"Choose MMCV package variants, inspect compiled ops availability,
Scanned 9/8/2026
Install to Claude Code
npx -y skills add VectorSpaceLab/AREX-Skill --skill ops-and-builds --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Ops And Builds?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/vectorspacelab-ops-and-builds)More formats (shields.io, HTML) on the badges page.
---
name: ops-and-builds
description: "Choose MMCV package variants, inspect compiled ops availability,
and troubleshoot install, build, CUDA, and backend failures."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
---
# MMCV Ops And Builds
Use this sub-skill when a task involves installing MMCV, deciding between `mmcv` and `mmcv-lite`, checking whether compiled `mmcv.ops` are available, selecting a wheel or source build, or diagnosing native-extension/CUDA/backend errors.
## Route Here
- The user needs NMS, ROI/rotated ROI pooling, deformable convolution, sparse convolution, 3D/point-cloud ops, attention kernels, focal losses, or other `mmcv.ops` APIs.
- The user reports `No module named mmcv.ops`, `No module named mmcv._ext`, shared-object load errors, undefined symbols, CUDA kernel errors, or build failures.
- The user asks which package to install for CPU-only, CUDA, MLU, MUSA, NPU, MPS, or lite/no-ops workflows.
- The user needs verification commands after installing or building MMCV.
## Route Elsewhere
- Pure image/video/visualization utilities belong in `../media-processing/`.
- Data transform pipelines belong in `../data-transforms/`.
- CNN layer builders that do not require compiled `mmcv.ops` belong in `../cnn-model-building/`.
## Start With Package Selection
1. Use `mmcv-lite` when the workflow does not import `mmcv.ops` and does not need compiled native kernels.
2. Use full `mmcv` when any compiled op is required; in MMCV 2.x, `mmcv` is the full package and `mmcv-lite` is the lite package.
3. Do not install both packages in one environment. If switching variants, uninstall the other variant first.
4. Treat compiled ops as conditional: successful `import mmcv` does not prove that `mmcv.ops` or `mmcv._ext` exists.
## Key References
- Read [references/install-and-build.md](references/install-and-build.md) for package choice, MIM/pip/source-build flows, environment variables, compatibility checks, and verification commands.
- Read [references/ops-reference.md](references/ops-reference.md) for ops families, representative APIs, and backend support cautions.
- Read [references/troubleshooting.md](references/troubleshooting.md) for symptom-to-cause recovery guidance.
- Run [scripts/check_mmcv_install.py](scripts/check_mmcv_install.py) to inspect installed package facts without relying on the original checkout.
## Diagnostic Commands
```bash
python scripts/check_mmcv_install.py
python scripts/check_mmcv_install.py --require-ops
python scripts/check_mmcv_install.py --require-cuda --require-ops
```
The checker reports import/package/Torch/CUDA facts, handles missing ops gracefully, and exits nonzero only when a required check fails.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!