openstef-foundation-models

Foundation model support for OpenSTEF — bringing pre-trained, ONNX-based forecasting models to the OpenSTEF ecosystem.

Installation

Pick exactly one ONNX runtime — [cpu] and [gpu] are mutually exclusive.

CPU (default — the meta-package openstef installs this for you):

pip install "openstef-foundation-models[cpu]"

GPU (CUDA):

pip install "openstef-foundation-models[gpu]"

Note: Do not install both [cpu] and [gpu] in the same environment — onnxruntime and onnxruntime-gpu collide. They're declared as conflicting extras so uv enforces the choice; pip does not, so pick one yourself.

Selecting a checkpoint

OpenSTEF publishes its checkpoints to the HuggingFace Hub. Pick a model size and variant from the catalog instead of hand-writing repo ids and filenames:

from openstef_foundation_models.models import Chronos2, CheckpointVariant
from openstef_foundation_models.presets import ForecastingWorkflowConfig

# Default: the base Chronos-2, dynamic shapes — runs on any provider.
config = ForecastingWorkflowConfig()

# The smaller model, static shapes — enables the CoreML provider on macOS.
config = ForecastingWorkflowConfig(
    checkpoint=Chronos2.SMALL.checkpoint(CheckpointVariant.STATIC),
)

# Let the host decide: static on macOS, dynamic elsewhere.
config = ForecastingWorkflowConfig(
    checkpoint=Chronos2.BASE.checkpoint(CheckpointVariant.recommended()),
)

Available sizes are Chronos2.BASE (chronos-2) and Chronos2.SMALL (chronos-2-small); variants are DYNAMIC (portable) and STATIC (frozen shapes, the macOS/CoreML path in the default provider fallback chain).