Skip to content

Commit

Permalink
praxis/paxml 1.0.0 release
Browse files Browse the repository at this point in the history
PiperOrigin-RevId: 523742465
  • Loading branch information
The paxml Authors authored and aireenmei committed Apr 12, 2023
1 parent 142e3b5 commit 89e112b
Show file tree
Hide file tree
Showing 3 changed files with 19 additions and 6 deletions.
19 changes: 16 additions & 3 deletions RELEASE.md
Original file line number Diff line number Diff line change
@@ -1,7 +1,20 @@
# Version: 0.4.0
# Version: 1.0.0
## Major Features and Improvements
## Breaking changes
## Deprecations
* **Fiddle** - Pax's layer library, Praxis, now has layers and BaseParameterizable configured with [Fiddle](https://github.com/google/fiddle), a Python-first configuration library. Fiddle reduces boilerplate, and adds productivity features including history tracking, graphviz visualization, support for aliasing objects, and more.
* **CLI Experiment and Data Injectability** - Pax users are now able to select which experiments to run without the need to recompile for each experiment. Using a CLI interface based on Fiddle, users can override subsets of the experiment’s canonical dataset.
* **CLU Metrics** - Through Praxis, Pax has adopted CLU metrics as its standard metric interface. This allows other Jax/Flax codebases that have CLU metrics to use them in Praxis.
* **Orbax Integration** - Pax has consolidated on the [Orbax checkpoint library](https://github.com/google/orbax/tree/main/checkpoint) as the standard checkpointing library. Orbax supports pjit, pmap, and Pathways checkpointing. Orbax facilitates checkpoint compatibility with other frameworks and provides common functionality throughout the JAX ecosystem.
* **Flax Interoperability** - Through Praxis, Pax now supports shape inference, `__call__` for forward propagation, and has adopted Linen’s AxisMetadata for its mid-level sharding APIs. These changes improve interoperability with other Flax-based libraries such as T5X.
* **Custom Training Loop Support** - Pax now provides limited support for customized training pipelines. Users can define a custom "train program" to encapsulate their training logic and use it in BaseExperiment. Customizable training loops pave the path for features such as multi-task support.
* **SeqIO** - Pax has adopted [SeqIO](https://github.com/google/seqio). Supporting SeqIO allows users with existing SeqIO Tasks, input pipelines and evaluation workflows to use them directly in Pax. We are working to make the Pax-SeqIO evaluation setup more flexible and robust.
* **Documentation** - We have added documentation and Jupyter Notebook tutorials to our [docs](https://github.com/google/paxml/tree/main/paxml/docs) folder. (Although you may notice some empty links in the doc, they are placeholders for upcoming docs.)
## Note
* Version: 1.0.0
* Build Date: 20230329
* Paxml commit: 033eb2421a6fc3e24f76bb19dd260c6776c5933b
* Praxis version: 1.0.0
* Praxis commit: 621c2ca7bfcd0e21ea118a3d8e40e29b48313c0c
# Version: 0.4.0
## Note
* Version: 0.4.0
* Build Date: 20230329
Expand Down
4 changes: 2 additions & 2 deletions paxml/pip_package/cloudbuild-release.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -17,8 +17,8 @@ steps:

substitutions:
_PYTHON_VERSION: '3.8'
_RELEASE_VERSION: '0.4.0' # or rX.Y
_PRAXIS_VERSION: '0.4.0' # or rX.Y
_RELEASE_VERSION: '1.0.0' # or rX.Y
_PRAXIS_VERSION: '1.0.0' # or rX.Y
_IMAGE_NAME: 'paxml_${_RELEASE_VERSION}_${_PYTHON_VERSION}'
_WHEEL_FOLDER: '/tmp/wheels'
options:
Expand Down
2 changes: 1 addition & 1 deletion setup.py
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ def _get_requirements():

setup(
name='paxml',
version='0.4.0', # use major/minor version number, e.g. "0.1.0"
version='1.0.0', # use major/minor version number, e.g. "0.1.0"
description=('Framework to configure and run machine learning experiments '
'on top of Jax.'),
author='PAX team',
Expand Down

0 comments on commit 89e112b

Please sign in to comment.