This repository contains the scripts and code to host an API for the Descartes ML model for generating article descriptions for Wikipedia articles.
Scripts for setting up a Cloud VPS instance, the necessary web serving architecture, and updating the API with code updates. For more details about the specific configuration files and their roles, see the original template.
Code for running the Flask app and model. A few components:
transformers: modified HuggingFace code that runs the underlying Descartes model.utils/utils.py: utilities for loading in the model and making predictions.wsgi_template.py: Flask app with code for taking article names, gathering model features, and returning model outputs.
This repository assumes two things already are in place:
- A Cloud VPS instance has been created. The current one has 8GB RAM and 8 VCPUs but likely the API can run on 4GB RAM and 4 VCPUs.
- The model binary / config has been uploaded to the server. Currently those dependencies are stored on a Cinder volume under
/srv/model-25lang-all/. - You can ssh onto the service once you have the correct Cloud VPS config via something like
ssh isaacj@android-machine-generated-desc.recommendation-api.eqiad1.wikimedia.cloud
Then, the cloudvps_setup.sh script can be copied onto the server and run (sudo. /cloudvps_setup.sh).
The instance takes about a minute to load in the model. Progress can be checked via sudo tail -n50 /var/log/uwsgi/uwsgi.log and the process is complete when uWSGI has spawned a worker.
Code updates can be incorporated by running release.sh
- If no responses are working, ssh onto the server and try
systemctl status modelto see the status of the service - Run
sudo tail -n100 /var/log/uwsgi/uwsgi.logto see if there are any errors being displayed
If the whole system is down for some reason, there are a few things to try:
- Restart the server manually via Horizon
- ssh onto the server and run
sudo /home/isaacj/release.shto generally reset things