Struct PretrainLoop 

pub struct PretrainLoop<S: StepFn, V: ValFn> { /* private fields */ }

Concrete driver for the 370M pretraining loop.

The model + autograd + optimizer are injected by the caller so this module does not take a hard dependency on a specific model crate — it is a pure driver around the contract’s invariants. Tests in this module use a deterministic synthetic StepFn that does not require the 370M scaffold at all.

Implementations

impl<S: StepFn, V: ValFn> PretrainLoop<S, V>

pub fn new(config: PretrainConfig, step_fn: S, val_fn: V) -> Self

Construct a loop with a fixed seed (GATE-TRAIN-006).

pub fn with_checkpoint_fn(self, ckpt: Box<dyn CheckpointFn>) -> Self

Attach a per-epoch APR checkpoint writer (task #111 step 7). Returns self for builder-style chaining.

pub fn train_step(&mut self, step: u64) -> Result<StepMetrics, PretrainAbort>

Execute a single training step. Records metrics into step_metrics and returns the metric record. Aborts on INV-TRAIN-007/008 violation.

pub fn run_epoch( &mut self, epoch: usize, ) -> Result<EpochArtifact, PretrainAbort>

Run one epoch: steps_per_epoch train steps, then validation + divergence check + epoch artifact.

pub fn check_convergence(&mut self, epoch: usize) -> bool

INV-TRAIN-004 convergence/early-stop check. Returns true if the loop should halt cleanly with early-stop status.

pub fn run(&mut self) -> RunStatus

Execute the full pretraining loop. Returns the terminal status.

pub fn step_metrics(&self) -> &[StepMetrics]

Accessors for test / CLI wiring.

pub fn epoch_artifacts(&self) -> &[EpochArtifact]

pub fn val_loss_history(&self) -> &[f32]