`build-CUDA (3.7, 1.8)` fails with `ModuleNotFoundError: No module named 'apex'`

[akihironitta]

🐛 Bug

See title and recent failures of the workflow:
https://github.com/PyTorchLightning/pytorch-lightning/actions/workflows/ci_dockers.yml

Seems like the workflow has been failing for quite a long time...

To Reproduce

Expected behavior

No failure.

Environment

• PyTorch Lightning Version (e.g., 1.5.0):
• PyTorch Version (e.g., 1.10):
• Python version (e.g., 3.9):
• OS (e.g., Linux):
• CUDA/cuDNN version:
• GPU models and configuration:
• How you installed PyTorch (conda, pip, source):
• If compiling from source, the output of torch.__config__.show():
• Any other relevant information:

Additional context

cc @carmocca @akihironitta @Borda @justusschock @awaelchli @rohitgr7

[Borda]

I guess that the apex is installed to another site-packages then default one