Ori Kronfeld, 29/10/25

🔍 Overview

MLFlow is a widely used UI MLOps tool for tracking details of ML experiments:

• hyperparameters of model and training like latent dimensions, batch keys, learning rates etc’..
• model and system metrics like loss curves and cpu/gpu performance
• downstream analysis of single cell like UMAPs, DE, correlation matrix, scib and so on

and much more can be logged, compared between experiments, shared between researchers, and reproduced at any time, which is crucial in debugging and optimizing sc ml models.

We recently integrated MLFlow into SCVI-Tools and installed it on-prem in our lab (via the workstations). In order to use it, you need to be connected to WISecure/VPN.

In what follows on this page, I will show how to integrate your code with MLFLOW, log additional downstream tasks, explore the UI, and show a complete code example as a reference.

📦 Setup

We made sure the integration of scvi + mlflow will be as simple as possible:

1. Install scvi-tools latest version with support for MLflow:

pip install scvi-tools[mlflow]