MMCV installation and missing-operator errors: check the build combination
Distinguish full and lite distributions, source-build failures and runtime extension loading.
The final exit code is not a diagnosis
Current MMCV 2.x documentation calls the full distribution mmcv and the no-CUDA-ops distribution mmcv-lite. They must not coexist in one environment. Its install guide shows a .whl for a matched prebuilt package and .tar.gz when installation takes the source-build path; that log distinction is more informative than a final exit code. Older 1.x material may call the full package mmcv-full, so check the node's required MMCV major version before copying a command. Current installation guide
python -m pip show mmcv mmcv-lite torch
Using the interpreter that starts the affected node, record its required MMCV major version, Python version, Torch version/CUDA build and platform. Check whether both mmcv and mmcv-lite are installed. A missing _ext is not proof that the top-level Python package is absent: the lite route lacks CUDA ops, and a full build may fail to load native extensions. Preserve the exact import or loader error.
For installation failure, preserve the selected .whl or .tar.gz artifact and the first concrete compiler/linker error, not only exit code 1. Match the node requirement to an upstream-supported Python, Torch/CUDA, MMCV and platform combination; if no wheel exists, source build is a separate project in an isolated copy. Do not mix full and lite variants, substitute a similarly named wheel, or remove every Torch-related package as the first action. Installation guide
Validate the trusted top-level package import, the exact operator required by the node and a minimal original-node workflow separately. Printing a version number verifies only the first checkpoint; it does not prove that compiled ops load or execute. Test changes in a recoverable environment copy and record the other retained workflows.
Related checks
- SAM and GroundingDINO failures: separate code, checkpoints and extensions
- Nunchaku still fails after installation: align the node, wheel and model
Source review completed 2026-09-26. No local package installation, source build, operator load, workflow or GPU run was performed.
Was this page helpful?
Anonymous. Only a yes/no counter is stored; no account, IP address or device details.
Sources & references
2026-09-26 reviewed current MMCV 2.x installation guidance for full/lite variants, wheel/source path and version matching; English source review only. No local installation or operator/GPU run.
01MMCV 2.x installationSource checked: 2026-09-26