Struct CheckpointManager 

pub struct CheckpointManager<T> { /* private fields */ }

Manager for activation checkpointing

Implementations

impl<T> CheckpointManager<T>

where T: Clone + Default + Send + Sync,

pub fn new(config: CheckpointingConfig) -> Self

Create a new checkpoint manager

pub fn with_memory_budget( memory_budget_mb: usize, avg_activation_mb: usize, ) -> Self

Create with automatic memory-based policy

pub fn should_checkpoint(&self, layer_index: usize, total_layers: usize) -> bool

Check if layer should be checkpointed according to policy

pub fn save_checkpoint( &self, layer_index: usize, activations: Vec<Tensor<T>>, ) -> Result<()>

Save a checkpoint for the given layer

pub fn get_checkpoint( &self, layer_index: usize, ) -> Result<Option<Checkpoint<T>>>

Retrieve a checkpoint for the given layer

pub fn restore_rng_state(&self, _rng_state: &[u8]) -> Result<()>

Restore RNG state previously captured by capture_rng_state.

The current scirs2_core RNG (a rand 0.10 ChaCha-backed generator) does not expose serializable PRNG state, so there is nothing that can be faithfully restored. This returns an honest UnsupportedOperation error instead of silently pretending the state was restored.

pub fn record_recomputation(&self, compute_time_us: u64)

Record a recomputation event

pub fn record_backward_pass(&self)

Record a backward pass

pub fn get_statistics(&self) -> Result<CheckpointStatistics>

Get current statistics

pub fn clear(&self) -> Result<()>

Clear all checkpoints

pub fn checkpoint_count(&self) -> usize

Get number of active checkpoints

pub fn total_memory_bytes(&self) -> usize

Get total memory used by checkpoints

Trait Implementations

impl<T> Clone for CheckpointManager<T>

where T: Clone + Default + Send + Sync,

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.