Struct CheckpointCoordinator 

pub struct CheckpointCoordinator {
    pub adapters: HashMap<usize, AdapterStatus>,
    pub poll_interval_secs: u64,
    /* private fields */
}

Coordinator for multi-node training checkpoint polling.

Fields

adapters: HashMap<usize, AdapterStatus>

Adapter statuses indexed by adapter_idx.

poll_interval_secs: u64

Poll interval in seconds.

Implementations

impl CheckpointCoordinator

pub fn new( cluster: ClusterConfig, placements: &[PlacementDecision], checkpoint_dirs: &HashMap<usize, PathBuf>, poll_interval_secs: u64, ) -> Self

Create a new coordinator from placement decisions.

pub fn poll_all(&mut self) -> Vec<PollResult>

Poll all adapters for their latest checkpoint metadata.

For local nodes, reads the file directly. For SSH nodes, executes ssh host cat <path> and parses JSON.

pub fn leaderboard(&self) -> Vec<LeaderboardEntry>

Generate leaderboard sorted by loss (ascending).

pub fn best_adapter(&self) -> Option<&AdapterStatus>

Find the best adapter (lowest loss).

pub fn format_leaderboard(&self) -> String

Format leaderboard as a human-readable string.

impl CheckpointCoordinator

pub fn pull_best_checkpoint(&self, dest: &Path) -> Result<PathBuf, String>

Pull the best adapter’s checkpoint from its node to a local directory (§3.4).

For local nodes, copies the checkpoint directory. For SSH nodes, uses scp -r to fetch the checkpoint.

Returns the local path where the checkpoint was saved.