cuPyNumeric Installation Guide
Guides safe cuPyNumeric installation and verifies that it actually works.
The skill explicitly forbids agent-run installation, requires environment isolation, and scopes itself to prebuilt conda/pip usage; verification, GPU resource use, and likely CUDA failures are disclosed. Points are deducted because the Miniforge bootstrap uses an unpinned latest URL, lacks hash or signature verification, exposes user-run curl/bash and rm -rf commands, and gives no complete rollback or dependency-security guidance.
The installation paths, prerequisite checks, smoke test, GPU check, and troubleshooting flow are fairly detailed, with expected outputs. Points are deducted because this is static-only review with no execution, several commands are Linux-shell dependent, some paths rely on local environment details, and the main text requires legate-based verification while the reference file also presents direct-python checks; failure coverage remains incomplete.
The audience, supported platforms, Python/CUDA/GPU requirements, CPU variant, pip/conda paths, and out-of-scope source-build and multi-node cases are reasonably clear. Points are deducted because Chinese-language support and mainland-China network reachability are not addressed; GitHub, conda-forge, NVIDIA documentation, and package-source availability are unassessed, and users must still resolve some GPU/container boundary conditions.
The material provides front matter, author, version 2.0.0, license statements, upstream and documentation references, layered sections, limitations, and troubleshooting; the skill card also states output expectations. Points are deducted for the missing recommended Examples section, absent changelog, unclear maintenance owner and update path, and the prompt-supplied repository license metadata of NOASSERTION, which leaves license governance requiring confirmation.
For common single-node installation, isolated environment selection, CPU/GPU variant handling, and post-install verification, the skill offers directly reusable commands and should cover the core task. Points are deducted because static review cannot establish success for the claimed versions and platforms, GPU verification is resource-intensive, and several conclusions still depend on user judgment or external documentation.
Expected values, diagnostic commands, reference verification material, and 24 evaluation cases provide some auditability for key claims. Points are deducted because no real CI workflow, committed executable test suite with results, or independent third-party corroboration is shown; BENCHMARK.md is repository self-reporting and is not independent execution evidence for this static review.
- Do not treat the repository benchmark PASS or uplift percentages as independent validation; no commands were executed for this review.
- The Miniforge bootstrap downloads and executes an unpinned remote script; users should review the source and consider hashes, signatures, or an internal mirror.
- The GPU verification workload allocates large matrices and runs for an extended period, potentially consuming substantial GPU memory, compute time, and temporary storage; confirm resource limits first.
- The front-matter license and the prompt-supplied NOASSERTION repository metadata create a governance uncertainty that should be resolved before publication.
What does this skill do, and when should you use it?
This Agent Skill from NVIDIA’s skills repository guides users through installing and verifying cuPyNumeric with conda or pip. It checks operating-system, Python, CUDA, GPU, and package-manager prerequisites before recommending an isolated environment. Verification includes a NumPy-compatible smoke test and, when a supported GPU is present, a separate check that confirms GPU use. Source builds, multi-node networking, MPI, and UCX are outside its scope.
Recommends read-only checks using conda, pip, nvidia-smi, nvcc, and version commands; provides bootstrap commands for Miniforge or Python/pip; gives conda and pip installation commands; creates temporary Python scripts and runs them through the legate launcher for smoke testing; performs GPU launch checks and nvidia-smi sampling; and provides troubleshooting directions for import errors, CUDA mismatches, inactive environments, and incorrect CPU/GPU variants. It does not execute installation commands for the user.
- A Python user installing cuPyNumeric on a Linux GPU server and verifying a Volta-or-newer GPU is used.
- A user on a CPU-only laptop who wants an isolated environment for cuPyNumeric.
- A container builder whose GPU is not visible during installation and who needs CONDA_OVERRIDE_CUDA.
- A user with an existing installation who needs to diagnose environment mismatches, CUDA errors, or missing GPU utilization.
What are this skill's strengths and limitations?
- Supports both conda and pip installation paths.
- Requires environment isolation and prohibits the Agent from running installers.
- Provides copyable smoke tests and a separate GPU-use verification procedure.
- Addresses CPU, GPU, container, and common environment-mismatch scenarios.
- The user must run all installation commands themselves.
- GPU use requires CUDA 12.2 or newer and Volta-or-newer hardware.
- Source builds, multi-node operation, MPI, and UCX are out of scope.
- On macOS aarch64, only pip wheels are supported; Windows requires WSL.
How do you install this skill?
Install the skills/cupynumeric-install/ directory from the repository into a compatible Agent Skills client. The README provides: npx skills add nvidia/skills --skill cupynumeric-install --yes. Alternatively, run npx skills add nvidia/skills and select the skill and destination interactively. The Skill does not document how to install the npx skills CLI itself.
How do you use this skill?
After the Skill is loaded, ask an Agent something like “Guide me through installing and verifying cuPyNumeric on this Linux GPU machine” or “Help diagnose why cuPyNumeric is not using my GPU.” The Agent should check prerequisites, choose conda or pip, print installation commands for the user to run, then perform the legate smoke test and, when relevant, the GPU-use checks. Source builds are not covered.