Repository navigation
Home
GEAR is a GPU-accelerated global solver for bounded continuous nonlinear programs. It computes certified lower bounds with linear bound propagation, the technique behind neural network verifiers such as alpha,beta-CROWN, and searches for feasible points with a GPU-batched augmented Lagrangian method. The paper is "GEAR: A GPU-Accelerated Global Solver for Bounded Continuous Nonlinear Programs via Linear Bound Propagation" (Zhou, Chen, Zhong, Hanasusanto, Zhang).
On 505 bounded continuous instances from GAMS World and MINLPLib, GEAR solves 421 within 180 seconds on one NVIDIA RTX 5090 (see Reproducing the Paper for the scoring rule).
| Page | Content |
|---|---|
| Installation | Requirements and setup |
| Quick Start | Solve one benchmark instance and read the output |
| Python API |
gear.compile, gear.solve, options, results |
| Your Own Problem | State a new problem from PyTorch; the file format the command line reads |
| Configuration | The default preset and the options you are likely to change |
| Dataset | The 505 benchmark instances |
| Reproducing the Paper | Re-run the 505 instances and compare with the released logs |
| Output and Logs | What a log contains |
| How GEAR Works | Conceptual overview |
| Troubleshooting | GPU memory, CUDA and Triton versions, first-run cache |
| Citing | BibTeX |
- Repository: https://github.com/Verified-Intelligence/GEAR
- Project page: https://verified-intelligence.github.io/GEAR/
- Gap atlas (all 505 instances, GEAR and 12 AMPL solvers): https://verified-intelligence.github.io/GEAR/atlas.html
GEAR targets nonlinear programs with a finite box on every variable and continuous variables, written as a computation graph of supported operators. A solve returns a primal value P (the objective at a returned feasible point) and a certified lower bound D. The tables in the paper count an instance as solved when both agree to a relative gap of 1e-4 and the returned point passes a strict feasibility check; see Quick Start.