CheckpointManager

class CheckpointManager(trainer)

Bases: object

Own checkpoint format, save/load backends, and trainer resume semantics.

load_optimizer_state(optimizer_state)

Load canonical optimizer state and fail if moments are not restored.

load_weight_and_resume(*, keywords='', replacement=None, rules=None, strict=False)

Load configured weights and restore training state when cfg.resume is true.

Pass rules (a list of (keywords, replacement) pairs) to apply several key rewrites in a single load; the scalar keywords/replacement form is kept for back-compat and is treated as a one-rule list.

resume_training_state(checkpoint)

Restore structured or legacy optimizer, scheduler, RNG, and cursor state.

save_epoch_checkpoint(*, is_best=False, step_count=0, save_freq=None)

Save an epoch/metric-oriented checkpoint. Must run on all ranks.

save_iteration_checkpoint(*, backend=None, is_best=False, step_count=0, save_freq=None, save_iter_checkpoints=False)

Save an iteration-oriented checkpoint. Must run on all ranks.