Module coordinator 

Checkpoint coordination for multi-node adapter training (GPU-SHARE Phase 3, §3.4).

The coordinator polls each node’s checkpoint directory for metadata, compares val_loss across adapters, maintains a leaderboard, and identifies the best adapter at end of training.

Remote nodes are polled via cat checkpoint_dir/best/metadata.json over SSH. Local nodes read the file directly.

Structs

AdapterStatus

Status of a single adapter across the cluster.

CheckpointCoordinator

Coordinator for multi-node training checkpoint polling.

CheckpointMetadata

Metadata written by save_adapter_checkpoint() in multi_adapter_pipeline.

LeaderboardEntry

Leaderboard entry ranking adapters by loss.

NodeHealth

Result of a node health check.

Enums

PollResult

Result of polling a single adapter’s checkpoint.

Functions

build_launch_command

Build a remote launch command for an adapter job on a node.

check_cluster_health

Check health of all nodes in a cluster (GPU-SHARE §3.6).

exec_launch

Execute a training job on a remote or local node.