Struct CheckpointManager 

pub struct CheckpointManager<T: Float + Debug + ScalarOperand + Send + Sync + 'static> { /* private fields */ }

Checkpoint management

Implementations

impl<T: Float + Debug + ScalarOperand + Send + Sync + 'static> CheckpointManager<T>

pub fn new(config: &TransformerBasedOptimizerConfig<T>) -> Result<Self>

Build a checkpoint manager whose auto-save cadence follows config.

The auto-save frequency used to be a fixed 100 steps regardless of the optimizer’s configuration; it now tracks StateConfig::from_optimizer_config, which derives it from performance_config.metrics_interval.

pub fn auto_save_config(&self) -> &AutoSaveConfig

The auto-save policy this manager was built with.

pub fn maybe_auto_save( &mut self, step: usize, snapshot: OptimizerStateSnapshot<T>, ) -> Result<Option<String>>

Save snapshot if step falls on the configured auto-save cadence.

Returns the new checkpoint id, or None when auto-save is disabled or step is not a save point. Auto-saves are named auto_<step> and are capped at AutoSaveConfig::max_auto_saves independently of the explicit-checkpoint limit, oldest evicted first, so a long run cannot grow the store without bound. step == 0 saves the initial state.

Without this, auto_save_config was populated at construction and never read by anything: the auto-save policy existed only as a value.

pub fn auto_save_count(&self) -> usize

Number of auto-saved checkpoints currently retained.

pub fn save_checkpoint( &mut self, name: String, snapshot: OptimizerStateSnapshot<T>, ) -> Result<String>

pub fn load_checkpoint( &self, checkpoint_id: &str, ) -> Result<OptimizerStateSnapshot<T>>

pub fn list_checkpoints(&self) -> Vec<CheckpointMetadata>

pub fn delete_checkpoint(&mut self, checkpoint_id: &str) -> Result<bool>

pub fn get_checkpoint_count(&self) -> usize