# Copyright 2024 The Flax Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import typing as tp
import flax.core.spmd as core_spmd
from flax.nnx import variablelib, graph
from flax.nnx.transforms.transforms import eval_shape
from flax.typing import (
Sharding,
)
import jax
from jax.sharding import PartitionSpec
# JAX version compatibility
if hasattr(jax.sharding, 'use_mesh'):
set_mesh = jax.sharding.use_mesh
else:
set_mesh = jax.set_mesh
A = tp.TypeVar('A')
F = tp.TypeVar('F', bound=tp.Callable[..., tp.Any])
PARTITION_NAME = 'partition_name'
# Transform axis change helpers
# ------------------------------------------------------------------------------
def add_axis(tree: A, index: int, transform_metadata: tp.Mapping) -> A:
axis_name, other_meta = _get_partition_name_and_metadata(transform_metadata)
def insert_field(fields, index, value):
iterable = list(fields)
while len(iterable) < index:
iterable.append(None)
iterable.insert(index, value)
return tuple(iterable)
def _add_axis(x: tp.Any):
if isinstance(x, variablelib.Variable):
metadata = x.get_metadata()
if 'sharding_names' in metadata and metadata['sharding_names']:
sharding = metadata['sharding_names']
x.sharding_names = insert_field(sharding, index, axis_name)
for k, v in other_meta.items():
if hasattr(x, k) and (t := getattr(x, k)) and isinstance(t, tuple):
setattr(x, k, insert_field(t, index, v))
assert isinstance(x, variablelib.Variable)
x.add_axis(index, axis_name)
return x
return jax.tree.map(
_add_axis, tree, is_leaf=lambda x: isinstance(x, variablelib.Variable)
)
def remove_axis(
tree: A, index: int, transform_metadata: tp.Mapping[tp.Any, tp.Any]
) -> A:
axis_name, other_meta = _get_partition_name_and_metadata(transform_metadata)
def remove_field(fields, index, value):
iterable = list(fields)
assert iterable.pop(index) == value
return tuple(iterable)
def _remove_axis(x: tp.Any):
if isinstance(x, variablelib.Variable):
if hasattr(x, 'sharding_names') and x.sharding_names is not None:
x.sharding_names = remove_field(x.sharding_names, index, axis_name)
for k, v in other_meta.items():
if hasattr(x, k) and (t := getattr(x, k)) and isinstance(t, tuple):
setattr(x, k, remove_field(t, index, v))
x.remove_axis(index, axis_name)
return x
return jax.tree.map(
_remove_axis,
tree,
is_leaf=lambda x: isinstance(x, variablelib.Variable),
)
def _get_partition_name_and_metadata(
transform_metadata: tp.Mapping[tp.Any, tp.Any],
) -> tuple[str, tp.Mapping[tp.Any, tp.Any]]:
if PARTITION_NAME not in transform_metadata:
raise ValueError(
'Trying to transform a Partitioned variable but "partition_name" '
f'is not specified in transform_metadata: {transform_metadata}'
)
other_meta = dict(transform_metadata) # shallow copy
other_meta.pop(PARTITION_NAME)
return transform_metadata[PARTITION_NAME], other_meta
# Annotation handling
# ------------------------------------------------------------------------------
[docs]def with_partitioning(
initializer: F,
sharding: Sharding,
mesh: tp.Optional[jax.sharding.Mesh] = None,
**metadata: tp.Any,
) -> F:
"""A wrapper over any initializer to add sharding annotation data to a `Variable`."""
return variablelib.with_metadata(
initializer,
sharding_names=sharding,
mesh=mesh,
**metadata,
)
def get_var_pspec(v: variablelib.Variable) -> PartitionSpec | None:
"""Given an `nnx.Variable`, return its `PartitionSpec`."""
metadata = v.get_metadata()
if 'sharding_names' in metadata and metadata['sharding_names']:
sharding = metadata['sharding_names']
if core_spmd.get_logical_axis_rules() or 'sharding_rules' in metadata:
context_rules = core_spmd.get_logical_axis_rules()
local_rules = metadata.get('sharding_rules', ())
rules = core_spmd.composite_rules(context_rules, local_rules)
return PartitionSpec(*core_spmd.from_sharding_rules(sharding, rules))
return PartitionSpec(*sharding)
elif hasattr(v, 'shape'):
return PartitionSpec()
return None
[docs]def get_partition_spec(tree: A) -> A:
"""Extracts a PartitionSpec tree from a PyTree containing ``Variable`` values."""
def f(x):
if isinstance(x, variablelib.Variable):
return x.replace(get_var_pspec(x))
elif hasattr(x, 'shape'):
return PartitionSpec()
return None
return jax.tree.map(
f, tree, is_leaf=lambda x: isinstance(x, variablelib.Variable)
)
[docs]def get_named_sharding(tree: A, mesh: jax.sharding.Mesh) -> A:
spec = get_partition_spec(tree)
sharding = jax.tree.map(lambda p: jax.sharding.NamedSharding(mesh, p), spec)
return sharding
# Other utilities
# ------------------------------------------------------------------------------
def get_abstract_model(init_fn, mesh):
with set_mesh(mesh):
abs_model = eval_shape(init_fn)
gdef, abs_state = graph.split(abs_model)
abs_state = jax.tree.map(
lambda a, s: jax.ShapeDtypeStruct(a.shape, a.dtype, sharding=s),
abs_state, get_named_sharding(abs_state, mesh)
)
return gdef, abs_state
