cupynumeric-install

cuPyNumeric Install is an NVIDIA-authored guidance skill for installing and verifying cuPyNumeric (a distributed, GPU-accelerated drop-in for NumPy built on Legate) for use from Python. It solves the problem of getting a working, correctly isolated cuPyNumeric environment via conda or pip and then proving the install actually works, including confirming genuine GPU usage. Building cuPyNumeric from source is explicitly out of scope.

A defining feature is its safety posture: the skill's mandatory rules state it must never run installers itself. It prints commands for the user to run, always isolates into a dedicated environment (never base conda, system Python, or shared global envs), and only performs read-only version detection before recommending anything. It documents prerequisites (GPU compute capability 7.0+, CUDA 12.2+, supported OS and Python versions, conda 24.1+), scoping questions, conda and pip install paths, an optional forced-GPU-variant override, and a nightly channel. Verification is thorough: a self-contained smoke test run through the legate launcher with expected outputs, a mandatory GPU-usage check (since a CPU-variant install on a GPU box still returns correct results), plus bundled reference commands for nvidia-smi sampling, verbose Legate startup, package-version checks, CPU-only fallback, and container sanity checks. A BENCHMARK.md documents that the skill passed NVSkills-Eval across security, correctness, discoverability, effectiveness, and efficiency dimensions.

Target users are data scientists, ML engineers, and HPC practitioners who want cuPyNumeric running reliably on a laptop, server, cloud, or container and want to confirm GPU acceleration is actually engaged.