v0.12.0

Added

• Add LatestDuration preservation policy to support duration-based (TTL)
  checkpoint retention.
• Add EveryNStepsClosest preservation policy, which preserves checkpoints at
  steps closest to multiples of interval_steps.
• Add 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.

Changed

• #v1 Rename 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.

Fixed

• Fix Pathways MTC Colocated worker initialization and save/restore PyTree bugs.
• Fix bug where TensorStore transaction was never used, due to byte_limiter
  always having a non-None value.

v0.11.40

Removed

• #v1 Remove numpy and pytorch layouts, as these are insecure and not worth the
  effort in maintaining. Safetensors is instead the recommended conversion case.
• #v1 Remove LeafHandler as a user-exposed layer (it remains as an internal layer).

Added

• #v1 Allow a context to be default-configured for all Checkpointer
  operations.

Changed

• Upgrade to TensorStore>=0.1.84 (and Python>=3.11 as a consequence.)

v0.11.39

Added

• Tensorstore non-atomic file I/O locking on local/POSIX-compatible filesystems
  to avoid unnecessary renames.

Release v0.11.37

Added

• Add colocated runtime helpers for Pathways MTC.
• #v1 Centralize StorageOptions into ArrayOptions and implement field-level
  merging.
• Add Patch for Pathways CPU ids.
• Add sidecar MTC backend support
• Add Colocated Python Sidecar MTC Checkpoint Manager
• Add Colocated Python Orchestration handling
• #v1 Add LeafHandler as a CheckpointableHandler, so that ordinary PyTree
  leaves can also be saved as individual checkpointables.
• Move MTC files to multi_tier_checkpointing and use local checkpoint engine

Changed

• Reduced default OCDBT target file size from 2GiB to 400MiB for GCS paths.

v0.11.36

Added

• #v1 Add auto loading functionality to load_pytree and pytree_metadata.

v0.11.35

Added

• Add PyTorch DCP (Distributed Checkpoint) to the benchmark suite, with GCS
  connector and fsspec options.
• Eliminate unadvertised flax dependency from tree/utils.

Fixed

• #v1 Fix import path for step in orbax.checkpoint.experimental.v1.path.

v0.11.34

Increment Orbax version.

PiperOrigin-RevId: 896170688

v0.11.33

Increment Orbax version.

PiperOrigin-RevId: 871625440

v0.11.32

directory compatibility in safetensors_layout

PiperOrigin-RevId: 858993617

v0.11.30

No public description

PiperOrigin-RevId: 837257237