---

name: "setup-build" description: "Guides legacy-compatible AdelaiDet installation, Detectron2/PyTorch/CUDA version selection, editable extension builds, and runtime smoke checks." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

setup-build

Use this sub-skill when a task involves installing AdelaiDet, selecting PyTorch/Detectron2/CUDA versions, building adet._C, checking custom operators, or diagnosing import/build failures.

Use this route for

• Creating a fresh environment for AdelaiDet.
• Choosing a Detectron2 wheel that matches PyTorch and CUDA.
• Building AdelaiDet editable from source.
• Verifying adet.config.get_cfg, Detectron2 registries, adet._C, BezierAlign, DefROIAlign, and multi-label NMS.
• Fixing build failures such as missing torch in build isolation, missing CUDA headers, missing THC headers, Pillow/rapidfuzz/OpenCV incompatibilities, or CUDA architecture problems.

Do not use this route for

• Choosing a model config after installation succeeds. Use ../train-eval/SKILL.md.
• Demo commands with an already working install. Use ../demo-visualize/SKILL.md.
• Dataset annotation conversion. Use ../data-prep/SKILL.md.
• ONNX and checkpoint conversion. Use ../export-convert/SKILL.md.

Read first

• references/setup-build.md for the known-good install recipe and version pins.
• references/runtime-checks.md for what each smoke check proves.
• ../../references/compatibility.md for package-version rationale.
• ../../references/troubleshooting.md for cross-cutting error fixes.

Skill-owned scripts

• ../../scripts/check_install.py — package/config/registry/extension smoke check, with optional --cuda-ops.

Standard workflow

1. Create the compatible Python/PyTorch/CUDA/Detectron2 environment described in references/setup-build.md.

2. Install AdelaiDet with python -m pip install --no-build-isolation -e <checkout>.

3. Pin compatibility fixes: Pillow<10, rapidfuzz<3, NumPy 1.23.x, and OpenCV headless 4.8.x.

4. Run:

   python ../../scripts/check_install.py --cuda-ops
   

5. If --cuda-ops fails, do not start training or inference; fix setup first or explicitly narrow the task to import/config-only analysis.

Decision points

• Need unmodified CUDA extension? Use PyTorch 1.10.x + CUDA 11.3. PyTorch 2.x is not the default.
• No GPU available? You may perform import/config-only checks without --cuda-ops, but record that DefROIAlign and native _C.ml_nms are not validated.
• Need ONNX/TensorRT/NCNN? First prove the base AdelaiDet install, then switch to export-convert for optional deployment runtimes.

Success signal

A healthy CUDA-capable environment reports:

• import adet succeeds.
• adet._C exposes BezierAlign, DefROIAlign, and ml_nms symbols.
• scripts/check_install.py --cuda-ops exits 0.
• CLI help for training/demo/visualization/export can import all required modules.