Enzyme.jl support on Julia 1.12+

[ChrisRackauckas-Claude]

Summary

Enzyme.jl does not currently support Julia 1.12+. This issue tracks the workarounds implemented and items to revisit once Enzyme v1.12 support lands.

Upstream Issue

See: EnzymeAD/Enzyme.jl#2699

Current Status

Enzyme.jl displays the following warning on Julia 1.12:

Warning: Enzyme.jl support for Julia 1.12 is presently in progress.
         For the time being we recommend using 1.11 or LTS (1.10).

Affected Areas with MWEs

All MWEs below fail on Julia 1.12+ but work on Julia 1.11 and earlier.

1. automatic_sensealg_choice - EnzymeVJP not selected (Julia 1.12+)

The automatic_sensealg_choice function tries Enzyme first for in-place VJP computation. On Julia 1.12+, this fails and falls back to ReverseDiffVJP.

MWE (fails on Julia 1.12+, works on Julia 1.11):

using Enzyme, OrdinaryDiffEq

function f!(du, u, p, t)
    du[1] = p[1] * u[1] - p[2] * u[1] * u[2]
    du[2] = -p[3] * u[2] + p[4] * u[1] * u[2]
    return nothing
end

u0 = [1.0, 1.0]
p = [1.5, 1.0, 3.0, 1.0]
du = zero(u0)
t0 = 0.0

# This fails on Julia 1.12+, works on Julia 1.11
Enzyme.autodiff(
    Enzyme.Reverse, f!,
    Enzyme.Duplicated(du, copy(u0)),
    Enzyme.Duplicated(copy(u0), zero(u0)),
    Enzyme.Duplicated(copy(p), zero(p)),
    Enzyme.Const(t0)
)

Test: test/automatic_sensealg_choice.jl now marks EnzymeVJP as @test_broken on Julia 1.12+.

2. NonlinearProblem gradient computation (Julia 1.12+)

The NonlinearProblem tests in test/steady_state.jl use Enzyme for gradient computation.

MWE (fails on Julia 1.12+, works on Julia 1.11):

using Enzyme, NonlinearSolve, SciMLSensitivity

u0 = [0.0]
p = [2.0, 1.0]
prob = NonlinearProblem((du, u, p) -> du[1] = u[1] - p[1] + p[2], u0, p)

function test_loss(p, prob, alg)
    _prob = remake(prob, p = p)
    sol = sum(solve(_prob, alg, sensealg = SteadyStateAdjoint(autojacvec = ReverseDiffVJP())))
    return sol
end

# This fails on Julia 1.12+, works on Julia 1.11
dp = Enzyme.make_zero(p)
dprob = Enzyme.make_zero(prob)
Enzyme.autodiff(
    Enzyme.Reverse, test_loss, Enzyme.Active,
    Enzyme.Duplicated(p, dp),
    Enzyme.Duplicated(prob, dprob),
    Enzyme.Const(NewtonRaphson())
)

Test: test/steady_state.jl NonlinearProblem Enzyme tests are now conditionally skipped on Julia 1.12+.

3. Lux Neural Network ODE (Julia 1.12+)

The automatic sensealg choice test uses a Lux neural network ODE which also triggers Enzyme failures.

MWE (fails on Julia 1.12+, works on Julia 1.11):

using Lux, ComponentArrays, OrdinaryDiffEq, Enzyme, Random

rng = Random.default_rng()
ann = Lux.Chain(Lux.Dense(1, 32, tanh), Lux.Dense(32, 32, tanh), Lux.Dense(32, 1))
ps, st = Lux.setup(rng, ann)
p = ComponentArray(ps)
θ, ax = getdata(p), getaxes(p)

function dxdt!(dx, x, p, t)
    ps = ComponentArray(p, ax)
    dx[1] = x[2]
    dx[2] = first(ann([t], ps, st))[1]^3
    return nothing
end

x0 = [-4.0f0, 0.0f0]
du = zero(x0)
t0 = 0.0f0

# This fails on Julia 1.12+, works on Julia 1.11
Enzyme.autodiff(
    Enzyme.Reverse, dxdt!,
    Enzyme.Duplicated(du, copy(x0)),
    Enzyme.Duplicated(copy(x0), zero(x0)),
    Enzyme.Duplicated(copy(θ), zero(θ)),
    Enzyme.Const(t0)
)

TODO when Enzyme 1.12 support lands

• Remove @test_broken from automatic_sensealg_choice.jl
• Re-enable Enzyme tests in steady_state.jl NonlinearProblem section
• Consider switching back to EnzymeVJP as default autojacvec on Julia 1.12+
• Review if any other tests need to use Enzyme instead of fallback backends

Related Changes

• PR Refactor AD backend test suite with matrix-based approach #1322: Refactor AD backend test suite with Julia 1.12 compatibility fixes

[ChrisRackauckas-Claude]

Status Update (2026-03-20)

Retested all three MWEs on Julia 1.12.4 with Enzyme v0.13.132. The upstream Enzyme Julia 1.12 roadmap issue (EnzymeAD/Enzyme.jl#2699) is now closed.

Results

MWE	Julia 1.12.4	Julia 1.11.9	Status
#1 Basic Enzyme autodiff (Lotka-Volterra)	PASS	PASS	Fixed
#2 Enzyme + NonlinearSolve gradient	SEGFAULT	PASS	Still broken
#3 Lux Neural Network ODE	PASS	PASS	Fixed

Details

• MWE Morris Global Sensitivity Analysis #1 and Update CI URLs to point to new caching infrastructure #3 now work on Julia 1.12. The @test_broken in automatic_sensealg_choice.jl and the Lux neural ODE test can be reverted.
• MWE Add Travis CI & AppVeyor testing of v0.6, add test badges to README.md #2 still crashes with a segfault in Enzyme's LLVM optimization pass (post_optimize! at compiler/optimize.jl:464). This happens even with a simple Enzyme.autodiff over NonlinearSolve.solve without SciMLSensitivity in the loop — it's an Enzyme + NonlinearSolve interaction issue on 1.12.

Remaining TODO

• Remove @test_broken from automatic_sensealg_choice.jl (can now be done)
• Re-enable Enzyme tests in steady_state.jl NonlinearProblem section — blocked on upstream: filed EnzymeAD/Enzyme.jl issue for the NonlinearSolve segfault
• Consider switching back to EnzymeVJP as default autojacvec on Julia 1.12+ (basic EnzymeVJP works now)
• Review if any other tests need to use Enzyme instead of fallback backends

Scope of this issue is now narrowed to the NonlinearProblem/Enzyme interaction on Julia 1.12.

[ChrisRackauckas-Claude]

Filed upstream: EnzymeAD/Enzyme.jl#3010 for the NonlinearSolve segfault on Julia 1.12.

[wsmoses]

@ChrisRackauckas-Claude please re-evaluate failures and re enable all tests with enzyme

[ChrisRackauckas-Claude]

Re-evaluation against current package set (2026-05-20)

Re-ran the three MWEs plus all @test_broken / file-level skips that mention Enzyme.

Env: Julia 1.12.6, Enzyme 0.13.147, NonlinearSolve 4.19.1, Lux 1.31.4, ModelingToolkit 11.26.3, SciMLBase 3.13.0.

MWE status

MWE	Julia 1.12.6	Notes
#1 Basic Enzyme + Lotka-Volterra	PASS	confirmed
#2 Enzyme over NonlinearSolve.solve	FAIL — bus error in LLVMRunJuliaPasses	EnzymeAD/Enzyme.jl#3010 still open
#3 Lux Neural Network ODE + Enzyme	PASS	confirmed

Tests run with file-level VERSION >= v"1.12" skip stripped

File	Result on 1.12.6	Skip claim	Verdict
test/adjoint_shapes.jl	2/2 Pass	"Enzyme has issues"	Remove skip — works (path routes through inplace_vjp→EnzymeVJP)
test/enzyme_closure.jl	4 Pass, 4 Broken	"Enzyme compatibility"	Remove skip — all EnzymeVJP cases pass; the 4 inner @test_broken are ReactantVJP-specific
test/automatic_sensealg_choice.jl	2/2 Pass	—	already a hard @test sensealg.autojacvec isa EnzymeVJP, currently passes; no change needed
test/layers.jl	6/6 Pass	"Zygote compatibility"	Remove skip — works
test/layers_dde.jl	3/3 Pass	"Zygote compatibility"	Remove skip — works
test/layers_sde.jl	2/2 Pass	"Zygote compatibility"	Remove skip — works
test/ensembles.jl	Segfault at line 10 (during using)	"Zygote compatibility"	Keep skip (not Enzyme-related)
test/sde_neural.jl	Segfault at using Zygote	"Zygote compatibility"	Keep skip (not Enzyme-related)
test/distributed.jl	Worker-deserialize failure under SafeTestsets; under @testset it'd need to be re-run — see note	"Zygote compatibility"	Out of scope for this issue (no Enzyme path)

Tests run as-is (inline conditionals)

File	Result	Notes
test/scc_nonlinearsolve.jl	7 Pass, 1 Broken	Broken is the inline @test_skip for Enzyme on Julia 1.11+. Manually retried Enzyme.gradient over the SCC loss → still segfaults on 1.12.6 in LLVMRunJuliaPasses. Keep skip. Upstream is Enzyme.jl#3021 + #3010 family.
test/parameter_initialization.jl	2 Pass, 1 Broken	Broken is Enzyme.gradient over remake(prob; p = repack(...))+solve with GaussAdjoint(autojacvec=EnzymeVJP()). Goes through NonlinearSolve init → hits Enzyme#3010. Keep @test_broken.
test/mtk.jl	1 Pass, 2 Broken	Both broken are Enzyme.gradient through DAE init via MTK + NonlinearSolve. Keep @test_broken.
test/desauty_dae_mwe.jl	17 Pass, 7 Broken	The 2 "Enzyme through init" broken cases hit the same NonlinearSolve+Enzyme blocker. Keep @test_broken. Other 5 broken are ForwardDiff/Tracker-related and out of scope here.
test/steady_state.jl	54 Pass, 2 Fail (Fails are iip concrete_solve rtol, unrelated to Enzyme), 16/16 in NonlinearProblem section	The Julia<1.12 Enzyme NonlinearProblem block is still gated correctly — it would segfault if exercised. Keep gating.

Concrete action items

Can be removed in a follow-up PR:

1. The if VERSION >= v"1.12" … @test_skip false … else wrapper around test/adjoint_shapes.jl — Enzyme MWE-style path works on 1.12.
2. The same wrapper around test/enzyme_closure.jl — all EnzymeVJP assertions pass on 1.12; the existing inner @test_brokens are ReactantVJP, leave those.
3. The same wrapper around test/layers.jl, test/layers_dde.jl, test/layers_sde.jl — these are Zygote-attributed but the Zygote 0.7 series now works on 1.12 for these cases.

Must stay until Enzyme.jl#3010 closes (Enzyme + NonlinearSolve on 1.12 segfaults):

1. Inline VERSION < v"1.11" gate in test/scc_nonlinearsolve.jl.
2. The Enzyme NonlinearProblem branch in test/steady_state.jl (already inline-gated).
3. @test_broken blocks for Enzyme.gradient-through-init in test/mtk.jl, test/desauty_dae_mwe.jl ("Enzyme through init" testset), test/parameter_initialization.jl ("Adjoint through Prob (Enzyme)" testset).

Out of scope for this issue (separate upstream blockers):

• test/ensembles.jl, test/sde_neural.jl, test/prob_kwargs.jl (Zygote/codegen on 1.12; not Enzyme).
• test/concrete_solve_derivatives.jl Tracker @test_brokens (issue Tracker gradient tests fail on Julia 1.12+ #1331).
• test/enzyme_closure.jl ReactantVJP @test_brokens (Reactant _copyto! limitation).

Net effect

Once Enzyme.jl#3010 lands, the remaining work is removing the four @test_broken/skip blocks gated on the NonlinearSolve+Enzyme interaction. Everything else this issue tracked is already either fixed in current Enzyme or unrelated to Enzyme.

[ChrisRackauckas-Claude]

Correction (2026-05-20)

In PR #1445 I attempted to remove the file-level 1.12 skips on test/adjoint_shapes.jl and test/enzyme_closure.jl based on my local Julia 1.12.6 runs (which passed). CI surfaced an LLVM/GC segfault in both files (LLVMRunJuliaPasses during Enzyme compilation, Core1 and Core6) — same code path, different LLVM JIT outcome between my local machine and the CI runners. My local "pass" was flaky.

I've reverted those two file unmaskings on the PR. Only test/layers.jl, test/layers_dde.jl, test/layers_sde.jl (the Zygote-attributed group) survive — those passed both locally and on CI's Core4 job.

Practical implication for this issue: adjoint_shapes.jl and enzyme_closure.jl should stay listed as still-broken on 1.12, alongside the NonlinearSolve/Enzyme @test_brokens already tracked. The LLVM/GC crash is a different blocker class from EnzymeAD/Enzyme.jl#3010 (which is about NonlinearSolve's mapreduce/break pattern). Worth a separate upstream issue if it can be reduced to a clean MWE; the enzyme_closure BacksolveAdjoint+EnzymeVJP adjoint_sensitivities call at line 148 is the failing entry point.

[ChrisRackauckas-Claude]

Status update (2026-05-21) — Enzyme 0.13.148 + FunctionWrappersWrappers 1.9.0

Big wins from upstream Enzyme.jl 0.13.148 (released 2026-05-21), which closes:

• EnzymeAD/Enzyme.jl#3012 — LLVM GC invariant violation with in-place NonlinearProblem on Julia 1.12
• EnzymeAD/Enzyme.jl#3026 — Intermittent GC segfault with SCCNonlinearProblem + custom rule

…and FunctionWrappersWrappers 1.9.0 (which adds Enzyme-aware rules / typeasserts).

Verified on Julia 1.12.6 + Enzyme 0.13.148

Item	Before	After
MWE2: Enzyme.autodiff over NonlinearSolve.solve	bus error	PASS
Enzyme.gradient over SCCNonlinearProblem	segfault	PASS
steady_state.jl enzyme_gradient block (NewtonRaphson/TrustRegion/Klement, iip+oop)	gated off (1.12)	4/4 PASS
steady_state.jl SteadyStateAdjoint(autojacvec = EnzymeVJP()) through adjoint_sensitivities	gated off (1.12)	PASS
Full test/steady_state.jl	54/2 fail	63 Pass / 1 Broken / 0 Fail

PRs landing the unmaskings

• Enable Enzyme SCC test on Julia 1.11+ (Enzyme 0.13.148) #1449 — drop @test_skip for Enzyme.gradient SCC on Julia 1.11+
• Enable Enzyme tests in steady_state.jl on Julia 1.12+ (Enzyme 0.13.148) #1450 — set ENZYME_AVAILABLE = true and lift using Enzyme out of the VERSION < v"1.12" gate in test/steady_state.jl

Still broken (tracked separately under #1359)

The three @test_broken Enzyme.gradient blocks that route through MTK's remake path still fail:

• test/mtk.jl — Enzyme.gradient through DAE init across 16 setups
• test/parameter_initialization.jl — "Adjoint through Prob (Enzyme)" testset
• test/desauty_dae_mwe.jl — "Enzyme through init" testset

These hit a different failure mode (EnzymeRuntimeActivityError / EnzymeMutabilityException from MTK's remake / SciMLStructures.replace), not the NonlinearSolve segfault that Enzyme 0.13.148 fixed. Tracked at #1359 ("Enzyme: Cannot differentiate through MTK DAE solve"). An investigation is in progress to characterize the remaining root cause on the current package stack.

Unrelated 1.12 items (out of scope for this issue)

• adjoint_shapes.jl, enzyme_closure.jl — file-level skips: local Julia 1.12 tests pass; CI's LLVM JIT was flaky on the previous Enzyme version. Need a CI re-test on Enzyme 0.13.148 before unmasking.
• test/ensembles.jl, test/sde_neural.jl — Zygote 1.12 issues, not Enzyme; tracked elsewhere.

This issue can stay open as the umbrella tracker, or it can be closed in favor of #1359 once we confirm the MTK piece is the only remaining blocker.