Snowflake ML release notes¶

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 infering a model signature.

New features¶

New Cortex features:

• New guardrails option on the Complete function for enabling Cortex Guard.

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.