Implementation:Mlflow Mlflow ML Package Versions

Knowledge Sources	Mlflow_Mlflow
Domains	CI/CD, Package Management, Testing, Configuration
Last Updated	2026-02-13 20:00 GMT

Overview

Source-of-truth YAML configuration defining version compatibility ranges, test commands, and dependency requirements for all ML frameworks supported by MLflow.

Description

ml-package-versions.yml is the master configuration file that drives MLflow's cross-version compatibility testing infrastructure. It defines entries for 40+ ML frameworks, each specifying package metadata, supported version ranges, test commands, and conditional dependencies.

Each top-level key in the YAML is a flavor name (e.g., sklearn, pytorch, openai, langchain) and contains the following structure:

package_info: Package metadata including:

pip_release -- The PyPI package name (e.g., "scikit-learn" for the sklearn flavor)
install_dev -- Shell commands to install the development version from source
module_name -- Python import name if different from flavor name (e.g., "torch" for pytorch)
repo -- GitHub repository URL for source-based Python version inference
genai -- Boolean flag indicating if the package is a GenAI integration

models: and autologging: Test categories, each containing:

minimum / maximum -- Supported version range boundaries
unsupported -- Version specifiers for known-broken versions (e.g., "== 4.52.1")
requirements -- Conditional dependencies keyed by version specifiers (e.g., ">= 0.0.0" for all versions, "< 2.3" for older versions)
run -- pytest commands to execute for this category
python / java / runs_on -- Version-specific overrides for runtime environments
test_every_n_versions -- Sampling rate for testing (e.g., 4 means test every 4th minor version)
test_tracing_sdk -- Whether to run tracing SDK test variants
allow_unreleased_max_version -- Permit testing versions not yet on PyPI (used for pyspark)
pre_test -- Commands to run before tests (e.g., apt-get installs, model downloads)

Covered Frameworks: The file covers traditional ML (sklearn, xgboost, lightgbm, catboost, statsmodels), deep learning (pytorch, pytorch-lightning, keras, tensorflow, transformers, sentence_transformers), NLP (spacy), time series (prophet, pmdarima), big data (spark), ONNX, GenAI model providers (openai, anthropic, gemini, mistral, groq, bedrock), GenAI frameworks (langchain, langgraph, llama_index, dspy, crewai, autogen, ag2, pydantic_ai, smolagents, strands, haystack, agno, semantic_kernel), and utility libraries (litellm, shap, h2o, paddle, johnsnowlabs).

Usage

This file is consumed by dev/set_matrix.py to generate CI test matrices and by dev/update_ml_package_versions.py to keep version ranges current. It is also used to auto-generate mlflow/ml_package_versions.py.

Code Reference

Source Location

Repository: Mlflow_Mlflow
File: mlflow/ml-package-versions.yml
Lines: 1-1063

Signature

# Top-level structure for each flavor entry
flavor_name:
  package_info:
    pip_release: "package-name"
    install_dev: |
      pip install git+https://github.com/org/repo.git
    module_name: "import_name"
    repo: https://github.com/org/repo/tree/HEAD
    genai: true  # optional, for GenAI packages
  models:
    minimum: "1.0.0"
    maximum: "2.0.0"
    unsupported: ["== 1.5.0"]
    requirements:
      ">= 0.0.0": ["dep1", "dep2"]
      "< 1.5": ["dep3<2.0"]
    run: |
      pytest tests/flavor/test_model_export.py
  autologging:
    minimum: "1.0.0"
    maximum: "2.0.0"
    run: |
      pytest tests/flavor/test_autolog.py
    test_tracing_sdk: true

Import

# Read by set_matrix.py
import yaml
with open("mlflow/ml-package-versions.yml") as f:
    config = yaml.safe_load(f)

# Or via the auto-generated Python module
from mlflow.ml_package_versions import _ML_PACKAGE_VERSIONS, FLAVOR_TO_MODULE_NAME

I/O Contract

Inputs

Name	Type	Required	Description
N/A	N/A	N/A	This is a static configuration file; updated by dev/update_ml_package_versions.py

Outputs

Name	Type	Description
YAML configuration	dict[str, FlavorConfig]	Parsed by set_matrix.py into FlavorConfig Pydantic models for test matrix generation
mlflow/ml_package_versions.py	Python module	Auto-generated derivative containing _ML_PACKAGE_VERSIONS and FLAVOR_TO_MODULE_NAME

Covered Frameworks

Traditional ML

Flavor	PyPI Package	Categories
sklearn	scikit-learn	models, autologging
xgboost	xgboost	models, autologging
lightgbm	lightgbm	models, autologging
catboost	catboost	models
statsmodels	statsmodels	models, autologging
prophet	prophet	models
pmdarima	pmdarima	models

Deep Learning

Flavor	PyPI Package	Categories
pytorch	torch	models, autologging
pytorch-lightning	pytorch-lightning	models, autologging
keras	keras	models, autologging
tensorflow	tensorflow	models, autologging
transformers	transformers	models, autologging
sentence_transformers	sentence-transformers	models

GenAI

Flavor	PyPI Package	Categories
openai	openai	models, autologging
langchain	langchain	models, autologging
langgraph	langgraph	models, autologging
llama_index	llama-index	models, autologging
dspy	dspy	models, autologging
anthropic	anthropic	autologging
gemini	google-genai	autologging
crewai	crewai	autologging
pydantic_ai	pydantic-ai	autologging
smolagents	smolagents	autologging
strands	strands-agents	autologging
bedrock	boto3	autologging

Usage Examples

Adding a New Flavor

new_framework:
  package_info:
    pip_release: "new-framework"
    genai: true
    repo: https://github.com/org/new-framework/tree/HEAD
    install_dev: |
      pip install git+https://github.com/org/new-framework
  autologging:
    minimum: "1.0.0"
    maximum: "2.0.0"
    requirements:
      ">= 0.0.0": ["openai"]
    run: |
      pytest tests/new_framework/test_autolog.py
    test_tracing_sdk: true

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment