holoscan-install-conda

holoscan-install-conda is an official NVIDIA install skill that guides installing the Holoscan SDK (v4.3+) — the Python runtime and/or C++ development headers — into a Conda environment on Linux x86_64 with a CUDA 13 driver. It solves the fiddly problem of correct channel selection and CUDA metapackage pinning: packages come from a mix of conda-forge (holoscan, libholoscan, libholoscan-dev) and rapidsai (rmm, ucxx), and the CUDA version must be pinned so the solver picks compatible runtime libraries. The skill routes CUDA 12 hosts to the container or wheel install skills instead.

The workflow first consults the official NVIDIA installation docs (treated as authoritative if they disagree), checks prerequisites (conda via Miniforge, nvidia-smi), then creates a Conda environment and installs the right package set for the user's goal — holoscan for Python-only, libholoscan-dev for C++ development, or both — plus rmm, ucxx, and cuda-version=13 from the rapidsai and conda-forge channels, optionally adding a cxx-compiler/cmake/ninja toolchain. It documents the roles of each package, verification commands (import holoscan, checking the include directory), and a runtime note to set ulimit -s 32768 to avoid segfaults. A published skill-card and BENCHMARK.md record NVSkills-Eval results across security, correctness, discoverability, effectiveness, and efficiency dimensions.

It targets developers and engineers setting up the Holoscan SDK for GPU-accelerated streaming AI applications in a Conda environment. Commands are standard installation operations (conda create/install, a wget of the official Miniforge installer, optional sudo for system dependencies) with no destructive or exfiltrating behavior.