You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Re-enable the moonlander / quadrotor :manual ADNLP tests (blocked on ADNLPModels#383) #626
test/ci/test_all_ocp.jl had its :moonlander and :quadrotor testsets disabled
in #622 (commit on branch v1.1.1). Re-enable them once the upstream cause is fixed.
Why they were disabled
Both testsets ran test_problem(moonlander(); adnlp_backend=:manual) (the :quadrotor
one via a long-standing copy-paste slip — it never actually tested quadrotor()).
moonlander is a free-final-time OCP with a rotation-matrix dynamics: cos(θ) / sin(θ)
with θ = x₅ an unbounded state. The :manual (and :optimized) ADNLP backend build
a SparseReverseADHessian, which records its ReverseDiff tape on an uninitialised Vector{ForwardDiff.Dual} and executes the model function on that memory. CTDirect's NLP
variable bounds are -Inf .* ones(nvar); once freed, that storage is recycled into the
uninitialised tape input, so θ is seeded with -Inf and cos(-Inf) throws DomainError: sincos(x) is only defined for finite xbefore any solve.
Deterministic on the Julia 1.12 CI runners, flaky on 1.10 — it depends only on which
dependency versions each CI run resolves (surfaced by the CTParser 0.9 / CTModels 0.18
bump in #622, not caused by it).
Moonlander tests now seem to fail due to local solutions, instead of the above error.
Behavior is not always deterministic: ubuntu tests seem fine, while windows can either pass or fail. Mac runners are apparently too busy at the moment.
I suspect we could ensure proper convergence by choosing a suitable initial guess (not necessarily a complex one).
However, maybe we want to keep this case to investigate the AD part a bit more, and more precisely why the manual patterns seems to give a better convergence than ADNLP built-in detection, despite being 'less' sparse ?
Activity
test: disable moonlander / quadrotor :manual ADNLP tests (#626)
ocots commentedon Aug 28, 2026
I have commented truck trailer too.
PierreMartinon commentedon Oct 15, 2026
Hi @ocots
I have re-enabled the truck trailer and quadrotor test problems that look fine with manual AD backend. Moonlander still fails. PR #641
ocots commentedon Oct 15, 2026
What is the error for the Moonlander problem?
PierreMartinon commentedon Oct 16, 2026
Moonlander tests now seem to fail due to local solutions, instead of the above error.
Behavior is not always deterministic: ubuntu tests seem fine, while windows can either pass or fail. Mac runners are apparently too busy at the moment.
I suspect we could ensure proper convergence by choosing a suitable initial guess (not necessarily a complex one).
However, maybe we want to keep this case to investigate the AD part a bit more, and more precisely why the manual patterns seems to give a better convergence than ADNLP built-in detection, despite being 'less' sparse ?
Ps. This has little to do with #383 :D