// the find
NVIDIA/cuopt-examples
NVIDIA cuOpt examples for decision optimization
A collection of Jupyter notebooks demonstrating NVIDIA's GPU-accelerated optimization engine (cuOpt) across vehicle routing, linear programming, mixed-integer programming, and portfolio optimization problems. It's an examples repo, not a library — the actual solver ships separately as a Docker image or cloud API. Aimed at operations researchers and supply chain engineers who want to see what GPU-accelerated solvers can do on real problem classes.
The problem coverage is genuinely broad: VRP with time windows, MILP, LP, QP for portfolio optimization, and factory AMR routing — not just toy TSP demos. Integration examples with CVXPY, PuLP, and Pyomo mean you can keep your existing model-building workflow and just swap in the GPU solver as a backend. The AI-accelerated routing directory combines a reinforcement learning agent with cuOpt as a local search refinement step, which is a non-obvious and interesting architectural choice. Benchmark tooling in `benchmark_apis/` includes C API drivers and format converters for AMPL, GAMS, and Julia — useful for anyone seriously comparing solvers.
This is a demo repo for a proprietary commercial product; the solver itself is NVIDIA-hosted or runs inside their Docker image, so you're not getting the source of what actually matters. The NVIDIA GPU requirement is a hard dependency — no fallback to CPU, no cloud-agnostic path, which rules out most CI environments and non-NVIDIA cloud instances. Notebook-heavy repos age badly: no automated testing, no way to know if a given notebook still runs without pulling the Docker image and executing it manually. The cuopt-agent subdirectory introduces a separate agentic framework with its own Docker stack and config format, making the repo feel like it's accumulating surface area without a coherent entry point for newcomers.