2021-03-05 11:07:50 -08:00
# Change log
Best viewed [here ](https://jax.readthedocs.io/en/latest/changelog.html ).
<!--
Remember to align the itemized text with the first line of an item within a list.
-->
2023-06-22 17:54:57 -07:00
## jax 0.4.14
2023-07-14 14:27:29 -07:00
* Changes
* `jax.jit` takes `donate_argnames` as an argument. It's semantics are similar
to `static_argnames` .
If neither donate_argnums nor donate_argnames is provided, no
arguments are donated. If donate_argnums is not provided but
donate_argnames is, or vice versa, JAX uses
`inspect.signature(fun)` to find any positional arguments that
correspond to donate_argnames (or vice versa). If both donate_argnums and donate_argnames are provided, inspect.signature is not used, and only actual
parameters listed in either donate_argnums or donate_argnames will
be donated.
2023-06-22 18:29:02 -07:00
* Deprecations
* Python 3.8 support has been dropped as per
https://jax.readthedocs.io/en/latest/deprecation.html
2023-06-23 01:22:49 -07:00
* JAX now requires NumPy 1.22 or newer as per
https://jax.readthedocs.io/en/latest/deprecation.html
2023-06-30 04:29:43 -07:00
* Passing optional arguments to {func}`jax.numpy.ndarray.at` by position is
no longer supported, after being deprecated in JAX version 0.4.7.
For example, instead of `x.at[i].get(True)` , use `x.at[i].get(indices_are_sorted=True)`
2023-07-11 10:26:33 -07:00
* The following `jax.Array` methods have been removed, after being deprecated
in JAX v0.4.5:
* `jax.Array.broadcast` : use {func}`jax.lax.broadcast` instead.
* `jax.Array.broadcast_in_dim` : use {func}`jax.lax.broadcast_in_dim` instead.
* `jax.Array.split` : use {func}`jax.numpy.split` instead.
2023-07-11 12:42:32 -07:00
* The following APIs have been removed after previous deprecation:
* `jax.ad` : use {mod}`jax.interpreters.ad` .
* `jax.curry` : use ``curry = lambda f: partial(partial, f)` `.
* `jax.partial_eval` : use {mod}`jax.interpreters.partial_eval` .
* `jax.pxla` : use {mod}`jax.interpreters.pxla` .
* `jax.xla` : use {mod}`jax.interpreters.xla` .
* `jax.ShapedArray` : use {class}`jax.core.ShapedArray` .
* `jax.interpreters.pxla.device_put` : use {func}`jax.device_put` .
* `jax.interpreters.pxla.make_sharded_device_array` : use {func}`jax.make_array_from_single_device_arrays` .
* `jax.interpreters.pxla.ShardedDeviceArray` : use {class}`jax.Array` .
* `jax.numpy.DeviceArray` : use {class}`jax.Array` .
2023-06-22 18:29:02 -07:00
2023-06-28 10:09:10 -07:00
* Breaking changes
2023-07-07 12:07:44 -07:00
* JAX now requires ml_dtypes version 0.2.0 or newer.
2023-06-28 10:09:10 -07:00
* To fix a corner case, calls to {func}`jax.lax.cond` with five
arguments will always resolve to the "common operands" `cond`
behavior (as documented) if the second and third arguments are
callable, even if other operands are callable as well. See
[#16413 ](https://github.com/google/jax/issues/16413 ).
2023-07-14 15:24:24 -07:00
* The deprecated config option `jax_jit_pjit_api_merge` , which did nothing,
has been removed.
2023-06-28 10:09:10 -07:00
2023-07-16 09:26:27 -07:00
* New features
* JAX now supports a configuration flag --jax_serialization_version
and a JAX_SERIALIZATION_VERSION environment variable to control the
serialization version ({jax-issue}`#16746` ).
2023-06-22 17:54:57 -07:00
## jaxlib 0.4.14
2023-06-22 18:29:02 -07:00
* Deprecations
* Python 3.8 support has been dropped as per
https://jax.readthedocs.io/en/latest/deprecation.html
2023-06-22 17:54:57 -07:00
## jax 0.4.13 (June 22, 2023)
2023-06-15 15:21:36 -07:00
* Changes
* `jax.jit` now allows `None` to be passed to `in_shardings` and
`out_shardings` . The semantics are as follows:
* For in_shardings, JAX will mark is as replicated but this behavior
can change in the future.
* For out_shardings, we will rely on the XLA GSPMD partitioner to
determine the output shardings.
* `jax.experimental.pjit.pjit` also allows `None` to be passed to
`in_shardings` and `out_shardings` . The semantics are as follows:
* If the mesh context manager is *not* provided, JAX has the freedom to
choose whatever sharding it wants.
* For in_shardings, JAX will mark is as replicated but this behavior
can change in the future.
* For out_shardings, we will rely on the XLA GSPMD partitioner to
determine the output shardings.
* If the mesh context manager is provided, None will imply that the value
will be replicated on all devices of the mesh.
2023-06-22 16:30:04 -04:00
* Executable.cost_analysis() works on Cloud TPU
* Added a warning if a non-allowlisted `jaxlib` plugin is in use.
* Added `jax.tree_util.tree_leaves_with_path` .
2023-06-26 18:15:04 -07:00
* `None` is not a valid input to
`jax.experimental.multihost_utils.host_local_array_to_global_array` or
`jax.experimental.multihost_utils.global_array_to_host_local_array` .
Please use `jax.sharding.PartitionSpec()` if you wanted to replicate your
input.
2023-06-15 15:21:36 -07:00
2023-06-12 15:31:38 -04:00
* Bug fixes
* Fixed incorrect wheel name in CUDA 12 releases (#16362 ); the correct wheel
is named `cudnn89` instead of `cudnn88` .
2023-06-08 16:53:22 -07:00
2023-06-10 09:27:42 +03:00
* Deprecations
* The `native_serialization_strict_checks` parameter to
{func}`jax.experimental.jax2tf.convert` is deprecated in favor of the
new `native_serializaation_disabled_checks` ({jax-issue}`#16347` ).
2023-06-22 17:54:57 -07:00
## jaxlib 0.4.13 (June 22, 2023)
2023-06-08 16:53:22 -07:00
2023-06-22 16:30:04 -04:00
* Changes
* Added Windows CPU-only wheels to the `jaxlib` Pypi release.
2023-06-20 10:08:38 -07:00
* Bug fixes
* `__cuda_array_interface__` was broken in previous jaxlib versions and is now
fixed ({jax-issue}`16440` ).
2023-06-22 16:30:04 -04:00
* Concurrent CUDA kernel tracing is now enabled by default on NVIDIA GPUs.
2023-06-20 10:08:38 -07:00
2023-06-08 16:53:22 -07:00
## jax 0.4.12 (June 8, 2023)
2023-05-31 23:48:50 -07:00
2023-06-02 22:38:15 -05:00
* Changes
* Added {class}`scipy.spatial.transform.Rotation` and {class}`scipy.spatial.transform.Slerp`
2023-06-02 04:10:46 -07:00
* Deprecations
2023-06-06 07:32:35 -07:00
* `jax.abstract_arrays` and its contents are now deprecated. See related
functionality in :mod:`jax.core` .
* `jax.numpy.alltrue` : use `jax.numpy.all` . This follows the deprecation
of `numpy.alltrue` in NumPy version 1.25.0.
* `jax.numpy.sometrue` : use `jax.numpy.any` . This follows the deprecation
of `numpy.sometrue` in NumPy version 1.25.0.
* `jax.numpy.product` : use `jax.numpy.prod` . This follows the deprecation
of `numpy.product` in NumPy version 1.25.0.
* `jax.numpy.cumproduct` : use `jax.numpy.cumprod` . This follows the deprecation
of `numpy.cumproduct` in NumPy version 1.25.0.
2023-06-08 13:45:43 -07:00
* `jax.sharding.OpShardingSharding` has been removed since it has been 3
months since it was deprecated.
2023-06-02 04:10:46 -07:00
2023-06-08 16:53:22 -07:00
## jaxlib 0.4.12 (June 8, 2023)
2023-05-31 23:48:50 -07:00
2023-06-05 09:42:12 -04:00
* Changes
2023-06-08 11:41:47 -07:00
* Includes PTX/SASS for Hopper (SM version 9.0+) GPUs. Previous
2023-06-05 09:42:12 -04:00
versions of jaxlib should work on Hopper but would have a long
JIT-compilation delay the first time a JAX operation was executed.
2023-06-08 11:41:47 -07:00
* Bug fixes
* Fixes incorrect source line information in JAX-generated Python tracebacks
under Python 3.11.
2023-06-08 13:21:43 -07:00
* Fixes crash when printing local variables of frames in JAX-generated Python
tracebacks (#16027 ).
2023-06-08 11:41:47 -07:00
2023-05-31 23:48:50 -07:00
## jax 0.4.11 (May 31, 2023)
2023-05-09 17:06:59 -07:00
2023-05-24 10:35:01 -07:00
* Deprecations
* The following APIs have been removed after a 3 month deprecation period, in
accordance with the {ref}`api-compatibility` policy:
2023-05-26 12:34:32 -07:00
* `jax.experimental.PartitionSpec` : use `jax.sharding.PartitionSpec` .
* `jax.experimental.maps.Mesh` : use `jax.sharding.Mesh`
* `jax.experimental.pjit.NamedSharding` : use `jax.sharding.NamedSharding` .
* `jax.experimental.pjit.PartitionSpec` : use `jax.sharding.PartitionSpec` .
* `jax.experimental.pjit.FROM_GDA` . Instead pass sharded `jax.Array` objects
2023-05-24 10:35:01 -07:00
as input and remove the optional `in_shardings` argument to `pjit` .
2023-05-26 12:34:32 -07:00
* `jax.interpreters.pxla.PartitionSpec` : use `jax.sharding.PartitionSpec` .
* `jax.interpreters.pxla.Mesh` : use `jax.sharding.Mesh`
2023-05-28 07:15:34 -07:00
* `jax.interpreters.xla.Buffer` : use `jax.Array` .
2023-05-26 12:34:32 -07:00
* `jax.interpreters.xla.Device` : use `jax.Device` .
2023-05-28 07:15:34 -07:00
* `jax.interpreters.xla.DeviceArray` : use `jax.Array` .
* `jax.interpreters.xla.device_put` : use `jax.device_put` .
* `jax.interpreters.xla.xla_call_p` : use `jax.experimental.pjit.pjit_p` .
2023-05-26 12:34:32 -07:00
* `axis_resources` argument of `with_sharding_constraint` is removed. Please
use `shardings` instead.
2023-05-24 10:35:01 -07:00
2023-05-31 23:48:50 -07:00
## jaxlib 0.4.11 (May 31, 2023)
2023-05-11 18:16:02 -07:00
2023-05-25 07:19:56 -07:00
* Changes
2023-06-07 14:41:34 -07:00
* Added `memory_stats()` method to `Device` s. If supported, this returns a
dict of string stat names with int values, e.g. `"bytes_in_use"` , or None if
the platform doesn't support memory statistics. The exact stats returned may
vary across platforms. Currently only implemented on Cloud TPU.
2023-05-25 07:19:56 -07:00
* Readded support for the Python buffer protocol (`memoryview` ) on CPU
devices.
2023-05-11 18:16:02 -07:00
## jax 0.4.10 (May 11, 2023)
## jaxlib 0.4.10 (May 11, 2023)
2023-05-09 17:06:59 -07:00
2023-05-11 14:46:06 -07:00
* Changes
* Fixed `'apple-m1' is not a recognized processor for this target (ignoring
processor)` issue that prevented previous release from running on Mac M1.
2023-05-09 17:06:59 -07:00
## jax 0.4.9 (May 9, 2023)
2023-03-29 14:25:26 -07:00
2023-04-07 08:28:46 -07:00
* Changes
2023-05-09 17:06:59 -07:00
* The flags experimental_cpp_jit, experimental_cpp_pjit and
2023-04-07 10:41:42 -07:00
experimental_cpp_pmap have been removed.
2023-04-07 08:28:46 -07:00
They are now always on.
2023-05-09 17:08:00 -07:00
* Accuracy of singular value decomposition (SVD) on TPU has been improved
(requires jaxlib 0.4.9).
2023-04-07 08:28:46 -07:00
2023-03-30 14:50:07 -07:00
* Deprecations
2023-03-31 18:07:08 -07:00
* `jax.experimental.gda_serialization` is deprecated and has been renamed to
`jax.experimental.array_serialization` .
Please change your imports to use `jax.experimental.array_serialization` .
2023-03-30 14:50:07 -07:00
* The `in_axis_resources` and `out_axis_resources` arguments of pjit have been
deprecated. Please use `in_shardings` and `out_shardings` respectively.
2023-03-30 13:18:28 -07:00
* The function `jax.numpy.msort` has been removed. It has been deprecated since
JAX v0.4.1. Use `jnp.sort(a, axis=0)` instead.
2023-04-13 15:04:48 -07:00
* `in_parts` and `out_parts` arguments have been removed from `jax.xla_computation`
since they were only used with sharded_jit and sharded_jit is long gone.
2023-04-14 08:19:31 -07:00
* `instantiate_const_outputs` argument has been removed from `jax.xla_computation`
since it has been unused for a very long time.
2023-03-30 14:50:07 -07:00
2023-05-09 17:06:59 -07:00
## jaxlib 0.4.9 (May 9, 2023)
2023-03-29 14:25:26 -07:00
## jax 0.4.8 (March 29, 2023)
2023-03-27 15:06:06 -07:00
2023-03-28 13:42:51 -07:00
* Breaking changes
* A major component of the Cloud TPU runtime has been upgraded. This enables
the following new features on Cloud TPU:
* {func}`jax.debug.print` , {func}`jax.debug.callback` , and
{func}`jax.debug.breakpoint()` now work on Cloud TPU
* Automatic TPU memory defragmentation
{func}`jax.experimental.host_callback` is no longer supported on Cloud TPU
with the new runtime component. Please file an issue on the [JAX issue
tracker](https://github.com/google/jax/issues) if the new `jax.debug` APIs
are insufficient for your use case.
The old runtime component will be available for at least the next three
months by setting the environment variable
`JAX_USE_PJRT_C_API_ON_TPU=false` . If you find you need to disable the new
runtime for any reason, please let us know on the [JAX issue
tracker](https://github.com/google/jax/issues).
2023-03-28 12:43:32 -07:00
* Changes
* The minimum jaxlib version has been bumped from 0.4.6 to 0.4.7.
2023-03-28 09:53:57 -07:00
* Deprecations
* CUDA 11.4 support has been dropped. JAX GPU wheels only support
CUDA 11.8 and CUDA 12. Older CUDA versions may work if jaxlib is built
from source.
2023-03-29 09:22:34 -07:00
* `global_arg_shapes` argument of pmap only worked with sharded_jit and has
been removed from pmap. Please migrate to pjit and remove global_arg_shapes
from pmap.
2023-03-28 09:53:57 -07:00
2023-03-27 15:06:06 -07:00
## jax 0.4.7 (March 27, 2023)
2023-03-09 14:46:11 -08:00
2023-03-17 12:57:18 -07:00
* Changes
* As per https://jax.readthedocs.io/en/latest/jax_array_migration.html#jax -array-migration
`jax.config.jax_array` cannot be disabled anymore.
* `jax.config.jax_jit_pjit_api_merge` cannot be disabled anymore.
2023-03-20 09:42:59 -07:00
* {func}`jax.experimental.jax2tf.convert` now supports the `native_serialization`
parameter to use JAX's native lowering to StableHLO to obtain a
StableHLO module for the entire JAX function instead of lowering each JAX
primitive to a TensorFlow op. This simplifies the internals and increases
the confidence that what you serialize matches the JAX native semantics.
See [documentation ](https://github.com/google/jax/blob/main/jax/experimental/jax2tf/README.md ).
As part of this change the config flag `--jax2tf_default_experimental_native_lowering`
has been renamed to `--jax2tf_native_serialization` .
2023-03-23 05:12:57 -07:00
* JAX now depends on `ml_dtypes` , which contains definitions of NumPy types
like bfloat16. These definitions were previously internal to JAX, but have
been split into a separate package to facilitate sharing them with other
projects.
2023-02-06 11:32:28 -05:00
* JAX now requires NumPy 1.21 or newer and SciPy 1.7 or newer.
2023-03-17 12:57:18 -07:00
2023-03-13 10:04:39 -07:00
* Deprecations
2023-03-15 08:45:09 -07:00
* The type `jax.numpy.DeviceArray` is deprecated. Use `jax.Array` instead,
for which it is an alias.
2023-03-21 05:13:21 -07:00
* The type `jax.interpreters.pxla.ShardedDeviceArray` is deprecated. Use
`jax.Array` instead.
2023-03-13 10:04:39 -07:00
* Passing additional arguments to {func}`jax.numpy.ndarray.at` by position is deprecated.
For example, instead of `x.at[i].get(True)` , use `x.at[i].get(indices_are_sorted=True)`
2023-03-16 15:46:57 -07:00
* `jax.interpreters.xla.device_put` is deprecated. Please use `jax.device_put` .
* `jax.interpreters.pxla.device_put` is deprecated. Please use `jax.device_put` .
2023-03-27 11:32:30 -07:00
* `jax.experimental.pjit.FROM_GDA` is deprecated. Please pass in sharded
jax.Arrays as input and remove the `in_shardings` argument to pjit since
it is optional.
2023-03-13 10:04:39 -07:00
2023-03-27 15:06:06 -07:00
## jaxlib 0.4.7 (March 27, 2023)
2023-03-09 14:46:11 -08:00
2023-03-23 05:12:57 -07:00
Changes:
* jaxlib now depends on `ml_dtypes` , which contains definitions of NumPy types
like bfloat16. These definitions were previously internal to JAX, but have
been split into a separate package to facilitate sharing them with other
projects.
2023-03-09 14:46:11 -08:00
## jax 0.4.6 (Mar 9, 2023)
2023-03-02 16:08:34 -08:00
2023-03-03 18:05:37 -08:00
* Changes
2023-03-05 13:14:57 -08:00
* `jax.tree_util` now contain a set of APIs that allow user to define keys for their
2023-03-03 18:05:37 -08:00
custom pytree node. This includes:
2023-03-08 14:09:35 -08:00
* `tree_flatten_with_path` that flattens a tree and return not only each leaf but
2023-03-03 18:05:37 -08:00
also their key paths.
* `tree_map_with_paths` that can map a function that takes the key path as argument.
2023-03-08 14:09:35 -08:00
* `register_pytree_with_keys` ` to register how the key path and leaves should looks
2023-03-03 18:05:37 -08:00
like in a custom pytree node.
* `keystr` that pretty-prints a key path.
2023-03-08 14:09:35 -08:00
* {func}`jax2tf.call_tf` has a new parameter `output_shape_dtype` (default `None` )
that can be used to declare the output shape and type of the result. This enables
{func}`jax2tf.call_tf` to work in the presence of shape polymorphism. ({jax-issue}`#14734` ).
2023-03-03 18:05:37 -08:00
* Deprecations
* The old key-path APIs in `jax.tree_util` are deprecated and will be removed 3 months
from Mar 10 2023:
2023-03-05 13:14:57 -08:00
* `register_keypaths` : use {func}`jax.tree_util.register_pytree_with_keys` instead.
2023-03-03 18:05:37 -08:00
* `AttributeKeyPathEntry` : use `GetAttrKey` instead.
* `GetitemKeyPathEntry` : use `SequenceKey` or `DictKey` instead.
2023-03-09 14:46:11 -08:00
## jaxlib 0.4.6 (Mar 9, 2023)
2023-03-02 16:08:34 -08:00
## jax 0.4.5 (Mar 2, 2023)
2023-02-16 13:48:13 -08:00
2023-02-17 17:10:27 -08:00
* Deprecations
* `jax.sharding.OpShardingSharding` has been renamed to `jax.sharding.GSPMDSharding` .
`jax.sharding.OpShardingSharding` will be removed in 3 months from Feb 17, 2023.
2023-02-23 16:15:09 -08:00
* The following `jax.Array` methods are deprecated and will be removed 3 months from
Feb 23 2023:
* `jax.Array.broadcast` : use {func}`jax.lax.broadcast` instead.
* `jax.Array.broadcast_in_dim` : use {func}`jax.lax.broadcast_in_dim` instead.
* `jax.Array.split` : use {func}`jax.numpy.split` instead.
2023-02-17 17:10:27 -08:00
2023-02-16 13:48:13 -08:00
## jax 0.4.4 (Feb 16, 2023)
2023-02-08 10:08:57 -08:00
2023-02-08 11:55:10 -08:00
* Changes
* The implementation of `jit` and `pjit` has been merged. Merging jit and pjit
changes the internals of JAX without affecting the public API of JAX.
Before, `jit` was a final style primitive. Final style means that the creation
of jaxpr was delayed as much as possible and transformations were stacked
on top of each other. With the `jit` -`pjit` implementation merge, `jit`
becomes an initial style primitive which means that we trace to jaxpr
as early as possible. For more information see
[this section in autodidax ](https://jax.readthedocs.io/en/latest/autodidax.html#on-the-fly-final-style-and-staged-initial-style-processing ).
Moving to initial style should simplify JAX's internals and make
development of features like dynamic shapes, etc easier.
You can disable it only via the environment variable i.e.
`os.environ['JAX_JIT_PJIT_API_MERGE'] = '0'` .
The merge must be disabled via an environment variable since it affects JAX
at import time so it needs to be disabled before jax is imported.
2023-02-13 10:53:21 -08:00
* `axis_resources` argument of `with_sharding_constraint` is deprecated.
Please use `shardings` instead. There is no change needed if you were using
`axis_resources` as an arg. If you were using it as a kwarg, then please
use `shardings` instead. `axis_resources` will be removed after 3 months
from Feb 13, 2023.
2023-02-13 15:53:08 -08:00
* added the {mod}`jax.typing` module, with tools for type annotations of JAX
functions.
2023-02-15 08:13:58 -08:00
* The following names have been deprecated:
* `jax.xla.Device` and `jax.interpreters.xla.Device` : use `jax.Device` .
2023-02-15 09:15:11 -08:00
* `jax.experimental.maps.Mesh` . Use `jax.sharding.Mesh`
instead.
2023-02-15 08:13:58 -08:00
* `jax.experimental.pjit.NamedSharding` : use `jax.sharding.NamedSharding` .
* `jax.experimental.pjit.PartitionSpec` : use `jax.sharding.PartitionSpec` .
* `jax.interpreters.pxla.Mesh` : use `jax.sharding.Mesh` .
* `jax.interpreters.pxla.PartitionSpec` : use `jax.sharding.PartitionSpec` .
2023-02-14 15:36:18 -08:00
* Breaking Changes
* the `initial` argument to reduction functions like :func:`jax.numpy.sum`
is now required to be a scalar, consistent with the corresponding NumPy API.
The previous behavior of broadcating the output against non-scalar `initial`
values was an unintentional implementation detail ({jax-issue}`#14446` ).
2023-02-08 11:55:10 -08:00
2023-02-16 13:48:13 -08:00
## jaxlib 0.4.4 (Feb 16, 2023)
2023-02-10 14:15:15 -05:00
* Breaking changes
* Support for NVIDIA Kepler series GPUs has been removed from the default
`jaxlib` builds. If Kepler support is needed, it is still possible to
build `jaxlib` from source with Kepler support (via the
`--cuda_compute_capabilities=sm_35` option to `build.py` ), however note
that CUDA 12 has completely dropped support for Kepler GPUs.
2023-02-08 10:08:57 -08:00
## jax 0.4.3 (Feb 8, 2023)
2023-01-26 23:01:06 +00:00
* Breaking changes
2023-02-08 10:08:57 -08:00
* Deleted {func}`jax.scipy.linalg.polar_unitary` , which was a deprecated JAX
2023-01-26 23:01:06 +00:00
extension to the scipy API. Use {func}`jax.scipy.linalg.polar` instead.
2023-01-20 19:04:46 +00:00
2023-02-07 12:07:00 -05:00
* Changes
* Added {func}`jax.scipy.stats.rankdata` .
2023-02-08 10:08:57 -08:00
## jaxlib 0.4.3 (Feb 8, 2023)
2023-02-23 11:56:48 -08:00
* `jax.Array` now has the non-blocking `is_ready()` method, which returns `True`
if the array is ready (see also {func}`jax.block_until_ready` ).
2023-01-20 19:04:46 +00:00
2023-01-26 01:09:26 +00:00
## jax 0.4.2 (Jan 24, 2023)
2022-12-13 10:50:25 -08:00
2023-01-12 22:57:49 -08:00
* Breaking changes
* Deleted `jax.experimental.callback`
2023-01-18 12:27:02 +02:00
* Operations with dimensions in presence of jax2tf shape polymorphism have
been generalized to work in more scenarios, by converting the symbolic
dimension to JAX arrays. Operations involving symbolic dimensions and
`np.ndarray` now can raise errors when the result is used as a shape value
({jax-issue}`#14106` ).
2023-01-23 09:37:58 -08:00
* jaxpr objects now raise an error on attribute setting in order to avoid
problematic mutations ({jax-issue}`14102` )
2023-01-12 22:57:49 -08:00
2023-01-12 08:44:53 +01:00
* Changes
* {func}`jax2tf.call_tf` has a new parameter `has_side_effects` (default `True` )
that can be used to declare whether an instance can be removed or replicated
by JAX optimizations such as dead-code elimination ({jax-issue}`#13980` ).
2023-01-18 12:27:02 +02:00
* Added more support for floordiv and mod for jax2tf shape polymorphism. Previously,
certain division operations resulted in errors in presence of symbolic dimensions
({jax-issue}`#14108` ).
2023-01-12 08:44:53 +01:00
2023-01-26 01:09:26 +00:00
## jaxlib 0.4.2 (Jan 24, 2023)
2023-01-20 19:04:46 +00:00
* Changes
* Set JAX_USE_PJRT_C_API_ON_TPU=1 to enable new Cloud TPU runtime, featuring
automatic device memory defragmentation.
2022-12-13 10:50:25 -08:00
## jax 0.4.1 (Dec 13, 2022)
2022-12-12 14:42:48 -08:00
2022-11-16 06:55:34 -08:00
* Changes
2022-11-29 15:01:47 -08:00
* Support for Python 3.7 has been dropped, in accordance with JAX's
{ref}`version-support-policy` .
2022-11-16 06:55:34 -08:00
* We introduce `jax.Array` which is a unified array type that subsumes
`DeviceArray` , `ShardedDeviceArray` , and `GlobalDeviceArray` types in JAX.
The `jax.Array` type helps make parallelism a core feature of JAX,
2022-12-01 20:50:55 +00:00
simplifies and unifies JAX internals, and allows us to unify `jit` and
`pjit` . `jax.Array` has been enabled by default in JAX 0.4 and makes some
breaking change to the `pjit` API. The [jax.Array migration
guide](https://jax.readthedocs.io/en/latest/jax_array_migration.html) can
help you migrate your codebase to `jax.Array` . You can also look at the
2022-12-10 19:59:14 -08:00
[Distributed arrays and automatic parallelization ](https://jax.readthedocs.io/en/latest/notebooks/Distributed_arrays_and_automatic_parallelization.html )
2022-11-16 06:55:34 -08:00
tutorial to understand the new concepts.
2022-12-01 19:28:02 -08:00
* `PartitionSpec` and `Mesh` are now out of experimental. The new API endpoints
are `jax.sharding.PartitionSpec` and `jax.sharding.Mesh` .
`jax.experimental.maps.Mesh` and `jax.experimental.PartitionSpec` are
deprecated and will be removed in 3 months.
2022-12-11 22:54:39 -08:00
* `with_sharding_constraint` s new public endpoint is
`jax.lax.with_sharding_constraint` .
2022-12-02 11:36:51 -08:00
* If using ABSL flags together with `jax.config` , the ABSL flag values are no
longer read or written after the JAX configuration options are initially
populated from the ABSL flags. This change improves performance of reading
`jax.config` options, which are used pervasively in JAX.
2022-12-06 08:23:40 -08:00
* The jax2tf.call_tf function now uses for TF lowering the first TF
device of the same platform as used by the embedding JAX computation.
Before, it was using the 0th device for the JAX-default backend.
* A number of `jax.numpy` functions now have their arguments marked as
positional-only, matching NumPy.
2022-12-07 10:08:18 -08:00
* `jnp.msort` is now deprecated, following the deprecation of `np.msort` in numpy 1.24.
It will be removed in a future release, in accordance with the {ref}`api-compatibility`
policy. It can be replaced with `jnp.sort(a, axis=0)` .
2022-11-15 09:02:23 -08:00
2022-12-13 10:50:25 -08:00
## jaxlib 0.4.1 (Dec 13, 2022)
2022-12-12 17:01:54 -08:00
2022-11-28 10:37:42 -08:00
* Changes
2022-11-29 15:01:47 -08:00
* Support for Python 3.7 has been dropped, in accordance with JAX's
{ref}`version-support-policy` .
2022-11-28 10:37:42 -08:00
* The behavior of `XLA_PYTHON_CLIENT_MEM_FRACTION=.XX` has been changed to allocate XX% of
the total GPU memory instead of the previous behavior of using currently available GPU memory
to calculate preallocation. Please refer to
[GPU memory allocation ](https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html ) for
more details.
2022-12-02 15:13:55 -08:00
* The deprecated method `.block_host_until_ready()` has been removed. Use
`.block_until_ready()` instead.
2022-11-28 10:37:42 -08:00
2022-12-12 17:01:54 -08:00
## jax 0.4.0 (Dec 12, 2022)
* The release was yanked.
## jaxlib 0.4.0 (Dec 12, 2022)
* The release was yanked.
2022-11-15 09:02:23 -08:00
## jax 0.3.25 (Nov 15, 2022)
2022-11-07 11:39:19 -05:00
* Changes
* {func}`jax.numpy.linalg.pinv` now supports the `hermitian` option.
2022-11-09 06:23:22 -08:00
* {func}`jax.scipy.linalg.hessenberg` is now supported on CPU only. Requires
jaxlib > 0.3.24.
* New functions {func}`jax.lax.linalg.hessenberg` ,
{func}`jax.lax.linalg.tridiagonal` , and
2022-11-14 07:38:39 -08:00
{func}`jax.lax.linalg.householder_product` were added. Householder reduction
is currently CPU-only and tridiagonal reductions are supported on CPU and
GPU only.
* The gradients of `svd` and `jax.numpy.linalg.pinv` are now computed more
economically for non-square matrices.
2022-11-11 14:00:33 -08:00
* Breaking Changes
* Deleted the `jax_experimental_name_stack` config option.
2022-11-11 15:23:44 -08:00
* Convert a string `axis_names` arguments to the
{class}`jax.experimental.maps.Mesh` constructor into a singleton tuple
instead of unpacking the string into a sequence of character axis names.
2022-11-04 09:41:29 -07:00
2022-11-15 09:02:23 -08:00
## jaxlib 0.3.25 (Nov 15, 2022)
2022-11-09 06:23:22 -08:00
* Changes
2022-11-14 07:38:39 -08:00
* Added support for tridiagonal reductions on CPU and GPU.
* Added support for upper Hessenberg reductions on CPU.
* Bugs
* Fixed a bug that meant that frames in tracebacks captured by JAX were
incorrectly mapped to source lines under Python 3.10+
2022-11-04 09:41:29 -07:00
## jax 0.3.24 (Nov 4, 2022)
2022-10-26 13:55:32 -07:00
* Changes
2022-10-06 01:43:54 +00:00
* JAX should be faster to import. We now import scipy lazily, which accounted
for a significant fraction of JAX's import time.
2022-10-28 23:53:30 +00:00
* Setting the env var `JAX_PERSISTENT_CACHE_MIN_COMPILE_TIME_SECS=$N` can be
used to limit the number of cache entries written to the persistent cache.
By default, computations that take 1 second or more to compile will be
cached.
2022-10-20 15:55:57 -07:00
* Added {func}`jax.scipy.stats.mode` .
2022-10-26 13:55:32 -07:00
* The default device order used by `pmap` on TPU if no order is specified now
matches `jax.devices()` for single-process jobs. Previously the
two orderings differed, which could lead to unnecessary copies or
out-of-memory errors. Requiring the orderings to agree simplifies matters.
* Breaking Changes
2022-10-24 15:27:33 -07:00
* {func}`jax.numpy.gradient` now behaves like most other functions in {mod}`jax.numpy` ,
and forbids passing lists or tuples in place of arrays ({jax-issue}`#12958` )
2022-10-27 09:01:28 -07:00
* Functions in {mod}`jax.numpy.linalg` and {mod}`jax.numpy.fft` now uniformly
require inputs to be array-like: i.e. lists and tuples cannot be used in place
of arrays. Part of {jax-issue}`#7737` .
2022-11-02 19:12:32 -07:00
* Deprecations
* `jax.sharding.MeshPspecSharding` has been renamed to `jax.sharding.NamedSharding` .
`jax.sharding.MeshPspecSharding` name will be removed in 3 months.
2022-10-12 11:09:15 -07:00
2022-11-04 09:41:29 -07:00
## jaxlib 0.3.24 (Nov 4, 2022)
2022-10-26 10:07:32 -07:00
* Changes
* Buffer donation now works on CPU. This may break code that marked buffers
for donation on CPU but relied on donation not being implemented.
2022-10-12 11:09:15 -07:00
## jax 0.3.23 (Oct 12, 2022)
2022-10-11 17:49:10 -07:00
* Changes
* Update Colab TPU driver version for new jaxlib release.
2022-10-11 21:37:53 +00:00
## jax 0.3.22 (Oct 11, 2022)
2022-10-05 01:01:31 +00:00
* Changes
2022-10-05 18:16:49 +00:00
* Add `JAX_PLATFORMS=tpu,cpu` as default setting in TPU initialization,
so JAX will raise an error if TPU cannot be initialized instead of falling
back to CPU. Set `JAX_PLATFORMS=''` to override this behavior and automatically
choose an available backend (the original default), or set `JAX_PLATFORMS=cpu`
to always use CPU regardless of if the TPU is available.
2022-10-11 08:15:47 -07:00
* Deprecations
* Several test utilities deprecated in JAX v0.3.8 are now removed from
{mod}`jax.test_util` .
2022-10-03 22:06:19 +00:00
2022-10-11 21:37:53 +00:00
## jaxlib 0.3.22 (Oct 11, 2022)
2022-10-03 22:06:19 +00:00
## jax 0.3.21 (Sep 30, 2022)
2022-10-11 06:24:38 -07:00
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.20...jax-v0.3.21 ).
2022-09-27 20:59:08 +00:00
* Changes
* The persistent compilation cache will now warn instead of raising an
exception on error ({jax-issue}`#12582` ), so program execution can continue
if something goes wrong with the cache. Set
`JAX_RAISE_PERSISTENT_CACHE_ERRORS=true` to revert this behavior.
2022-09-28 18:49:22 +00:00
## jax 0.3.20 (Sep 28, 2022)
2022-09-28 13:33:52 +00:00
* Bug fixes:
* Adds missing `.pyi` files that were missing from the previous release ({jax-issue}`#12536` ).
* Fixes an incompatibility between `jax` 0.3.19 and the libtpu version it pinned ({jax-issue}`#12550` ). Requires jaxlib 0.3.20.
* Fix incorrect `pip` url in `setup.py` comment ({jax-issue}`#12528` ).
2022-09-26 12:38:32 -07:00
2022-09-28 18:49:22 +00:00
## jaxlib 0.3.20 (Sep 28, 2022)
2022-10-11 06:24:38 -07:00
* [GitHub commits ](https://github.com/google/jax/compare/jaxlib-v0.3.15...jaxlib-v0.3.20 ).
2022-09-28 13:33:52 +00:00
* Bug fixes
* Fixes support for limiting the visible CUDA devices via
`jax_cuda_visible_devices` in distributed jobs. This functionality is needed for
the JAX/SLURM integration on GPU ({jax-issue}`#12533` ).
2022-09-27 11:00:27 -07:00
## jax 0.3.19 (Sep 27, 2022)
2022-10-11 06:24:38 -07:00
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.18...jax-v0.3.19 ).
2022-09-28 13:33:52 +00:00
* Fixes required jaxlib version.
2022-09-26 12:38:32 -07:00
2022-09-26 14:10:57 +00:00
## jax 0.3.18 (Sep 26, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.17...jax-v0.3.18 ).
2022-09-02 08:23:57 -07:00
* Changes
* Ahead-of-time lowering and compilation functionality (tracked in
{jax-issue}`#7733` ) is stable and public. See [the
overview](https://jax.readthedocs.io/en/latest/aot.html) and the API docs
for {mod}`jax.stages` .
2022-09-23 09:59:46 -07:00
* Introduced {class}`jax.Array` , intended to be used for both `isinstance` checks
and type annotations for array types in JAX. Notice that this included some subtle
changes to how `isinstance` works for {class}`jax.numpy.ndarray` for jax-internal
objects, as {class}`jax.numpy.ndarray` is now a simple alias of {class}`jax.Array` .
2022-09-09 07:05:30 -07:00
* Breaking changes
* `jax._src` is no longer imported into the from the public `jax` namespace.
This may break users that were using JAX internals.
2022-09-13 15:51:30 -07:00
* `jax.soft_pmap` has been deleted. Please use `pjit` or `xmap` instead.
`jax.soft_pmap` is undocumented. If it were documented, a deprecation period
would have been provided.
2022-08-31 11:37:09 -07:00
## jax 0.3.17 (Aug 31, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.16...jax-v0.3.17 ).
2022-08-22 11:53:02 -07:00
* Bugs
* Fix corner case issue in gradient of `lax.pow` with an exponent of zero
({jax-issue}`12041` )
2022-08-16 12:07:27 -07:00
* Breaking changes
* {func}`jax.checkpoint` , also known as {func}`jax.remat` , no longer supports
the `concrete` option, following the previous version's deprecation; see
[JEP 11830 ](https://jax.readthedocs.io/en/latest/jep/11830-new-remat-checkpoint.html ).
2022-08-15 17:05:27 -07:00
* Changes
2022-08-17 10:43:50 -07:00
* Added {func}`jax.pure_callback` that enables calling back to pure Python functions from compiled functions (e.g. functions decorated with `jax.jit` or `jax.pmap` ).
2022-08-16 11:35:13 -07:00
* Deprecations:
* The deprecated `DeviceArray.tile()` method has been removed. Use {func}`jax.numpy.tile`
({jax-issue}`#11944` ).
2022-08-25 07:27:54 -07:00
* `DeviceArray.to_py()` has been deprecated. Use `np.asarray(x)` instead.
2022-08-11 17:10:21 -07:00
## jax 0.3.16
2022-07-25 09:47:44 -07:00
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.15...main ).
2022-08-06 14:49:09 +00:00
* Breaking changes
* Support for NumPy 1.19 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to NumPy 1.20 or newer.
2022-07-25 09:47:44 -07:00
* Changes
2022-08-02 22:28:52 -07:00
* Added {mod}`jax.debug` that includes utilities for runtime value debugging such at {func}`jax.debug.print` and {func}`jax.debug.breakpoint` .
* Added new documentation for [runtime value debugging ](debugging/index )
2022-07-25 09:47:44 -07:00
* Deprecations
* {func}`jax.mask` {func}`jax.shapecheck` APIs have been removed.
See {jax-issue}`#11557` .
2022-07-26 03:39:17 -07:00
* {mod}`jax.experimental.loops` has been removed. See {jax-issue}`#10278`
2022-08-10 08:26:47 -07:00
for an alternative API.
2022-07-06 09:06:16 -07:00
* {func}`jax.tree_util.tree_multimap` has been removed. It has been deprecated since
JAX release 0.3.5, and {func}`jax.tree_util.tree_map` is a direct replacement.
2022-08-01 14:58:41 -07:00
* Removed `jax.experimental.stax` ; it has long been a deprecated alias of
{mod}`jax.example_libraries.stax` .
2022-08-02 14:41:50 -07:00
* Removed `jax.experimental.optimizers` ; it has long been a deprecated alias of
{mod}`jax.example_libraries.optimizers` .
2022-08-10 08:26:47 -07:00
* {func}`jax.checkpoint` , also known as {func}`jax.remat` , has a new
implementation switched on by default, meaning the old implementation is
deprecated; see [JEP 11830 ](https://jax.readthedocs.io/en/latest/jep/11830-new-remat-checkpoint.html ).
2022-07-25 09:47:44 -07:00
## jax 0.3.15 (July 22, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.14...jax-v0.3.15 ).
2022-06-23 14:48:53 -07:00
* Changes
2022-06-27 11:04:50 -07:00
* `JaxTestCase` and `JaxTestLoader` have been removed from `jax.test_util` . These
classes have been deprecated since v0.3.1 ({jax-issue}`#11248` ).
2022-06-23 10:00:54 -04:00
* Added {class}`jax.scipy.gaussian_kde` ({jax-issue}`#11237` ).
2022-06-24 10:03:27 -07:00
* Binary operations between JAX arrays and built-in collections (`dict` , `list` , `set` , `tuple` )
now raise a `TypeError` in all cases. Previously some cases (particularly equality and inequality)
would return boolean scalars inconsistent with similar operations in NumPy ({jax-issue}`#11234` ).
2022-07-28 11:33:01 -07:00
* Several {mod}`jax.tree_util` routines accessed as top-level JAX package imports are now
deprecated, and will be removed in a future JAX release in accordance with the
{ref}`api-compatibility` policy:
2022-07-07 11:41:46 -07:00
* {func}`jax.treedef_is_leaf` is deprecated in favor of {func}`jax.tree_util.treedef_is_leaf`
* {func}`jax.tree_flatten` is deprecated in favor of {func}`jax.tree_util.tree_flatten`
* {func}`jax.tree_leaves` is deprecated in favor of {func}`jax.tree_util.tree_leaves`
* {func}`jax.tree_structure` is deprecated in favor of {func}`jax.tree_util.tree_structure`
* {func}`jax.tree_transpose` is deprecated in favor of {func}`jax.tree_util.tree_transpose`
* {func}`jax.tree_unflatten` is deprecated in favor of {func}`jax.tree_util.tree_unflatten`
2022-07-19 13:57:49 -07:00
* The `sym_pos` argument of {func}`jax.scipy.linalg.solve` is deprecated in favor of `assume_a='pos'` ,
following a similar deprecation in {func}`scipy.linalg.solve` .
2022-06-21 12:41:26 -07:00
2022-07-25 09:47:44 -07:00
## jaxlib 0.3.15 (July 22, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jaxlib-v0.3.14...jaxlib-v0.3.15 ).
2022-06-21 12:41:26 -07:00
2022-06-27 09:21:54 -04:00
## jax 0.3.14 (June 27, 2022)
2022-06-23 14:48:53 -07:00
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.13...jax-v0.3.14 ).
2022-05-19 21:06:25 +00:00
* Breaking changes
2022-06-13 13:58:15 -07:00
* {func}`jax.experimental.compilation_cache.initialize_cache` does not support
`max_cache_size_ bytes` anymore and will not get that as an input.
2022-06-17 16:38:56 +00:00
* `JAX_PLATFORMS` now raises an exception when platform initialization fails.
2022-05-17 11:23:10 -07:00
* Changes
2022-06-27 09:21:54 -04:00
* Fixed compatibility problems with NumPy 1.23.
2022-05-17 11:23:10 -07:00
* {func}`jax.numpy.linalg.slogdet` now accepts an optional `method` argument
that allows selection between an LU-decomposition based implementation and
an implementation based on QR decomposition.
* {func}`jax.numpy.linalg.qr` now supports `mode="raw"` .
2022-05-17 13:20:38 -07:00
* `pickle` , `copy.copy` , and `copy.deepcopy` now have more complete support when
used on jax arrays ({jax-issue}`#10659` ). In particular:
- `pickle` and `deepcopy` previously returned `np.ndarray` objects when used
on a `DeviceArray` ; now `DeviceArray` objects are returned. For `deepcopy` ,
the copied array is on the same device as the original. For `pickle` the
deserialized array will be on the default device.
- Within function transformations (i.e. traced code), `deepcopy` and `copy`
previously were no-ops. Now they use the same mechanism as `DeviceArray.copy()` .
2022-06-03 17:20:09 -07:00
- Calling `pickle` on a traced array now results in an explicit
2022-05-17 13:20:38 -07:00
`ConcretizationTypeError` .
2022-05-31 14:03:27 -04:00
* The implementation of singular value decomposition (SVD) and
symmetric/Hermitian eigendecomposition should be significantly faster on
TPU, especially for matrices above 1000x1000 or so. Both now use a spectral
divide-and-conquer algorithm for eigendecomposition (QDWH-eig).
2022-06-01 09:14:47 -07:00
* {func}`jax.numpy.ldexp` no longer silently promotes all inputs to float64,
instead it promotes to float32 for integer inputs of size int32 or smaller
({jax-issue}`#10921` ).
2022-05-25 16:01:16 -07:00
* Add a `create_perfetto_link` option to {func}`jax.profiler.start_trace` and
{func}`jax.profiler.start_trace` . When used, the profiler will generate a
link to the Perfetto UI to view the trace.
2022-05-31 15:42:46 -07:00
* Changed the semantics of {func}`jax.profiler.start_server(...)` to store the
keepalive globally, rather than requiring the user to keep a reference to
it.
2022-06-03 15:11:29 -04:00
* Added {func}`jax.random.generalized_normal` .
* Added {func}`jax.random.ball` .
2022-06-03 17:20:09 -07:00
* Added {func}`jax.default_device` .
2022-06-02 22:15:53 -07:00
* Added a `python -m jax.collect_profile` script to manually capture program
traces as an alternative to the Tensorboard UI.
2022-06-09 10:34:25 -07:00
* Added a `jax.named_scope` context manager that adds profiler metadata to
Python programs (similar to `jax.named_call` ).
2022-06-09 15:21:49 -07:00
* In scatter-update operations (i.e. :attr:`jax.numpy.ndarray.at` ), unsafe implicit
dtype casts are deprecated, and now result in a `FutureWarning` .
In a future release, this will become an error. An example of an unsafe implicit
cast is `jnp.zeros(4, dtype=int).at[0].set(1.5)` , in which `1.5` previously was
silently truncated to `1` .
2022-06-13 13:58:15 -07:00
* {func}`jax.experimental.compilation_cache.initialize_cache` now supports gcs
2022-05-19 21:06:25 +00:00
bucket path as input.
2022-06-11 14:24:19 -04:00
* Added {func}`jax.scipy.stats.gennorm` .
2022-06-27 09:21:54 -04:00
* {func}`jax.numpy.roots` is now better behaved when `strip_zeros=False` when
coefficients have leading zeros ({jax-issue}`#11215` ).
2022-05-15 12:13:54 -07:00
2022-06-27 09:21:54 -04:00
## jaxlib 0.3.14 (June 27, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jaxlib-v0.3.10...jaxlib-v0.3.14 ).
2022-06-01 20:27:42 -04:00
* x86-64 Mac wheels now require Mac OS 10.14 (Mojave) or newer. Mac OS 10.14
was released in 2018, so this should not be a very onerous requirement.
2022-06-27 09:21:54 -04:00
* The bundled version of NCCL was updated to 2.12.12, fixing some deadlocks.
* The Python flatbuffers package is no longer a dependency of jaxlib.
2022-05-15 12:13:54 -07:00
2022-05-16 12:17:07 -07:00
## jax 0.3.13 (May 16, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.12...jax-v0.3.13 ).
2022-05-15 18:35:15 -07:00
## jax 0.3.12 (May 15, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.11...jax-v0.3.12 ).
* Changes
* Fixes [#10717 ](https://github.com/google/jax/issues/10717 ).
2022-05-15 12:13:54 -07:00
## jax 0.3.11 (May 15, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.10...jax-v0.3.11 ).
2022-05-11 11:45:28 -07:00
* Changes
* {func}`jax.lax.eigh` now accepts an optional `sort_eigenvalues` argument
that allows users to opt out of eigenvalue sorting on TPU.
2022-05-12 07:15:55 -07:00
* Deprecations
2022-05-11 13:06:19 -07:00
* Non-array arguments to functions in {mod}`jax.lax.linalg` are now marked
keyword-only. As a backward-compatibility step passing keyword-only
arguments positionally yields a warning, but in a future JAX release passing
keyword-only arguments positionally will fail.
However, most users should prefer to use {mod}`jax.numpy.linalg` instead.
2022-05-12 07:15:55 -07:00
* {func}`jax.scipy.linalg.polar_unitary` , which was a JAX extension to the
scipy API, is deprecated. Use {func}`jax.scipy.linalg.polar` instead.
2022-04-12 18:12:35 -07:00
2022-05-03 14:52:12 -07:00
## jax 0.3.10 (May 3, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.9...jax-v0.3.10 ).
## jaxlib 0.3.10 (May 3, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jaxlib-v0.3.7...jaxlib-v0.3.10 ).
* Changes
* [TF commit ](https://github.com/tensorflow/tensorflow/commit/207d50d253e11c3a3430a700af478a1d524a779a )
fixes an issue in the MHLO canonicalizer that caused constant folding to
take a long time or crash for certain programs.
2022-04-29 20:15:52 -07:00
2022-05-02 19:50:38 -07:00
## jax 0.3.9 (May 2, 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.8...jax-v0.3.9 ).
* Changes
* Added support for fully asynchronous checkpointing for GlobalDeviceArray.
2022-04-29 20:15:52 -07:00
## jax 0.3.8 (April 29 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.3.7...jax-v0.3.8 ).
2022-04-18 13:37:16 -07:00
* Changes
2022-04-29 16:42:08 -07:00
* {func}`jax.numpy.linalg.svd` on TPUs uses a qdwh-svd solver.
* {func}`jax.numpy.linalg.cond` on TPUs now accepts complex input.
* {func}`jax.numpy.linalg.pinv` on TPUs now accepts complex input.
* {func}`jax.numpy.linalg.matrix_rank` on TPUs now accepts complex input.
2022-04-20 22:38:50 +08:00
* {func}`jax.scipy.cluster.vq.vq` has been added.
2022-04-18 13:37:16 -07:00
* `jax.experimental.maps.mesh` has been deleted.
Please use `jax.experimental.maps.Mesh` . Please see https://jax.readthedocs.io/en/latest/_autosummary/jax.experimental.maps.Mesh.html#jax .experimental.maps.Mesh
for more information.
2022-04-26 11:26:56 -07:00
* {func}`jax.scipy.linalg.qr` now returns a length-1 tuple rather than the raw array when
`mode='r'` , in order to match the behavior of `scipy.linalg.qr` ({jax-issue}`#10452` )
2022-04-19 16:05:29 -04:00
* {func}`jax.numpy.take_along_axis` now takes an optional `mode` parameter
2022-04-21 08:51:35 -07:00
that specifies the behavior of out-of-bounds indexing. By default,
invalid values (e.g., NaN) will be returned for out-of-bounds indices. In
previous versions of JAX, invalid indices were clamped into range. The
previous behavior can be restored by passing `mode="clip"` .
2022-04-28 06:01:22 -07:00
* {func}`jax.numpy.take` now defaults to `mode="fill"` , which returns
invalid values (e.g., NaN) for out-of-bounds indices.
* Scatter operations, such as `x.at[...].set(...)` , now have `"drop"` semantics.
2022-04-27 12:26:17 -07:00
This has no effect on the scatter operation itself, but it means that when
differentiated the gradient of a scatter will yield zero cotangents for
out-of-bounds indices. Previously out-of-bounds indices were clamped into
range for the gradient, which was not mathematically correct.
2022-04-20 10:04:37 -07:00
* {func}`jax.numpy.take_along_axis` now raises a `TypeError` if its indices
are not of an integer type, matching the behavior of
{func}`numpy.take_along_axis` . Previously non-integer indices were silently
cast to integers.
2022-04-21 13:15:03 +08:00
* {func}`jax.numpy.ravel_multi_index` now raises a `TypeError` if its `dims` argument
is not of an integer type, matching the behavior of
{func}`numpy.ravel_multi_index` . Previously non-integer `dims` was silently
cast to integers.
* {func}`jax.numpy.split` now raises a `TypeError` if its `axis` argument
is not of an integer type, matching the behavior of
{func}`numpy.split` . Previously non-integer `axis` was silently
cast to integers.
* {func}`jax.numpy.indices` now raises a `TypeError` if its dimensions
are not of an integer type, matching the behavior of
{func}`numpy.indices` . Previously non-integer dimensions were silently
cast to integers.
* {func}`jax.numpy.diag` now raises a `TypeError` if its `k` argument
is not of an integer type, matching the behavior of
{func}`numpy.diag` . Previously non-integer `k` was silently
cast to integers.
2022-04-29 14:20:50 -04:00
* Added {func}`jax.random.orthogonal` .
2022-04-21 12:12:40 -07:00
* Deprecations
* Many functions and objects available in {mod}`jax.test_util` are now deprecated and will raise a
warning on import. This includes `cases_from_list` , `check_close` , `check_eq` , `device_under_test` ,
`format_shape_dtype_string` , `rand_uniform` , `skip_on_devices` , `with_config` , `xla_bridge` , and
`_default_tolerance` ({jax-issue}`#10389` ). These, along with previously-deprecated `JaxTestCase` ,
`JaxTestLoader` , and `BufferDonationTestCase` , will be removed in a future JAX release.
2022-04-21 13:37:56 -07:00
Most of these utilites can be replaced by calls to standard python & numpy testing utilities found
in e.g. {mod}`unittest` , {mod}`absl.testing` , {mod}`numpy.testing` , etc. JAX-specific functionality
such as device checking can be replaced through the use of public APIs such as {func}`jax.devices` .
Many of the deprecated utilities will still exist in {mod}`jax._src.test_util` , but these are not
public APIs and as such may be changed or removed without notice in future releases.
2022-04-13 12:11:10 -07:00
2022-04-15 12:02:05 -04:00
## jax 0.3.7 (April 15, 2022)
2022-03-21 08:33:11 -07:00
* [GitHub
2022-04-15 12:02:05 -04:00
commits](https://github.com/google/jax/compare/jax-v0.3.6...jax-v0.3.7).
* Changes:
* Fixed a performance problem if the indices passed to
{func}`jax.numpy.take_along_axis` were broadcasted ({jax-issue}`#10281` ).
* {func}`jax.scipy.special.expit` and {func}`jax.scipy.special.logit` now
require their arguments to be scalars or JAX arrays. They also now promote
integer arguments to floating point.
2022-04-15 15:20:02 -04:00
* The `DeviceArray.tile()` method is deprecated, because numpy arrays do not have a
`tile()` method. As a replacement for this, use {func}`jax.numpy.tile`
({jax-issue}`#10266` ).
2022-04-07 16:38:09 -04:00
2022-04-18 08:17:14 -04:00
## jaxlib 0.3.7 (April 15, 2022)
2022-04-15 15:20:02 -04:00
* Changes:
* Linux wheels are now built conforming to the `manylinux2014` standard, instead
of `manylinux2010` .
2022-04-07 16:38:09 -04:00
2022-04-12 18:12:35 -07:00
## jax 0.3.6 (April 12, 2022)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.3.5...jax-v0.3.6).
* Changes:
2022-04-15 12:02:05 -04:00
* Upgraded libtpu wheel to a version that fixes a hang when initializing a TPU
pod. Fixes [#10218 ](https://github.com/google/jax/issues/10218 ).
2022-04-14 12:38:18 -07:00
* Deprecations:
2022-04-14 16:12:48 -04:00
* {mod}`jax.experimental.loops` is being deprecated. See {jax-issue}`#10278`
2022-04-21 12:12:40 -07:00
for an alternative API.
2022-04-07 16:38:09 -04:00
## jax 0.3.5 (April 7, 2022)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.3.4...jax-v0.3.5).
2022-03-21 08:33:11 -07:00
* Changes:
* added {func}`jax.random.loggamma` & improved behavior of {func}`jax.random.beta`
2022-03-23 20:39:39 +00:00
and {func}`jax.random.dirichlet` for small parameter values ({jax-issue}`#9906` ).
2022-03-25 15:02:45 -07:00
* the private `lax_numpy` submodule is no longer exposed in the `jax.numpy` namespace ({jax-issue}`#10029` ).
2022-03-29 10:52:47 -07:00
* added array creation routines {func}`jax.numpy.frombuffer` , {func}`jax.numpy.fromfunction` ,
and {func}`jax.numpy.fromstring` ({jax-issue}`#10049` ).
2022-03-30 08:05:34 -07:00
* `DeviceArray.copy()` now returns a `DeviceArray` rather than a `np.ndarray` ({jax-issue}`#10069` )
2022-03-27 12:31:12 +01:00
* added {func}`jax.scipy.linalg.rsf2csf`
2022-04-18 13:48:19 -07:00
* `jax.experimental.sharded_jit` has been deprecated and will be removed soon.
2022-03-23 20:39:39 +00:00
* Deprecations:
* {func}`jax.nn.normalize` is being deprecated. Use {func}`jax.nn.standardize` instead ({jax-issue}`#9899` ).
2022-04-01 14:51:54 -07:00
* {func}`jax.tree_util.tree_multimap` is deprecated. Use {func}`jax.tree_util.tree_map` instead ({jax-issue}`#5746` ).
2022-04-06 23:56:41 +00:00
* `jax.experimental.sharded_jit` is deprecated. Use `pjit` instead.
2022-03-17 15:35:26 -07:00
2022-04-07 16:38:09 -04:00
## jaxlib 0.3.5 (April 7, 2022)
2022-04-04 14:34:00 -07:00
* Bug fixes
* Fixed a bug where double-precision complex-to-real IRFFTs would mutate their
input buffers on GPU ({jax-issue}`#9946` ).
2022-04-06 23:56:41 +00:00
* Fixed incorrect constant-folding of complex scatters ({jax-issue}`#10159` )
2022-03-17 15:35:26 -07:00
2022-03-18 14:16:00 -07:00
## jax 0.3.4 (March 18, 2022)
* [GitHub
2022-03-21 08:33:11 -07:00
commits](https://github.com/google/jax/compare/jax-v0.3.3...jax-v0.3.4).
2022-03-18 14:16:00 -07:00
2022-03-17 15:35:26 -07:00
## jax 0.3.3 (March 17, 2022)
2022-02-10 20:32:20 -08:00
* [GitHub
2022-03-18 14:16:00 -07:00
commits](https://github.com/google/jax/compare/jax-v0.3.2...jax-v0.3.3).
2022-03-16 10:17:42 -07:00
## jax 0.3.2 (March 16, 2022)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.3.1...jax-v0.3.2).
2022-02-17 15:23:44 -05:00
* Changes:
* The functions `jax.ops.index_update` , `jax.ops.index_add` , which were
deprecated in 0.2.22, have been removed. Please use
[the `.at` property on JAX arrays ](https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.ndarray.at.html )
instead, e.g., `x.at[idx].set(y)` .
2022-03-02 08:58:22 -08:00
* Moved `jax.experimental.ann.approx_*_k` into `jax.lax` . These functions are
optimized alternatives to `jax.lax.top_k` .
2022-03-02 15:23:34 -08:00
* {func}`jax.numpy.broadcast_arrays` and {func}`jax.numpy.broadcast_to` now require scalar
or array-like inputs, and will fail if they are passed lists (part of {jax-issue}`#7737` ).
2022-03-16 10:17:42 -07:00
* The standard jax[tpu] install can now be used with Cloud TPU v4 VMs.
2022-03-16 14:25:19 -07:00
* `pjit` now works on CPU (in addition to previous TPU and GPU support).
2022-02-17 15:23:44 -05:00
2022-02-18 16:23:02 -08:00
2022-03-16 10:17:42 -07:00
## jaxlib 0.3.2 (March 16, 2022)
2022-02-18 16:23:02 -08:00
* Changes
2022-03-14 13:45:49 -07:00
* ``XlaComputation.as_hlo_text()` ` now supports printing large constants by
passing boolean flag ``print_large_constants=True` `.
2022-03-08 09:14:14 -08:00
* Deprecations:
* The ``.block_host_until_ready()` ` method on JAX arrays has been deprecated.
Use ``.block_until_ready()` ` instead.
2022-02-18 16:23:02 -08:00
## jax 0.3.1 (Feb 18, 2022)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.3.0...jax-v0.3.1).
2022-02-17 14:58:58 -08:00
* Changes:
* `jax.test_util.JaxTestCase` and `jax.test_util.JaxTestLoader` are now deprecated.
The suggested replacement is to use `parametrized.TestCase` directly. For tests that
rely on custom asserts such as `JaxTestCase.assertAllClose()` , the suggested replacement
is to use standard numpy testing utilities such as {func}`numpy.testing.assert_allclose()` ,
which work directly with JAX arrays ({jax-issue}`#9620` ).
* `jax.test_util.JaxTestCase` now sets `jax_numpy_rank_promotion='raise'` by default
({jax-issue}`#9562` ). To recover the previous behavior, use the new
`jax.test_util.with_config` decorator:
```python
@jtu .with_config(jax_numpy_rank_promotion='allow')
class MyTestCase(jtu.JaxTestCase):
...
```
* Added {func}`jax.scipy.linalg.schur` , {func}`jax.scipy.linalg.sqrtm` ,
{func}`jax.scipy.signal.csd` , {func}`jax.scipy.signal.stft` ,
{func}`jax.scipy.signal.welch` .
2022-02-10 20:32:20 -08:00
2022-02-10 11:58:30 -08:00
## jax 0.3.0 (Feb 10, 2022)
2022-02-01 17:59:49 -08:00
* [GitHub
2022-02-10 20:32:20 -08:00
commits](https://github.com/google/jax/compare/jax-v0.2.28...jax-v0.3.0).
2022-02-01 17:59:49 -08:00
2022-02-10 11:58:30 -08:00
* Changes
* jax version has been bumped to 0.3.0. Please see the [design doc ](https://jax.readthedocs.io/en/latest/design_notes/jax_versioning.html )
for the explanation.
## jaxlib 0.3.0 (Feb 10, 2022)
* Changes
* Bazel 5.0.0 is now required to build jaxlib.
* jaxlib version has been bumped to 0.3.0. Please see the [design doc ](https://jax.readthedocs.io/en/latest/design_notes/jax_versioning.html )
for the explanation.
## jax 0.2.28 (Feb 1, 2022)
2021-12-08 12:07:42 -08:00
* [GitHub
2022-02-10 11:58:30 -08:00
commits](https://github.com/google/jax/compare/jax-v0.2.27...jax-v0.2.28).
2022-01-20 09:49:40 -08:00
* `jax.jit(f).lower(...).compiler_ir()` now defaults to the MHLO dialect if no
`dialect=` is passed.
[JAX] Change signature of .mhlo() method on compiler IR objects to return an ir.Module object instead of its string representation.
It isn't free to pretty-print IR, so it's best to avoid it unless necessary. In addition, by returning an IR object, the user is now free to, say, print it with different options.
For example, one can now write things like:
```
In [1]: import numpy as np, jax, jax.numpy as jnp
In [2]: m = jax.jit(lambda x: x + jnp.array(np.arange(1000))).lower(7.).compiler_ir(dialect='mhlo')
In [3]: m.operation.print(large_elements_limit=10)
module @jit__lambda_.4 {
func public @main(%arg0: tensor<f32>) -> tensor<1000xf32> {
%0 = mhlo.constant opaque<"_", "0xDEADBEEF"> : tensor<1000xi32>
%1 = "mhlo.convert"(%0) : (tensor<1000xi32>) -> tensor<1000xf32>
%2 = "mhlo.broadcast_in_dim"(%arg0) {broadcast_dimensions = dense<> : tensor<0xi64>} : (tensor<f32>) -> tensor<1000xf32>
%3 = mhlo.add %2, %1 : tensor<1000xf32>
return %3 : tensor<1000xf32>
}
}
```
Fixes https://github.com/google/jax/issues/9226
PiperOrigin-RevId: 422855649
2022-01-19 11:01:03 -08:00
* The `jax.jit(f).lower(...).compiler_ir(dialect='mhlo')` now returns an MLIR
`ir.Module` object instead of its string representation.
2022-01-18 11:38:32 -08:00
2022-01-26 22:19:11 +00:00
## jaxlib 0.1.76 (Jan 27, 2022)
2022-01-18 11:38:32 -08:00
* New features
* Includes precompiled SASS for NVidia compute capability 8.0 GPUS
(e.g. A100). Removes precompiled SASS for compute capability 6.1 so as not
to increase the number of compute capabilities: GPUs with compute capability
6.1 can use the 6.0 SASS.
2022-01-26 22:19:11 +00:00
* With jaxlib 0.1.76, JAX uses the MHLO MLIR dialect as its primary target compiler IR
by default.
2022-01-18 11:38:32 -08:00
* Breaking changes
* Support for NumPy 1.18 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to a supported NumPy version.
2022-01-26 22:19:11 +00:00
* Bug fixes
* Fixed a bug where apparently identical pytreedef objects constructed by different routes
do not compare as equal (#9066 ).
* The JAX jit cache requires two static arguments to have identical types for a cache hit (#9311 ).
2022-01-18 11:38:32 -08:00
## jax 0.2.27 (Jan 18 2022)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.26...jax-v0.2.27 ).
2021-12-08 12:07:42 -08:00
2021-11-24 12:58:16 +02:00
* Breaking changes:
2022-01-04 12:11:00 -08:00
* Support for NumPy 1.18 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to a supported NumPy version.
2021-11-24 12:58:16 +02:00
* The host_callback primitives have been simplified to drop the
2021-11-24 12:58:16 +02:00
special autodiff handling for hcb.id_tap and id_print.
From now on, only the primals are tapped. The old behavior can be
obtained (for a limited time) by setting the ``JAX_HOST_CALLBACK_AD_TRANSFORMS` `
environment variable, or the ```--flax_host_callback_ad_transforms` `` flag.
Additionally, added documentation for how to implement the old behavior
using JAX custom AD APIs ({jax-issue}`#8678` ).
2022-01-13 13:03:41 -08:00
* Sorting now matches the behavior of NumPy for ``0.0`` and ``NaN` ` regardless of the
bit representation. In particular, ``0.0`` and ``-0.0` ` are now treated as equivalent,
where previously ``-0.0`` was treated as less than ``0.0``. Additionally all ``NaN` `
representations are now treated as equivalent and sorted to the end of the array.
Previously negative ``NaN`` values were sorted to the front of the array, and ``NaN` `
values with different internal bit representations were not treated as equivalent, and
were sorted according to those bit patterns ({jax-issue}`#9178` ).
2022-01-13 15:54:07 -08:00
* {func}`jax.numpy.unique` now treats ``NaN`` values in the same way as ` np.unique` in
NumPy versions 1.21 and newer: at most one ``NaN` ` value will appear in the uniquified
output ({jax-issue}`9184` ).
2021-11-24 12:58:16 +02:00
* Bug fixes:
* host_callback now supports ad_checkpoint.checkpoint ({jax-issue}`#8907` ).
2021-11-24 12:58:16 +02:00
2021-12-14 13:36:46 -08:00
* New features:
* add `jax.block_until_ready` ({jax-issue}`#8941 )
2021-12-14 17:43:40 -08:00
* Added a new debugging flag/environment variable `JAX_DUMP_IR_TO=/path` .
If set, JAX dumps the MHLO/HLO IR it generates for each computation to a
file under the given path.
2022-01-10 20:57:56 -08:00
* Added `jax.ensure_compile_time_eval` to the public api ({jax-issue}`#7987` ).
2022-01-13 12:28:25 +02:00
* jax2tf now supports a flag jax2tf_associative_scan_reductions to change
the lowering for associative reductions, e.g., jnp.cumsum, to behave
like JAX on CPU and GPU (to use an associative scan). See the jax2tf README
for more details ({jax-issue}`#9189` ).
2021-12-14 13:36:46 -08:00
2021-12-08 11:13:37 -08:00
2021-12-08 12:07:42 -08:00
## jaxlib 0.1.75 (Dec 8, 2021)
2021-12-08 11:13:37 -08:00
* New features:
* Support for python 3.10.
2021-11-17 11:43:19 -05:00
2021-12-08 12:07:42 -08:00
## jax 0.2.26 (Dec 8, 2021)
2021-11-10 13:41:25 -08:00
* [GitHub
2021-12-08 12:07:42 -08:00
commits](https://github.com/google/jax/compare/jax-v0.2.25...jax-v0.2.26).
2021-09-07 18:36:47 +03:00
2021-11-22 13:32:25 -08:00
* Bug fixes:
* Out-of-bounds indices to `jax.ops.segment_sum` will now be handled with
`FILL_OR_DROP` semantics, as documented. This primarily afects the
reverse-mode derivative, where gradients corresponding to out-of-bounds
indices will now be returned as 0. (#8634 ).
2021-09-07 18:36:47 +03:00
* jax2tf will force the converted code to use XLA for the code fragments
under jax.jit, e.g., most jax.numpy functions ({jax-issue}`#7839` ).
2021-11-10 13:41:25 -08:00
2021-11-17 11:43:19 -05:00
## jaxlib 0.1.74 (Nov 17, 2021)
2021-11-16 17:36:28 -05:00
* Enabled peer-to-peer copies between GPUs. Previously, GPU copies were bounced via
the host, which is usually slower.
* Added experimental MLIR Python bindings for use by JAX.
2021-10-18 19:55:51 -07:00
2021-11-10 13:41:25 -08:00
## jax 0.2.25 (Nov 10, 2021)
2021-10-12 18:21:11 -07:00
* [GitHub
2021-11-10 13:41:25 -08:00
commits](https://github.com/google/jax/compare/jax-v0.2.24...jax-v0.2.25).
2021-10-19 08:04:59 -07:00
2021-10-25 15:34:57 -07:00
* New features:
* (Experimental) `jax.distributed.initialize` exposes multi-host GPU backend.
2021-11-02 11:39:41 -07:00
* `jax.random.permutation` supports new `independent` keyword argument
({jax-issue}`#8430` )
2021-10-19 17:30:16 -07:00
* Breaking changes
* Moved `jax.experimental.stax` to `jax.example_libraries.stax`
* Moved `jax.experimental.optimizers` to `jax.example_libraries.optimizers`
2021-10-29 14:44:27 -07:00
* New features:
* Added `jax.lax.linalg.qdwh` .
2021-10-19 17:30:16 -07:00
2021-10-19 08:04:59 -07:00
## jax 0.2.24 (Oct 19, 2021)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.2.22...jax-v0.2.24).
2021-10-12 18:21:11 -07:00
2021-10-13 17:11:50 -06:00
* New features:
* `jax.random.choice` and `jax.random.permutation` now support
multidimensional arrays and an optional `axis` argument ({jax-issue}`#8158` )
2021-10-14 11:54:28 -07:00
* Breaking changes:
* `jax.numpy.take` and `jax.numpy.take_along_axis` now require array-like inputs
(see {jax-issue}`#7737` )
2021-10-13 17:11:50 -06:00
2021-10-18 19:55:51 -07:00
## jaxlib 0.1.73 (Oct 18, 2021)
* Multiple cuDNN versions are now supported for jaxlib GPU `cuda11` wheels.
* cuDNN 8.2 or newer. We recommend using the cuDNN 8.2 wheel if your cuDNN
installation is new enough, since it supports additional functionality.
* cuDNN 8.0.5 or newer.
* Breaking changes:
* The install commands for GPU jaxlib are as follows:
```bash
pip install --upgrade pip
# Installs the wheel compatible with CUDA 11 and cuDNN 8.2 or newer.
2021-10-19 06:24:27 -07:00
pip install --upgrade "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_releases.html
2021-10-18 19:55:51 -07:00
# Installs the wheel compatible with Cuda 11 and cudnn 8.2 or newer.
2021-10-19 06:24:27 -07:00
pip install jax[cuda11_cudnn82] -f https://storage.googleapis.com/jax-releases/jax_releases.html
2021-10-18 19:55:51 -07:00
# Installs the wheel compatible with Cuda 11 and cudnn 8.0.5 or newer.
2021-10-19 06:24:27 -07:00
pip install jax[cuda11_cudnn805] -f https://storage.googleapis.com/jax-releases/jax_releases.html
2021-10-18 19:55:51 -07:00
```
2021-10-12 18:43:40 -07:00
2021-10-12 18:21:11 -07:00
## jax 0.2.22 (Oct 12, 2021)
2021-09-15 12:10:30 -07:00
* [GitHub
2021-10-12 18:21:11 -07:00
commits](https://github.com/google/jax/compare/jax-v0.2.21...jax-v0.2.22).
2021-07-19 13:11:38 -04:00
* Breaking Changes
* Static arguments to `jax.pmap` must now be hashable.
Unhashable static arguments have long been disallowed on `jax.jit` , but they
were still permitted on `jax.pmap` ; `jax.pmap` compared unhashable static
arguments using object identity.
This behavior is a footgun, since comparing arguments using
object identity leads to recompilation each time the object identity
changes. Instead, we now ban unhashable arguments: if a user of `jax.pmap`
wants to compare static arguments by object identity, they can define
`__hash__` and `__eq__` methods on their objects that do that, or wrap their
objects in an object that has those operations with object identity
semantics. Another option is to use `functools.partial` to encapsulate the
unhashable static arguments into the function object.
* `jax.util.partial` was an accidental export that has now been removed. Use
`functools.partial` from the Python standard library instead.
2021-10-06 09:31:42 -04:00
* Deprecations
* The functions `jax.ops.index_update` , `jax.ops.index_add` etc. are
deprecated and will be removed in a future JAX release. Please use
[the `.at` property on JAX arrays ](https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.ndarray.at.html )
instead, e.g., `x.at[idx].set(y)` . For now, these functions produce a
`DeprecationWarning` .
2021-10-06 10:07:41 -07:00
* New features:
2021-10-06 09:31:42 -04:00
* An optimized C++ code-path improving the dispatch time for `pmap` is now the
default when using jaxlib 0.1.72 or newer. The feature can be disabled using
the `--experimental_cpp_pmap` flag (or `JAX_CPP_PMAP` environment variable).
2021-10-06 16:28:36 -07:00
* `jax.numpy.unique` now supports an optional `fill_value` argument ({jax-issue}`#8121` )
2021-09-23 10:58:55 -07:00
2021-10-12 17:37:29 -07:00
## jaxlib 0.1.72 (Oct 12, 2021)
2021-10-12 18:43:40 -07:00
* Breaking changes:
* Support for CUDA 10.2 and CUDA 10.1 has been dropped. Jaxlib now supports
CUDA 11.1+.
2021-10-13 11:11:24 -04:00
* Bug fixes:
* Fixes https://github.com/google/jax/issues/7461, which caused wrong
outputs on all platforms due to incorrect buffer aliasing inside the XLA
compiler.
2021-10-12 17:37:29 -07:00
2021-09-23 10:58:55 -07:00
## jax 0.2.21 (Sept 23, 2021)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.2.20...jax-v0.2.21).
2021-09-15 12:10:30 -07:00
* Breaking Changes
2021-09-16 16:29:06 -04:00
* `jax.api` has been removed. Functions that were available as `jax.api.*`
were aliases for functions in `jax.*` ; please use the functions in
`jax.*` instead.
2021-07-19 13:11:38 -04:00
* `jax.partial` , and `jax.lax.partial` were accidental exports that have now
been removed. Use `functools.partial` from the Python standard library
instead.
2021-09-15 12:50:44 -07:00
* Boolean scalar indices now raise a `TypeError` ; previously this silently
returned wrong results ({jax-issue}`#7925` ).
2021-09-16 14:10:08 -07:00
* Many more `jax.numpy` functions now require array-like inputs, and will error
if passed a list ({jax-issue}`#7747` {jax-issue}`#7802` {jax-issue}`#7907` ).
See {jax-issue}`#7737` for a discussion of the rationale behind this change.
2021-09-21 09:05:56 -07:00
* When inside a transformation such as `jax.jit` , `jax.numpy.array` always
stages the array it produces into the traced computation. Previously
`jax.numpy.array` would sometimes produce a on-device array, even under
a `jax.jit` decorator. This change may break code that used JAX arrays to
perform shape or index computations that must be known statically; the
workaround is to perform such computations using classic NumPy arrays
instead.
2021-09-27 11:38:36 -07:00
* `jnp.ndarray` is now a true base-class for JAX arrays. In particular, this
means that for a standard numpy array `x` , `isinstance(x, jnp.ndarray)` will
now return `False` ({jax-issue}`7927` ).
2021-09-16 14:10:08 -07:00
* New features:
* Added {func}`jax.numpy.insert` implementation ({jax-issue}`#7936` ).
2021-09-15 12:10:30 -07:00
2021-09-02 15:38:47 -07:00
## jax 0.2.20 (Sept 2, 2021)
2021-08-12 21:17:53 -07:00
* [GitHub
2021-09-15 12:10:30 -07:00
commits](https://github.com/google/jax/compare/jax-v0.2.19...jax-v0.2.20).
2021-08-26 11:12:16 -07:00
* Breaking Changes
* `jnp.poly*` functions now require array-like inputs ({jax-issue}`#7732` )
* `jnp.unique` and other set-like operations now require array-like inputs
({jax-issue}`#7662` )
2021-08-12 21:17:53 -07:00
2021-09-01 10:43:20 -07:00
## jaxlib 0.1.71 (Sep 1, 2021)
* Breaking changes:
* Support for CUDA 11.0 and CUDA 10.1 has been dropped. Jaxlib now supports
CUDA 10.2 and CUDA 11.1+.
2021-08-12 21:17:53 -07:00
## jax 0.2.19 (Aug 12, 2021)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.2.18...jax-v0.2.19).
2021-07-29 09:18:01 -04:00
* Breaking changes:
* Support for NumPy 1.17 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to a supported NumPy version.
2021-08-10 06:48:55 -07:00
* The `jit` decorator has been added around the implementation of a number of
operators on JAX arrays. This speeds up dispatch times for common
operators such as `+` .
This change should largely be transparent to most users. However, there is
one known behavioral change, which is that large integer constants may now
produce an error when passed directly to a JAX operator
(e.g., `x + 2**40` ). The workaround is to cast the constant to an
explicit type (e.g., `np.float64(2**40)` ).
[jax2tf] Expand shape polymorphism support to use dimension polynomials as values.
The goal of this change is to support shape polymorphism for operations
such as average (which needs to divide by the size of a dimension) or
indexing (which needs to normalize indices by comparing them with 0 and
adding dimension size for negative indices). In both of these cases
the size of a dimenion needs to be used as a value in the array
computation. In general, the size of a dimension is used only to
customize primitives.
This change introduces `core.dim_as_value` which must be used on
a dimension size before using it as a value in the array computation.
E.g.,
```
def average(x):
return jnp.sum(x, axis=0) / core.dim_as_value(x.shape[0])
```
This function is the identity function if the dimension size is
constant, otherwise it uses a new primitive `shape_poly.dim_as_value_p`.
Note that this does not change fundamentally the flavor of shape
polymorphism supported in jax2tf: intermediate shapes and their values
may depend on the input shapes, but never does a shape depend on the
input values. In fact, one could have expressed the `dim_as_value`
already:
```
def dim_as_value(d):
jnp.sum(jnp.broadcast_to(jnp.array(1), shape=(d,)))
```
We were able to suppot `jnp.mean`, `jnp.average`, `jnp.take`,
`lax.dynamic_slice`, `lax.dynamic_update_slice` by using
`core.dim_as_value` internally, but to fully roll-up the solution
we need to make `core.dim_as_value` a public API and teach the
users how to use it when they want to use shape polymorphism.
Alternatively, perhaps there is a way to automatically convert
dimension polynomials to values when passed to the lax primitives.
2021-07-16 20:01:22 +03:00
* New features:
* Improved the support for shape polymorphism in jax2tf for operations that
need to use a dimension size in array computation, e.g., `jnp.mean` .
({jax-issue}`#7317` )
2021-08-12 21:17:53 -07:00
* Bug fixes:
* Some leaked trace errors from the previous release ({jax-issue}`#7613` )
2021-08-09 16:37:15 -07:00
## jaxlib 0.1.70 (Aug 9, 2021)
2021-07-21 11:56:24 -07:00
* Breaking changes:
* Support for Python 3.6 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to a supported Python version.
2021-07-29 09:18:01 -04:00
* Support for NumPy 1.17 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to a supported NumPy version.
[jax2tf] Expand shape polymorphism support to use dimension polynomials as values.
The goal of this change is to support shape polymorphism for operations
such as average (which needs to divide by the size of a dimension) or
indexing (which needs to normalize indices by comparing them with 0 and
adding dimension size for negative indices). In both of these cases
the size of a dimenion needs to be used as a value in the array
computation. In general, the size of a dimension is used only to
customize primitives.
This change introduces `core.dim_as_value` which must be used on
a dimension size before using it as a value in the array computation.
E.g.,
```
def average(x):
return jnp.sum(x, axis=0) / core.dim_as_value(x.shape[0])
```
This function is the identity function if the dimension size is
constant, otherwise it uses a new primitive `shape_poly.dim_as_value_p`.
Note that this does not change fundamentally the flavor of shape
polymorphism supported in jax2tf: intermediate shapes and their values
may depend on the input shapes, but never does a shape depend on the
input values. In fact, one could have expressed the `dim_as_value`
already:
```
def dim_as_value(d):
jnp.sum(jnp.broadcast_to(jnp.array(1), shape=(d,)))
```
We were able to suppot `jnp.mean`, `jnp.average`, `jnp.take`,
`lax.dynamic_slice`, `lax.dynamic_update_slice` by using
`core.dim_as_value` internally, but to fully roll-up the solution
we need to make `core.dim_as_value` a public API and teach the
users how to use it when they want to use shape polymorphism.
Alternatively, perhaps there is a way to automatically convert
dimension polynomials to values when passed to the lax primitives.
2021-07-16 20:01:22 +03:00
* The host_callback mechanism now uses one thread per local device for
2021-07-21 11:56:24 -07:00
making the calls to the Python callbacks. Previously there was a single
thread for all devices. This means that the callbacks may now be called
interleaved. The callbacks corresponding to one device will still be
called in sequence.
## jax 0.2.18 (July 21 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.17...jax-v0.2.18 ).
2021-07-09 15:19:24 -04:00
2021-07-15 09:50:30 -04:00
* Breaking changes:
* Support for Python 3.6 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
Please upgrade to a supported Python version.
2021-07-15 16:39:18 -04:00
* The minimum jaxlib version is now 0.1.69.
* The `backend` argument to {py:func}`jax.dlpack.from_dlpack` has been
removed.
2021-07-15 09:50:30 -04:00
2021-07-20 13:12:13 -04:00
* New features:
* Added a polar decomposition ({py:func}`jax.scipy.linalg.polar` ).
2021-07-12 01:11:17 -07:00
* Bug fixes:
* Tightened the checks for lax.argmin and lax.argmax to ensure they are
2021-08-02 17:57:09 -07:00
not used with an invalid `axis` value, or with an empty reduction dimension.
2021-07-12 01:11:17 -07:00
({jax-issue}`#7196` )
2021-07-15 16:39:18 -04:00
2021-07-09 21:16:48 -07:00
## jaxlib 0.1.69 (July 9 2021)
* Fix bugs in TFRT CPU backend that results in incorrect results.
2021-07-09 15:19:24 -04:00
## jax 0.2.17 (July 9 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.16...jax-v0.2.17 ).
* Bug fixes:
* Default to the older "stream_executor" CPU runtime for jaxlib < = 0.1.68
to work around #7229 , which caused wrong outputs on CPU due to a concurrency
problem.
2021-07-02 10:42:29 -07:00
* New features:
* New SciPy function {py:func}`jax.scipy.special.sph_harm` .
AWN-enabled reduction over named axes in reverse-mode AD
Previously, reverse-mode AD operators inside JAX maps always meant "compute
a gradient (or VJP, etc.) for each axis index in the map". For instance,
`vmap(grad(f))` is the standard JAX spelling of the per-example gradient of `f`.
In batching tracer terms, this "elementwise" behavior means that, if any inputs
to a function being transposed are mapped, the cotangents of all inputs, even
unmapped ones, would also be mapped. But a user might want them to be unmapped
(if, for instance, they're interested in a total gradient rather than a
per-example gradient). They could always reduce (`psum`) the cotangents
afterwards, but computing mapped cotangents in the first place would likely be
an unacceptable waste of memory and can't necessarily be optimized away.
If we want to fuse these reductions into reverse-mode autodiff itself, we need
the backward_pass logic and/or transpose rules to know about whether primal
values are mapped or unmapped. This is made possible by avals-with-names,
which encodes that information in the avals of the primal jaxpr.
Putting things together, **this change adds an option to reverse-mode AD APIs
that indicates which named axes should be reduced over in the backward pass in
situations where they were broadcasted over in the forward pass**. All other
named axes will be treated in the current elementwise way. This has the effect
of making APIs like `grad` behave akin to collectives like `psum`: they act
collectively over axes that are named explicitly, and elementwise otherwise.
Since avals-with-names is currently enabled only in `xmap`, this behavior is
only available in that context for now. It's also missing some optimizations:
- reductions aren't fused into any first-order primitives (e.g. a `pdot`
should have a named contracting axis added rather than being followed by a
`psum`; this can be implemented by putting these primitives into
`reducing_transposes`)
- reductions are performed eagerly, even over axes that are mapped to
hardware resources (the optimal thing to do would be to reduce eagerly
over any vectorized axis component while delaying the reduction over any
hardware-mapped component until the end of the overall backward pass; this
would require a way to represent these partially-reduced values)
PiperOrigin-RevId: 383685336
2021-07-08 12:05:56 -07:00
* Reverse-mode autodiff functions ({func}`jax.grad` ,
{func}`jax.value_and_grad` , {func}`jax.vjp` , and
{func}`jax.linear_transpose` ) support a parameter that indicates which named
axes should be summed over in the backward pass if they were broadcasted
over in the forward pass. This enables use of these APIs in a
non-per-example way inside maps (initially only
{func}`jax.experimental.maps.xmap` ) ({jax-issue}`#6950` ).
2021-06-23 14:42:36 -07:00
2021-06-28 13:52:19 -07:00
2021-06-23 14:42:36 -07:00
## jax 0.2.16 (June 23 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.15...jax-v0.2.16 ).
2021-06-23 11:55:40 -07:00
## jax 0.2.15 (June 23 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.14...jax-v0.2.15 ).
2021-06-10 17:01:22 +02:00
* New features:
2021-06-28 13:52:19 -07:00
* [#7042 ](https://github.com/google/jax/pull/7042 ) Turned on TFRT CPU backend
with significant dispatch performance improvements on CPU.
2021-06-10 12:42:40 +02:00
* The {func}`jax2tf.convert` supports inequalities and min/max for booleans
({jax-issue}`#6956` ).
2021-07-02 10:42:29 -07:00
* New SciPy function {py:func}`jax.scipy.special.lpmn_values` .
2021-06-10 13:15:53 -07:00
2021-06-10 17:01:22 +02:00
* Breaking changes:
2021-06-10 12:12:13 -04:00
* Support for NumPy 1.16 has been dropped, per the
[deprecation policy ](https://jax.readthedocs.io/en/latest/deprecation.html ).
2021-06-10 13:15:53 -07:00
2021-06-10 17:01:22 +02:00
* Bug fixes:
* Fixed bug that prevented round-tripping from JAX to TF and back:
`jax2tf.call_tf(jax2tf.convert)` ({jax-issue}`#6947` ).
2021-06-23 11:55:40 -07:00
2021-06-28 13:52:19 -07:00
## jaxlib 0.1.68 (June 23 2021)
* Bug fixes:
* Fixed bug in TFRT CPU backend that gets nans when transfer TPU buffer to
CPU.
2021-06-10 13:15:53 -07:00
## jax 0.2.14 (June 10 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.13...jax-v0.2.14 ).
2021-05-11 11:11:37 +03:00
* New features:
2021-06-01 22:50:12 -07:00
* The {func}`jax2tf.convert` now has support for `pjit` and `sharded_jit` .
2021-06-02 15:22:50 -04:00
* A new configuration option JAX_TRACEBACK_FILTERING controls how JAX filters
tracebacks.
* A new traceback filtering mode using `__tracebackhide__` is now enabled by
default in sufficiently recent versions of IPython.
2021-06-04 11:02:50 +03:00
* The {func}`jax2tf.convert` supports shape polymorphism even when the
unknown dimensions are used in arithmetic operations, e.g., `jnp.reshape(-1)`
2021-06-10 13:15:53 -07:00
({jax-issue}`#6827` ).
[jax2tf] Add support for generating HLO OpMetadata in the TF graph
The goal is to ensure that the HLO that
jax2tf->TF/XLA generates has the same metadata as what JAX generates.
This includes `op_type`, `op_name`, and source information, which are
used for debugging and profiling.
In order to ensure that this metadata is carried from the JAX tracing
time to TF/XLA, we save the metadata in custom TF op attributes. These
attributes are automatically preserved through SavedModel. This relies
on a separate change in TF/XLA to look for these custom attributes
and override its default.
For the source information, we use pretty much the same code that
xla.py uses. HLO OpMetadata has room for only one source location.
JAX (xla.py) picks the top-most user frame, which is obtained by
filtering out the stack frames in the JAX source tree. When used
with jax2tf we also need to filter out stack frames in the
TensorFlow source tree.
The hardest part is to generate the `op_name`, which is a hierarchical
name with components separated by '/', e.g., `jax2tf(top_func)/while/cond/le`.
We carry the current `name_stack` in thread-local state. Unfortunately, there
is no easy way to share the exact code that achieves this in xla.py. At the
same time it is not crucial that we have exactly identical name stacks as in
JAX.
I attempted to also carry this state in the JAX `MainTrace`, but could not
fully control the name stack. E.g., when calling a jitted-function we
have to reuse the current `MainTrace` although we want to push an element
on the name stack.
For now this option is not yet enabled until we make the necessary
changes in TensorFlow.
2021-05-25 13:33:35 +02:00
* The {func}`jax2tf.convert` generates custom attributes with location information
in TF ops. The code that XLA generates after jax2tf
has the same location information as JAX/XLA.
2021-07-02 10:42:29 -07:00
* New SciPy function {py:func}`jax.scipy.special.lpmn` .
[jax2tf] Add support for generating HLO OpMetadata in the TF graph
The goal is to ensure that the HLO that
jax2tf->TF/XLA generates has the same metadata as what JAX generates.
This includes `op_type`, `op_name`, and source information, which are
used for debugging and profiling.
In order to ensure that this metadata is carried from the JAX tracing
time to TF/XLA, we save the metadata in custom TF op attributes. These
attributes are automatically preserved through SavedModel. This relies
on a separate change in TF/XLA to look for these custom attributes
and override its default.
For the source information, we use pretty much the same code that
xla.py uses. HLO OpMetadata has room for only one source location.
JAX (xla.py) picks the top-most user frame, which is obtained by
filtering out the stack frames in the JAX source tree. When used
with jax2tf we also need to filter out stack frames in the
TensorFlow source tree.
The hardest part is to generate the `op_name`, which is a hierarchical
name with components separated by '/', e.g., `jax2tf(top_func)/while/cond/le`.
We carry the current `name_stack` in thread-local state. Unfortunately, there
is no easy way to share the exact code that achieves this in xla.py. At the
same time it is not crucial that we have exactly identical name stacks as in
JAX.
I attempted to also carry this state in the JAX `MainTrace`, but could not
fully control the name stack. E.g., when calling a jitted-function we
have to reuse the current `MainTrace` although we want to push an element
on the name stack.
For now this option is not yet enabled until we make the necessary
changes in TensorFlow.
2021-05-25 13:33:35 +02:00
2021-05-11 11:11:37 +03:00
* Bug fixes:
2021-06-02 21:03:16 +03:00
* The {func}`jax2tf.convert` now ensures that it uses the same typing rules
for Python scalars and for choosing 32-bit vs. 64-bit computations
as JAX ({jax-issue}`#6883` ).
2021-05-11 11:11:37 +03:00
* The {func}`jax2tf.convert` now scopes the `enable_xla` conversion parameter
properly to apply only during the just-in-time conversion
({jax-issue}`#6720` ).
2021-05-12 02:29:51 -07:00
* The {func}`jax2tf.convert` now converts `lax.dot_general` using the
`XlaDot` TensorFlow op, for better fidelity w.r.t. JAX numerical precision
({jax-issue}`#6717` ).
2021-06-04 11:02:50 +03:00
* The {func}`jax2tf.convert` now has support for inequality comparisons and
2021-06-10 13:15:53 -07:00
min/max for complex numbers ({jax-issue}`#6892` ).
2021-05-12 02:29:51 -07:00
2021-06-28 13:52:19 -07:00
## jaxlib 0.1.67 (May 17 2021)
2021-05-11 16:49:32 -07:00
## jaxlib 0.1.66 (May 11 2021)
2021-05-12 02:29:51 -07:00
2021-05-03 16:32:00 -07:00
* New features:
* CUDA 11.1 wheels are now supported on all CUDA 11 versions 11.1 or higher.
NVidia now promises compatibility between CUDA minor releases starting with
CUDA 11.1. This means that JAX can release a single CUDA 11.1 wheel that
is compatible with CUDA 11.2 and 11.3.
There is no longer a separate jaxlib release for CUDA 11.2 (or higher); use
the CUDA 11.1 wheel for those versions (cuda111).
2021-05-04 09:28:04 -04:00
* Jaxlib now bundles `libdevice.10.bc` in CUDA wheels. There should be no need
to point JAX to a CUDA installation to find this file.
* Added automatic support for static keyword arguments to the {func}`jit`
implementation.
* Added support for pretransformation exception traces.
* Initial support for pruning unused arguments from {func}`jit` -transformed
computations.
Pruning is still a work in progress.
* Improved the string representation of {class}`PyTreeDef` objects.
* Added support for XLA's variadic ReduceWindow.
* Bug fixes:
* Fixed a bug in the remote cloud TPU support when large numbers of arguments
are passed to a computation.
* Fix a bug that meant that JAX garbage collection was not triggered by
{func}`jit` transformed functions.
2021-05-03 16:32:00 -07:00
## jax 0.2.13 (May 3 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.12...jax-v0.2.13 ).
2021-04-09 07:10:02 -07:00
* New features:
* When combined with jaxlib 0.1.66, {func}`jax.jit` now supports static
keyword arguments. A new `static_argnames` option has been added to specify
keyword arguments as static.
2021-04-20 09:18:26 -07:00
* {func}`jax.nonzero` has a new optional `size` argument that allows it to
2021-04-21 16:00:14 -07:00
be used within `jit` ({jax-issue}`#6501` )
* {func}`jax.numpy.unique` now supports the `axis` argument ({jax-issue}`#6532` ).
2021-04-28 12:22:32 +03:00
* {func}`jax.experimental.host_callback.call` now supports `pjit.pjit` ({jax-issue}`#6569` ).
2021-05-03 11:27:07 -04:00
* Added {func}`jax.scipy.linalg.eigh_tridiagonal` that computes the
eigenvalues of a tridiagonal matrix. Only eigenvalues are supported at
present.
2021-05-04 09:28:04 -04:00
* The order of the filtered and unfiltered stack traces in exceptions has been
changed. The traceback attached to an exception thrown from JAX-transformed
code is now filtered, with an `UnfilteredStackTrace` exception
containing the original trace as the `__cause__` of the filtered exception.
Filtered stack traces now also work with Python 3.6.
* If an exception is thrown by code that has been transformed by reverse-mode
automatic differentiation, JAX now attempts to attach as a `__cause__` of
the exception a `JaxStackTraceBeforeTransformation` object that contains the
stack trace that created the original operation in the forward pass.
Requires jaxlib 0.1.66.
2021-04-09 07:10:02 -07:00
* Breaking changes:
2021-04-20 17:56:41 -07:00
* The following function names have changed. There are still aliases, so this
should not break existing code, but the aliases will eventually be removed
so please change your code.
* `host_id` --> {func}`~jax.process_index`
* `host_count` --> {func}`~jax.process_count`
* `host_ids` --> `range(jax.process_count())`
* Similarly, the argument to {func}`~jax.local_devices` has been renamed from
`host_id` to `process_index` .
2021-04-08 10:32:35 -04:00
* Arguments to {func}`jax.jit` other than the function are now marked as
keyword-only. This change is to prevent accidental breakage when arguments
are added to `jit` .
2021-04-09 07:10:02 -07:00
* Bug fixes:
* The {func}`jax2tf.convert` now works in presence of gradients for functions
with integer inputs ({jax-issue}`#6360` ).
2021-04-28 12:22:32 +03:00
* Fixed assertion failure in {func}`jax2tf.call_tf` when used with captured
`tf.Variable` ({jax-issue}`#6572` ).
2021-04-01 10:11:52 -07:00
2021-04-07 17:51:20 -07:00
## jaxlib 0.1.65 (April 7 2021)
2021-04-01 10:11:52 -07:00
## jax 0.2.12 (April 1 2021)
2021-04-07 11:24:31 +03:00
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.11...v0.2.12 ).
2021-03-17 17:33:25 +00:00
* New features
* New profiling APIs: {func}`jax.profiler.start_trace` ,
{func}`jax.profiler.stop_trace` , and {func}`jax.profiler.trace`
2021-03-23 10:31:02 -04:00
* {func}`jax.lax.reduce` is now differentiable.
2021-03-24 13:49:05 -04:00
* Breaking changes:
* The minimum jaxlib version is now 0.1.64.
2021-03-17 17:33:25 +00:00
* Some profiler APIs names have been changed. There are still aliases, so this
should not break existing code, but the aliases will eventually be removed
so please change your code.
* `TraceContext` --> {func}`~jax.profiler.TraceAnnotation`
* `StepTraceContext` --> {func}`~jax.profiler.StepTraceAnnotation`
* `trace_function` --> {func}`~jax.profiler.annotate_function`
2022-06-24 12:18:11 -07:00
* Omnistaging can no longer be disabled. See [omnistaging ](https://github.com/google/jax/blob/main/docs/design_notes/omnistaging.md )
2021-03-25 16:44:58 -07:00
for more information.
2021-03-29 09:26:19 -07:00
* Python integers larger than the maximum `int64` value will now lead to an overflow
in all cases, rather than being silently converted to `uint64` in some cases ({jax-issue}`#6047` ).
2021-03-30 10:05:03 -07:00
* Outside X64 mode, Python integers outside the range representable by `int32` will now lead to an
`OverflowError` rather than having their value silently truncated.
2021-03-29 17:21:56 +03:00
* Bug fixes:
* `host_callback` now supports empty arrays in arguments and results ({jax-issue}`#6262` ).
2021-04-01 09:29:22 -07:00
* {func}`jax.random.randint` clips rather than wraps of out-of-bounds limits, and can now generate
integers in the full range of the specified dtype ({jax-issue}`#5868` )
2021-03-24 13:49:05 -04:00
2021-03-23 19:33:04 -07:00
## jax 0.2.11 (March 23 2021)
* [GitHub
commits](https://github.com/google/jax/compare/jax-v0.2.10...jax-v0.2.11).
2021-03-05 11:07:50 -08:00
* New features:
2021-03-23 19:33:04 -07:00
* [#6112 ](https://github.com/google/jax/pull/6112 ) added context managers:
`jax.enable_checks` , `jax.check_tracer_leaks` , `jax.debug_nans` ,
`jax.debug_infs` , `jax.log_compiles` .
* [#6085 ](https://github.com/google/jax/pull/6085 ) added `jnp.delete`
2021-03-23 10:31:02 -04:00
2021-03-05 11:07:50 -08:00
* Bug fixes:
2021-03-23 19:33:04 -07:00
* [#6136 ](https://github.com/google/jax/pull/6136 ) generalized
`jax.flatten_util.ravel_pytree` to handle integer dtypes.
* [#6129 ](https://github.com/google/jax/issues/6129 ) fixed a bug with handling
some constants like `enum.IntEnums`
* [#6145 ](https://github.com/google/jax/pull/6145 ) fixed batching issues with
incomplete beta functions
* [#6014 ](https://github.com/google/jax/pull/6014 ) fixed H2D transfers during
tracing
* [#6165 ](https://github.com/google/jax/pull/6165 ) avoids OverflowErrors when
converting some large Python integers to floats
2021-03-05 11:07:50 -08:00
* Breaking changes:
2021-03-16 12:13:41 -04:00
* The minimum jaxlib version is now 0.1.62.
2021-03-05 11:07:50 -08:00
2021-03-18 16:11:40 -07:00
## jaxlib 0.1.64 (March 18 2021)
2021-03-17 10:14:52 -07:00
## jaxlib 0.1.63 (March 17 2021)
2021-03-09 17:55:40 -08:00
2021-03-05 11:07:50 -08:00
## jax 0.2.10 (March 5 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.9...jax-v0.2.10 ).
* New features:
* {func}`jax.scipy.stats.chi2` is now available as a distribution with logpdf and pdf methods.
* {func}`jax.scipy.stats.betabinom` is now available as a distribution with logpmf and pmf methods.
* Added {func}`jax.experimental.jax2tf.call_tf` to call TensorFlow functions
from JAX ({jax-issue}`#5627` )
2021-06-18 08:55:08 +03:00
and [README ](https://github.com/google/jax/blob/main/jax/experimental/jax2tf/README.md#calling-tensorflow-functions-from-jax )).
2021-03-05 11:07:50 -08:00
* Extended the batching rule for `lax.pad` to support batching of the padding values.
* Bug fixes:
* {func}`jax.numpy.take` properly handles negative indices ({jax-issue}`#5768` )
* Breaking changes:
* JAX's promotion rules were adjusted to make promotion more consistent and
invariant to JIT. In particular, binary operations can now result in weakly-typed
values when appropriate. The main user-visible effect of the change is that
some operations result in outputs of different precision than before; for
example the expression `jnp.bfloat16(1) + 0.1 * jnp.arange(10)`
previously returned a `float64` array, and now returns a `bfloat16` array.
JAX's type promotion behavior is described at {ref}`type-promotion` .
* {func}`jax.numpy.linspace` now computes the floor of integer values, i.e.,
rounding towards -inf rather than 0. This change was made to match NumPy
1.20.0.
* {func}`jax.numpy.i0` no longer accepts complex numbers. Previously the
function computed the absolute value of complex arguments. This change was
made to match the semantics of NumPy 1.20.0.
* Several {mod}`jax.numpy` functions no longer accept tuples or lists in place
of array arguments: {func}`jax.numpy.pad` , :func`jax.numpy.ravel` ,
{func}`jax.numpy.repeat` , {func}`jax.numpy.reshape` .
In general, {mod}`jax.numpy` functions should be used with scalars or array arguments.
2021-03-09 17:55:40 -08:00
## jaxlib 0.1.62 (March 9 2021)
2021-03-05 11:07:50 -08:00
* New features:
* jaxlib wheels are now built to require AVX instructions on x86-64 machines
by default. If you want to use JAX on a machine that doesn't support AVX,
you can build a jaxlib from source using the `--target_cpu_features` flag
to `build.py` . `--target_cpu_features` also replaces
`--enable_march_native` .
## jaxlib 0.1.61 (February 12 2021)
## jaxlib 0.1.60 (Febuary 3 2021)
* Bug fixes:
* Fixed a memory leak when converting CPU DeviceArrays to NumPy arrays. The
memory leak was present in jaxlib releases 0.1.58 and 0.1.59.
* `bool` , `int8` , and `uint8` are now considered safe to cast to
`bfloat16` NumPy extension type.
## jax 0.2.9 (January 26 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.8...jax-v0.2.9 ).
* New features:
* Extend the {mod}`jax.experimental.loops` module with support for pytrees. Improved
error checking and error messages.
* Add {func}`jax.experimental.enable_x64` and {func}`jax.experimental.disable_x64` .
These are context managers which allow X64 mode to be temporarily enabled/disabled
within a session.
* Breaking changes:
* {func}`jax.ops.segment_sum` now drops segment IDs that are out of range rather
than wrapping them into the segment ID space. This was done for performance
reasons.
## jaxlib 0.1.59 (January 15 2021)
## jax 0.2.8 (January 12 2021)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.7...jax-v0.2.8 ).
* New features:
* Add {func}`jax.closure_convert` for use with higher-order custom
derivative functions. ({jax-issue}`#5244` )
* Add {func}`jax.experimental.host_callback.call` to call a custom Python
function on the host and return a result to the device computation.
({jax-issue}`#5243` )
* Bug fixes:
* `jax.numpy.arccosh` now returns the same branch as `numpy.arccosh` for
complex inputs ({jax-issue}`#5156` )
2021-08-02 17:57:09 -07:00
* `host_callback.id_tap` now works for `jax.pmap` also. There is an
2021-03-05 11:07:50 -08:00
optional parameter for `id_tap` and `id_print` to request that the
device from which the value is tapped be passed as a keyword argument
to the tap function ({jax-issue}`#5182` ).
* Breaking changes:
* `jax.numpy.pad` now takes keyword arguments. Positional argument `constant_values`
has been removed. In addition, passing unsupported keyword arguments raises an error.
* Changes for {func}`jax.experimental.host_callback.id_tap` ({jax-issue}`#5243` ):
* Removed support for `kwargs` for {func}`jax.experimental.host_callback.id_tap` .
(This support has been deprecated for a few months.)
* Changed the printing of tuples for {func}`jax.experimental.host_callback.id_print`
to use '(' instead of '['.
* Changed the {func}`jax.experimental.host_callback.id_print` in presence of JVP
to print a pair of primal and tangent. Previously, there were two separate
print operations for the primals and the tangent.
* `host_callback.outfeed_receiver` has been removed (it is not necessary,
and was deprecated a few months ago).
* New features:
* New flag for debugging `inf` , analagous to that for `NaN` ({jax-issue}`#5224` ).
## jax 0.2.7 (Dec 4 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.6...jax-v0.2.7 ).
* New features:
* Add `jax.device_put_replicated`
* Add multi-host support to `jax.experimental.sharded_jit`
2021-08-02 17:57:09 -07:00
* Add support for differentiating eigenvalues computed by `jax.numpy.linalg.eig`
2021-03-05 11:07:50 -08:00
* Add support for building on Windows platforms
* Add support for general in_axes and out_axes in `jax.pmap`
* Add complex support for `jax.numpy.linalg.slogdet`
* Bug fixes:
* Fix higher-than-second order derivatives of `jax.numpy.sinc` at zero
* Fix some hard-to-hit bugs around symbolic zeros in transpose rules
* Breaking changes:
* `jax.experimental.optix` has been deleted, in favor of the standalone
`optax` Python package.
* indexing of JAX arrays with non-tuple sequences now raises a `TypeError` . This type of indexing
has been deprecated in Numpy since v1.16, and in JAX since v0.2.4.
See {jax-issue}`#4564` .
## jax 0.2.6 (Nov 18 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.5...jax-v0.2.6 ).
* New Features:
* Add support for shape-polymorphic tracing for the jax.experimental.jax2tf converter.
2021-06-18 08:55:08 +03:00
See [README.md ](https://github.com/google/jax/blob/main/jax/experimental/jax2tf/README.md ).
2021-03-05 11:07:50 -08:00
* Breaking change cleanup
* Raise an error on non-hashable static arguments for jax.jit and
xla_computation. See [cb48f42 ](https://github.com/google/jax/commit/cb48f42 ).
* Improve consistency of type promotion behavior ({jax-issue}`#4744` ):
* Adding a complex Python scalar to a JAX floating point number respects the precision of
the JAX float. For example, `jnp.float32(1) + 1j` now returns `complex64` , where previously
it returned `complex128` .
* Results of type promotion with 3 or more terms involving uint64, a signed int, and a third type
are now independent of the order of arguments. For example:
`jnp.result_type(jnp.uint64, jnp.int64, jnp.float16)` and
`jnp.result_type(jnp.float16, jnp.uint64, jnp.int64)` both return `float16` , where previously
the first returned `float64` and the second returned `float16` .
* The contents of the (undocumented) `jax.lax_linalg` linear algebra module
are now exposed publicly as `jax.lax.linalg` .
* `jax.random.PRNGKey` now produces the same results in and out of JIT compilation
({jax-issue}`#4877` ).
This required changing the result for a given seed in a few particular cases:
* With `jax_enable_x64=False` , negative seeds passed as Python integers now return a different result
outside JIT mode. For example, `jax.random.PRNGKey(-1)` previously returned
`[4294967295, 4294967295]` , and now returns `[0, 4294967295]` . This matches the behavior in JIT.
* Seeds outside the range representable by `int64` outside JIT now result in an `OverflowError`
rather than a `TypeError` . This matches the behavior in JIT.
To recover the keys returned previously for negative integers with `jax_enable_x64=False`
outside JIT, you can use:
```
key = random.PRNGKey(-1).at[0].set(0xFFFFFFFF)
```
* DeviceArray now raises `RuntimeError` instead of `ValueError` when trying
to access its value while it has been deleted.
## jaxlib 0.1.58 (January 12ish 2021)
* Fixed a bug that meant JAX sometimes return platform-specific types (e.g.,
`np.cint` ) instead of standard types (e.g., `np.int32` ). (#4903 )
* Fixed a crash when constant-folding certain int16 operations. (#4971 )
* Added an `is_leaf` predicate to {func}`pytree.flatten` .
## jaxlib 0.1.57 (November 12 2020)
* Fixed manylinux2010 compliance issues in GPU wheels.
* Switched the CPU FFT implementation from Eigen to PocketFFT.
* Fixed a bug where the hash of bfloat16 values was not correctly initialized
and could change (#4651 ).
* Add support for retaining ownership when passing arrays to DLPack (#4636 ).
* Fixed a bug for batched triangular solves with sizes greater than 128 but not
a multiple of 128.
* Fixed a bug when performing concurrent FFTs on multiple GPUs (#3518 ).
* Fixed a bug in profiler where tools are missing (#4427 ).
* Dropped support for CUDA 10.0.
## jax 0.2.5 (October 27 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.4...jax-v0.2.5 ).
* Improvements:
* Ensure that `check_jaxpr` does not perform FLOPS. See {jax-issue}`#4650` .
* Expanded the set of JAX primitives converted by jax2tf.
2021-06-18 08:55:08 +03:00
See [primitives_with_limited_support.md ](https://github.com/google/jax/blob/main/jax/experimental/jax2tf/primitives_with_limited_support.md ).
2021-03-05 11:07:50 -08:00
## jax 0.2.4 (October 19 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.3...jax-v0.2.4 ).
* Improvements:
* Add support for `remat` to jax.experimental.host_callback. See {jax-issue}`#4608` .
* Deprecations
* Indexing with non-tuple sequences is now deprecated, following a similar deprecation in Numpy.
In a future release, this will result in a TypeError. See {jax-issue}`#4564` .
## jaxlib 0.1.56 (October 14, 2020)
## jax 0.2.3 (October 14 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.2...jax-v0.2.3 ).
* The reason for another release so soon is we need to temporarily roll back a
new jit fastpath while we look into a performance degradation
## jax 0.2.2 (October 13 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.1...jax-v0.2.2 ).
## jax 0.2.1 (October 6 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.2.0...jax-v0.2.1 ).
* Improvements:
* As a benefit of omnistaging, the host_callback functions are executed (in program
order) even if the result of the {py:func}`jax.experimental.host_callback.id_print` /
{py:func}`jax.experimental.host_callback.id_tap` is not used in the computation.
## jax (0.2.0) (September 23 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.77...jax-v0.2.0 ).
* Improvements:
2021-03-25 16:44:58 -07:00
* Omnistaging on by default. See {jax-issue}`#3370` and
2022-06-24 12:18:11 -07:00
[omnistaging ](https://github.com/google/jax/blob/main/docs/design_notes/omnistaging.md )
2021-03-05 11:07:50 -08:00
## jax (0.1.77) (September 15 2020)
* Breaking changes:
* New simplified interface for {py:func}`jax.experimental.host_callback.id_tap` (#4101 )
## jaxlib 0.1.55 (September 8, 2020)
* Update XLA:
* Fix bug in DLPackManagedTensorToBuffer (#4196 )
## jax 0.1.76 (September 8, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.75...jax-v0.1.76 ).
## jax 0.1.75 (July 30, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.74...jax-v0.1.75 ).
* Bug Fixes:
* make jnp.abs() work for unsigned inputs (#3914 )
* Improvements:
* "Omnistaging" behavior added behind a flag, disabled by default (#3370 )
## jax 0.1.74 (July 29, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.73...jax-v0.1.74 ).
* New Features:
* BFGS (#3101 )
2021-08-02 17:57:09 -07:00
* TPU support for half-precision arithmetic (#3878 )
2021-03-05 11:07:50 -08:00
* Bug Fixes:
* Prevent some accidental dtype warnings (#3874 )
* Fix a multi-threading bug in custom derivatives (#3845 , #3869 )
* Improvements:
* Faster searchsorted implementation (#3873 )
* Better test coverage for jax.numpy sorting algorithms (#3836 )
## jaxlib 0.1.52 (July 22, 2020)
* Update XLA.
## jax 0.1.73 (July 22, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.72...jax-v0.1.73 ).
* The minimum jaxlib version is now 0.1.51.
* New Features:
* jax.image.resize. (#3703 )
* hfft and ihfft (#3664 )
* jax.numpy.intersect1d (#3726 )
* jax.numpy.lexsort (#3812 )
* `lax.scan` and the `scan` primitive support an `unroll`
parameter for loop unrolling when lowering to XLA
({jax-issue}`#3738` ).
* Bug Fixes:
* Fix reduction repeated axis error (#3618 )
* Fix shape rule for lax.pad for input dimensions of size 0. (#3608 )
* make psum transpose handle zero cotangents (#3653 )
* Fix shape error when taking JVP of reduce-prod over size 0 axis. (#3729 )
* Support differentiation through jax.lax.all_to_all (#3733 )
* address nan issue in jax.scipy.special.zeta (#3777 )
* Improvements:
* Many improvements to jax2tf
* Reimplement argmin/argmax using a single pass variadic reduction. (#3611 )
* Enable XLA SPMD partitioning by default. (#3151 )
* Add support for 0d transpose convolution (#3643 )
* Make LU gradient work for low-rank matrices (#3610 )
* support multiple_results and custom JVPs in jet (#3657 )
* Generalize reduce-window padding to support (lo, hi) pairs. (#3728 )
* Implement complex convolutions on CPU and GPU. (#3735 )
* Make jnp.take work for empty slices of empty arrays. (#3751 )
* Relax dimension ordering rules for dot_general. (#3778 )
* Enable buffer donation for GPU. (#3800 )
* Add support for base dilation and window dilation to reduce window op… (#3803 )
## jaxlib 0.1.51 (July 2, 2020)
* Update XLA.
* Add new runtime support for host_callback.
## jax 0.1.72 (June 28, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.71...jax-v0.1.72 ).
* Bug fixes:
* Fix an odeint bug introduced in the previous release, see
{jax-issue}`#3587` .
## jax 0.1.71 (June 25, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.70...jax-v0.1.71 ).
* The minimum jaxlib version is now 0.1.48.
* Bug fixes:
* Allow `jax.experimental.ode.odeint` dynamics functions to close over
values with respect to which we're differentiating
{jax-issue}`#3562` .
## jaxlib 0.1.50 (June 25, 2020)
* Add support for CUDA 11.0.
* Drop support for CUDA 9.2 (we only maintain support for the last four CUDA
versions.)
* Update XLA.
## jaxlib 0.1.49 (June 19, 2020)
* Bug fixes:
* Fix build issue that could result in slow compiles
(< https: / / github . com / tensorflow / tensorflow / commit / f805153a25b00d12072bd728e91bb1621bfcf1b1 > )
## jaxlib 0.1.48 (June 12, 2020)
* New features:
* Adds support for fast traceback collection.
* Adds preliminary support for on-device heap profiling.
* Implements `np.nextafter` for `bfloat16` types.
* Complex128 support for FFTs on CPU and GPU.
* Bugfixes:
* Improved float64 `tanh` accuracy on GPU.
* float64 scatters on GPU are much faster.
* Complex matrix multiplication on CPU should be much faster.
* Stable sorts on CPU should actually be stable now.
* Concurrency bug fix in CPU backend.
## jax 0.1.70 (June 8, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.69...jax-v0.1.70 ).
* New features:
* `lax.switch` introduces indexed conditionals with multiple
branches, together with a generalization of the `cond`
primitive
{jax-issue}`#3318` .
## jax 0.1.69 (June 3, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.68...jax-v0.1.69 ).
## jax 0.1.68 (May 21, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.67...jax-v0.1.68 ).
* New features:
* {func}`lax.cond` supports a single-operand form, taken as the argument
to both branches
{jax-issue}`#2993` .
* Notable changes:
* The format of the `transforms` keyword for the {func}`jax.experimental.host_callback.id_tap`
primitive has changed {jax-issue}`#3132` .
## jax 0.1.67 (May 12, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.66...jax-v0.1.67 ).
* New features:
* Support for reduction over subsets of a pmapped axis using `axis_index_groups`
{jax-issue}`#2382` .
* Experimental support for printing and calling host-side Python function from
compiled code. See [id_print and id_tap ](https://jax.readthedocs.io/en/latest/jax.experimental.host_callback.html )
({jax-issue}`#3006` ).
* Notable changes:
* The visibility of names exported from {mod}`jax.numpy` has been
tightened. This may break code that was making use of names that were
previously exported accidentally.
## jaxlib 0.1.47 (May 8, 2020)
* Fixes crash for outfeed.
## jax 0.1.66 (May 5, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.65...jax-v0.1.66 ).
* New features:
* Support for `in_axes=None` on {func}`pmap`
{jax-issue}`#2896` .
## jaxlib 0.1.46 (May 5, 2020)
* Fixes crash for linear algebra functions on Mac OS X (#432 ).
* Fixes an illegal instruction crash caused by using AVX512 instructions when
an operating system or hypervisor disabled them (#2906 ).
## jax 0.1.65 (April 30, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.64...jax-v0.1.65 ).
* New features:
* Differentiation of determinants of singular matrices
{jax-issue}`#2809` .
* Bug fixes:
* Fix {func}`odeint` differentiation with respect to time of ODEs with
time-dependent dynamics {jax-issue}`#2817` ,
also add ODE CI testing.
* Fix {func}`lax_linalg.qr` differentiation
{jax-issue}`#2867` .
## jaxlib 0.1.45 (April 21, 2020)
* Fixes segfault: {jax-issue}`#2755`
* Plumb is_stable option on Sort HLO through to Python.
## jax 0.1.64 (April 21, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.63...jax-v0.1.64 ).
* New features:
* Add syntactic sugar for functional indexed updates
{jax-issue}`#2684` .
* Add {func}`jax.numpy.linalg.multi_dot` {jax-issue}`#2726` .
* Add {func}`jax.numpy.unique` {jax-issue}`#2760` .
* Add {func}`jax.numpy.rint` {jax-issue}`#2724` .
* Add {func}`jax.numpy.rint` {jax-issue}`#2724` .
* Add more primitive rules for {func}`jax.experimental.jet` .
* Bug fixes:
* Fix {func}`logaddexp` and {func}`logaddexp2` differentiation at zero {jax-issue}`#2107` .
* Improve memory usage in reverse-mode autodiff without {func}`jit`
{jax-issue}`#2719` .
* Better errors:
* Improves error message for reverse-mode differentiation of {func}`lax.while_loop`
{jax-issue}`#2129` .
## jaxlib 0.1.44 (April 16, 2020)
* Fixes a bug where if multiple GPUs of different models were present, JAX
would only compile programs suitable for the first GPU.
* Bugfix for `batch_group_count` convolutions.
* Added precompiled SASS for more GPU versions to avoid startup PTX compilation
hang.
## jax 0.1.63 (April 12, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.62...jax-v0.1.63 ).
* Added `jax.custom_jvp` and `jax.custom_vjp` from {jax-issue}`#2026` , see the [tutorial notebook ](https://jax.readthedocs.io/en/latest/notebooks/Custom_derivative_rules_for_Python_code.html ). Deprecated `jax.custom_transforms` and removed it from the docs (though it still works).
* Add `scipy.sparse.linalg.cg` {jax-issue}`#2566` .
* Changed how Tracers are printed to show more useful information for debugging {jax-issue}`#2591` .
* Made `jax.numpy.isclose` handle `nan` and `inf` correctly {jax-issue}`#2501` .
* Added several new rules for `jax.experimental.jet` {jax-issue}`#2537` .
* Fixed `jax.experimental.stax.BatchNorm` when `scale` /`center` isn't provided.
* Fix some missing cases of broadcasting in `jax.numpy.einsum` {jax-issue}`#2512` .
* Implement `jax.numpy.cumsum` and `jax.numpy.cumprod` in terms of a parallel prefix scan {jax-issue}`#2596` and make `reduce_prod` differentiable to arbitray order {jax-issue}`#2597` .
* Add `batch_group_count` to `conv_general_dilated` {jax-issue}`#2635` .
* Add docstring for `test_util.check_grads` {jax-issue}`#2656` .
* Add `callback_transform` {jax-issue}`#2665` .
* Implement `rollaxis` , `convolve` /`correlate` 1d & 2d, `copysign` ,
`trunc` , `roots` , and `quantile` /`percentile` interpolation options.
## jaxlib 0.1.43 (March 31, 2020)
* Fixed a performance regression for Resnet-50 on GPU.
## jax 0.1.62 (March 21, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.61...jax-v0.1.62 ).
* JAX has dropped support for Python 3.5. Please upgrade to Python 3.6 or newer.
* Removed the internal function `lax._safe_mul` , which implemented the
convention `0. * nan == 0.` . This change means some programs when
differentiated will produce nans when they previously produced correct
values, though it ensures nans rather than silently incorrect results are
produced for other programs. See #2447 and #1052 for details.
* Added an `all_gather` parallel convenience function.
* More type annotations in core code.
## jaxlib 0.1.42 (March 19, 2020)
* jaxlib 0.1.41 broke cloud TPU support due to an API incompatibility. This
release fixes it again.
* JAX has dropped support for Python 3.5. Please upgrade to Python 3.6 or newer.
## jax 0.1.61 (March 17, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.60...jax-v0.1.61 ).
* Fixes Python 3.5 support. This will be the last JAX or jaxlib release that
supports Python 3.5.
## jax 0.1.60 (March 17, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.59...jax-v0.1.60 ).
* New features:
* {py:func}`jax.pmap` has `static_broadcast_argnums` argument which allows
the user to specify arguments that should be treated as compile-time
constants and should be broadcasted to all devices. It works analogously to
`static_argnums` in {py:func}`jax.jit` .
* Improved error messages for when tracers are mistakenly saved in global state.
* Added {py:func}`jax.nn.one_hot` utility function.
* Added {mod}`jax.experimental.jet` for exponentially faster
higher-order automatic differentiation.
* Added more correctness checking to arguments of {py:func}`jax.lax.broadcast_in_dim` .
* The minimum jaxlib version is now 0.1.41.
## jaxlib 0.1.40 (March 4, 2020)
* Adds experimental support in Jaxlib for TensorFlow profiler, which allows
tracing of CPU and GPU computations from TensorBoard.
* Includes prototype support for multihost GPU computations that communicate via
NCCL.
* Improves performance of NCCL collectives on GPU.
* Adds TopK, CustomCallWithoutLayout, CustomCallWithLayout, IGammaGradA and
RandomGamma implementations.
* Supports device assignments known at XLA compilation time.
## jax 0.1.59 (February 11, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/jax-v0.1.58...jax-v0.1.59 ).
* Breaking changes
* The minimum jaxlib version is now 0.1.38.
* Simplified {py:class}`Jaxpr` by removing the `Jaxpr.freevars` and
`Jaxpr.bound_subjaxprs` . The call primitives (`xla_call` , `xla_pmap` ,
`sharded_call` , and `remat_call` ) get a new parameter `call_jaxpr` with a
fully-closed (no `constvars` ) jaxpr. Also, added a new field `call_primitive`
to primitives.
* New features:
* Reverse-mode automatic differentiation (e.g. `grad` ) of `lax.cond` , making it
now differentiable in both modes ({jax-issue}`#2091` )
* JAX now supports DLPack, which allows sharing CPU and GPU arrays in a
zero-copy way with other libraries, such as PyTorch.
* JAX GPU DeviceArrays now support `__cuda_array_interface__` , which is another
zero-copy protocol for sharing GPU arrays with other libraries such as CuPy
and Numba.
* JAX CPU device buffers now implement the Python buffer protocol, which allows
zero-copy buffer sharing between JAX and NumPy.
* Added JAX_SKIP_SLOW_TESTS environment variable to skip tests known as slow.
## jaxlib 0.1.39 (February 11, 2020)
* Updates XLA.
## jaxlib 0.1.38 (January 29, 2020)
* CUDA 9.0 is no longer supported.
* CUDA 10.2 wheels are now built by default.
## jax 0.1.58 (January 28, 2020)
* [GitHub commits ](https://github.com/google/jax/compare/46014da21...jax-v0.1.58 ).
* Breaking changes
* JAX has dropped Python 2 support, because Python 2 reached its end of life on
January 1, 2020. Please update to Python 3.5 or newer.
* New features
> > * Forward-mode automatic differentiation (`jvp`) of while loop
> ({jax-issue}`#1980`)
> * New NumPy and SciPy functions:
2021-03-09 17:55:40 -08:00
>
2021-03-05 11:07:50 -08:00
> * {py:func}`jax.numpy.fft.fft2`
> * {py:func}`jax.numpy.fft.ifft2`
> * {py:func}`jax.numpy.fft.rfft`
> * {py:func}`jax.numpy.fft.irfft`
> * {py:func}`jax.numpy.fft.rfft2`
> * {py:func}`jax.numpy.fft.irfft2`
> * {py:func}`jax.numpy.fft.rfftn`
> * {py:func}`jax.numpy.fft.irfftn`
> * {py:func}`jax.numpy.fft.fftfreq`
> * {py:func}`jax.numpy.fft.rfftfreq`
> * {py:func}`jax.numpy.linalg.matrix_rank`
> * {py:func}`jax.numpy.linalg.matrix_power`
> * {py:func}`jax.scipy.special.betainc`
> * Batched Cholesky decomposition on GPU now uses a more efficient batched
> kernel.
### Notable bug fixes
* With the Python 3 upgrade, JAX no longer depends on `fastcache` , which should
help with installation.
