checkpoint

Checkpoint resolution for foundation-model forecasters.

A checkpoint bundles an ONNX weights file with a CheckpointMetadata JSON document, written next to the weights, describing the model-family specifics (IO tensor names, native quantile grid, context/horizon sizing, resolution). Keeping these specifics in data rather than code keeps the inference backends and forecasters generic.

Checkpoints can be resolved from a local path (LocalCheckpoint) or downloaded from the HuggingFace Hub (HubCheckpoint). Both expose a common resolve() returning a ResolvedCheckpoint.

Classes

CheckpointMetadata(**data)	Metadata describing a foundation-model checkpoint.
HubCheckpoint(**data)	A checkpoint resolved from a HuggingFace Hub repository.
LocalCheckpoint(**data)	A checkpoint resolved from a local filesystem path.
ResolvedCheckpoint(**data)	A checkpoint whose weights and metadata have been materialized locally.