• Enable MLflow client integration
• Model experiments
  • Associating a candidate to a CI/CD job
• Model registry
  • Models
    • Creating a model
    • Fetching a model
    • Updating a model
    • Deleting a model
  • Logging candidates to a model
  • Model version
    • Creating a model version
    • Updating a model
    • Fetching a model version
    • Getting latest versions of a model
    • Logging metrics and parameters to a model version
    • Logging artifacts to a model version
    • Linking a model version to a CI/CD job
• Supported MLflow client methods and caveats
• Limitations

MLflow client compatibility

MLflow is a popular open source tool for Machine Learning Experiment Tracking. GitLab Model experiment tracking and GitLab Model registry are compatible with the MLflow client. The setup requires minimal changes to existing code.

GitLab plays the role of a MLflow server. Running mlflow server is not necessary.

Enable MLflow client integration

Prerequisites:

• A personal, project, or group access token with at least the Developer role and the api permission.
• The project ID. To find the project ID:
  1. On the left sidebar, select Search or go to and find your project.
  2. Select Settings > General.

To use MLflow client compatibility from a local environment:

1. Set the tracking URI and token environment variables on the host that runs the code. This can be your local environment, CI pipeline, or remote host. For example:

   export MLFLOW_TRACKING_URI="<your gitlab endpoint>/api/v4/projects/<your project id>/ml/mlflow"
   export MLFLOW_TRACKING_TOKEN="<your_access_token>"
   

2. If the training code contains the call to mlflow.set_tracking_uri(), remove it.

Model experiments

When running the training code, MLflow client can be used to create experiments, runs, models, model versions, log parameters, metrics, metadata and artifacts on GitLab.

After experiments are logged, they are listed under /<your project>/-/ml/experiments.

Runs are registered as candidates, which can be explored by selecting an experiment, model, or model version.

Associating a candidate to a CI/CD job

If your training code is being run from a CI/CD job, GitLab can use that information to enhance candidate metadata. To associate a candidate to a CI/CD job:

1. In the Project CI variables, include the following variables:
   • MLFLOW_TRACKING_URI: "<your gitlab endpoint>/api/v4/projects/<your project id>/ml/mlflow"
   • MLFLOW_TRACKING_TOKEN: <your_access_token>

2. In your training code within the run execution context, add the following code snippet:

    with mlflow.start_run(run_name=f"Candidate {index}"):
      # Your training code
   
      # Start of snippet to be included
      if os.getenv('GITLAB_CI'):
        mlflow.set_tag('gitlab.CI_JOB_ID', os.getenv('CI_JOB_ID'))
      # End of snippet to be included
   

Model registry

You can also manage models and model versions by using the MLflow client. Models are registered under /<your project>/-/ml/models.

Models

Creating a model

client = MlflowClient()
model_name = '<your_model_name>'
description = 'Model description'
model = client.create_registered_model(model_name, description=description)

Notes

• create_registered_model argument tags is ignored.
• name must be unique within the project.
• name cannot be the name of an existing experiment.

Fetching a model

client = MlflowClient()
model_name = '<your_model_name>'
model = client.get_registered_model(model_name)

Updating a model

client = MlflowClient()
model_name = '<your_model_name>'
description = 'New description'
client.update_registered_model(model_name, description=description)

Deleting a model

client = MlflowClient()
model_name = '<your_model_name>'
client.delete_registered_model(model_name)

Logging candidates to a model

Every model has an associated experiment with the same name. To log a candidate/run to the model, use the experiment with the name of the model:

client = MlflowClient()
model_name = '<your_model_name>'
exp = client.get_experiment_by_name(model_name)
run = client.create_run(exp.experiment_id)

Model version

Creating a model version

client = MlflowClient()
model_name = '<your_model_name>'
description = 'Model version description'
model_version = client.create_model_version(model_name, source="", description=description)

Notes

• Argument run_id is ignored. Every model version behaves as a Candidate/Run. Creating a mode version from a run is not yet supported.
• Argument source is ignored. GitLab will create a package location for the model version files.
• Argument tags is ignored.
• Argument run_link is ignored.
• Argument await_creation_for is ignored.

Updating a model

client = MlflowClient()
model_name = '<your_model_name>'
version = '<your_version>'
description = 'New description'
client.update_model_version(model_name, version, description=description)

Fetching a model version

client = MlflowClient()
model_name = '<your_model_name>'
version = '<your_version>'
client.get_model_version(model_name, version)

Getting latest versions of a model

client = MlflowClient()
model_name = '<your_model_name>'
client.get_latest_versions(model_name)

Notes

• Argument stages is ignored.
• Versions are ordered by last created.

Logging metrics and parameters to a model version

Every model version is also a candidate/run, allowing users to log parameters and metrics. The run ID can either be found at the Model version page in GitLab, or by using the MLflow client:

client = MlflowClient()
model_name = '<your_model_name>'
version = '<your_version>'
model_version = client.get_model_version(model_name, version)
run_id = model_version.run_id

# Your training code

client.log_metric(run_id, '<metric_name>', '<metric_value>')
client.log_param(run_id, '<param_name>', '<param_value>')
client.log_batch(run_id, metric_list, param_list, tag_list)

Logging artifacts to a model version

GitLab creates a package that can be used by the MLflow client to upload files.

client = MlflowClient()
model_name = '<your_model_name>'
version = '<your_version>'
model_version = client.get_model_version(model_name, version)
run_id = model_version.run_id

# Your training code

client.log_artifact(run_id, '<local/path/to/file.txt>', artifact_path="")
client.log_figure(run_id, figure, artifact_file="my_plot.png")
client.log_dict(run_id, my_dict, artifact_file="my_dict.json")
client.log_image(run_id, image, artifact_file="image.png")

Artifacts will then be available under https/<your project>/-/ml/models/<model_id>/versions/<version_id>.

Linking a model version to a CI/CD job

Similar to candidates, it is also possible to link a model version to a CI/CD job:

client = MlflowClient()
model_name = '<your_model_name>'
version = '<your_version>'
model_version = client.get_model_version(model_name, version)
run_id = model_version.run_id

# Your training code

if os.getenv('GITLAB_CI'):
    client.set_tag(model_version.run_id, 'gitlab.CI_JOB_ID', os.getenv('CI_JOB_ID'))

Supported MLflow client methods and caveats

GitLab supports these methods from the MLflow client. Other methods might be supported but were not tested. More information can be found in the MLflow Documentation. The MlflowClient counterparts of the methods below are also supported with the same caveats.

Other MLflowClient methods:

Limitations

• The API GitLab supports is the one defined at MLflow version 2.7.1.
• MLflow client methods not listed above are not supported.
• During creation of experiments and runs, ExperimentTags are stored, even though they are not displayed.