1 lightning.finite_horizon.resume

lightning.finite_horizon.resume

Verified epoch-boundary recovery for diagnostic Q_H fits.

The immutable diagnostic publication remains owned by aria_nbv.lightning.finite_horizon.experiment. This module owns the smaller mutable seam used while a fit is active: one last-good checkpoint per completed training epoch and a content-bound reference that can reopen it.

1.1 Classes

Name Description
QhDiagnosticResumeRef Content-bound full Lightning checkpoint for one compatible fit.

1.1.1 QhDiagnosticResumeRef

lightning.finite_horizon.resume.QhDiagnosticResumeRef(
    checkpoint_path,
    checkpoint_sha256,
    execution_identity_sha256,
)

Content-bound full Lightning checkpoint for one compatible fit.

1.1.1.1 Attributes

Name Type Description
checkpoint_path Path Trusted local full-state Lightning checkpoint.
checkpoint_sha256 str SHA-256 of the exact checkpoint bytes.
execution_identity_sha256 str Fit identity embedded by the checkpoint’s originating experiment. The receiving fit independently recomputes this identity before asking Lightning to restore any state.

1.1.1.2 Methods

Name Description
from_checkpoint Build a verified reference from a trusted local checkpoint.
1.1.1.2.1 from_checkpoint
lightning.finite_horizon.resume.QhDiagnosticResumeRef.from_checkpoint(
    checkpoint_path,
)

Build a verified reference from a trusted local checkpoint.

1.2 Functions

Name Description
load_trusted_qh_resume_checkpoint Deserialize a provenance-verified local checkpoint for validation.
validate_qh_resume_checkpoint_payload Validate one already trusted full-state Lightning checkpoint payload.

1.2.1 load_trusted_qh_resume_checkpoint

lightning.finite_horizon.resume.load_trusted_qh_resume_checkpoint(path)

Deserialize a provenance-verified local checkpoint for validation.

Resume checkpoints contain Python and NumPy RNG tuples, which cannot be reconstructed by weights_only=True. Callers must establish trusted local provenance and an expected SHA-256 before crossing this pickle boundary. This function validates structure after deserialization; it does not make an untrusted pickle safe.

1.2.2 validate_qh_resume_checkpoint_payload

lightning.finite_horizon.resume.validate_qh_resume_checkpoint_payload(payload)

Validate one already trusted full-state Lightning checkpoint payload.

Back to top