Model Registry & Serving Guides#

A model goes from training into the registry, out through a deployment, and stays under monitoring. The guides below follow that path.

Start here

Register a trained model with its metrics, then deploy it in one call.

mr = project.get_model_registry()
model = mr.python.create_model(
    name="fraud_detector",
    metrics={{"f1": 0.92}},
)
model.save("model_dir")
deployment = model.deploy()

Model registry · Deployment creation · Model monitoring

Register

Model registry Save a model with metrics, a schema and an input example, one version per save.
Frameworks TensorFlow, PyTorch, scikit-learn, LLM and plain Python models.
Import from Hugging Face Bring a Hub model into the registry without training it here.
Evaluation images Attach plots and confusion matrices to a model version.

Serve

Deployment creation Deploy a registered model and check its state.
Predictor and transformer Custom inference code and pre/post-processing on KServe.
Logging and batching Log requests to a feature group, batch them for throughput.
Resources and autoscaling CPU, memory, GPU and replica bounds, with scheduling constraints.
Reach the endpoint API protocol, REST access from outside, troubleshooting.

Observe

Model monitoring Compare logged inference data against the training dataset on a schedule.
Provenance Trace a model back to its training data and features.
Vector database Similarity search over embeddings stored in the feature store.
Agents Agent tasks as jobs and served interactive agents.