Skip to main content

· 2 min read
MLflow maintainers

Major New Features​

  • 💡Type Hint-Based Model Signature: Define your model's signature in the most Pythonic way. MLflow now supports defining a model signature based on the type hints in your PythonModel's predict function, and validating input data payloads against it. (#14182, #14168, #14130, #14100, #14099, @serena-ruan)

  • 🧠 Bedrock / Groq Tracing Support: MLflow Tracing now offers a one-line auto-tracing experience for Amazon Bedrock and Groq LLMs. Track LLM invocation within your model by simply adding mlflow.bedrock.tracing or mlflow.groq.tracing call to the code. (#14018, @B-Step62, #14006, @anumita0203)

  • 🗒ï¸? Inline Trace Rendering in Jupyter Notebook: MLflow now supports rendering a trace UI within the notebook where you are running models. This eliminates the need to frequently switch between the notebook and browser, creating a seamless local model debugging experience. (#13955, @daniellok-db)

  • âš¡ï¸?Faster Model Validation with uv Package Manager: MLflow has adopted uv, a new Rust-based, super-fast Python package manager. This release adds support for the new package manager in the mlflow.models.predict API, enabling faster model environment validation. Stay tuned for more updates! (#13824, @serena-ruan)

  • 🖥ï¸? New Chat Panel in Trace UI: THe MLflow Trace UI now shows a unified chat panel for LLM invocations. The update allows you to view chat messages and function calls in a rich and consistent UI across LLM providers, as well as inspect the raw input and output payloads. (#14211, @TomuHirata)

Other Features:

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

· 2 min read
MLflow maintainers

MLflow 2.20.0rc0 is a release candidate for 2.20.0. To install, run the following command:

pip install mlflow==2.20.0rc0

Major New Features​

  • 💡Type Hint-Based Model Signature: Define your model's signature in the most Pythonic way. MLflow now supports defining a model signature based on the type hints in your PythonModel's predict function, and validating input data payloads against it. (#14182, #14168, #14130, #14100, #14099, @serena-ruan)

  • 🧠 Bedrock / Groq Tracing Support: MLflow Tracing now offers a one-line auto-tracing experience for Amazon Bedrock and Groq LLMs. Track LLM invocation within your model by simply adding mlflow.bedrock.tracing or mlflow.groq.tracing call to the code. (#14018, @B-Step62, #14006, @anumita0203)

  • 🗒ï¸? Inline Trace Rendering in Jupyter Notebook: MLflow now supports rendering a trace UI within the notebook where you are running models. This eliminates the need to frequently switch between the notebook and browser, creating a seamless local model debugging experience. (#13955, @daniellok-db)

  • âš¡ï¸?Faster Model Validation with uv Package Manager: MLflow has adopted uv, a new Rust-based, super-fast Python package manager. This release adds support for the new package manager in the mlflow.models.predict API, enabling faster model environment validation. Stay tuned for more updates! (#13824, @serena-ruan)

  • 🖥ï¸? New Chat Panel in Trace UI: THe MLflow Trace UI now shows a unified chat panel for LLM invocations. The update allows you to view chat messages and function calls in a rich and consistent UI across LLM providers, as well as inspect the raw input and output payloads. (#14211, @TomuHirata)

Other Features:​

Please try it out and report any issues on the issue tracker!

· 2 min read
MLflow maintainers

2.19.0 (2024-12-11)​

We are excited to announce the release of MLflow 2.19.0! This release includes a number of significant features, enhancements, and bug fixes.

Major New Features​

  • ChatModel enhancements - ChatModel now adopts ChatCompletionRequest and ChatCompletionResponse as its new schema. The predict_stream interface uses ChatCompletionChunk to deliver true streaming responses. Additionally, the custom_inputs and custom_outputs fields in ChatModel now utilize AnyType, enabling support for a wider variety of data types. Note: In a future version of MLflow, ChatParams (and by extension, ChatCompletionRequest) will have the default values for n, temperature, and stream removed. (#13782, #13857, @stevenchen-db)

  • Tracing improvements - MLflow Tracing now supports both automatic and manual tracing for DSPy, LlamaIndex and Langchain flavors. Tracing feature is also auto-enabled for mlflow evaluation for all supported flavors. (#13790, #13793, #13795, #13897, @B-Step62)

  • New Tracing Integrations - MLflow Tracing now supports CrewAI and Anthropic, enabling a one-line, fully automated tracing experience. (#13903, @TomeHirata, #13851, @gabrielfu)

  • Any Type in model signature - MLflow now supports AnyType in model signature. It can be used to host any data types that were not supported before. (#13766, @serena-ruan)

Other Features:

  • [Tracking] Add update_current_trace API for adding tags to an active trace. (#13828, @B-Step62)
  • [Deployments] Update databricks deployments to support AI gateway & additional update endpoints (#13513, @djliden)
  • [Models] Support uv in mlflow.models.predict (#13824, @serena-ruan)
  • [Models] Add type hints support including pydantic models (#13924, @serena-ruan)
  • [Tracking] Add the trace.search_spans() method for searching spans within traces (#13984, @B-Step62)

Bug fixes:

  • [Tracking] Allow passing in spark connect dataframes in mlflow evaluate API (#13889, @WeichenXu123)
  • [Tracking] Fix mlflow.end_run inside a MLflow run context manager (#13888, @WeichenXu123)
  • [Scoring] Fix spark_udf conditional check on remote spark-connect client or Databricks Serverless (#13827, @WeichenXu123)
  • [Models] Allow changing max_workers for built-in LLM-as-a-Judge metrics (#13858, @B-Step62)
  • [Models] Support saving all langchain runnables using code-based logging (#13821, @serena-ruan)
  • [Model Registry] return empty array when DatabricksSDKModelsArtifactRepository.list_artifacts is called on a file (#14027, @shichengzhou-db)
  • [Tracking] Stringify param values in client.log_batch() (#14015, @B-Step62)
  • [Tracking] Remove deprecated squared parameter (#14028, @B-Step62)
  • [Tracking] Fix request/response field in the search_traces output (#13985, @B-Step62)

Documentation updates:

  • [Docs] Add Ollama and Instructor examples in tracing doc (#13937, @B-Step62)

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

· 5 min read
MLflow maintainers

We are excited to announce the release of MLflow 2.18.0! This release includes a number of significant features, enhancements, and bug fixes.

Python Version Update​

Python 3.8 is now at an end-of-life point. With official support being dropped for this legacy version, MLflow now requires Python 3.9 as a minimum supported version.

Note: If you are currently using MLflow's ChatModel interface for authoring custom GenAI applications, please ensure that you have read the future breaking changes section below.

Major New Features​

  • 🦺 Fluent API Thread/Process Safety - MLflow's fluent APIs for tracking and the model registry have been overhauled to add support for both thread and multi-process safety. You are now no longer forced to use the Client APIs for managing experiments, runs, and logging from within multiprocessing and threaded applications. (#13456, #13419, @WeichenXu123)

  • 🧩 DSPy flavor - MLflow now supports logging, loading, and tracing of DSPy models, broadening the support for advanced GenAI authoring within MLflow. Check out the MLflow DSPy Flavor documentation to get started! (#13131, #13279, #13369, #13345, @chenmoneygithub, #13543, #13800, #13807, @B-Step62, #13289, @michael-berk)

  • 🖥ï¸? Enhanced Trace UI - MLflow Tracing's UI has undergone a significant overhaul to bring usability and quality of life updates to the experience of auditing and investigating the contents of GenAI traces, from enhanced span content rendering using markdown to a standardized span component structure. (#13685, #13357, #13242, @daniellok-db)

  • 🚄 New Tracing Integrations - MLflow Tracing now supports DSPy, LiteLLM, and Google Gemini, enabling a one-line, fully automated tracing experience. These integrations unlock enhanced observability across a broader range of industry tools. Stay tuned for upcoming integrations and updates! (#13801, @TomeHirata, #13585, @B-Step62)

  • 📊 Expanded LLM-as-a-Judge Support - MLflow now enhances its evaluation capabilities with support for additional providers, including Anthropic, Bedrock, Mistral, and TogetherAI, alongside existing providers like OpenAI. Users can now also configure proxy endpoints or self-hosted LLMs that follow the provider API specs by using the new proxy_url and extra_headers options. Visit the LLM-as-a-Judge documentation for more details! (#13715, #13717, @B-Step62)

  • â?° Environment Variable Detection - As a helpful reminder for when you are deploying models, MLflow now detects and reminds users of environment variables set during model logging, ensuring they are configured for deployment. In addition to this, the mlflow.models.predict utility has also been updated to include these variables in serving simulations, improving pre-deployment validation. (#13584, @serena-ruan)

Breaking Changes to ChatModel Interface​

  • ChatModel Interface Updates - As part of a broader unification effort within MLflow and services that rely on or deeply integrate with MLflow's GenAI features, we are working on a phased approach to making a consistent and standard interface for custom GenAI application development and usage. In the first phase (planned for release in the next few releases of MLflow), we are marking several interfaces as deprecated, as they will be changing. These changes will be:

    • Renaming of Interfaces:
      • ChatRequest → ChatCompletionRequest to provide disambiguation for future planned request interfaces.
      • ChatResponse → ChatCompletionResponse for the same reason as the input interface.
      • metadata fields within ChatRequest and ChatResponse → custom_inputs and custom_outputs, respectively.
    • Streaming Updates:
      • predict_stream will be updated to enable true streaming for custom GenAI applications. Currently, it returns a generator with synchronous outputs from predict. In a future release, it will return a generator of ChatCompletionChunks, enabling asynchronous streaming. While the API call structure will remain the same, the returned data payload will change significantly, aligning with LangChain’s implementation.
    • Legacy Dataclass Deprecation:
      • Dataclasses in mlflow.models.rag_signatures will be deprecated, merging into unified ChatCompletionRequest, ChatCompletionResponse, and ChatCompletionChunks.

Other Features:

Here is the updated section with links to each PR ID and author:

markdown Copy code Other Features:

  • [Evaluate] Add Huggingface BLEU metrics to MLflow Evaluate (#12799, @nebrass)
  • [Models / Databricks] Add support for spark_udf when running on Databricks Serverless runtime, Databricks Connect, and prebuilt Python environments (#13276, #13496, @WeichenXu123)
  • [Scoring] Add a model_config parameter for pyfunc.spark_udf for customization of batch inference payload submission (#13517, @WeichenXu123)
  • [Tracing] Standardize retriever span outputs to a list of MLflow Documents (#13242, @daniellok-db)
  • [UI] Add support for visualizing and comparing nested parameters within the MLflow UI (#13012, @jescalada)
  • [UI] Add support for comparing logged artifacts within the Compare Run page in the MLflow UI (#13145, @jescalada)
  • [Databricks] Add support for resources definitions for LangChain model logging (#13315, @sunishsheth2009)
  • [Databricks] Add support for defining multiple retrievers within dependencies for Agent definitions (#13246, @sunishsheth2009)

Bug fixes:

  • [Database] Cascade deletes to datasets when deleting experiments to fix a bug in MLflow's gc command when deleting experiments with logged datasets (#13741, @daniellok-db)
  • [Models] Fix a bug with LangChain's pyfunc predict input conversion (#13652, @serena-ruan)
  • [Models] Fix signature inference for subclasses and Optional dataclasses that define a model's signature (#13440, @bbqiu)
  • [Tracking] Fix an issue with async logging batch splitting validation rules (#13722, @WeichenXu123)
  • [Tracking] Fix an issue with LangChain's autologging thread-safety behavior (#13672, @B-Step62)
  • [Tracking] Disable support for running Spark autologging in a threadpool due to limitations in Spark (#13599, @WeichenXu123)
  • [Tracking] Mark role and index as required for chat schema (#13279, @chenmoneygithub)
  • [Tracing] Handle raw response in OpenAI autolog (#13802, @harupy)
  • [Tracing] Fix a bug with tracing source run behavior when running inference with multithreading on LangChain models (#13610, @WeichenXu123)

Documentation updates:

  • [Docs] Add docstring warnings for upcoming changes to ChatModel (#13730, @stevenchen-db)
  • [Docs] Add a contributor's guide for implementing tracing integrations (#13333, @B-Step62)
  • [Docs] Add guidance in the use of model_config when logging models as code (#13631, @sunishsheth2009)
  • [Docs] Add documentation for the use of custom library artifacts with the code_paths model logging feature (#13702, @TomeHirata)
  • [Docs] Improve SparkML log_model documentation with guidance on how to return probabilities from classification models (#13684, @WeichenXu123)

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

· One min read
MLflow maintainers

MLflow 2.17.2 includes several major features and improvements

Features:

Bug fixes:

Documentation updates:

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

· One min read
MLflow maintainers

2.17.1 (2024-10-25)​

MLflow 2.17.1 includes several major features and improvements

Features:

Bug fixes:

  • [Tracking] Fix tool span inputs/outputs format in LangChain autolog (#13527, @B-Step62)
  • [Models] Fix code_path handling for LlamaIndex flavor (#13486, @B-Step62)
  • [Models] Fix signature inference for subclass and optional dataclasses (#13440, @bbqiu)
  • [Tracking] Fix error thrown in set_retriever_schema's behavior when it's called twice (#13422, @sunishsheth2009)
  • [Tracking] Fix dependency extraction from RunnableCallables (#13423, @aravind-segu)

Documentation updates:

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

· 4 min read
MLflow maintainers

2.17.0 (2024-10-11)​

We are excited to announce the release of MLflow 2.17.0! This release includes several enhancements to extend the functionality of MLflow's ChatModel interface to further extend its versatility for handling custom GenAI application use cases. Additionally, we've improved the interface within the tracing UI to provide a structured output for retrieved documents, enhancing the ability to read the contents of those documents within the UI. We're also starting the work on improving both the utility and the versatility of MLflow's evaluate functionality for GenAI, initially with support for callable GenAI evaluation metrics.

Major Features and notifications​

  • ChatModel enhancements - As the GenAI-focused 'cousin' of PythonModel, ChatModel is getting some sizable functionality extensions. From native support for tool calling (a requirement for creating a custom agent), simpler conversions to the internal dataclass constructs needed to interface with ChatModel via the introduction of from_dict methods to all data structures, the addition of a metadata field to allow for full input payload customization, handling of the new refusal response type, to the inclusion of the interface type to the response structure to allow for greater integration compatibility. (#13191, #13180, #13143, @daniellok-db, #13102, #13071, @BenWilson2)

  • Callable GenAI Evaluation Metrics - As the intial step in a much broader expansion of the functionalities of mlflow.evaluate for GenAI use cases, we've converted the GenAI evaluation metrics to be callable. This allows you to use them directly in packages that support callable GenAI evaluation metrics, as well as making it simpler to debug individual responses when prototyping solutions. (#13144, @serena-ruan)

  • Audio file support in the MLflow UI - You can now directly 'view' audio files that have been logged and listen to them from within the MLflow UI's artifact viewer pane.

  • MLflow AI Gateway is no longer deprecated - We've decided to revert our deprecation for the AI Gateway feature. We had renamed it to the MLflow Deployments Server, but have reconsidered and reverted the naming and namespace back to the original configuration.

Features:

Bug fixes:

  • [Tracking] Fix tracing for LangGraph (#13215, @B-Step62)
  • [Tracking] Fix an issue with presigned_url_artifact requests being in the wrong format (#13366, @WeichenXu123)
  • [Models] Update Databricks dependency extraction functionality to work with the langchain-databricks partner package. (#13266, @B-Step62)
  • [Model Registry] Fix retry and credential refresh issues with artifact downloads from the model registry (#12935, @rohitarun-db)
  • [Tracking] Fix LangChain autologging so that langchain-community is not required for partner packages (#13172, @B-Step62)
  • [Artifacts] Fix issues with file removal for the local artifact repository (#13005, @rzalawad)

Documentation updates:

  • [Docs] Add guide for building custom GenAI apps with ChatModel (#13207, @BenWilson2)
  • [Docs] Add updates to the MLflow AI Gateway documentation (#13217, @daniellok-db)
  • [Docs] Remove MLflow AI Gateway deprecation status (#13153, @BenWilson2)
  • [Docs] Add contribution guide for MLflow tracing integrations (#13333, @B-Step62)
  • [Docs] Add documentation regarding the run_id parameter within the search_trace API (#13251, @B-Step62)

Please try it out and report any issues on the issue tracker.

· 2 min read
MLflow maintainers

2.16.1 (2024-09-13)​

MLflow 2.16.1 is a patch release that includes some minor feature improvements and addresses several bug fixes.

Features:

  • [Tracing] Add Support for an Open Telemetry compatible exporter to configure external sinks for MLflow traces (#13118, @B-Step62)
  • [Model Registry, AWS] Add support for utilizing AWS KMS-based encryption for the MLflow Model Registry (#12495, @artjen)
  • [Model Registry] Add support for using the OSS Unity Catalog server as a Model Registry (#13034, #13065, #13066, @rohitarun-db)
  • [Models] Introduce path-based transformers logging to reduce memory requirements for saving large transformers models (#13070, @B-Step62)

Bug fixes:

  • [Tracking] Fix a data payload size issue with Model.get_tags_dict by eliminating the return of the internally-used config field (#13086, @harshilprajapati96)
  • [Models] Fix an issue with LangChain Agents where sub-dependencies were not being properly extracted (#13105, @aravind-segu)
  • [Tracking] Fix an issue where the wrong checkpoint for the current best model in auto checkpointing was being selected (#12981, @hareeen)
  • [Tracking] Fix an issue where local timezones for trace initialization were not being taken into account in AutoGen tracing (#13047, @B-Step62)

Documentation updates:

  • [Docs] Added RunLLM chat widget to MLflow's documentation site (#13123, @likawind)

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

· 3 min read
MLflow maintainers

We are excited to announce the release of MLflow 2.16.0. This release includes many major features and improvements!

Major features:​

  • LlamaIndex Enhancements🦙 - to provide additional flexibility to the LlamaIndex integration, we now have support for the models-from-code functionality for logging, extended engine-based logging, and broadened support for external vector stores.

  • LangGraph Support - We've expanded the LangChain integration to support the agent framework LangGraph. With tracing and support for logging using the models-from-code feature, creating and storing agent applications has never been easier!

  • AutoGen Tracing - Full automatic support for tracing multi-turn agent applications built with Microsoft's AutoGen framework is now available in MLflow. Enabling autologging via mlflow.autogen.autolog() will instrument your agents built with AutoGen.

  • Plugin support for AI Gateway - You can now define your own provider interfaces that will work with MLflow's AI Gateway (also known as the MLflow Deployments Server). Creating an installable provider definition will allow you to connect the Gateway server to any GenAI service of your choosing.

Features:

  • [UI] Add updated deployment usage examples to the MLflow artifact viewer (#13024, @serena-ruan, @daniellok-db)
  • [Models] Support logging LangGraph applications via the models-from-code feature (#12996, @B-Step62)
  • [Models] Extend automatic authorization pass-through support for Langgraph agents (#13001, @aravind-segu)
  • [Models] Expand the support for LangChain application logging to include UCFunctionToolkit dependencies (#12966, @aravind-segu)
  • [Models] Support saving LlamaIndex engine directly via the models-from-code feature (#12978, @B-Step62)
  • [Models] Support models-from-code within the LlamaIndex flavor (#12944, @B-Step62)
  • [Models] Remove the data structure conversion of input examples to ensure enhanced compatibility with inference signatures (#12782, @serena-ruan)
  • [Models] Add the ability to retrieve the underlying model object from within pyfunc model wrappers (#12814, @serena-ruan)
  • [Models] Add spark vector UDT type support for model signatures (#12758, @WeichenXu123)
  • [Tracing] Add tracing support for AutoGen (#12913, @B-Step62)
  • [Tracing] Reduce the latency overhead for tracing (#12885, @B-Step62)
  • [Tracing] Add Async support for the trace decorator (#12877, @MPKonst)
  • [Deployments] Introduce a plugin provider system to the AI Gateway (Deployments Server) (#12611, @gabrielfu)
  • [Projects] Add support for parameter submission to MLflow Projects run in Databricks (#12854, @WeichenXu123)
  • [Model Registry] Introduce support for Open Source Unity Catalog as a model registry service (#12888, @artjen)

Bug fixes:

  • [Tracking] Reduce the contents of the model-history tag to only essential fields (#12983, @harshilprajapati96)
  • [Models] Fix the behavior of defining the device to utilize when loading transformers models (#12977, @serena-ruan)
  • [Models] Fix evaluate behavior for LlamaIndex (#12976, @B-Step62)
  • [Models] Replace pkg_resources with importlib.metadata due to package deprecation (#12853, @harupy)
  • [Tracking] Fix error handling for OpenAI autolog tracing (#12841, @B-Step62)
  • [Tracking] Fix a condition where a deadlock can occur when connecting to an SFTP artifact store (#12938, @WeichenXu123)
  • [Tracking] Fix an issue where code_paths dependencies were not properly initialized within the system path for LangChain models (#12923, @harshilprajapati96)
  • [Tracking] Fix a type error for metrics value logging (#12876, @beomsun0829)
  • [Tracking] Properly catch NVML errors when collecting GPU metrics (#12903, @chenmoneygithub)
  • [Deployments] Improve Gateway schema support for the OpenAI provider (#12781, @danilopeixoto)
  • [Model Registry] Fix deletion of artifacts when downloading from a non-standard DBFS location during UC model registration (#12821, @smurching)

Documentation updates:

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.