gRPC

Note

For routing, gRPC currently supports problems of up to about 5500 locations when a travel time matrix is included, or about 8000 locations when it is not. For a larger problem, use a local cuopt.routing.Solve() or the legacy HTTP server.

NVIDIA cuOpt can run LP, MIP, QP, and routing (VRP, TSP, PDP) solves on a remote GPU host through cuopt_grpc_server. There are two ways to reach that server:

Remote execution (zero code change)

Set CUOPT_REMOTE_HOST and CUOPT_REMOTE_PORT on the client machine. The Python solver APIs, the C API cuOptSolve, and cuopt_cli forward LP, MIP, and QP solves automatically. Modeling-language integrations that call one of those APIs, including AMPL, GAMS, PuLP, JuMP, Pyomo, and CVXPY, forward the same way. No client code changes are required. This does not work for routing at this time (tracked in #1633). See Quick Start and Third-Party Modeling Languages.

gRPC clients (explicit client)

Your program opens a gRPC connection and manages jobs itself. cuOpt ships two Python clients. Both are asynchronous job APIs: submit, then wait or poll.

• The gRPC linear programming client (cuopt.grpc.linear_programming.Client) manages LP, MIP, and QP jobs.

• The gRPC routing client (cuopt.grpc.routing.RoutingClient) manages VRP, TSP, and PDP jobs.

• A custom client can speak CuOptRemoteService directly (gRPC API (Reference)).

In this section, remote execution always means the zero-code-change path above. gRPC linear programming client and gRPC routing client are the two Python clients. gRPC client means any program that opens the connection itself, including a custom CuOptRemoteService client.

When to Choose Which Path

• Remote execution — drop-in remote LP, MIP, and QP solves with no code changes, including modeling-language integrations that call the Python solver APIs, cuOptSolve, or cuopt_cli.

• gRPC linear programming client — explicit LP, MIP, and QP job control: submit now, wait or poll later, cancel, stream solver logs, stream MIP incumbents.

• gRPC routing client — explicit VRP, TSP, and PDP job control: submit now, wait or poll later, cancel.

• Custom ``CuOptRemoteService`` client — non-Python (or fully custom) integrations that speak the protos directly. See gRPC API (Reference).

Start with Quick Start (install, server, and a minimal LP). Use gRPC Linear Programming Client for the gRPC linear programming client and gRPC Routing Client for the gRPC routing client. See Advanced Configuration for TLS, Docker, environment variables, and troubleshooting, and Examples for additional patterns.

In this section

• Quick Start
  • How Remote Execution Works
  • Install NVIDIA cuOpt
  • Run the gRPC Server (GPU Server)
  • Minimal Python Example
  • Minimal cuopt_cli Example (LP)
  • Next Steps
• gRPC Linear Programming Client
  • Prerequisites
  • Connect and Solve
  • Variable Names
  • TLS / mTLS
  • Next Steps
• gRPC Linear Programming Client Examples
  • Log Streaming
  • Incumbent Streaming (MIP)
  • See Also
• gRPC Linear Programming Client API Reference
  • Client
  • Supporting Types
  • Exceptions
  • See Also
• gRPC Routing Client
  • Prerequisites
  • Connect and Solve
  • Limitations and Roadmap
  • Next Steps
• gRPC Routing Client API Reference
  • Client
  • Supporting Types
  • Exceptions
  • Settings
  • Solution Fields
  • See Also
• Advanced Configuration
  • Configuration Parameters
  • Usage
  • Limitations and Scope
  • Troubleshooting
  • Further Reading
• Examples
  • Where to Find Examples
  • More Samples
  • REST vs gRPC
• gRPC API (Reference)
  • Service: CuOptRemoteService
  • Messages and Constraints
  • Further Reading
• gRPC Server Behavior
  • Process Model
  • Job Lifecycle (Summary)
  • Job States
  • Logs, Capacity, and Workers
  • Fault Tolerance and Cancellation
  • Further Reading

See System Requirements for GPU, CUDA, and OS requirements.