Currently, the entrypoint used for CMS containers only allows serving from local model files. This can be restrictive in cases where using an MLflow URI is desired. Given that the functionality is already supported by the backend and exposed through the CLI, the entrypoint script can simply be extended to use the --mlflow-model-uri option if the associated environment variable is provided on deployment (e.g. through the compose file) instead of the model path.
The added flexibility will benefit upstream apps like the CogStack Model Gateway, which allows deploying CMS instances, but currently has to hardcode the complete entrypoint command (in order to explicitly set the desired flags), leading to potential inconsistencies when changes are introduced on the CMS side.
Currently, the entrypoint used for CMS containers only allows serving from local model files. This can be restrictive in cases where using an MLflow URI is desired. Given that the functionality is already supported by the backend and exposed through the CLI, the entrypoint script can simply be extended to use the
--mlflow-model-urioption if the associated environment variable is provided on deployment (e.g. through the compose file) instead of the model path.The added flexibility will benefit upstream apps like the CogStack Model Gateway, which allows deploying CMS instances, but currently has to hardcode the complete entrypoint command (in order to explicitly set the desired flags), leading to potential inconsistencies when changes are introduced on the CMS side.