Zygote causes segfaults on Julia 1.12+

[ChrisRackauckas-Claude]

Summary

Several tests that use Zygote for outer differentiation cause segfaults on Julia 1.12+. These tests have been temporarily skipped on Julia 1.12+ until upstream fixes are available.

Affected Test Files

1. test/ensembles.jl

• Uses AutoZygote() with Optimization.jl
• Causes segfault during ensemble solve differentiation on Julia 1.12+
• Workaround: Using Mooncake instead of Zygote on Julia 1.12+

2. test/parameter_initialization.jl

• Uses Zygote.gradient() and Zygote.ignore()
• Uses ZygoteVJP() for sensitivity algorithms
• Tests gradient computation through parameter initialization
• Workaround: Tests skipped on Julia 1.12+

3. test/mtk.jl

• Uses Zygote.gradient() and Zygote.ChainRules.ChainRulesCore.ignore_derivatives()
• Uses ZygoteVJP() for sensitivity algorithms
• Tests MTK integration with gradient computation
• Workaround: Tests skipped on Julia 1.12+

MWEs

Ensemble Segfault (Julia 1.12+, works on Julia 1.11)

using SciMLSensitivity, OrdinaryDiffEq, Optimization, OptimizationOptimisers
using Zygote

function prob_func(prob, i, repeat)
    remake(prob, u0 = 0.5 .+ i / 100 .* prob.u0)
end

function model(p)
    prob = ODEProblem((u, p, t) -> 1.01u .* p, p[1:1], (0.0, 1.0), p[2:2])
    ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)
    sim = solve(ensemble_prob, Tsit5(), EnsembleSerial(), saveat = 0.1, trajectories = 100)
    return sim
end

loss = (p, _) -> sum(abs2, 1.0 .- Array(model(p)))

# This causes segfault on Julia 1.12+
solve(
    OptimizationProblem(
        OptimizationFunction(loss, AutoZygote()),
        [1.0, 3.0]
    ),
    Adam(0.1); maxiters = 10
)

Related Issues

• Enzyme.jl support on Julia 1.12+ #1323 - Enzyme Julia 1.12+ issues
• FluxML/Zygote.jl - upstream Zygote issues

Current Status

Tests are temporarily skipped on Julia 1.12+ with @test_skip. Once upstream fixes are available (either in Zygote or by switching to alternative AD backends like Mooncake), these tests should be re-enabled.