transforms

class flax.nnx.Jit(*args, **kwargs)

class flax.nnx.Remat(*args, **kwargs)

class flax.nnx.Scan(*args, **kwargs)

class flax.nnx.Vmap(*args, **kwargs)

flax.nnx.grad(f=<flax.typing.Missing object>, *, argnums=0, has_aux=False, holomorphic=False, allow_int=False, reduce_axes=())

Lifted version of jax.grad that can handle Modules / graph nodes as arguments.

The differentiable state of each graph node is defined by the wrt filter, which by default is set to nnx.Param. Internally the State of graph nodes is extracted, filtered according to wrt filter, and passed to the underlying jax.grad function. The gradients of graph nodes are of type State.

Example:

>>> from flax import nnx
>>> import jax.numpy as jnp
...
>>> m = nnx.Linear(2, 3, rngs=nnx.Rngs(0))
>>> x = jnp.ones((1, 2))
>>> y = jnp.ones((1, 3))
...
>>> loss_fn = lambda m, x, y: jnp.mean((m(x) - y) ** 2)
>>> grad_fn = nnx.grad(loss_fn)
...
>>> grads = grad_fn(m, x, y)
>>> jax.tree.map(jnp.shape, grads)
State({
  'bias': VariableState(
    type=Param,
    value=(3,)
  ),
  'kernel': VariableState(
    type=Param,
    value=(2, 3)
  )
})

Parameters

• fun – Function to be differentiated. Its arguments at positions specified by argnums should be arrays, scalars, graph nodes or standard Python containers. Argument arrays in the positions specified by argnums must be of inexact (i.e., floating-point or complex) type. It should return a scalar (which includes arrays with shape () but not arrays with shape (1,) etc.)

• argnums – Optional, integer or sequence of integers. Specifies which positional argument(s) to differentiate with respect to (default 0).

• has_aux – Optional, bool. Indicates whether fun returns a pair where the first element is considered the output of the mathematical function to be differentiated and the second element is auxiliary data. Default False.

• holomorphic – Optional, bool. Indicates whether fun is promised to be holomorphic. If True, inputs and outputs must be complex. Default False.

• allow_int – Optional, bool. Whether to allow differentiating with respect to integer valued inputs. The gradient of an integer input will have a trivial vector-space dtype (float0). Default False.

• reduce_axes – Optional, tuple of axis names. If an axis is listed here, and fun implicitly broadcasts a value over that axis, the backward pass will perform a psum of the corresponding gradient. Otherwise, the gradient will be per-example over named axes. For example, if 'batch' is a named batch axis, grad(f, reduce_axes=('batch',)) will create a function that computes the total gradient while grad(f) will create one that computes the per-example gradient.

flax.nnx.jit(fun=<class 'flax.typing.Missing'>, *, in_shardings=None, out_shardings=None, static_argnums=None, static_argnames=None, donate_argnums=None, donate_argnames=None, keep_unused=False, device=None, backend=None, inline=False, abstracted_axes=None)

Lifted version of jax.jit that can handle Modules / graph nodes as arguments.

Parameters

• fun –

  Function to be jitted. fun should be a pure function, as side-effects may only be executed once.

  The arguments and return value of fun should be arrays, scalars, or (nested) standard Python containers (tuple/list/dict) thereof. Positional arguments indicated by static_argnums can be anything at all, provided they are hashable and have an equality operation defined. Static arguments are included as part of a compilation cache key, which is why hash and equality operators must be defined.

  JAX keeps a weak reference to fun for use as a compilation cache key, so the object fun must be weakly-referenceable. Most Callable objects will already satisfy this requirement.

• in_shardings –

  Pytree of structure matching that of arguments to fun, with all actual arguments replaced by resource assignment specifications. It is also valid to specify a pytree prefix (e.g. one value in place of a whole subtree), in which case the leaves get broadcast to all values in that subtree.

  The in_shardings argument is optional. JAX will infer the shardings from the input jax.Array’s and defaults to replicating the input if the sharding cannot be inferred.

  The valid resource assignment specifications are:

  • Sharding, which will decide how the value

    will be partitioned. With this, using a mesh context manager is not required.

  • None, will give JAX 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.

  The size of every dimension has to be a multiple of the total number of resources assigned to it. This is similar to pjit’s in_shardings.

• out_shardings –

  Like in_shardings, but specifies resource assignment for function outputs. This is similar to pjit’s out_shardings.

  The out_shardings argument is optional. If not specified, jax.jit() will use GSPMD’s sharding propagation to figure out what the sharding of the output(s) should be.

• static_argnums –

  An optional int or collection of ints that specify which positional arguments to treat as static (compile-time constant). Operations that only depend on static arguments will be constant-folded in Python (during tracing), and so the corresponding argument values can be any Python object.

  Static arguments should be hashable, meaning both __hash__ and __eq__ are implemented, and immutable. Calling the jitted function with different values for these constants will trigger recompilation. Arguments that are not arrays or containers thereof must be marked as static.

  If neither static_argnums nor static_argnames is provided, no arguments are treated as static. If static_argnums is not provided but static_argnames is, or vice versa, JAX uses inspect.signature(fun) to find any positional arguments that correspond to static_argnames (or vice versa). If both static_argnums and static_argnames are provided, inspect.signature is not used, and only actual parameters listed in either static_argnums or static_argnames will be treated as static.

• static_argnames – An optional string or collection of strings specifying which named arguments to treat as static (compile-time constant). See the comment on static_argnums for details. If not provided but static_argnums is set, the default is based on calling inspect.signature(fun) to find corresponding named arguments.

• donate_argnums –

  Specify which positional argument buffers are “donated” to the computation. It is safe to donate argument buffers if you no longer need them once the computation has finished. In some cases XLA can make use of donated buffers to reduce the amount of memory needed to perform a computation, for example recycling one of your input buffers to store a result. You should not reuse buffers that you donate to a computation, JAX will raise an error if you try to. By default, no argument buffers are donated.

  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.

  For more details on buffer donation see the FAQ.

• donate_argnames – An optional string or collection of strings specifying which named arguments are donated to the computation. See the comment on donate_argnums for details. If not provided but donate_argnums is set, the default is based on calling inspect.signature(fun) to find corresponding named arguments.

• keep_unused – If False (the default), arguments that JAX determines to be unused by fun may be dropped from resulting compiled XLA executables. Such arguments will not be transferred to the device nor provided to the underlying executable. If True, unused arguments will not be pruned.

• device – This is an experimental feature and the API is likely to change. Optional, the Device the jitted function will run on. (Available devices can be retrieved via jax.devices().) The default is inherited from XLA’s DeviceAssignment logic and is usually to use jax.devices()[0].

• backend – This is an experimental feature and the API is likely to change. Optional, a string representing the XLA backend: 'cpu', 'gpu', or 'tpu'.

• inline – Specify whether this function should be inlined into enclosing jaxprs (rather than being represented as an application of the xla_call primitive with its own subjaxpr). Default False.

Returns

A wrapped version of fun, set up for just-in-time compilation.

flax.nnx.remat(f=<flax.typing.Missing object>, *, prevent_cse=True, static_argnums=(), policy=None)

flax.nnx.scan(f=<class 'flax.typing.Missing'>, *, length=None, reverse=False, unroll=1, _split_transpose=False, in_axes=(<class 'flax.nnx.transforms.iteration.Carry'>, 0), out_axes=(<class 'flax.nnx.transforms.iteration.Carry'>, 0), transform_metadata=FrozenDict({}))

flax.nnx.value_and_grad(f=<class 'flax.typing.Missing'>, *, argnums=0, has_aux=False, holomorphic=False, allow_int=False, reduce_axes=())

flax.nnx.vmap(f=<class 'flax.typing.Missing'>, *, in_axes=0, out_axes=0, axis_name=None, axis_size=None, spmd_axis_name=None, transform_metadata=FrozenDict({}))

Reference-aware version of jax.vmap.

Parameters

• f – Function to be mapped over additional axes.

• in_axes – An integer, None, or sequence of values specifying which input array axes to map over (see jax.vmap). In addition to integers and None, StateAxes can be used to control how graph nodes like Modules are vectorized by specifying the axes to be applied to substates of the graph node given a Filter.

• out_axes – An integer, None, or pytree indicating where the mapped axis should appear in the output (see jax.vmap).

• axis_name – Optional, a hashable Python object used to identify the mapped axis so that parallel collectives can be applied.

• axis_size – Optional, an integer indicating the size of the axis to be mapped. If not provided, the mapped axis size is inferred from arguments.

Returns

Batched/vectorized version of f with arguments that correspond to those of f, but with extra array axes at positions indicated by in_axes, and a return value that corresponds to that of f, but with extra array axes at positions indicated by out_axes.

Example:

>>> from flax import nnx
>>> from jax import random, numpy as jnp
...
>>> model = nnx.Linear(2, 3, rngs=nnx.Rngs(0))
>>> x = jnp.ones((5, 2))
...
>>> @nnx.vmap(in_axes=(None, 0), out_axes=0)
... def forward(model, x):
...   return model(x)
...
>>> y = forward(model, x)
>>> y.shape
(5, 3)

>>> class LinearEnsemble(nnx.Module):
...   def __init__(self, num, rngs):
...     self.w = nnx.Param(jax.random.uniform(rngs(), (num, 2, 3)))
...
>>> model = LinearEnsemble(5, rngs=nnx.Rngs(0))
>>> x = jnp.ones((2,))
...
>>> @nnx.vmap(in_axes=(0, None), out_axes=0)
... def forward(model, x):
...   return jnp.dot(x, model.w.value)
...
>>> y = forward(model, x)
>>> y.shape
(5, 3)

To control control how graph node substates are vectorized, StateAxes can be passed to in_axes and out_axes specifying the axes to be applied to each substate given a filter. The following example shows how to share the parameters between the ensemble members which keeping different batch statistics and dropout random state:

>>> class Foo(nnx.Module):
...   def __init__(self):
...     self.a = nnx.Param(jnp.arange(4))
...     self.b = nnx.BatchStat(jnp.arange(4))
...
>>> state_axes = nnx.StateAxes({nnx.Param: 0, nnx.BatchStat: None})
>>> @nnx.vmap(in_axes=(state_axes,), out_axes=0)
... def mul(foo):
...   return foo.a * foo.b
...
>>> foo = Foo()
>>> y = mul(foo)
>>> y
Array([[0, 0, 0, 0],
       [0, 1, 2, 3],
       [0, 2, 4, 6],
       [0, 3, 6, 9]], dtype=int32)

flax.nnx.eval_shape(f, *args, **kwargs)

flax.nnx.cond(pred, true_fun, false_fun, *operands, **kwargs)