I updated the issue.
The problem is related to the release of scipy 1.8.0 on February 5. I worked around the problem by going back to 1.7.3. I think any models trained in an environment with Scipy 1.8.0 will fail when deployed to a production environment.