Skip to main content
Version: v1.2.0

MLflow scorer framework builds

The mlflow-scorer runtime serves MLflow models inside H2O MLOps. Most images ship with one ML framework pre-installed (for example, sklearn, pytorch, or tensorflow) and a specific Python version. H2O.ai does not maintain a public catalog of every framework, framework version, and Python version combination. Instead, H2O.ai builds each framework image on demand when requested.

The minimal variant is the exception: it ships no ML framework, installs the model's dependencies at startup, and is available out of the box without a build request. In H2O MLOps, it appears on the deployment page as [Py 3.X] Dynamic MLflow Model Scorer, where X is the Python version.

This page explains which frameworks and Python versions are available, how to request a build, what information to include in the request, and what level of validation each image receives.

note

For the list of MLflow model formats H2O MLOps accepts and how to package them, see MLflow model support.

Supported frameworks​

The following table lists the supported mlflow-scorer variants. The image size is approximate and varies with the framework version you request.

VariantApproximate sizeDescription
sklearn~450 MBscikit-learn classical ML models.
xgboost~550 MBXGBoost gradient-boosted models.
lightgbm~500 MBLightGBM gradient-boosted models.
catboost~600 MBCatBoost gradient-boosted models.
pytorch~2.4 GBPyTorch neural networks.
tensorflow~1.9 GBTensorFlow and Keras deep learning models.
onnx~350 MBModels exported to the Open Neural Network Exchange (ONNX) format and served through ONNX Runtime.
minimal~250 MBNo ML framework pre-installed. The runtime installs the model's dependencies from requirements.txt at startup. Available out of the box; no build request required.

H2O.ai can build an image for any released version of these frameworks, subject to compatibility with the available Python versions.

Every image targets the linux/amd64 platform and installs CPU-only framework builds.

note

Every image, including minimal, includes fallback copies of pandas, opencv-python-headless, mlflow-skinny, and pyyaml in /opt/fallback-libs. The runtime uses these only when the model's requirements.txt does not pin those packages.

Supported Python versions​

mlflow-scorer images support the following Python versions:

  • 3.10
  • 3.11
  • 3.12
  • 3.13

Each image tag includes both the framework version and the Python version (for example, h2oai/mlflow-scorer:pytorch-2.6.0-py3.13).

What "supported" means​

H2O.ai's continuous integration runs smoke tests on only two configurations to catch internal regressions: sklearn 1.7.1 and pytorch 2.10.0, both on Python 3.12. On-demand builds cover every other framework, framework version, and Python-version combination. These builds receive no automated regression testing.

In practice:

  • H2O.ai delivers on-demand builds on a best-effort basis. No formal service level agreement (SLA) applies to build turnaround or to ongoing maintenance of a specific image.
  • Test every on-demand image end-to-end with your own model before promoting it to production. Confirm scoring outputs, latency, memory footprint, and behavior under your expected load.
  • Treat each delivered image as immutable. If you need a patched version (for example, a security fix or a new framework patch release), request a new build.

How to request a build​

Contact H2O.ai support to request an mlflow-scorer build. Ask your contact to confirm the turnaround time. Include the following information in your request:

  • Framework name: One of sklearn, xgboost, lightgbm, catboost, pytorch, tensorflow, or onnx.
  • Framework version: The exact version you need (for example, 2.5.1).
  • Python version: One of 3.10, 3.11, 3.12, or 3.13.
  • Special requirements: For example, extra system libraries or specific versions of fallback libraries (such as pandas or opencv-python-headless).

Use the image after delivery​

After H2O.ai delivers your image, an administrator registers it in H2O MLOps as a runtime image version. You can register the image through the Helm chart (for bootstrap-managed runtimes) or through the Python client (for API-managed runtimes). For step-by-step instructions, see:

Feedback