---

name: native-build-and-cpp description: "Route cuML source-build, libcuml, C++/CUDA API, and native test workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

native-build-and-cpp

Use this sub-skill when the task is about building cuML from source, choosing between libcuml and cuml source-build targets, validating the C++/CUDA toolchain, using public C++ APIs, selecting CTest or pytest partitions for source changes, or reviewing native/Python contribution pitfalls.

Do not use this sub-skill for ordinary installed-package estimator workflows, cuml.dask operating recipes, cuml.accel activation, release automation, CI secrets, devcontainer maintenance, or benchmark execution unless the user explicitly asks for a source-build benchmark.

Route

1. If the user only needs to run cuML algorithms from Python, route back to the root package workflow or python-estimators; source builds are only required for local source edits, C++ API work, package-build validation, ABI/linkage debugging, or native tests/examples.
2. Before any source build, run the bundled non-mutating probe in scripts/source_build_probe.py to check GPU visibility, compute capability hints, nvcc, compiler, CMake, Ninja, Python, and Cython readiness. The probe reports evidence only and does not configure or build.
3. Use references/source-build-and-cpp.md to choose the smallest build path: libcuml for C++/CUDA library and C++ tests/examples, cuml for the Python package after matching libcuml, prims for primitive tests, and benchmarks only on explicit request.
4. Use references/source-build-and-cpp.md for manual CMake flags, build.sh target/flag taxonomy, C++ example patterns, and native test partition commands.
5. Use references/troubleshooting.md when CUDA hardware, toolkit, RAPIDS dependency pins, CMake, linker, source-package, or native-test failures appear.

Safety and verification boundaries

• CUDA runtime validation needs an NVIDIA GPU. Building may configure without a GPU, but C++ examples, CTest GPU tests, Python estimator tests, and health checks require a compatible CUDA device.
• Match the CUDA major/minor and RAPIDS package version family across libcuml, cuml, RAFT, RMM, cuVS, cuDF, pylibraft, and optional Dask packages. Do not mix stable/nightly or CUDA 12/13 package suffixes.
• Prefer focused configure/build/test commands over broad CI scripts. Use benchmarks and multi-GPU tests only when the requested change requires them and the environment has the extra dependencies.