Checkpoints

A checkpoint_dir is managed by a manifest.json. Reading that manifest tells a caller which checkpoints can be resumed from. The tier rules are under Parameter tiers and resuming.

Name	Summary
kd.load_checkpoint_manifest(directory)	Load + fully verify a directory's manifest, returning entries in order.
kd.CheckpointManifestEntry(filename, kind[, ...])	A frozen, JSON-safe manifest entry (kd-ckptman-v1).
kd.CheckpointManifestError	Raised on any kd-ckptman-v1 manifest contract violation.
kd.KIND_FINAL	—
kd.FINAL_STATUS_COMPLETED	—
kd.resolve_checkpoint(checkpoint_dir, iteration)	The archive under checkpoint_dir for iteration (the latest when None).
kd.eligible_checkpoint_iterations(checkpoint_dir)	The iterations that can be resumed from checkpoint_dir, ascending, unique.
kd.ResolvedCheckpoint(path, iteration)	The archive file a segment restores and the iteration it holds.
kd.CheckpointSelectionError	No eligible archive answers the request.

load_checkpoint_manifest

load_checkpoint_manifest(
    directory: str | Path,
) -> tuple[CheckpointManifestEntry, ...]

Load + fully verify a directory's manifest, returning entries in order.

CheckpointManifestEntry dataclass

A frozen, JSON-safe manifest entry (kd-ckptman-v1).

from_dict classmethod

from_dict(data: object) -> CheckpointManifestEntry

Strictly reconstruct an entry, rejecting any wire drift.

to_dict

to_dict() -> dict[str, Any]

Return a JSON-safe dict, iterating the frozen scheme key table.

CheckpointManifestError

Bases: RecordSchemaError

Raised on any kd-ckptman-v1 manifest contract violation.

KIND_FINAL module-attribute

KIND_FINAL: Final[str] = 'final'

FINAL_STATUS_COMPLETED module-attribute

FINAL_STATUS_COMPLETED: Final[str] = 'completed'

resolve_checkpoint

resolve_checkpoint(
    checkpoint_dir: Path | str, iteration: int | None = None
) -> ResolvedCheckpoint

The archive under checkpoint_dir for iteration (the latest when None).

eligible_checkpoint_iterations

eligible_checkpoint_iterations(
    checkpoint_dir: Path | str,
) -> tuple[int, ...]

The iterations that can be resumed from checkpoint_dir, ascending, unique.

ResolvedCheckpoint dataclass

The archive file a segment restores and the iteration it holds.

CheckpointSelectionError

Bases: ValueError

No eligible archive answers the request.