tinker_cookbook.checkpoint_utils.save_checkpoint

tinker_cookbook.checkpoint_utils.save_checkpoint( training_client, name, log_path, loop_state, kind, ttl_seconds, store)

Save model checkpoint (synchronous wrapper around save_checkpoint_async).

Parameters:

• training_client ( tinker. TrainingClient) – Training client to save from.
• name ( str) – Name for the checkpoint (used in the tinker:// path).
• log_path ( str) – Directory containing checkpoints.jsonl.
• loop_state ( dict[str, Any]) – Training loop state dict (may include batch, epoch, etc.).
• kind ( Literal['state', 'sampler', 'both']) – Which checkpoint types to save ("state", "sampler", or "both").
• ttl_seconds ( int | None) – Server-side retention. None keeps the checkpoint indefinitely.
• store ( TrainingRunStore | None) – If provided, write the checkpoint record via Storage protocol.

Returns: Dict mapping "state_path" and/or "sampler_path" to tinker:// paths.

save_checkpoint(
training_client=training_client,
name="step-100",
log_path="./logs",
loop_state={"epoch": 0, "batch": 100},
)