Repository navigation
Highlights
-
Correctness. This release fixes a series of cases where a solve could report a certified (
gap_certified=True) but wrong optimum. Several came from an outside soundness audit, among them:- the sign of nested powers in products (#2200),
- the odd-power hull near zero (#2198),
sin/cosenclosures after argument reduction (#2190),- exponents and Hessians that were close to special values but treated as exactly special (#2191, #2192),
- exact collection of cancelling like terms (#2090, #2097–#2100),
- the bit semantics of
.nlcomplementarity flags (#2056).
Each fix ships with a regression test.
-
Complementarity and ReLU networks.
- New
relu_complementaritystrategy. discopt.mpec.solve_mpec(method="native")andsolve_mpec_multistart.- Partition-based (Tsay et al.) and extended convex-hull (Anderson et al.) ReLU encodings, plus Anderson's ideal cuts.
- Smooth ReLU-like activations (softplus, squareplus, Swish, GELU) with a sharpness
beta.
- New
-
Infeasibility diagnosis for MILPs.
compute_iis(include_integrality=True)can name integrality as part of the conflict. It uses SAT, parity and exact certificates, andportfolio=runs several deletion orders in parallel. The browser page explains an IIS. -
GDP. Feasibility-first fallback, the LOA proposer's proved bound, and certification of GDPlib
syngas(hull and big-M) andcstr(big-M). -
Defaults.
- The open branch-and-bound frontier now has a memory budget (
DISCOPT_FRONTIER_MEMORY_MB, 2 GiB). pounce-solver>=0.16.
- The open branch-and-bound frontier now has a memory budget (
-
Tooling. Relaxation soundness fuzzers (
scripts/fuzz/), and the developer/adversarycommand now has a soundness-fuzzing mode.
Known issues: on macOS/arm64, MINLPLib nous1 no longer certifies within 150 s with the new default DISCOPT_DUAL_PROOF_PROPAGATION (#2202); importing cyipopt and torch in one macOS process aborts (duplicate libomp). Details are below.
Added
-
Faster, smaller irreducible infeasible subsystems for MILPs (#2093, #2108,
#2110–#2114, #2118, #2122).compute_iis(include_integrality=True)treats an
integrality restriction as an IIS member and gains several candidate generators
and confirmation routes, all decided exactly: an in-house CDCL SAT solver
(discopt._rust.PySatSolver) for clause, cardinality and pseudo-Boolean rows; a
GF(2) parity certificate for integer equality rows; a staged-oracle ladder for
big-M models; and an exact decider that projects continuous columns out first
(about 2000x faster on MIPcc27otra).portfolio=/workers=run several
deletion orders in parallel against onetotal_time_limitand return the
smallest result; under a deadline the base set is confirmed up front, so a run
that times out still returns a valid (if large) subsystem. -
The browser page shows and explains an IIS (#2109): member table, a
constraint–variable graph, and links back to the declaring line. -
Smooth ReLU-like activations with a sharpness parameter (#2135, #2139).
DenseLayer.betaforSOFTPLUS,SQUAREPLUS,SWISH(alias"silu") and
GELU; the torch reader keepsSoftplus(beta). Each activation is defined once
(value, exact interval image, symbolic form) indiscopt/ml/activations.py. -
Stronger MILP encodings for trained ReLU networks (#2134, #2140):
strategy="relu_partition"(Tsay et al. 2021) andstrategy="relu_extended"
(Anderson et al. 2020), plus theAndersonCutSeparatorideal-cut callback. -
GDP primal heuristics. A feasibility-first fallback for the disjunct
configuration search, default ON (DISCOPT_GDP_CONFIG_FEAS_FALLBACK; #2144,
#2173); the root LOA proposer's proved bound now reaches the tree, so GDPlib
syngascertifies under hull and big-M at 120 s (#2143, #2166). -
Relaxation soundness fuzzers (#2200):
scripts/fuzz/fuzz_relax_atoms.pyand
scripts/fuzz/fuzz_relax_trees.pysearch for a point that violates a relaxation
bound and confirm it in mpmath. The developer/adversarycommand
(discopt-dev adversary) is tracked in the repository again and documents the
method as its soundness-fuzzing mode. -
The benchmark harness refuses to run on a broken external solver (#2186). Each
configured solver must solve a probe with a known optimum first; a missing or
demo-licensed binary no longer shows up as a solver column of timeouts. -
A
discopt-expertsubagent, the generalist beside the domain
specialists. It knows the codebase as a whole: the repository map, the order
Model.solveandsolve_modelroute a model, how astatus,
algorithm_routeorgap_certifiedtraces back to code, how incumbents are
verified, how to debug a solve, and the rules inCLAUDE.mdfor changing
discopt. It routes each question to the specialist that owns it, and for
POUNCE internals to pounce'spounce-expert(installed from pounce). It reads
the source rather than answering from memory. It complements thediscopt
skill, which covers using discopt. Ships withdiscopt install-skillsand
uvx knowhow add jkitchin/discopt. -
relu_complementaritystrategy for ReLU networks (#2167, slice 1).
NNFormulation(m, net, strategy="relu_complementarity")encodes each unstable ReLU
neuron asy - s == zwith a declared complementarity relation0 <= y ⊥ s >= 0
(new roleComplementarityRole.FROM_ACTIVATION,parent= layer,index= neuron)
instead of an indicator binary and big-M rows; the relations are onnn.relations.
They are registered but not lowered, soModel.solve()refuses until
solve_mpec(m, nn.relations, method=...)picks"sos1"/"gdp"(exact; reproduces
relu_bigm's certified objective) or"scholtes"(local). The strategy needs no
input_bounds. Checked bypython/tests/test_2167_relu_complementarity.py. -
Multistart local mode for complementarity models (#2167, slice 2).
discopt.mpec.solve_mpec_multistart(model, pairs, starts)and
NNFormulation.solve_local(n_starts=...), which uses forward-pass warm starts from
NNFormulation.local_starts. Each start runs a Scholtes continuation and then an
active-set polish that leaves the point exactly complementary. A point becomes
result.xonly afteraccept_local_incumbentverifies it. The result is always
local (local_optimal/local_infeasible, no bound), andresult.multistart
logs every start (a newSolveResultfield). -
Native complementarity lowering and branch (#2167, slice 3).
solve_mpec(method="native")encodes each pair asf·g <= 0overf, g >= 0.
This is exact, has no binaries, and needs no finite bounds. With
DISCOPT_COMPL_BRANCH=1(default off) the spatial tree splits a violated pair
intoub(f)=0/ub(g)=0before its other branching rules. -
The GAMS link reads a
discopt.optoption file. Withoptfile = 1, the link
reads GAMS's usual solver option file and accepts three keys:toland
constr_viol_tol, forwarded to every POUNCE NLP the solve runs (as
Model.solve(pounce_options=...)), andfrontier_memory_mb, the open-frontier memory
budget for that solve only. An unknown key, a bad value, a repeated key or a missing
file ends the job as9 Setup Failurewith the offending line in the log, rather
than solving at defaults the job did not ask for. Previously the link logged that
optfilewas ignored. Seedocs/gams_solver_link.md. Checked by
python/tests/test_gams_option_file.py. -
The browser page explains infeasibility. A new seventh example, an infeasible
two-mill blending LP, callscompute_iis(). AnyIISResulta script leaves in a
top-level variable is now drawn below the output: a summary (member counts, model
size, whether the IIS is proven minimal), a table of members with the variable
elements each touches and a button that selects its declaring line in the editor, and
a constraint–variable graph that also shows, dashed, the constraints outside the IIS
that share its variables. The table holds everything the graph shows; membership is
never shown by colour alone. Each run now gets a fresh namespace, so an earlier run's
IIS is never redrawn. Data:crates/discopt-wasm/web/iis_view.py(served with the
page, not in the wheel); drawing:web/iis_render.js. Checked by
python/tests/test_wasm_iis_view.py, the Pyodide harness,node --test, and a new
Chromium harness (crates/discopt-wasm/tests/iis_page.mjs, axe in both themes,
320 px reflow) in theBrowser pageworkflow. The example dropdown no longer scrolls
the page sideways on a 320 px screen. -
Exact lowering of box-constrained MCP relations (#2058).
F(x) ⊥ z ∈ [l, u]
(Model.mcp,mpec.box_mcp) used to be represented but refused at solve time.
Every exact lowering (mpec.gdp,mpec.sos1) and Scholtes now reduce it to the
existing NCP halves: a fixed box gives rows only; a one-sided box gives one
half; a free variable givesF == 0; a two-sided box splits
F == w⁺ - w⁻with(w⁺, z - l)and(w⁻, u - z), wherew±are bounded
by the interval enclosure ofF. A residual with a non-finite enclosure, an
empty box, array bounds or declared residual bounds other than the box's own
are refused with aValueError, as is a box pinned at±inf; a refused call
(any relation of it, any element) leaves no rows behind and registers nothing.
w±are enclosed overzrestricted to the relation's own[l, u].
Model.mcp(..., method="gdp"|"sos1")lowers on the spot.
Fixed
Correctness first. Every item below is a case where a solve could report
optimal (or infeasible) with gap_certified=True and a wrong objective or
bound; each has a regression test that fails before its fix.
-
Nested powers inside a product keep their sign (#2200). The canonicalizer
folded(b^q)^etob^(q*e)and(x*y)^etox^e*y^ewith no sign check, so
x*(x**2)**0.5becamex**2(true value-x**2forx < 0) and
1/(x**2)**0.5became1/x. Such a factor now folds as|b|, the exact identity
wherever the expression is real. Found by the new relaxation soundness fuzzers. -
The odd-power hull no longer uses the secant when the tangency is near 0
(#2198, #2199). Fort**pwith oddpover a box whose lower end is a tiny
negative number, the tangent search started at1e-12and missed a root near
1e-16; the secant it fell back to lies above the graph. The trigger is
ordinary:(1 - x)**3withxin[0, 1]has the rounded enclosure
[-2.2e-16, 1]. -
Sound
sin/cos/tanenclosures after argument reduction (#2190, #2193).
Reducing by the rounded2*piwas not enclosed, sosin(1e15)came back as the
point 0.850 (true 0.858). Endpoint values now come from the unreduced
arguments, and extrema near a reduction boundary are counted conservatively. The
same rounding was fixed at four other trig sites. -
Exponents and Hessians that are nearly special are no longer treated as
exactly special (#2191, #2192, #2195; #2194, #2197).np.isclosedecided
whether an exponent was 0, 1 or an integer, and|H| <= 1e-300counted as a zero
Hessian, sox**(1 - 2**-17)could be certified affine or a nonconvex model
convex. Exact tests replace both, in the convexity layer and on the LP/QP routing
path. -
Like terms are collected exactly (#2090, #2097, #2098, #2099, #2100; #2094,
#2106).(2**53*x + x) - 2**53*xis exactlyx, but every float consumer read
it as0*x. A row whose coefficients cancel catastrophically is now rewritten
from an exactFractioncollection, including array rows, sums that are exactly
zero beside a large term, andverify_pointcalled beforesolve. Collecting
terms of a shared sub-expression no longer walks it as a tree (exponential time
on deep sharing). -
.nlcomplementarity rows use AMPL's flag semantics (#2056, #2057). The
bound bits were swapped and the relation was taken againstxinstead of its
bound. -
Binary bounds are rounded after tightening (#2081, #2087), so a round-off of
1.6e-13can no longer invert a selector's box; LOA no longer stops on an
unresolved re-proposal (#2082). -
The default GDP hull path takes the Neumaier–Shcherbina safe bound on every
ill-conditioned OBBT probe (#2088, in #2086), which removes a false certificate on
clay0303; spatial OBBT edges are safe bounds, and a box whose bounds cross by
round-off is widened instead of panicking. -
compute_iisno longer returns a feasible subsystem (#2093). Finite bounds
of a partly free vector were applied in every probe but never reported; size-1
variables also kept their shape on the POUNCE tape again. -
Fewer sound certificates declined. The HiGHS MILP route accepts the
fixed-integer LP's Neumaier–Shcherbina verdict (#2183, #2188) and certifies its
round-off refusals on ReLU networks (#2187, #2189); the in-house Rust MILP engine
certifies trained ReLU big-M MILPs that it used to leave undecided (#2163, #2165). -
Deadlines. The per-node interval objective bound stops at the deadline
(#2071, #2073); every step of a native GDP solve, including lowering, is charged
totime_limit(#2079, #2080:stranded_gasoverran 120 s by 97 s, now by 2 s);
underdeterministic=True, per-NLP wall caps are role-2 budgets, so the answer
reproduces (#2120, #2124). -
GDP robustness (#1917). LOA/RIC sub-NLPs keep a sparse Jacobian and Hessian
and refuse ani32KKT overflow instead of panicking (#2078); the hull splits an
unboundablelog(sum)operand instead of refusing the model (#2076); a raising
disjunct-configuration constructor is logged at WARNING with its traceback, not
DEBUG (#2138). -
A failed pickle closes its children-first scope (#2180, #2181), which had
left later deep expressions pickling recursively untilRecursionError. -
The HiGHS MILP route no longer certifies a feasible model infeasible (#2156).
HiGHS's MIP answeredkInfeasibleon feasible models with continuous columns of width
0 < xu - xl < 1e-6, and the route certified it, because each cross-check (the
presolve-free solve, the other form, the coefficient-tightened form) ran the same MIP
code. With every integer fixed at a feasible point, HiGHS still answered
kInfeasibleafter zero LP iterations: domain propagation decided it. A HiGHS-only
infeasible label on such a model is now reported aserror, uncertified, with the
reason. A Farkas proof of the root relaxation is still certified, and nooptimal
is affected. On the 30-network panel below, the unmergedrelu_extendedencoding
had 8 false certified-infeasible results and now has none. The 89 correct
certificates on the same panel are all kept. Checked by
python/tests/test_2156_highs_sub_tolerance_infeasible.py, which includes a
canary that fails once HiGHS stops mislabelling a shipped 13-row sub-model. -
relu_extendedandrelu_partitioncertify trained ReLU networks (#2156).
Networks with many weights near 1e-6 to 1e-8 gaverelu_extendedauxiliaries whose
whole range was below HiGHS's feasibility tolerance. HiGHS then reported feasible
models infeasible, with presolve on and off, or returned a wrong optimum. A partition
whose range is at mostNEGLIGIBLE_PARTITION_RTOL(1e-4) of its neuron's range is now
folded into the neuron's widest partition. Any partition gives the same MILP, so the
integer-feasible set is unchanged, and the LP loosens by at most that fraction per
merged partition. On the issue's networkrelu_extendednow certifies −6.37458671,
agreeing withrelu_bigm; before, it returnedfeasiblewith gap 0.69. On a
30-network panel of 2-32-32-1 tiny-weight nets,relu_extendedwent from 0/30
certified (24 wrong or no point, 8 of theminfeasible) to 30/30 certified and equal
to big-M;relu_partitionwith P = 4 went from 0/30 to 30/30. Checked by
python/tests/test_2156_relu_negligible_partitions.py. -
The in-house simplex certifies LPs with huge-but-finite bounds (#2125). A bounded,
feasible multi-row LP with a finite variable bound in[1e10, 1e20)used to return
SolveStatus.ERRORfromlp_simplex.solve_lp. Measured on the issue's GDPlib
batch_processingmbigm LPs atu >= 3e16: 7-8 of 40 agreed with HiGHS before, and
40 of 40 now. On a seeded multi-row sweep it was 72 of 192, and is now 192 of 192.
Three fixes, each held to the original LP's audit:- the row audit lets a row's own slack absorb its rounding;
- a cold solve drops the huge sides and verifies the relaxation's answer;
- an active huge bound is fixed and certified from duals.
One case is still refused rather than certified: a vertex whose huge interior value
sits in a binding row.DISCOPT_LP_HUGE_BOUND_RELAX=0restores the previous engine
(bit-identical on 257 LPs measured); an unrecognized value is an error. -
The in-house MILP driver bounds nodes of translation-invariant masters
(#2115). Its node bound (Neumaier-Shcherbina from the node dual) was-inf
whenever an open column -- a free variable arrives splitx+ - x--- had a
reduced cost zero only up to rounding. On MINLPLibm6's final OA master (42 of
86 columns free) it abstained on 3194 of ~29k nodes, so withoutpyscipoptthe
#2069 confirmation could not close the master andm6endedfeasibleat 60 s
(bound 62.83, optimum 82.2569). A new last rung of the ladder
(lp/simplex/project.rs) builds an enclosure of a nearby dual whose reduced
costs on those columns are exactly zero and evaluates the bound rigorously over
it.m6now certifies in 14.8 s andm7in 44 s with SCIP blocked; panel of 91
instances cert-clean, certified 63 -> 64. Opt-out for A/B only:
DISCOPT_NS_PROJECT=0. -
Rounding residue of
yno longer passes as a real sub-tolreduced cost
(#2115 follow-up). The #1595 re-pricing judgedd_jonly against its own
terms, so a single-entry open slack (d_j = -a y_i) always looked real however
smally_iwas; the simplex entered it, refused the resultingc^T d = 0ray,
and returnedNumericalfor an LP its dual had solved (13 LPs onm6's
master, 6 onm7's). The noise test now also uses||a_j||_1 ||y||_inf, and
such columns go through the #1953 suspect verification. Panel cert-clean and
outcome-neutral; shipped as a verdict fix. Opt-out:
DISCOPT_SUBTOL_NOISE_GLOBAL=0. -
A convex node's NLP objective is no longer imported as its lower bound
(#2067). NLP-B&B (and the spatial loop's convex mode) used the objective of
the node NLP's approximate minimiser as the node bound; it sits above the
relaxation's optimum by the solver's residuals, and QPLIB_8153 certified a
bound 1.57e-7 past a reference point feasible to 2.7e-15. Each convex node now
imports a Neumaier-Shcherbina bound on its LP linearisation
(discopt.solvers.convex_node_bound), valid at any point of the box; the NLP
objective stays the incumbent value of an integral node, whose proved bound
floors the tree bound (safe_fathom_floorin the tree stats; a
point_values=argument onPyTreeManager.import_results). Opt-out for A/B
only:DISCOPT_CONVEX_NODE_SAFE_BOUND=0. -
Lowering-derived auxiliary bounds are re-derived at every solve (#2061
review). The splitw±of a box MCP, the SOS1 aux and the GDP operand lift
of an NCP pair are bounded from an interval enclosure over the model's box at
lowering time. Widening a variable afterwards (x.ub = 50after declaring
x <= 1) left that bound in place, and the stale big-M cut the optimum: a
certifiedoptimal-10 against a true -50. The NCP lifts had the same defect
before #2058.Model.solve,solve_modelandsavenow recompute every such
bound (mpec.refresh_lowering_bounds; an unchanged model is left exactly as
it was). A lowering read from a file carries no derivation, so widening a
bound or moving a parameter it reads afterloadis refused at solve and save.
An enclosure that reads discopt's default±9.999e19bounds now counts as
unbounded rather than as a finite big-M. -
Model.if_else's auxiliary box is re-derived at every solve and save
(#2061 review).w = m.if_else(x >= 0.5, x, 2 * x)onx in [0, 1], then
x.ub = 10,maximize(w)was certifiedoptimalat 1 against a true 10:
w's bounds were the branches' enclosure at declaration. Bounds the user sets
onware kept (the new derivation is intersected with them). A model read
from a file has itsif_elsedisjunctions recognised; a widening that the
saved box no longer encloses is refused at solve and save. -
A GDP-lowered complementarity whose side has no finite enclosure (typically a
box relation whosezis declared free) is still refused at solve, but the
error now names the relation and the side, and says to givezits own
bounds; it used to stop atconstraint '<unnamed>'.
Changed
-
The
pounce-solverfloor is now 0.16 (was 0.14 in v0.11.1; #2132 raised it to
0.15 during this cycle), for both the core dependency and thepounceextra.
POUNCE 0.16.0 fixes a false infeasibility verdict at a point that satisfies the
model (most visible under L-BFGS), makes declared complementarity much faster on
large models (wide paired rows are lifted into a variable) and complementary to
tolrather thansqrt(tol), and classifies two more convex infeasibility cases
correctly. 0.15.0 added the declared-complementarity driver and capped the
slack-row inertia perturbation. Neither has an in-memory solve report, so
_pounce_report.pykeeps its temporary-file route. The Pyodide wheel in
crates/discopt-wasm/web/wheels/is still 0.14.0; it is built in the POUNCE
repository and is updated separately. -
Flags graduated to default ON this cycle (the measurement behind each is in
its PR and indocs/dev/flag-retirement-audit.md;=0opts out):DISCOPT_ROUTE_ABANDON_CROSS_CHECKandDISCOPT_OA_SOC_OUTER
(portfol_classical050_1now certifies, #2059, #2064);
DISCOPT_SUBNLP_REDUCED,DISCOPT_HEUR_FIXING_SCREEN,
DISCOPT_ROOT_WARMSTART_FIRST,DISCOPT_GDP_LOA_PROPOSER,
DISCOPT_GDP_LOA_PROPOSER_BOUNDandDISCOPT_ROOT_PUMP_SKIP_WITH_INCUMBENT
(#2143, #2166);DISCOPT_DUAL_PROOF_PROPAGATIONand
DISCOPT_HULL_EXACT_PERSPECTIVE(GDPlibcstrcertifies under big-M, #2142,
#2169);DISCOPT_HEUR_WALL_COST(#2176, #2175);
DISCOPT_GDP_CONFIG_ROOT_SKI(#2182, #2184). -
The root-pump skip requires a closed gap (#2170, #2175). It skipped both root
feasibility pumps whenever any incumbent existed; it now requires that incumbent
to meet the solve's own gap test. -
DISCOPT_OBBT_NS_STRICT(#2077), a documented opt-out (default OFF):
tighten OBBT/DBBT only to the Neumaier–Shcherbina safe bound. -
Bound-neutral speedups. The nonlinear FBBT scan skips rows a rule
structurally cannot match (#2159); curvature proofs that must abstain are skipped
when the Hessian at a point of the box is provably indefinite (#2160); the
multiple-big-M GDP LP relaxation is sparse and memoized (#2075). -
Measured and not adopted, so absent from this release: per-LP caps on
persistent OBBT probes (#2065, #2068),DISCOPT_HULL_PLAIN_EQUALITY(#2137),
DISCOPT_SUBNLP_FBBT_BOX(#2141),DISCOPT_ROOT_FIXPOINT_EVEN_SPLIT(#2162),
DISCOPT_NLP_LADDER_SCOPED(#2161, retired in #2177) and
DISCOPT_KERNEL_PRESOLVE(#2168, #2171). Their measurements are in
docs/dev/flag-retirement-audit.md. -
The open branch-and-bound frontier has a memory budget, on by default (#2092).
DISCOPT_FRONTIER_MEMORY_MB(default 2048 MiB;0= no budget) becomes an open-node
cap for the native spatial kernel and the Python-tree loops. A search whose gap never
closes keeps every frontier node live, and on 1-hour runs that reached 28–42 GiB. A
solve with nomax_nodescan therefore now end withnode_limitonce its frontier
reaches the budget (never before 1,000 open nodes), with the incumbent and a valid
bound; it stops, it never prunes. Raise the budget on a machine with more memory, or
set0for the old unbounded behaviour. The same release cuts what each open node
stores (measured: about 37% less frontier growth in the native kernel, 48% less per
open node in the Python tree): sibling nodes share their parent's box, and in the
Python tree also the parent solution and basis. -
.nlcomplementarity rows with flag 3 (both bounds) or flag 0 (free
variable) are now imported (#2058) as a box MCP and asbody == 0
respectively, instead of refused. Flags 1/2 are unchanged. -
OA convexity certification uses a Taylor form for one-variable rows (#2062). A row whose
curvature depends on a single variable is now tested piece by piece with a Taylor expansion
off''with a Lagrange remainder (new_relax/convexity/interval_jet.py) before #2054's
interval-Hessian bisection. The PFR rows(1-k)(e^t-1)/(e^{(1-k)t}-1)and
(1-k) t e^t/(e^{(1-k)t}-1)now certify in ~50 pieces instead of 2048-8192, so they get OA
cuts. On a 40-model reactor-choice panel undergdp_method="loa": certified 20 -> 25,
0 lost, LOA iterations 93 -> 48. Default ON;DISCOPT_OA_CONVEXITY_TAYLOR=0restores the
previous behaviour.
Known issues
- On macOS/arm64, MINLPLib
nous1no longer certifies within 150 s (#2202).
WithDISCOPT_DUAL_PROOF_PROPAGATIONon (the new default) the tree stalls at a
valid bound of 1.3457 (optimum 1.5671); withDISCOPT_DUAL_PROOF_PROPAGATION=0it
certifies in about 11 s. Linux is not affected (25.5 s with the flag on). No bound
or answer is wrong. Bisected to #2169. - On macOS, importing
cyipoptand thentorchin one process aborts it
(OMP: Error #15: two copies oflibomp). This is an interaction between those two
packages' bundled OpenMP runtimes, not discopt code, but it affects a process that
usesnlp_solver="ipopt"and the torch reader together, in either import order.
Run the two in separate processes until the runtimes are reconciled.