Hi everyone,
I am trying to deploy a custom Python ML package to UiPath AI Center on-premise (v25.10) running on RKE2 Kubernetes v1.34 on RHEL 8.10. The package is configured with Python 3.12.
Since our Kubernetes pods do not have internet access due to enterprise firewall restrictions, I am bundling all required Python wheels and the model files directly inside the ZIP package. The requirements.txt is configured to install from a local packages folder using --find-links and --no-index so that no internet connection is needed during deployment.
The ZIP structure follows AI Center’s validation requirement of having a unique root folder containing main.py, requirements.txt, the model files, and the pre-downloaded wheel files.
However, when the ML Skill is deployed, the pod goes into CrashLoopBackOff with the following logs:
contentDir is not empty, reusing content
installationDir is empty, skipping pip install
ModuleNotFoundError: No module named 'transformers'
It appears AI Center is skipping the pip install step entirely, which means our bundled wheels are never installed and the package fails to load.
My questions are:
- What exactly is
installationDirand where does AI Center expect to findrequirements.txtafter extracting the ZIP? - Is bundling wheels inside the ZIP using
--find-linksand--no-indexa supported approach for air-gapped/offline environments? - Is there a way to force AI Center to re-run pip install and clear the content cache on-premise?
Any guidance would be greatly appreciated. Thank you!