checkpoints

What problem does it solve?

Checkpointing solves the challenge of restarting interrupted training runs and exporting sampler weights for evaluation by providing structured saving, loading, and lifecycle management through a CheckpointRecord.

Core Features & Use Cases

• Save full training state with save_state to resume training, including weights and optimizer state.
• Save only sampler weights with save_weights_for_sampler for sampling/export, or use save_weights_and_get_sampling_client to obtain a ready-to-use client.
• Manage checkpoints via REST API and CLI (list, publish, set TTL, delete) to maintain lifecycle and provenance.
• Serialize and deserialize checkpoint records with CheckpointRecord for robust tracking and reproducibility.

Quick Start

Save the full training state with save_state for resumption, or export only the sampler weights with save_weights_for_sampler for fast evaluation.