---

name: build-and-install description: "Guides package users through installing, building, and diagnosing MinkowskiEngine with PyPI, source, Docker, CPU/CUDA, BLAS, compiler, and import-check triggers." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Build and Install Router

Use this operating sub-skill when the user asks to install, build, rebuild, or diagnose MinkowskiEngine. Trigger terms include pip install, setup.py, Docker, CPU_ONLY, force_cuda, CUDA_HOME, nvcc, cublas_v2.h, BLAS not found, openblas, mkl, MAX_JOBS, OMP_NUM_THREADS, undefined symbol, pkg_resources, and import checks such as ME.is_cuda_available().

Route the request

• Choose an install/build command: read references/build-reference.md, then use python scripts/build_command_helper.py --help or a dry-run helper invocation to print a command. Prefer CPU-only commands unless the user explicitly needs CUDA and has a PyTorch CUDA build plus a matching CUDA toolkit with nvcc.
• Diagnose build or import failures: read references/troubleshooting.md and match the observed symptom text before recommending a rebuild.
• Verify a completed install: use the import checks in references/build-reference.md. Do not claim CUDA support unless MinkowskiEngine.is_cuda_available() returns True in the user's installed environment.
• Handle package usage questions after install: route sparse tensor and quantization tasks to ../sparse-tensor-data/SKILL.md, layer/network tasks to ../layers-and-networks/SKILL.md, and training/demo adaptation to ../training-and-demos/SKILL.md.

Operating guardrails

1. Keep CUDA conditional. A CPU-only build is a valid install route; CUDA instructions are documented but require user-side verification.
2. Do not point the user to the original repository documentation for missing details. Use the bundled references here.
3. Prefer safe dry-run planning before running build commands. The bundled helper prints commands only and never installs packages.
4. When a path is needed, ask the user for their active environment, source tree, CUDA toolkit, or BLAS locations instead of inventing local paths.