Fixed Issues in Gurobi Optimizer 14.0#

This page lists the issues fixed in Gurobi 14.0.

Fixed Issues in Gurobi 14.0.0#

Optimizer#

Wrong Answers and Numerical Issues#

  • Fixed numerical issues in product detection that could cause wrong answers

  • Fixed numerical issue in product detection on variables with very large or infinite bounds

  • Fixed numerical issue in presolve that could lead to declaring a linear constraint redundant and accepting solutions that significantly violate this constraint

  • Fixed numerical issue in presolve coefficient strengthening that could lead to a wrong answer

  • Fixed numerical issue in the biconnected component presolve reduction that could lead to a wrong answer

  • Fixed numerical issue in the parallel row presolve reduction with combining rows of very different magnitude

  • Fixed numerical issue in the Euclidean presolve reduction with rounding values that should be exact integers

  • Fixed numerical issue when removing variables with tiny domains from constraints that could lead to a wrong answer when presolve is turned off

  • Fixed issue with declaring a model to be unbounded in presolve if dual reductions are enabled and a MIP start exists

  • Fixed very rare issue in a presolve reduction on small knapsack constraints that could lead to wrong answers

  • Fixed numerical issue in presolve equality constraint detection that could lead to a wrong infeasibility conclusion even though a feasible solution within tolerances exists

  • Fixed numerical issue with translating indicator constraints to SOS1 that could lead to violations larger than the feasibility tolerance

  • Fixed numerical issue in reduced cost strengthening for continuous variables with slightly negative solution values

  • Fixed issue in clique-based reduced cost strengthening that could lead to wrong answers

  • Fixed numerical issue in BQP cut separator for variables with large domains

  • Fixed issue in network cut separation with bound-shifted binary variables that could lead to wrong answers

  • Fixed issue in strong branching that could lead to a wrong answer

  • Fixed potential wrong answer if non-basic strong branching LP solve fails

  • Fixed issue in a heuristic that could have produced solutions with unnecessarily large solution values for unbounded variables

  • Fixed issue with returning a wrong objective value for maximization models that are solved to optimality by presolve

  • Fixed issue with reporting a wrong DualVio attribute value for a non-optimal solution of a model with unbounded variables

  • Fixed issue that could lead to accepting an infeasible MIP start for a model with semi-continuous or semi-integer variables

  • Fixed issue with potentially discarding a MIP start for multi-scenario models

  • Fixed issue with potentially losing solution pool solutions for models with maximization sense

  • Fixed issue with models with disconnected components and unbounded LP relaxation that could lead to a wrong answer

Crashes#

  • Fixed very rare segfault in simplex triggered by numerical issues in a model

  • Fixed very rare segfault in simplex LU factorization

  • Fixed segfault in clique cut separator for large models

  • Fixed very rare segfault in relax-and-lift cuts

  • Fixed potential segfault when searching for weak symmetry in presolve

  • Fixed potential segfault in node presolve for clique propagation

  • Fixed potential segfault in sub-MIP heuristics if the sub-MIP is empty

  • Fixed potential segfault when accessing non-existing reduced costs from a heuristic

  • Fixed potential segfault in NoRel heuristic for QUBO problems

  • Fixed potential segfault in NoRel heuristic

  • Fixed potential segfault if creating the LP relaxation of a MIP runs out of memory

  • Fixed potential segfault if default names are used for quadratic constraints

  • Fixed potential segfault when removing variables from a model that contains bilinear constraints

  • Fixed potential segfault when using the recording feature on models with quadratic constraints

  • Fixed potential segfault when copying a model in which the IISQConstrForce or IISGenConstrForce flags are set

Memory#

  • Fixed memory leak in a heuristic in an error situation such as running out of memory

  • Fixed potential memory leak in the piecewise linear dual simplex algorithm

  • Fixed potential memory leak in MIP solution cleanup

  • Fixed issue in the network cut separator that could lead to excessive memory consumption

  • Fixed very rare issue with not checking for out of memory errors at one particular place in uncrushing solutions

  • Fixed various issues with index overflows when dealing with models with more than 2 billion non-zero values

  • Fixed issue with a potential integer overflow in the dependent row reduction of LP presolve

  • Fixed potential integer overflow in wide split cuts

  • Fixed buffer overflows when opening files with very long file names

Time Limits and Hangs#

  • Fixed issue with the time limit not being respected when simplex is stuck in a long pivot rejection loop for models with numerical issues

  • Fixed issue with overshooting the time limit in simplex when initializing norms

  • Fixed issue with sifting not correctly stopping at the iteration limit

  • Fixed issue with overshooting the time limit in the parallel row detection code for implicitly piecewise-linear constraints

  • Fixed issue with overshooting time limit when the model has large disconnected components

  • Fixed issue with models with disconnected components where a solve interrupt signal could get ignored

  • Fixed performance issue in clique merging presolve and tweaked work limits to avoid spending too much time for large models

  • Fixed performance issue in disconnected components presolve reduction that could lead to spending too much time for a component solve

  • Fixed various infinite loops in error situations

  • Fixed potential deadlock in parallel simplex, barrier, and PDHG when a thread creation fails

  • Fixed issue with not terminating previous threads when a thread creation fails in the partition heuristic

API Behavior#

  • Fixed issue with dropping the user callback if querying one of the LP sensitivity analysis attributes leads to a failure in a simplex solve to recalculate the duals

  • Fixed issue with GRBoptimizeasync not returning the error code if optimization fails

  • Fixed issue with ignoring an error return code of one particular heuristic

  • Fixed issue with using R instead of C for a default column name

  • Fixed recording feature for machines with big-endian integer encoding

Miscellaneous#

  • Fixed non-determinism in MIP when it solves LP sub-problems with multiple threads at the root node

  • Fixed issue that led to starting a solve from scratch instead of resuming an interrupted MIQP solve

  • Fixed performance issue with a semi-variable to SOS1 translation that would lead to a wrong translation of an original solution into the presolved space

  • Fixed issue that could lead to segfaults or wrong memory counting if a separate model is being solved inside the user callback

  • Fixed very rare issue in weak symmetry cut separator when the dual LP solution is not available

MINLP and Nonconvex MIQCP#

  • Fixed numerical issue in bound propagation of nonlinear constraints that could lead to pruning feasible solutions

  • Fixed numerical issue in the LP relaxation of bilinear constraints that involve variables with tiny domains

  • Fixed wrong dual bound for an MINLP with unbounded root node relaxation

  • Fixed issue with a solution not being accepted after MINLP presolve crushed the problem to nothing, thus leading to a wrong answer

  • Fixed performance issue in LP relaxation coefficient updates for MINLPs in which variables are involved in many nonlinear constraints

  • Fixed issue that could lead to wrong answers on bilinear models with non-positive variables

  • Fixed very rare issue with strong branching on bilinear or nonlinear models that could lead to a wrong answer

Barrier#

  • Fixed issue in logging with very long delay between log messages during barrier ordering

  • Fixed uninitialized memory read in barrier SOCP code

  • Fixed issue in barrier ordering that could lead to bad ordering decisions and slow performance

  • Fixed direction vector calculation in the iterative barrier solver for models with cones that could lead to bad performance

  • Fixed issue with a wrong sign in the objective function value of a QCP solve for maximization problems

Nonlinear Barrier#

  • Fixed potential segfault in error cases in nonlinear barrier

PDHG#

  • Fixed wrong violation calculation that could lead to wrong convergence conclusions and early stopping

  • Fixed potential segfault when freeing data in PDHG/GPU

  • Fixed potential memory leak in PDHG/GPU

Crossover#

  • Fixed numerics in crossover that could lead to a large number of simplex iterations in postsolve

Multi-Objective Optimization#

  • Fixed issue with reporting wrong individual objective values for multi-objective models if presolve solved the model

  • Fixed reading of objective constants in LP reader for multi-objective models

  • Fixed issue in multi-objective optimization where the ObjPassNRuntime and ObjPassNWork attributes were not set for the final objective

Concurrent Optimization#

  • Fixed issue with using too many threads in concurrent optimization with user environments

  • Fixed inconsistent status and objective bounds when two concurrent MIP solvers disagree on the feasibility status of the model

  • Fixed very rare race condition in logging and termination of concurrent MIP when accessing primal and dual bounds

Remote and Distributed Optimization#

  • Fixed issue with distributed optimization not working with WLS licenses

  • Fixed logging issue with distributed MIP reporting node counts in non-increasing order

  • Fixed issue in distributed MIP with reporting a wrong objective value

  • Fixed logging issue in client that could lead to many Optimization still in progress messages

  • Fixed worker deadlock after a temporary network connection outage

  • Fixed deadlock in the client when GRBsync is called after the remote worker crashed

  • Fixed buffer overflow when reading a .bas basis file that contains duplicate entries in a remote optimization

  • Fixed potential deadlock in an error situation of distributed MIP and distributed concurrent MIP

  • Fixed issue with ignoring an optimization error in distributed MIP

  • Fixed issue with distributed MIP and distributed concurrent MIP that could lead to wrong MIP_OBJBND, MIPSOL_OBJBND, and MIPNODE_OBJBND callback query results

  • Fixed wrong attribute values returned following calls to convertToFixed (C, C++, Java, .NET, Python) in remote environments

Tuning Tool#

  • Fixed issue with multi-objective environments not being respected in distributed tuning

  • Fixed small memory leak in remote tuning

  • Fixed issue with starting the tuner cleanup even though the TuneCleanup parameter was set to 0.0

Licensing#

  • Fixed race condition in WLS license check when running parallel environments

Python API#

  • Fixed very rare spurious error when adding an explicit zero MQuadExpr of size 1 to another MQuadExpr

  • Fixed small memory leak in Env.getParam

  • Fixed missing type check in Model.setAttr that could lead to not raising an exception but instead applying invalid model updates if the method is called for an object that does not match the type of the attribute

C++ API#

.NET API#

Java API#

  • Fixed potential segfault after querying a string attribute from a batch optimization job

  • Fixed potential exception when querying the version number

  • Fixed issue with not checking for a NULL pointer argument when retrieving double attributes, for example with GRBVar.get, GRBConstr.get, or GRBModel.get

  • Fixed very rare issue in ranged row counting when a ranged row is added that has both sides being infinite

MATLAB API#

  • Fixed potential buffer overflow in general constraint handling

  • Fixed potential segfault if querying general nonlinear or general piecewise linear constraints fails

  • Fixed issue with not raising an exception if a string parameter is used with a wrong parameter name

Compute Server, Cluster Manager, and Instant Cloud#

The list of fixed issues for Compute Server, Cluster Manager, and Instant Cloud can be found here.