I have used automl.classification to create a automl job in sdkv2. But the model trained is having sdkv1 packages why ??? and how can i resolve this i want my model to use sdkv2 packages.

Question

I have used automl.classification to create a automl job in sdkv2. But the model trained is having sdkv1 packages why ??? and how can i resolve this i want my model to use sdkv2 packages.

Aryan Rastogi 0

I have used below code to create a automl job. Models trained from it are using sdkv1 libraries why ??
though below code is sdkv2.

I want to evaluate my model on new testing data in my local environment but getting package issue with sdkv2 packages how to resolve it ????

automl_config = automl.classification(
                            experiment_name=config_dict["experiment_name"],
                            compute=cpu_cluster_name,
                            training_data=train_data,
                            validation_data=val_data,
                            target_column_name='AI_EVA_Retraining_JG',
                            primary_metric='accuracy',
                        )

    automl_config.set_featurization(
        mode='auto'
    )

    automl_config.set_limits(
        timeout_minutes=720
    )

Prashanth Veeragoni 4,360 Reputation points Microsoft External Staff Moderator

2025-05-08T07:50:04.1766667+00:00
Hi Aryan Rastogi,

I understand that you're facing an issue where AutoML models trained using Azure Machine Learning SDK v2 (azure-ai-ml) are still depending on SDK v1 (azureml-*) packages, which is problematic when you try to evaluate them locally using SDK v2. This confusion is common during the transition from SDK v1 to SDK v2.

Why Your AutoML Models Use SDK v1 Internals:

Even though your AutoML configuration code is SDK v2, the actual AutoML training backend still uses v1 runtime environments internally as of the current SDK versions (2025). Microsoft hasn't fully migrated the AutoML backend to native SDK v2 yet. This means:

·   The training environments and scoring scripts generated are based on SDK v1 packages like azureml-defaults, azureml-train-automl-runtime, etc.

·   When you try to load or run the model locally using SDK v2, dependency mismatches occur (like missing SDK v1 modules).

How to Fix:

Option 1: Download and Use the Model in a Compatible (SDK v1) Environment

1. Find the best model:

from azure.ai.ml.entities import Job job = ml_client.jobs.get('<your-automl-job-name>') print(job.outputs)

2.  Download the model or register it, and then:

·   Use a local conda environment with SDK v1 (azureml-sdk==1.48.0, etc.).

·   Or deploy using Azure endpoints which already support this setup.

Option 2: Use mlflow for Cross-SDK Compatibility

AutoML models are typically MLflow-tracked. You can:

1.  Get the MLflow URI of the best model:

from azure.ai.ml import MLClient from azure.identity import DefaultAzureCredential ml_client = MLClient(DefaultAzureCredential(), "<subscription_id>", "<resource_group>", "<workspace_name>") job = ml_client.jobs.get("automl_job_name") best_model_uri = job.outputs['best_model'].uri

2.  Use MLflow to load and evaluate model locally:

import mlflow model = mlflow.pyfunc.load_model(best_model_uri) predictions = model.predict(X_test)

This bypasses SDK v1 dependency issues by using MLflow as the abstraction layer.

Below are some documents you can refer:

Upgrade AutoML to SDK v2: Upgrade AutoML to SDK v2 - Azure Machine Learning

Upgrade Deployment Endpoints to SDK v2: Upgrade Deployment Endpoints to SDK v2 - Azure Machine Learning

Transitioning from Azure ML v1 to v2: A Comprehensive Guide: Transitioning from Azure ML v1 to v2: A Comprehensive Guide – datalabs.io

Hope this helps. do let me know if you have any further queries.

Thank you!
Aryan Rastogi 0 Reputation points

2025-05-08T08:24:04.9033333+00:00

Hi, Prashant Veeragoni I understood the options mentioned by you. I have one more issue that in option 2 as you mentioned using mlflow. I did try that but it also throws dependency issues in sdkv2 local environment
Prashanth Veeragoni 4,360 Reputation points Microsoft External Staff Moderator

2025-05-08T08:52:53.7533333+00:00
Hi Aryan Rastogi,

Thanks for the update. You are right — even mlflow.pyfunc.load_model() can fail locally if the AutoML-generated model depends on SDK v1-based environments (e.g., azureml-defaults, azureml-automl-runtime), which are not compatible with SDK v2 environments.

For this issue you need to clean SDK v2-Compatible Way

You have two options to do so

Option A: Manually Patch the Conda Environment (Recommended):

1.Download the model locally:

mlflow artifacts download --artifact-uri <your-best-model-uri> --dst-dir ./model_dir

2.Go to the downloaded folder and edit the conda.yaml:

Remove:

- azureml-defaults - azureml-train-automl-runtime - inference-schema[numpy-support]

Replace them with:

- azure-ai-ml - pandas - scikit-learn - cloudpickle - mlflow

3.Create a new conda environment manually using the patched conda.yaml:

conda env create -f conda.yaml conda activate <your-env-name>

Now you can load the model with:

import mlflow model = mlflow.pyfunc.load_model('./model_dir') predictions = model.predict(X_test)

Option B: Re-register Model in SDK v2 with Custom Environment

1. Create a custom environment in SDK v2:

from azure.ai.ml.entities import Environment my_env = Environment( name="clean-eval-env", conda_file="patched_conda.yaml", image="mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:20230306.v1" )

Register the original model path but with this new environment and scoring script if needed.

Reference link: AutoML Job Output – Understanding Model Environment

Hope this helps.

Thank you!
Prashanth Veeragoni 4,360 Reputation points Microsoft External Staff Moderator

2025-05-09T07:56:21.2366667+00:00

Hi Aryan Rastogi,

Following up to see if the above suggestion was helpful. And, if you have any further queries do let me know.

Thank you!

Your answer

Aryan Rastogi 0 Reputation points

2025-05-08T08:24:04.9033333+00:00

Hi, Prashant Veeragoni I understood the options mentioned by you. I have one more issue that in option 2 as you mentioned using mlflow. I did try that but it also throws dependency issues in sdkv2 local environment
Prashanth Veeragoni 4,360 Reputation points Microsoft External Staff Moderator

2025-05-08T08:52:53.7533333+00:00

Hi Aryan Rastogi,

Thanks for the update. You are right — even mlflow.pyfunc.load_model() can fail locally if the AutoML-generated model depends on SDK v1-based environments (e.g., azureml-defaults, azureml-automl-runtime), which are not compatible with SDK v2 environments.

For this issue you need to clean SDK v2-Compatible Way

You have two options to do so

Option A: Manually Patch the Conda Environment (Recommended):

1.Download the model locally:

mlflow artifacts download --artifact-uri <your-best-model-uri> --dst-dir ./model_dir

2.Go to the downloaded folder and edit the conda.yaml:

Remove:

- azureml-defaults - azureml-train-automl-runtime - inference-schema[numpy-support]

Replace them with:

- azure-ai-ml - pandas - scikit-learn - cloudpickle - mlflow

3.Create a new conda environment manually using the patched conda.yaml:

conda env create -f conda.yaml conda activate <your-env-name>

Now you can load the model with:

import mlflow model = mlflow.pyfunc.load_model('./model_dir') predictions = model.predict(X_test)

Option B: Re-register Model in SDK v2 with Custom Environment

1. Create a custom environment in SDK v2:

from azure.ai.ml.entities import Environment my_env = Environment( name="clean-eval-env", conda_file="patched_conda.yaml", image="mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:20230306.v1" )

Register the original model path but with this new environment and scoring script if needed.

Reference link: AutoML Job Output – Understanding Model Environment

Hope this helps.

Thank you!
Prashanth Veeragoni 4,360 Reputation points Microsoft External Staff Moderator

2025-05-09T07:56:21.2366667+00:00

Hi Aryan Rastogi,

Following up to see if the above suggestion was helpful. And, if you have any further queries do let me know.

Thank you!

Share via

I have used automl.classification to create a automl job in sdkv2. But the model trained is having sdkv1 packages why ??? and how can i resolve this i want my model to use sdkv2 packages.

Your answer