NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #2213 most downloaded on PyPI
Orbax Checkpoint
Latest stable
0.12.9
Last release 19 days ago
03 Nov 2026
Ships fairly regularly
a new release about every 3 weeks
Some releases are documented
notes for 15 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
4 years old
116 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
Add support to GCS Rapid Bucket.
FileOptions.skip_sync_file_validations=True (V0 and #v1) now checks for and
removes an existing destination in the background directory-creation thread
(with create_directories_asynchronously=True) instead of only skipping the
checks. With force=False / overwrite=False, an existing destination would
raise an exception when the save is awaited (e.g. wait_until_finished()).Consolidate GCS kvstore spec generation and support gcs_grpc with OCDBT kvstack.
logger parameter to v1.training.Checkpointer and log
RestoreStepStatistics in load_checkpointables.delete and delete_async free functions and
training.Checkpointer.delete / delete_async for deleting an entire
checkpoint or a single named checkpointable.regulator_options (MemoryOptions.RegulatorOptions) and
expected_surge_bytes to v1.options.MemoryOptions for dynamic D2H transfer
memory regulation.kvstore spec generation and support gcs_grpc with OCDBT
kvstack.CheckpointManager.wait_until_finished() to only accumulate
wait_for_prev_duration_secs when called from the main thread.Your coding agent can read these notes before it upgrades. Set up the MCP server →
CheckpointManager.save() with async_checkpointer will skip validating existing folder if skip_sync_file_validations is set to True.
skip_sync_file_validations is set to True.Add v1.training.pathways.Snapshotter to public API and docs, which provides a simple mechanism for scale-up/scale-down recovery with large-scale Pathw
CheckpointLayout.AUTO to detect the checkpoint layout from
filesystem markers when loading.v1.training.pathways.Snapshotter to public API and docs, which provides
a simple mechanism for scale-up/scale-down recovery with large-scale Pathways
training.Add FileOptions.skip_sync_file_validations to skip synchronized file validation during async checkpointing.
FileOptions.skip_sync_file_validations to skip synchronized file
validation during async checkpointing.Nothing published for this version
Deprecate temporary_path_class in CheckpointManagerOptions in favor of atomicity_options with AtomicityMode.
AtomicityOptions and AtomicityMode (AUTO, COMMIT_FILE,
ATOMIC_RENAME) to configure the checkpoint save atomicity protocol in
both v0 and v1.commit_success.txt on all newly finalized checkpoints across all
storage backends, enabling cross-backend reading compatibility.allow_legacy_atomic_rename option to AtomicityOptions to opt into
legacy fallback validation when reading legacy AtomicRename checkpoints
missing commit_success.txt.multi_tier_checkpointing_backup_interval_steps parametertemporary_path_class in CheckpointManagerOptions in favor of
atomicity_options with AtomicityMode.SafetensorsOptions.max_over_read_ratio, picked per file from the planned
reads by default, fetched as concurrent ranged reads of
SafetensorsOptions.read_chunk_bytes each, bounded by
MemoryOptions.read_concurrent_bytes), replacing the transient-array
resharding load. Supports arbitrary shardings, including inner-dimension
partitions; a warning reports shardings whose fragmented byte layout forces
heavy over-read.Nothing published for this version
Update MTC w/ Pathways to Support Scale Elasticity.
gcs_grpc driver.Add LatestDuration preservation policy to support duration-based (TTL) checkpoint retention.
LatestDuration preservation policy to support duration-based (TTL)
checkpoint retention.EveryNStepsClosest preservation policy, which preserves checkpoints at
steps closest to multiples of interval_steps.SerializationStatusCallback for save operations to track transfer and
disk write events. While primarily intended as lightweight progress handles,
this callback system allows the potential to inject custom logic for
advanced monitoring and custom actions during checkpointing.save/load_pytree to save/load. Eliminate most user-facing
"pytree" terminology in favor of "state" as a more specific term.
Add deprecations.py for handling deprecated public functions.Upgrade to TensorStore>=0.1.84 (and Python>=3.11 as a consequence.)
LeafHandler as a user-exposed layer (it remains as an internal layer).Checkpointer
operations.Tensorstore non-atomic file I/O locking on local/POSIX-compatible filesystems to avoid unnecessary renames.
Nothing published for this version
Add colocated runtime helpers for Pathways MTC.
StorageOptions into ArrayOptions and implement field-level
merging.LeafHandler as a CheckpointableHandler, so that ordinary PyTree
leaves can also be saved as individual checkpointables.### Added - #v1 Add auto loading functionality to load_pytree and pytree_metadata.
load_pytree and pytree_metadata.Add PyTorch DCP (Distributed Checkpoint) to the benchmark suite, with GCS connector and fsspec options.
step in orbax.checkpoint.experimental.v1.path.Increment Orbax version. PiperOrigin-RevId: 896170688
Increment Orbax version.
PiperOrigin-RevId: 896170688
Increment Orbax version. PiperOrigin-RevId: 871625440
Increment Orbax version.
PiperOrigin-RevId: 871625440
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per quarter.