xlm.utils.checkpoint_paths
Checkpoint path helpers for Lightning training resume paths.
is_distcp_sharded_checkpoint_dir(path)
Return True if path looks like a Lightning state_dict_type: sharded checkpoint directory.
Such checkpoints contain one *.distcp file per rank (flat layout under e.g. last.ckpt/).
is_consolidatable_lightning_sharded_dir(path)
Return True if path can be passed to Lightning's distributed checkpoint consolidation.
Requires shard files plus meta.pt (rank-0 metadata written by Lightning).
is_usable_lightning_train_checkpoint_path(path)
Return True if path can be passed to Trainer.fit(..., ckpt_path=...).
Accepts a regular .ckpt file or a directory with at least one *.distcp shard.
resolve_explicit_resume_checkpoint_path(path)
Validate resume_checkpoint_path from config.
Raises:
| Type | Description |
|---|---|
ValueError
|
If the path does not exist, or is a directory without |
find_auto_resume_checkpoint(checkpointing_dir)
Return on_exception.ckpt or last.ckpt under checkpointing_dir if usable.
Prefers on_exception.ckpt when present. Paths may be files or sharded directories.