Struct TrainingStateWriter 

pub struct TrainingStateWriter { /* private fields */ }

Lightweight state writer for the training loop (Producer)

Used by the trainer to write atomic state updates.

Implementations

impl TrainingStateWriter

pub fn new<P: AsRef<Path>>( experiment_dir: P, experiment_id: &str, model_name: &str, ) -> Self

Create a new state writer

pub fn with_console_progress(self, enabled: bool) -> Self

Enable inline console progress on stdout.

When enabled, update_step() periodically prints a formatted progress line so users see output without needing a separate apr monitor process.

pub fn set_epochs(&mut self, total_epochs: usize, steps_per_epoch: usize)

Set total epochs and steps for a training phase

Call this at the start of each training phase to ensure the TUI displays correct progress. This resets epoch/step counters.

pub fn set_config( &mut self, optimizer_name: &str, batch_size: usize, model_path: &str, checkpoint_path: &str, )

Set run configuration (optimizer, batch size, paths)

pub fn set_gpu(&mut self, device_name: &str, vram_total_gb: f32)

Set GPU device info for the training snapshot.

When CUDA is active, populates the gpu field with device name and total VRAM. Updated dynamically by the training loop.

pub fn start(&mut self) -> Result<()>

Mark training as started

pub fn update_step( &mut self, epoch: usize, step: usize, loss: f32, learning_rate: f32, gradient_norm: f32, tokens_per_second: f32, accuracy: f32, ) -> Result<()>

Update training step

Warnings

Logs a warning to stderr if step > steps_per_epoch, which indicates a configuration mismatch (likely set_epochs() not called at phase start).

pub fn emit_epoch_summary( &self, epoch: usize, total_epochs: usize, train_loss: f32, train_acc: f32, val_loss: f32, val_acc: f32, epoch_secs: f32, lr: f32, is_best: bool, )

Emit an epoch-end summary line to the console.

pub fn emit_info(&self, msg: &str)

Emit a one-shot informational message through the monitoring framework.

pub fn update_gpu(&mut self, gpu: GpuTelemetry) -> Result<()>

Update GPU telemetry

pub fn update_sample(&mut self, sample: SamplePeek) -> Result<()>

Update sample peek

pub fn complete(&mut self) -> Result<()>

Mark training as completed

pub fn fail(&mut self, error: &str) -> Result<()>

Mark training as failed

pub fn state_path(&self) -> &Path

Get the state file path