Checkpoint
One row of a checkpoints.jsonl manifest.
Attributes
attributelabelstrThe name field - e.g. "final", "epoch-3", or a step string.
attributestepint | None= NoneTraining step / batch index, if recorded (batch in tinker).
attributeepochint | None= Noneattributeweights_pathstr | None= NoneTraining-state checkpoint URI (state_path in tinker).
attributesampler_pathstr | None= NoneInference-ready sampler URI; what you pass to a sampling client.
attributerawdict= field(default_factory=dict)Untouched manifest row, for fields not modeled above.
attributehas_pathboolFunctions
funcfrom_manifest_row(cls, row) -> CheckpointparamclsparamrowdictReturns
evsys_sdk.checkpoint.Checkpointfuncpick_final(checkpoints) -> Checkpoint | NonePick the one to evaluate against.
Strategy: prefer an explicit name == "final" row that exposes a
path, else the last row that exposes a path, else None.
paramcheckpointslist['Checkpoint']Returns
evsys_sdk.checkpoint.Checkpoint | Nonefunc__init__(self, label, step=None, epoch=None, weights_path=None, sampler_path=None, raw=dict()) -> Noneparamselfparamlabelstrparamstepint | None= Noneparamepochint | None= Noneparamweights_pathstr | None= Noneparamsampler_pathstr | None= Noneparamrawdict= dict()Returns
None