Snowflake ML release notes¶

Bug fixes¶

Model Registry bug fixes:

• Remove redundant pip dependency warnings when artifact_repository_map is provided for warehouse model deployments.

New features¶

New modeling features:

• Support for scikit-learn versions earlier than 1.8.

New ML Jobs features:

• Support for configuring the runtime image via the runtime_environment parameter at submission time. You may specify an image tag or a full image URL.

  Examples for @remote decorator and submit_file function:

  @remote(compute_pool, stage_name = 'payload_stage', runtime_environment = '1.8.0')
  
  submit_file('/path/to/repo/test.py', compute_pool, stage_name = 'payload_stage',
    runtime_environment = '/mydb/myschema/myrepo/myimage:latest')
  

  Copy

New Model Registry features:

• Ability to mark model methods as volatile or immutable. Volatile methods may return different results when called multiple times with the same input, while immutable methods always return the same result for the same input. Methods in supported model types are immutable by default, while methods in custom models are volatile by default. Use the Volatility enum to specify the volatility of model methods when logging a model as follows:

  from snowflake.ml.model.volatility import Volatility
  
  options = {
      "embed_local_ml_library": True,
      "relax_version": True,
      "save_location": "/path/to/my/directory",
      "function_type": "TABLE_FUNCTION",
      "volatility": Volatility.IMMUTABLE,
      "method_options": {
          "predict": {
              "case_sensitive": False,
              "max_batch_size": 100,
              "function_type": "TABLE_FUNCTION",
              "volatility": Volatility.VOLATILE,
          },
  }
  

  Copy

Behavior changes¶

Model Registry behavior changes:

• Drop support for deprecated conversational task type for Hugging Face models. This task type has been deprecated by Hugging Face for some time and is due for removal from their API imminently.

New features¶

New ML Jobs features:

• The additional_payloads argument of the MLJob.submit_* methods has been renamed imports to better reflect its purpose. additional_payloads has been deprecated and will be removed in a future release.

New features¶

New Model Registry features:

• You can now log a HuggingFace model without having to load the model in memory using huggingface_pipeline.HuggingFacePipelineModel. Requires the huggingface_hub package. To disable downloading from the HuggingFace repository, pall download_snapshot=False when instantiating huggingface_pipeline.HuggingFacePipelineModel.

• You can now use XGBoost’s enable_categorical=True models to with pandas DataFrames

• When listing services, the PrivateLink inference endpoint in shown in the ModelVersion list.

Bug fixes¶

Model Registry bug fixes:

• Fixed an issue where the string representation of dictionary-type output columns was being incorrectly created during structured output deserialization, losing the original data type.

• Fixed an inference server performance issue for wide (500+ features) and JSON inputs.

New features¶

New Model Registry features:

• You can now log text-generation models with signatures compatible with OpenAI chat completion compatible signature, as shown in the following example:

  from snowflake.ml.model import openai_signatures
  import pandas as pd
  
  mv = snowflake_registry.log_model(
      model=generator,
      model_name=...,
      ...,
      signatures=openai_signatures.OPENAI_CHAT_SIGNATURE,
  )
  
  # create a pd.DataFrame with openai.client.chat.completions arguments like below:
  x_df = pd.DataFrame.from_records(
      [
          {
              "messages": [
                  {"role": "system", "content": "Complete the sentence."},
                  {
                      "role": "user",
                      "content": "A descendant of the Lost City of Atlantis, who swam to Earth while saying, ",
                  },
              ],
              "max_completion_tokens": 250,
              "temperature": 0.9,
              "stop": None,
              "n": 3,
              "stream": False,
              "top_p": 1.0,
              "frequency_penalty": 0.1,
              "presence_penalty": 0.2,
          }
      ],
  )
  

  Copy

New Model Monitoring features:

• Model monitors now support segment columns to enable filtered analysis, specified in the segment_columns field in the model monitor source options. Segment columns must exist in the source table and be of string type. add_segment_column and drop_segment_column methods are provided to add or remove segment columns in existing model monitors.

New features¶

New Model Registry features:

• Made image_repo argument optional in ModelVersion.create_service. If not specified, a default image repository is used.

Bug fixes¶

ML Jobs bug fixes:

• Fixed TypeError: SnowflakeCursor.execute() got an unexpected keyword argument '_force_qmark_paramstyle' inside stored procedure.

• Fixed Error: Unable to retrieve head IP address when not all instances start within the timeout period.

New features¶

New Model Registry features:

• Added progress bars for ModelVersion.create_service and ModelVersion.log_model.

• Logs from ModelVersion.create_service are now written to a file. The location of the log file is shown in the console.

Bug fixes¶

DataConnector bug fixes:

• Fixed a problem that caused errors mentioning self._session.

Model Registry bug fixes:

• Fixed a bug when passing None to array (pd.dtype('O')) in model signature and pandas data handler.

Bug fixes¶

Model Registry bug fixes:

• Fix a bug with setting the PAD token when the HuggingFace text-generation model had multiple EOS tokens. The handler now picks the first EOS token as PAD token.

New features¶

New DataConnector features:

• DataConnector objects can now be pickled.

New Dataset features:

• Dataset objects can now be pickled.

New Model Registry features:

• Models hosted on Snowpark Container Services now support wide input (500+ features).

Behavior changes¶

ML Jobs behavior changes:

• Removed scope parameter from list_jobs method.

• Added optional database and schema parameters to list_jobs method.

• The list_jobs method now returns a pandas DataFrame rather than a Snowpark DataFrame.

• The list_jobs method now returns the following columns: name, status, message, database_name, schema_name, owner, compute_pool, target_instances, created_time, and completed_time.

Model registry behavior changes:

• Set relax_version to false when pip_requirements is specified in log_model call.

• UserWarning is raised only on specified target_platforms to address spurious warnings

Bug fixes¶

Model registry bug fixes:

• Fixed failure in converting Snowpark DataFrame to pandas DataFrame when QUOTED_IDENTIFIERS_IGNORE_CASE parameter is enabled

• Fixed duplicate UserWarning log entries during model packaging

New features¶

New model registry features:

• New APIs for representing target platforms (snowflake.ml.model.target_platform.TargetPlatform), target platform constant,s and tasks (snowflake.ml.model.task.Task).

• The target_platform argument in the log_model method now accepts a TargetPlatformMode constant, which can be WAREHOUSE_ONLY, SNOWPARK_CONTAINER_SERVICES_ONLY, or BOTH_WAREHOUSE_AND_SNOWPARK_CONTAINER_SERVICES.

New ML Jobs features:

• Less-frequently-used job submission arguments have been moved to **kwargs.

• Platform metrics are enabled by default.

With this release, single-node ML Job APIs are now stable and have been designated Generally Available.

New features¶

New model registry features:

• Added service container information to logs

Behavior changes¶

ML Jobs behavior changes:

• Argument num_instances has been renamed to target_instances in job submission APIs and is now required.

Bug fixes¶

Model Registry bug fixes:

• Fixed a bug in listing and deleting container services.

• Fixed a bug with logging scikit-learn pipelines where the explain function was not created.

• Logging a container-only model no longer checks to make sure the required version of snowflake-ml-python is available in the Snowflake conda channel.

Explainability bug fixes:

• Minimum streamlit version has been decreased to 1.30 to improve compatibility.

Modeling bug fixes:

• xgboost is now a required dependency again (it was optional in v1.8.4).

New features¶

ML Jobs new features:

• Job decorator now has min_instances argument that makes a job wait for the specified number of workers to be ready before starting.

Behavior changes¶

ML Jobs behavior changes:

• The id property is now the job’s fully-qualified name. A new property, name, has been introduced to represent the ML Job name.

• The list_jobs method now returns the ML Job name instead of the job ID.

Model Registry behavior changes:

• In log_model, enabling explainability when the model is deployed only to Snowpark Container Services is now an error instead of a warning and will prevent the log operation from completing.

Bug fixes¶

Model Registry bug fixes:

• Fixed a bug in which logging PyTorch and TensorFlow models that caused UnboundLocalError: local variable 'multiple_inputs' referenced before assignment.

New features¶

New Model Registry features:

• Automatically enable explainability for models that can be deployed to a warehouse.

New Explainability features:

• New visualization functions in snowflake.ml.monitoring plot explanations in notebooks.

• Support for categorical transforms in scikit-learn pipelines.

New Modeling features:

• Support categorical types for xgboost.DMatrix inputs in XGBoost models.

New features¶

New Model Registry features:

• Default to a CUDA container image, if available, when logging a GPU-capable model for deployment to Container Runtime for ML.

• Model versions have a run_job method that runs inference methods as a single-node Snowpark Container Services job. This method is available for all models, including those that are not deployed to Container Runtime for ML.

• The target platform defaults to a Snowflake warehouse when logging a partitioned model.

New features¶

The ML Jobs API, which allows you to run code on Container Runtime for ML from your local workstation, is available in preview. Accordingly, documentation for this API is available in the Snowflake ML API Reference, and changes to the API appear in these release notes. New features in the ML Jobs API might not appear here until they are publicly announced, but they do appear in the API reference.

New Model Registry features:

• You can specify the path to write the model versions files that are stored in the model’s Snowflake stage using the save_location option in the log_model method.

• When logging models in Container Runtime for ML, model dependencies are now included in pip_requirements by default.

Bug fixes¶

Model Registry bug fixes:

• Fix unsupported model type error when logging a scikit-learn model with a score_samples inference method.

• Fix failure of inference service creation on an existing suspended service.

New features¶

New Model Registry features:

• Creating a copy of a model version with log_model with unsupported arguments now raises an exception.

Behavior changes¶

Model Registry behavior changes:

• Automatically-inferred signatures in transformers.Pipeline have been changed to use the FeatureGroupSpec task class, including:

  • Signature for fill-mask tasks:

    Before v1.8.0v1.8.0 and later

      ModelSignature(
          inputs=[
              FeatureSpec(name="inputs", dtype=DataType.STRING),
          ],
          outputs=[
              FeatureSpec(name="outputs", dtype=DataType.STRING),
          ],
    )
    

    Copy

  • Signature for token classification tasks:

    Before v1.8.0v1.8.0 and later

    ModelSignature(
        inputs=[
            FeatureSpec(name="inputs", dtype=DataType.STRING),
        ],
        outputs=[
            FeatureSpec(name="outputs", dtype=DataType.STRING),
        ],
    )
    

    Copy

  • Signature for question-answering tasks:

    Before v1.8.0v1.8.0 and later

    ModelSignature(
        inputs=[
            FeatureSpec(name="question", dtype=DataType.STRING),
            FeatureSpec(name="context", dtype=DataType.STRING),
        ],
        outputs=[
            FeatureSpec(name="outputs", dtype=DataType.STRING),
        ],
    )
    

    Copy

  • Signature for question-answering tasks when top_k is greater than 1:

    Before v1.8.0v1.8.0 and later

    ModelSignature(
        inputs=[
            FeatureSpec(name="question", dtype=DataType.STRING),
            FeatureSpec(name="context", dtype=DataType.STRING),
        ],
        outputs=[
            FeatureSpec(name="outputs", dtype=DataType.STRING),
        ],
    )
    

    Copy

  • Signature for text-classification tasks when top_k is None:

    Before v1.8.0v1.8.0 and later

    ModelSignature(
        inputs=[
            FeatureSpec(name="text", dtype=DataType.STRING),
            FeatureSpec(name="text_pair", dtype=DataType.STRING),
        ],
        outputs=[
            FeatureSpec(name="label", dtype=DataType.STRING),
            FeatureSpec(name="score", dtype=DataType.DOUBLE),
        ],
    )
    

    Copy

  • Signature for text-classification tasks when top_k is not None:

    Before v1.8.0v1.8.0 and later

    ModelSignature(
        inputs=[
            FeatureSpec(name="text", dtype=DataType.STRING),
            FeatureSpec(name="text_pair", dtype=DataType.STRING),
        ],
        outputs=[
            FeatureSpec(name="outputs", dtype=DataType.STRING),
        ],
    )
    

    Copy

  • Signature for text-generation tasks:

    Before v1.8.0v1.8.0 and later

    ModelSignature(
        inputs=[FeatureSpec(name="inputs", dtype=DataType.STRING)],
        outputs=[
            FeatureSpec(name="outputs", dtype=DataType.STRING),
        ],
    )
    

    Copy

• PyTorch and TensorFlow models now expect a single tensor input and output by default when they are logged to the Model Registry. To use multiple tensors (previous behavior), set options={"multiple_inputs": True}.

  Example with single tensor input:

  import torch
  
  class TorchModel(torch.nn.Module):
      def __init__(self, n_input: int, n_hidden: int, n_out: int, dtype: torch.dtype = torch.float32) -> None:
          super().__init__()
          self.model = torch.nn.Sequential(
              torch.nn.Linear(n_input, n_hidden, dtype=dtype),
              torch.nn.ReLU(),
              torch.nn.Linear(n_hidden, n_out, dtype=dtype),
              torch.nn.Sigmoid(),
          )
  
      def forward(self, tensor: torch.Tensor) -> torch.Tensor:
          return cast(torch.Tensor, self.model(tensor))
  
  # Sample usage:
  data_x = torch.rand(size=(batch_size, n_input))
  
  # Log model with single tensor
  reg.log_model(
      model=model,
      ...,
      sample_input_data=data_x
  )
  
  # Run inference with single tensor
  mv.run(data_x)
  

  Copy

  For multiple tensor inputs or outputs, use:

  reg.log_model(
      model=model,
      ...,
      sample_input_data=[data_x_1, data_x_2],
      options={"multiple_inputs": True}
  )
  

  Copy

• enable_explainability now defaults to False when the model can be deployed to Snowpark Container Services.

Bug fixes¶

Modeling bug fixes:

• Fix a bug in some metrics that allowed an unsupported version of numpy to be installed automatically in the stored procedure, resulting in a numpy error on execution.

Model Registry bug fixes:

• Fix a bug that leads to incorrect Model does not have _is_inference_api error message when assigning a supported model as a property of a CustomModel.

• Fix a bug where inference does not work when models with more than 500 input features are deployed to SPCS.

New features¶

New Model Registry features:

• Support for using a single torch.Tensor, tensorflow.Tensor and tensorflow.Variable as input or output data.

• Support for xgboost.DMatrix datatype for XGBoost models.

Bug fixes¶

Model Registry bug fixes:

• Fixed a compatibility issue where, when using snowflake-ml-python 1.7.0 or later to save a tensorflow.keras model with keras 2.x, the model could not be run in Snowflake. This issue occurred when relax_version is set to True (or default) and a new version of snowflake-ml-python is available. If you have logged an affected model, you can recover it by loading it using ModelVerison.load and logging it again with the latest version of snowflake-ml-python.

• Removed the validation that prevents data that does not have non-null values from being passed to ModelVersion.run.

New features¶

New Model Registry features:

• Support for Hugging Face model configurations with auto-mapping functionality.

• Support for keras 3.x models with tensorflow and pytorch backends.

New Model Explainability features:

• Support for native and snowflake-ml-python sklearn pipelines.

Bug fixes¶

Registry bug fixes:

• Fixed an issue in which Hugging Face pipelines were loaded using an incorrect data type.

• Fixed an issue in which only one row was actually used when inferring a model signature.

New features¶

New Cortex features:

• New guardrails option on the Complete function.

Dependency upgrades¶

• fsspec and s3fs must be 2024.6.1 or later and less than 2026.

• mlflow must be 2.16.0 or later and less than 3.

New features¶

New Cortex features:

• Cortex functions now have “snake_case” names. For example, ClassifyText is now classify_text. The old “CamelCase” names still work, but will be removed in a future release.

New Model Registry features:

• Registry now supports more than 500,000 features.

• Added user_files argument to Registry.log_model for including images or other files with the model.

• Added support for handling Hugging Face model configurations with auto-mapping functionality.

New Data features:

• Added the DataConnector.from_sql constructor.

Bug fixes¶

Registry bug fixes:

• Fixed a bug that occurred when providing a non-range index pandas DataFrame as the input to ModelVersion.run.

• Improved random model registry name generation to avoid collisions.

• Fixed an issue when inferring a signature or running inference with Snowpark DataFrame that has a column whose type is ARRAY and contains a NULL value.

• ModelVersion.run now accepts a fully-qualified service name.

• Fixed an error in log_model for any scikit-learn models with only data preprocessing, including preprocessing-only pipeline models.

Monitoring bug fixes:

• Fixed an issue with creating monitors using fully-qualified names.