Deploying inference pipelines