Struct SmartCheckpointManager 

pub struct SmartCheckpointManager { /* private fields */ }

Smart checkpoint manager with differential checkpointing

Implementations

impl SmartCheckpointManager

pub fn new(config: CheckpointConfig, checkpoint_dir: PathBuf) -> Result<Self>

pub fn with_differential_threshold(self, threshold: f32) -> Self

Set the absolute change below which an element is treated as unchanged by differential checkpointing. The default, 0.0, records every element that differs at all (lossless).

pub fn baseline_state(&self) -> &HashMap<String, Tensor>

Model state the next differential checkpoint will be diffed against.

pub fn should_checkpoint( &self, step: usize, performance_metrics: &PerformanceMetrics, ) -> bool

pub fn create_checkpoint( &mut self, step: usize, model_state: &HashMap<String, Tensor>, ) -> Result<CheckpointInfo>

pub fn load_checkpoint(&self, step: usize) -> Result<HashMap<String, Tensor>>

Restore the model state recorded at step.

Differential checkpoints are replayed on top of the nearest preceding full checkpoint, so any step in the retained history can be restored.

pub fn get_latest_checkpoint(&self) -> Option<&CheckpointInfo>

pub fn get_checkpoint_history(&self) -> &[CheckpointInfo]