I would recommend updating this post to include discussion of uv as a depenedency manager instead of conda. MLflow added support for uv in version 2.20.0 (docs here, PR here). 

The current section 4.5.1 is somewhat misleading given the availability of non-conda managers in MLflow. This isn't really true:
"...the Custom PyFunc Model's required dependencies are specified using conda_env. Given that MLflow continues to use Conda internally to capture environment dependencies, users are reminded to make a note of Anaconda license requirements."