TimeSeriesFoundationModel

class autogluon.cloud.TimeSeriesFoundationModel(model_id: str, **kwargs)

Pretrained time series foundation model for zero-shot forecasting on AWS SageMaker.

Wraps pretrained models like Chronos-2 and Chronos-Bolt and runs prediction as a managed SageMaker job, with no training required. See the foundation model tutorial for the supported model_id values and a full walkthrough.

Predictions can be produced in three modes:

• Batch — predict() runs a one-off SageMaker training job and writes forecasts to S3. Best for one-shot inference.

• Real-time — deploy() provisions a real-time endpoint; call TimeSeriesEndpoint.predict() for low-latency inference, then TimeSeriesEndpoint.delete_endpoint() to tear it down.

• Serverless — deploy() with inference_mode="serverless" provisions a SageMaker Serverless Inference endpoint that scales to zero. Requires a cached model artifact (see cache_model_artifact()).

Parameters:

• model_id – ID of the foundation model from the model registry. See Available models in the foundation model tutorial for the list of supported values.

• cloud_output_path –

  S3 location where intermediate artifacts are stored. Accepts:

  • s3://bucket — a unique timestamped subfolder ag-<timestamp> is appended.

  • s3://bucket/prefix — used verbatim. Re-running with the same prefix will overwrite previously written artifacts.

  • None (default) — use the bucket saved in ~/.autogluon/cloud.yaml (set by autogluon.cloud.bootstrap() / autogluon.cloud.register()) and append a timestamped subfolder. Raises if no bucket is configured.

• role – ARN of the SageMaker execution role used to run training and inference jobs. If None, falls back to role_arn in ~/.autogluon/cloud.yaml (set by autogluon.cloud.bootstrap() / autogluon.cloud.register()), and finally to sagemaker.get_execution_role().

• hyperparameters – Default hyperparameters applied to inference and (when supported) training.

• model_artifact_uri – S3 URI of a pre-bundled model.tar.gz produced by cache_model_artifact(). When set, deploys skip the runtime HuggingFace download and load weights from the bundled artifact.

• backend – Cloud backend to use.

Methods

cache_model_artifact	Download model weights from HuggingFace, bundle them with the FM serve script into a SageMaker-compatible model.tar.gz, and upload to S3.
deploy	Deploy model to an inference endpoint.
from_dict	Restore from to_dict() output.
from_json	Restore from a to_json() string.
predict	Run batch prediction for time series.
to_dict	Serialize the model identity.
to_json	Serialize to_dict() output as a JSON string.