cuOpt Installation Guide
Install and verify NVIDIA cuOpt for Python, C, or REST deployment.
The skill asks for interface, CUDA version, package manager, and environment before advising installation, uses no identified credentials, and generally limits actions to user-run commands. However, it lacks explicit execution confirmation, rollback guidance, and supply-chain verification; Docker uses --gpus all and binds 0.0.0.0, while external effects and data flow are not fully disclosed, so points were deducted.
Installation paths, dependency explanations, troubleshooting, and verification steps are broadly coherent. Reliability is reduced by the 26.08.00 frontmatter versus 26.2.* CUDA 12 pins, mostly CUDA 12 server guidance, mutable latest Docker tagging, and the inability to reproduce key paths during static review.
The skill clearly covers Python, C, REST, pip, conda, Docker, local GPU, and remote-server scenarios, and redirects source builds to another skill. It provides no Chinese-language guidance, and its core package index and image dependencies rely on overseas services without mainland-China reachability guidance, so points were deducted.
It includes a name, version, license, author, references, verification material, benchmark, and skill card. It is nevertheless missing the recommended Instructions and Examples sections, has an author-format schema finding, and lacks a clear skill-level changelog and maintenance/update path, so points were deducted.
The material supplies directly usable installation, startup, health-check, and troubleshooting commands, and the supplied evaluation report shows one external task passing. Static evidence does not establish actual runtime correctness, and version, platform coverage, and some verification commands still require user review, so full marks are not justified.
Committed evaluation cases, a benchmark report, and concrete verification recipes make several claims auditable. Coverage is thin, the external benchmark contains only one task, no real CI workflow with key-path tests is shown, and independent reproduction was not performed, so the score is limited.
- Before execution, verify CUDA, driver, GPU compute capability, Python environment, and package versions; do not mix pip and conda in one environment.
- The Docker example uses latest and binds to 0.0.0.0; production deployments should pin image versions and restrict network exposure.
- NVIDIA’s package index and Docker registry may be affected by mainland-China network reachability; prepare a mirror, proxy, or offline plan.
What does this skill do, and when should you use it?
This is an installation-focused Agent Skill for users who want to use NVIDIA cuOpt rather than build it from source. It covers the Python API, C API, and REST server, with pip, conda, and Docker workflows. It guides selection based on CUDA version, GPU capability, package manager, and deployment environment, then provides verification and troubleshooting steps. Source builds and code changes belong to the separate `cuopt-developer` skill.
Asks for or confirms the interface, CUDA version, package manager, and environment; checks the stated NVIDIA GPU, CUDA, and driver requirements; provides pip and conda commands for the Python and C packages; provides pip, conda, and Docker commands for the REST server; verifies Python installation by importing cuOpt, printing its version, and creating a DataModel, or verifies the server with a health request; diagnoses missing Python modules, unavailable CUDA, and Python/C package dependency confusion.
- A Python user installing the cuOpt API with pip on a local or cloud machine equipped with an NVIDIA GPU.
- A C or C++ user who needs the cuOpt library and headers without installing the Python bindings.
- A team deploying cuOpt as an HTTP-accessible REST service.
- An operations or platform user working with Docker, Kubernetes, or a remote server and needing a container startup path.
What are this skill's strengths and limitations?
- Covers Python, C, and REST interfaces in one installation workflow.
- Provides pip, conda, and Docker paths with basic verification commands.
- Clearly documents CUDA-suffix matching, Python-to-C package dependencies, and common failure modes.
- The complete C verification commands are delegated to `references/verification_examples.md`, whose contents are not included here.
- Examples focus on CUDA 12/13, Linux-style shell commands, and NVIDIA GPU environments; other platforms are not described.
- The CUDA 12 pip examples pin `26.2.*`, while the CUDA 13 examples do not pin a version, so the version policy is asymmetric.
- The source does not state pricing, performance gains, or comprehensive test coverage.
How do you install this skill?
Install the NVIDIA skills collection with npx skills add nvidia/skills, then select cuopt-install and an installation destination in the interactive prompts. Prepare an environment with an NVIDIA GPU of Compute Capability 7.0 or newer, CUDA 12.x or 13.x, and a compatible NVIDIA driver. Install cuOpt through one of the pip, conda, or Docker workflows in the skill. Do not run both pip and conda for the same interface, and do not install both Python pip variants.
How do you use this skill?
Ask an Agent that has loaded the skill, for example: “I need to install and verify the cuOpt Python API on an NVIDIA GPU machine with CUDA 12, preferably with pip.” The Agent will confirm the interface, CUDA version, package manager, and environment before giving the relevant commands. Use cuopt-developer instead for source builds or code contributions.
How does this skill compare with similar options?
The source distinguishes this skill from cuopt-developer: cuopt-install installs and verifies released cuOpt packages, while cuopt-developer is for building from source, contributing, or modifying the codebase.