checkpoints

CheckpointManager - write the checkpoints.jsonl manifest that :mod:evsys_sdk.checkpoint already knows how to read.

The reader half lives in evsys_sdk/checkpoint.py (:func:~evsys_sdk.checkpoint.read_manifest, :meth:~evsys_sdk.checkpoint.Checkpoint.pick_final). This is the writer counterpart used by :class:~evsys_sdk.training.loop.TrainingLoop.

The format is intentionally identical to the manifest tinker_cookbook used to emit, so downstream consumers (e.g. :meth:~evsys_sdk.inference.tinker.TinkerInference.from_run_result) keep working unchanged when an algorithm is ported to the native loop.

Schema per row (one JSON object per line)::

{"name": "step_500", "batch": 499, "epoch": 3, "state_path": "tinker://.../weights/step_500", "sampler_path": "tinker://.../sampler_weights/step_500"}

state_path is the full training state (weights + optimizer) used for resume. sampler_path is the inference-ready weights snapshot used by the eval client.

attributelogger

= logging.getLogger(__name__)

attribute__all__

= ['CheckpointManager', 'ManifestRow']