Struct CheckpointManager 

pub struct CheckpointManager {
    pub dir: PathBuf,
    pub max_to_keep: usize,
    pub format: CheckpointFormat,
    /* private fields */
}

Manages writing and reading checkpoints under a directory.

Keeps a rolling window of the most-recent max_to_keep checkpoints, deleting older files automatically after each save.

Fields

dir: PathBuf

max_to_keep: usize

format: CheckpointFormat

Implementations

impl CheckpointManager

pub fn new( dir: impl AsRef<Path>, max_to_keep: usize, format: CheckpointFormat, ) -> Result<Self, CheckpointError>

Create a new manager, creating dir if it does not already exist.

pub fn save( &mut self, ckpt: &OptimizerCheckpoint, ) -> Result<PathBuf, CheckpointError>

Save ckpt to disk and prune old checkpoints. Returns the saved path.

pub fn load_latest(&self) -> Result<OptimizerCheckpoint, CheckpointError>

Load the most recently saved checkpoint.

pub fn load_at_step( &self, step: u64, ) -> Result<OptimizerCheckpoint, CheckpointError>

Load the checkpoint saved at a specific training step.

pub fn list(&self) -> Vec<(u64, &Path)>

Return (step, path) pairs for all retained checkpoints.

pub fn count(&self) -> usize

Number of checkpoints currently retained.