Skip to content

Releases: google/flax

0.12.6

20 Mar 21:09
@cgarciae cgarciae
v0.12.6
1572c84

Choose a tag to compare

View all tags
0.12.6 Latest
Latest

What's Changed

Full Changelog: v0.12.5...v0.12.6

Contributors

samanklesaria, vfdev-5, and 3 other contributors
Assets 2
Loading
0 Join discussion

0.12.5

04 Mar 19:18
@cgarciae cgarciae
v0.12.5
8d7f486

Choose a tag to compare

View all tags
0.12.5

What's Changed

Full Changelog: v0.12.4...v0.12.5

Contributors

samanklesaria, cgarciae, and copybara-service
Loading
0 Join discussion

0.12.4

12 Feb 19:09
@cgarciae cgarciae
v0.12.4
a5c7683

Choose a tag to compare

View all tags
0.12.4

This release fixes an issue with nnx.List and nnx.Sequential passing all its elements as static when being treated as a pytree.

What's Changed

New Contributors

Full Changelog: v0.12.3...v0.12.4

Contributors

samanklesaria, vfdev-5, and 5 other contributors
Loading
0 Join discussion

0.12.3

27 Jan 21:03
@cgarciae cgarciae
v0.12.3
7c487b1

Choose a tag to compare

View all tags
0.12.3

What's Changed

New Contributors

Full Changelog: v0.12.2...v0.12.3

Contributors

samanklesaria, vfdev-5, and 6 other contributors
Loading
0 Join discussion

Version 0.12.2

18 Dec 22:35
@IvyZX IvyZX
v0.12.2
d179184

Choose a tag to compare

View all tags
Version 0.12.2

What's Changed

New Contributors

Full Changelog: v0.12.1...v0.12.2

Contributors

samanklesaria, vfdev-5, and 5 other contributors
Loading
0 Join discussion

v0.12.1

19 Nov 21:55
@cgarciae cgarciae
v0.12.1
697f4e5

Choose a tag to compare

View all tags
v0.12.1

Deprecations

Variable.value

Variable.value is now deprecated. Consider the following example:

import jax.numpy as jnp
import jax
from flax import nnx

my_param = nnx.Param({'a': 0.0})

@nnx.jit
def f(m):
    m.value['a'] = 1.0
    return m

Running f(my_param) produces Param(value={'a': 0.0}), not Param(value={'a': 1.0}) as before. This is because getting the value parameter new returns a copy of the pytree values (like dict / list). Instead, use the __setitem__ method to update the value:

@nnx.jit
def f(m):
    m['a'] = 1.0
    return m

nnx.Data and nnx.Static

nnx.Data and nnx.Static annotations are now deprecated. To create nnx.Pytree or nnx.Module dataclasses use the new nnx.dataclass with nnx.data and nnx.static as field descriptors.

# old
@dataclasses.dataclass
class Foo(nnx.Pytree):
  a: nnx.Data[int]
  b: nnx.Static[str]

# new
@nnx.dataclass
class Foo(nnx.Pytree):
  a: int = nnx.data()
  b: str = nnx.static()

Pull Requests

New Contributors

Full Changelog: v0.12.0...v0.12.1

Contributors

samanklesaria, jburnim, and 11 other contributors
Loading
0 Join discussion

0.12.0

25 Sep 23:58
@cgarciae cgarciae
v0.12.0
34e1d68

Choose a tag to compare

View all tags
0.12.0

Flax 0.12.0 includes many updates and some important breaking changes to the NNX API.

Breaking Changes

Pytree Strict Attributes

nnx.Pytree and therefore nnx.Module are now stricter with regards to attributes that contain Arrays and changing the status of attributes. For example, the code below now fails:

from flax import nnx
import jax
import jax.numpy as jnp

class Foo(nnx.Module):
  def __init__(self, use_bias, rngs):
    self.layers = [  # ERROR
      nnx.Linear(3, 3, rngs=rngs) for _ in range(5)
    ]
    self.bias = None # status = static
    if use_bias:
      self.bias = nnx.Param(rngs.params.uniform(3,)) # ERROR

This happens for two reasons:

  1. JAX pytree structures that contain Arrays now have to be marked with nnx.data. Alternatively, if the container pytree is a list or a dict, you can use nnx.List or nnx.Dict, which additionally allow mixed "data" and "static" elements.
  2. Attributes will no longer automatically change their status—this now has to be done explicitly using nnx.data or nnx.static. Additionally, assigning Arrays or structures with Arrays to static attributes is now an error, as they will not automatically change to data.

To fix the above you can just create layers as a List Module which is automatically recognized as data, and be explicit about bias being a data attribute on the first assignment by using nnx.data:

class Foo(nnx.Module):
  def __init__(self, use_bias, rngs):
    self.layers = nnx.List([  # nnx.data also works but List is recommended
      nnx.Linear(3, 3, rngs=rngs) for _ in range(5)
    ])
    self.bias = nnx.data(None)
    if use_bias:
      self.bias = nnx.Param(rngs.params.uniform(3,))

For more information check the Module & Pytree guide.

Eager Sharding

Variables will now eagerly shard their values when sharding_names metadata is provided. A mesh is required—it can be provided either via passing a mesh metadata attribute or setting the global mesh context via jax.set_mesh. This simplifies the process of sharding a Variable to construction time:

jax.config.update('jax_num_cpu_devices', 8)
mesh = jax.make_mesh((2, 4), ('data', 'model'))

with jax.set_mesh(mesh):
  variable = nnx.Param(jnp.ones((16, 32)), sharding_names=(None, 'model'))
  
print(variable.value.sharding)

Eager sharding will also occur when using the nnx.with_partitioning initializer decorator and will automatically extend to the Optimizer. This means that both model and optimizer will be sharded at construction without the need for the somewhat cumbersome nnx.get_partition_spec + jax.lax.with_sharding_constraint + nnx.update pattern:

with jax.set_mesh(mesh):
  linear = nnx.Linear(
    in_features=16, out_features=16, use_bias=False,
    kernel_init=nnx.with_partitioning(
      nnx.initializers.lecun_normal(), (None, 'model')
    ),
    rngs=nnx.Rngs(0),
  )
  optimizer = nnx.Optimizer(linear, optax.adam(1e-3), wrt=nnx.Param)
  
print(linear.kernel.value.sharding)
print(optimizer.opt_state[0].mu.kernel.value.sharding)

For projects that currently rely on other means for sharding, eager sharding can be turned off by passing eager_sharding=False to the Variable constructor, either directly or through initializer decorators like nnx.with_partitioning:

linear = nnx.Linear(
  in_features=16, out_features=16, use_bias=False,
  kernel_init=nnx.with_partitioning(
    nnx.initializers.lecun_normal(), (None, 'model'), eager_sharding=False
  ),
  rngs=nnx.Rngs(0),
)
optimizer = nnx.Optimizer(linear, optax.adam(1e-3), wrt=nnx.Param)
  
print(linear.kernel.value.sharding)
print(optimizer.opt_state[0].mu.kernel.value.sharding)

Eager sharding can also be turned off globally via the flax_always_shard_variable config flag or the FLAX_ALWAYS_SHARD_VARIABLE environment variable:

import flax
flax.config.update('flax_always_shard_variable', False)

For more information, check out the Variable eager sharding FLIP.

In-Place Operators No Longer Allowed

In-place operators will now raise an error. This is done as part of the push for Variables to be compatible with Tracer semantics:

w = nnx.Variable(jnp.array(0))
w += 1  # ERROR

The fix is to simply operate on the .value property instead:

w.value += 1

All Changes

New Contributors

Full Changelog: v0.11.2...v0.12.0

Contributors

vfdev-5, ywrt, and 4 other contributors
Loading
5 Join discussion

0.11.2

28 Aug 17:55
@cgarciae cgarciae
v0.11.2
405f656

Choose a tag to compare

View all tags
0.11.2

What's Changed

nnx.merge now doesn't create a copy of the Variables in the incoming states by default, meaning that the new merged structures holds references to the incoming Variables. This enables new patterns, for example its now possible to create models with the same state but with different runtime behavior:

model = SomeModel(...)
# create eval model
eval_model = nnx.merge(*nnx.split(model))  # same Variables, different structure
eval_model.eval()

model and eval_model share the same Variables and are therefore kept in sync but have different runtime behavior, this avoids having to constantly mutate a single model back and forth between different runtime modes which can be error prone / cause unwanted recompilation.

To keep the old behavior use nnx.merge(..., copy=True).

PRs

New Contributors

Full Changelog: v0.11.1...v0.11.2

Contributors

vfdev-5, codinfox, and 5 other contributors
Loading
1 Join discussion

v0.11.1

08 Aug 21:25
@cgarciae cgarciae
v0.11.1
a0a291c

Choose a tag to compare

View all tags
v0.11.1

What's Changed

New Contributors

Full Changelog: v0.11.0...v0.11.1

Contributors

SobhanMP, cgarciae, and 2 other contributors
Loading

v0.11.0

29 Jul 21:04
@cgarciae cgarciae
v0.11.0
c9fc00b

Choose a tag to compare

View all tags
v0.11.0

v0.11.0 - Pytrees, MutableArrays, and more!

This version of Flax introduces some changes to improve interop with native JAX and adds support for the new jax.experimental.MutableArray. More on this soon! However, some breaking changes to align with the JAX way of doing things were necessary. Most code should remain intact, however, the following changes deviate from the current behavior:

  • Rngs in standard layers: all standard layers no longer hold a shared reference to the rngs object given in the constructor, instead they now keep a fork-ed copy of the Rngs or RngStream objects. This impacts Using Rngs in NNX Transforms and Loading Checkpoints with RNGs.
  • Optimizer Updates: the Optimizer abstraction no longer holds a reference to the model to avoid reference sharing, instead the model must be provided as the first argument to update.
  • Modules as Pytrees: Modules are now pytrees! This avoid unnecessary use of split and merge when interacting trivially with raw JAX transforms (state must still be manually propagated if not using MutableArrays, and referential transparency is still an issue). This affects when operating on Pytrees containing NNX Objects with jax.tree.* APIs.

Checkout the full NNX 0.10 to NNX 0.11 migration guide.

In the near future we'll share more information about new ways of using NNX with JAX transforms directly by leveraging the new Pytree and MutableArray support. Stay tuned!

What's Changed

New Contributors

Full Changelog: v0.10.7...v0.11.0

Contributors

vfdev-5, cgarciae, and 3 other contributors
Loading
7 Join discussion