Struct TensorCheckpointScheduler 

pub struct TensorCheckpointScheduler {
    pub config: SchedulerConfig,
    pub checkpoints: Vec<CheckpointRecord>,
    pub next_checkpoint_id: u64,
    pub last_checkpoint_step: u64,
    pub last_checkpoint_tick: u64,
    pub best_loss: f64,
    pub stats: SchedulerStats,
}

Manages automatic scheduling of tensor checkpoint saves.

Evaluates a configurable list of CheckpointTriggers on each call to advance and records a checkpoint whenever the first matching trigger fires.

Fields

config: SchedulerConfig

Configuration (triggers + retention policy).

checkpoints: Vec<CheckpointRecord>

Ordered (oldest-first) list of saved checkpoint records.

next_checkpoint_id: u64

Next checkpoint ID to assign (starts at 0, increments after use).

last_checkpoint_step: u64

Training step at which the last checkpoint was saved.

last_checkpoint_tick: u64

Tick value at which the last checkpoint was saved.

best_loss: f64

Best (lowest) loss seen so far; initialised to f64::MAX.

stats: SchedulerStats

Running statistics.

Implementations

impl TensorCheckpointScheduler

pub fn new(config: SchedulerConfig) -> Self

Create a new scheduler with the given configuration.

pub fn should_checkpoint( &self, current_step: u64, current_loss: f64, current_tick: u64, ) -> Option<CheckpointTrigger>

Evaluate all configured triggers and return the first one that fires, or None if no trigger is currently satisfied.

Trigger semantics:

• StepInterval { every_n_steps: n }: fires when current_step > 0 AND (current_step − last_checkpoint_step) >= n.
• LossImprovement { min_delta: d }: fires when best_loss == f64::MAX (first observation) OR (best_loss − current_loss) >= d.
• TickInterval { every_n_ticks: n }: fires when current_tick > 0 AND (current_tick − last_checkpoint_tick) >= n.

pub fn record_checkpoint( &mut self, step: u64, loss: f64, trigger: CheckpointTrigger, current_tick: u64, ) -> u64

Record a checkpoint save event.

Updates internal state (step/tick cursors, best loss) and statistics, then prunes the oldest record if the history exceeds max_checkpoints.

Returns the newly assigned checkpoint ID.

pub fn advance( &mut self, step: u64, loss: f64, current_tick: u64, ) -> Option<u64>

Advance the scheduler by one observation.

Evaluates triggers; if one fires, records a checkpoint and returns Some(checkpoint_id). Returns None when no checkpoint is needed.

pub fn latest_checkpoint(&self) -> Option<&CheckpointRecord>

Return a reference to the most recently saved checkpoint record, if any.

pub fn stats(&self) -> &SchedulerStats

Return a reference to the cumulative statistics.