●Change8

v2.20.0

📦 mlflowView on GitHub →
✨ 11 features🐛 8 fixes🔧 18 symbols

Summary

MLflow 2.20.0 introduces type‑hint‑based model signatures, Bedrock/Groq tracing, inline Jupyter trace rendering, uv‑powered model validation, a new chat panel, and several API enhancements along with numerous bug fixes.

✨ New Features

  • Type Hint-Based Model Signature: define model signature from type hints in PythonModel.predict and validate input payloads.
  • Bedrock / Groq Tracing Support: one-line auto-tracing for Amazon Bedrock and Groq LLMs via mlflow.bedrock.tracing or mlflow.groq.tracing.
  • Inline Trace Rendering in Jupyter Notebook: render trace UI directly inside notebooks.
  • Faster Model Validation with uv Package Manager: mlflow.models.predict now supports uv for rapid environment validation.
  • New Chat Panel in Trace UI: unified chat panel for LLM invocations showing messages, function calls, and raw payloads.
  • Introduced ChatAgent base class for defining custom python agents.
  • Supported Tool Calling in DSPy Tracing.
  • Applied timeout override to within-request local scoring server for Spark UDF inference.
  • Supported dictionary type for inference parameters.
  • Made context parameter optional when calling PythonModel instances.
  • Set default task for ChatModel.

🐛 Bug Fixes

  • Tracking: fixed filename encoding issue in log_image.
  • Models: fixed faithfulness metric handling for custom override parameters.
  • Artifacts: updated presigned URL list_artifacts to return an empty list instead of raising an exception.
  • Tracking: fixed rename permission in model registry.
  • Tracking: removed hard dependency on langchain package in autologging.
  • Tracking: fixed constraint name for MSSQL in migration 0584bdc529eb.
  • Scoring: fixed uninitialized loaded_model variable.
  • Model Registry: returned empty array when DatabricksSDKModelsArtifactRepository.list_artifacts is called on a file.

Affected Symbols

PythonModel.predictmlflow.bedrock.tracingmlflow.groq.tracingmlflow.models.predictChatAgentDSPy tracing tool callingSpark UDF scoring server timeout overrideinference params dictionary handlingPythonModel.context parameterChatModel default tasklog_imagefaithfulness metric implementationlist_artifacts (presigned URL handling)model registry rename permission logicautologging langchain dependencyMSSQL migration constraint nameloaded_model variable in scoringDatabricksSDKModelsArtifactRepository.list_artifacts