---

name: environment-and-installation description: "Install and preflight the legacy VoxFormer/OpenMMLab CUDA stack, build native operators, and diagnose version, toolchain, import, and optional preprocessing-environment failures." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Environment and installation

Use this route when a request mentions installing VoxFormer, OpenMMLab versions, CUDA/MMCV compatibility, mmdetection3d native operators, deform_attn_3d, deform3D configs, import failures, or the optional MobileStereoNet environment.

This is a legacy, CUDA-first application checkout rather than a pip package. The standard and deform3D model families require an NVIDIA-capable PyTorch runtime for truthful execution; CPU is useful for parsing and layout checks but is not a substitute for model execution or native CUDA operators.

Route

1. Read install-matrix.md and install the pinned Python 3.8-era stack in the documented order. Keep the core VoxFormer environment separate from optional preprocessing environments.

2. Run the safe, read-only checker:

   python skills/disco/voxformer/sub-skills/environment-and-installation/scripts/check_environment.py --help
   python skills/disco/voxformer/sub-skills/environment-and-installation/scripts/check_environment.py --repo-root <VOXFORMER_ROOT>
   

   A missing package, CUDA compiler, or extension is a diagnostic result, not permission to silently substitute a different version or CPU execution.

3. Follow native-build-and-smoke.md to build mmdetection3d from tag v0.17.1, optionally build the custom deform3dattn_custom_cn extension, and run import/backend smoke tests.

4. For a failure, use the symptom-first recovery in troubleshooting.md. Re-check after each layer: Python/Torch, MMCV, OpenMMLab, native ops, then project imports.

5. After environment readiness, hand data artifacts to dataset-preparation, choose a model and config through model-configuration, and only then use training-and-evaluation.

Core versus optional variants

• Core baseline: Python 3.8, PyTorch 1.9.1 with the documented cu111 runtime, torchvision 0.10.1, mmcv-full 1.4.0, mmdet 2.14.0, mmsegmentation 0.14.1, mmdetection3d v0.17.1 built from source, and timm 0.6.13. Standard configs still depend on CUDA and MMCV's compiled deformable-attention operator.
• Custom deform3D: the deform_attn_3d CUDA extension and the voxformer-*-deform3D.py configs. This is a separate native-build gate, not a CPU fallback. The repository's custom wrapper contains an intentional path placeholder and raises before import in an unmodified checkout; never report that stock import as working.
• Optional MobileStereoNet: a dependency-conflicting legacy image-to-depth environment (Python 3.6, PyTorch 1.4.0, torchvision 0.5.0, CUDA 10.0 per the repository's preprocessing notes). Isolate it and do not install it into the core environment. It has no claimed CPU substitute.

Operating contract

Inputs: a VoxFormer checkout, a user-owned environment, and any requested variant (standard, deform3D, or optional preprocessing). Paths must be supplied by the user as placeholders; do not invent a checkout or Conda prefix.

Outputs: pinned package/version observations, CUDA and compiler readiness, readiness of mmdet3d.ops and the custom extension when requested, and a clear handoff to data/config or train/test routes. The checker never installs, builds, downloads, trains, evaluates, or modifies files.

Stop conditions: stop before training or evaluation when the required CUDA runtime, MMCV native operator, mmdetection3d native ops, or selected custom extension is unavailable. Report the exact blocked layer and route to the fallback only when a documented standard variant is genuinely applicable.

Scope boundaries

This route does not explain SemanticKITTI conversion or artifact generation; use dataset-preparation. It does not define model architecture, config semantics, or variant selection; use model-configuration. It does not run full training, testing, or evaluation; use training-and-evaluation. The project entry router is voxformer.