NVIDIA cuOpt

The NVIDIA cuOpt Python API provides GPU-accelerated interfaces for three primary optimization domains:

1. Routing Optimization: Solving complex vehicle routing and logistics problems.
2. Convex Optimization: Solving Linear Programming (LP), Quadratic Programming (QP), Quadratically Constrained Quadratic Programming (QCQP), and Second-Order Cone Programming (SOCP) problems.
3. Mixed-Integer Programming (MIP): Solving optimization problems involving both continuous and integer variables.

Users can access these capabilities through the cuopt Python package.

NVIDIA cuOpt provides GPU-accelerated combinatorial optimization through several distinct interfaces depending on your deployment and programming needs:

• C (libcuopt): Low-level C API for high-performance integration.
• Python (cuopt): High-level Python API for rapid development and data science workflows.
• gRPC Remote Execution: Interface for executing optimization tasks remotely via gRPC.
• Server (cuopt-server): A microserver implementation for hosting cuOpt services.
• Command Line Interface (cuopt-cli): A CLI for interacting with cuOpt services.
• Third-Party Modeling Languages: Support for integrating with external modeling languages.

NVIDIA® cuOpt™ is a GPU-accelerated optimization engine designed for large-scale problems including:

• Linear Programming (LP)
• Quadratic Programming (QP)
• Vehicle Routing Problems (VRP), including TSP and PDP
• Mixed Integer Linear Programming (MILP) (beta)
• Quadratically Constrained Quadratic Programming (QCQP) (beta)
• Second-Order Cone Programming (SOCP) (beta)

It is optimized for near real-time solutions with millions of variables and constraints. The core engine is written in C++ and provides C, C++, Python, and Server APIs.

NVIDIA cuOpt provides a C API for GPU-accelerated optimization. This interface allows developers to solve complex optimization problems efficiently using GPU hardware. The C API is divided into two primary functional areas:

1. Convex Optimization: Supports Linear Programming (LP), Quadratic Programming (QP), Quadratically Constrained Quadratic Programming (QCQP), and Second-Order Cone Programming (SOCP).
2. Mixed-Integer Programming (MIP): Supports solving problems with both continuous and integer variables.

The NVIDIA cuOpt self-hosted server is a REST (HTTP/JSON) service designed for integrations that communicate via HTTP. It provides a way to offload combinatorial optimization tasks to a GPU-accelerated service.

Important Distinction:

• REST API: Uses HTTP/JSON for communication. Use the Server API and Client API documentation for this service.
• gRPC API: Uses a different protocol for remote execution (supported via Python, C API, cuopt_cli, or custom clients connecting to cuopt_grpc_server). The gRPC interface is distinct from the HTTP REST surface.

libcuopt is a C++ library designed for GPU-accelerated optimization. It provides high-performance solvers specifically focused on:

• Vehicle Routing Problem (VRP)
• Linear Programming (LP)
• Mixed Integer Linear Programming (MILP)

NVIDIA® cuOpt™ is a GPU-accelerated optimization library designed to solve large-scale problems with millions of variables and constraints. It is built on a C++ core engine with various API wrappers (Python, C, gRPC, etc.).

Supported Problem Types:

• Routing: Traveling Salesperson Problem (TSP), Vehicle Routing Problem (VRP), and Pickup and Delivery Problems (PDP).
• Linear Programming (LP): Optimization of linear objective functions subject to linear constraints.
• Quadratic Programming (QP): Optimization of quadratic objective functions.
• Mixed Integer Linear Programming (MILP) (Beta): Optimization where some variables are restricted to integer values.
• Quadratically Constrained Quadratic Programming (QCQP) (Beta)
• Second-Order Cone Programming (SOCP) (Beta)

The cuopt_cli is a command-line interface designed for LP/MILP (Linear Programming / Mixed-Integer Linear Programming) solvers. It allows users to solve optimization problems by providing model files in standard formats.

Supported Input Formats: The CLI automatically dispatches files to the appropriate parser based on their extension (case-insensitive). Supported extensions include:

• .lp (supports optional .gz or .bz2 compression)
• .mps (supports optional .gz or .bz2 compression)
• .qps (supports optional .gz or .bz2 compression)

Solver Capabilities: The cuOpt MIP solver is currently in beta. It is optimized for finding high-quality feasible solutions rapidly using GPU-accelerated primal heuristics. Note that proving the optimality of feasible solutions is a feature currently under active development.

The cuOpt Routing solver is available in two deployment modes:

• Self-Hosted Service: Deploy cuOpt in your own infrastructure for full control and integration into existing systems. See cuopt-server/quick-start for details.
• Python SDK: A direct Python package for seamless integration into Python applications and workflows. See cuopt-python/quick-start for details.

Both modes provide identical routing optimization capabilities.

cuOpt provides GPU-accelerated solvers for Linear Programming (LP), Quadratic Programming (QP), Quadratically Constrained Quadratic Programming (QCQP), and Second-Order Cone Programming (SOCP). You can access these solvers through several interfaces:

• Third-Party Modeling Languages: Use existing workflows in languages like AMPL, CVXPY, GAMS, JuMP, PuLP, or Pyomo.
  • Note: QCQP/SOCP support is currently in beta and is only available via CVXPY among the listed modeling languages.
• C API: Direct low-level access for integration into C-compatible applications.
• Python SDK: A high-level, intuitive API for Python applications.
• Self-Hosted Service: Deploy solvers as a service within your own infrastructure.