlined]
 [11] macro expansion
    @ ~/.julia/packages/Enzyme/KdRPO/src/sugar.jl:671 [inlined]
 [12] gradient(::EnzymeCore.ForwardMode{…}, ::typeof(f), ::Vector{…}; chunk::Nothing, shadows::Tuple{…})
    @ Enzyme ~/.julia/packages/Enzyme/KdRPO/src/sugar.jl:573
 [13] gradient(::EnzymeCore.ForwardMode{false, EnzymeCore.FFIABI, false, false, false}, ::typeof(f), ::Vector{Float64})
    @ Enzyme ~/.julia/packages/Enzyme/KdRPO/src/sugar.jl:573
 [14] top-level scope
    @ REPL[5]:1
Some type information was truncated. Use `show(err)` to see complete types.

Environment:

julia> versioninfo()
Julia Version 1.11.7
Commit f2b3dbda30a (2025-09-08 12:10 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-16.0.6 (ORCJIT, apple-m1)
Threads: 1 default, 0 interactive, 1 GC (on 8 virtual cores)

(ppl) pkg> st
Status `~/ppl/Project.toml`
  [7da242da] Enzyme v0.13.75

[wsmoses]

yeah this is literally complaining that you have a type unstable array copy, is there any way for the original code to be made type stable?

[penelopeysm]

As far as Julia is concerned it is type stable though:

julia> using Test; @inferred vcat(Any[1])
1-element Vector{Any}:
 1

I understand that there's an underlying abstract type and that makes it harder for tools like Enzyme, but also as an average Julia user that's not really on me to fix.

I looked further and it only really happens with some combination of arguments, like vcat(Any[1]) fails but vcat(Any[1], Real[]) is alright.

As far as Turing is concerned marking it as inactive is one way, I think it might also be possible to avoid vcat, though I'll investigate separately :)

[wsmoses]

fixed by #2602