|
| 1 | +# Copyright 2025 Google LLC |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# https://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | +"""Experimental split by mesh axis API.""" |
| 15 | + |
| 16 | +from typing import Any, Sequence |
| 17 | + |
| 18 | +import jax |
| 19 | +from pathwaysutils import jax as pw_jax |
| 20 | +from pathwaysutils import lru_cache |
| 21 | + |
| 22 | + |
| 23 | +@lru_cache.lru_cache(maxsize=16384) |
| 24 | +def _cached_named_sharding( |
| 25 | + mesh: jax.sharding.Mesh, |
| 26 | + spec: jax.sharding.PartitionSpec, |
| 27 | + memory_kind: str | None = None, |
| 28 | +): |
| 29 | + return jax.sharding.NamedSharding(mesh, spec, memory_kind=memory_kind) |
| 30 | + |
| 31 | + |
| 32 | +@lru_cache.lru_cache(maxsize=1024) |
| 33 | +def _get_per_mesh_shardings( |
| 34 | + meshes: tuple[jax.sharding.Mesh, ...], |
| 35 | + spec: jax.sharding.PartitionSpec, |
| 36 | + memory_kind: str | None = None, |
| 37 | +) -> Sequence[jax.sharding.NamedSharding]: |
| 38 | + """Returns per-mesh shardings.""" |
| 39 | + return [ |
| 40 | + _cached_named_sharding(mesh, spec, memory_kind=memory_kind) |
| 41 | + for mesh in meshes |
| 42 | + ] |
| 43 | + |
| 44 | + |
| 45 | +def split_by_mesh_axis( |
| 46 | + arrays: Any, |
| 47 | + mesh_axis: str, |
| 48 | + mesh_axis_indices_or_sections: int | Sequence[int] | None = None, |
| 49 | + *, |
| 50 | + donate: bool = False, |
| 51 | +) -> Sequence[Any]: |
| 52 | + """Splits arrays by a mesh axis, and returns arrays on each split mesh. |
| 53 | +
|
| 54 | + Args: |
| 55 | + arrays: PyTree of JAX arrays with NamedSharding whose mesh is identical. |
| 56 | + mesh_axis: Mesh axis to split the arrays by. |
| 57 | + mesh_axis_indices_or_sections: If it is an integer, N, the mesh axis will be |
| 58 | + divided into N equal submeshes along `mesh_axis`. If it is a 1-D sequence, |
| 59 | + the entries indicate the boundary on the mesh axis along `mesh_axis`. For |
| 60 | + example, [2, 3] for splitting first mesh axis results in three output |
| 61 | + arrays (per each input array) on `mesh[:2], mesh[2:3], mesh[3:]`, |
| 62 | + respectively. If it is None, it will be the same as `N = |
| 63 | + mesh.axis_size[mesh.axis_names.index(mesh_axis)]`. Note: the sequence must |
| 64 | + be monotonoically increasing and should not contain the start or end |
| 65 | + boundaries. |
| 66 | + donate: Whether to donate input arrays. By default, input arrays are |
| 67 | + aliased. |
| 68 | +
|
| 69 | + Returns: |
| 70 | + A sequence of PyTrees whose structure is the same as `arrays`. |
| 71 | + Each element `i` has arrays with their shards filtered out to match |
| 72 | + mesh corresponding mesh constructed according to |
| 73 | + `mesh_axis_indices_or_sections`. An array's shape remains the same if the |
| 74 | + array is replicated along `mesh_axis`, or is shrunk by a split factor |
| 75 | + computed from `mesh_axis_indices_or_sections` if the array is partitioned |
| 76 | + along `mesh_axis`. |
| 77 | + """ |
| 78 | + flat_arrays, treedef = jax.tree.flatten(arrays) |
| 79 | + |
| 80 | + if not flat_arrays: |
| 81 | + return arrays |
| 82 | + |
| 83 | + sharding = flat_arrays[0].sharding |
| 84 | + if not isinstance(sharding, jax.sharding.NamedSharding): |
| 85 | + raise ValueError(f"Array must have a NamedSharding. Got {sharding=}") |
| 86 | + mesh = sharding.mesh |
| 87 | + mesh_axis_idx = mesh.axis_names.index(mesh_axis) |
| 88 | + sharded_dim_idxs = [] |
| 89 | + for array in flat_arrays: |
| 90 | + sharding = array.sharding |
| 91 | + if not isinstance(sharding, jax.sharding.NamedSharding): |
| 92 | + raise ValueError(f"Array must have a NamedSharding. Got {sharding=}") |
| 93 | + if mesh != sharding.mesh: |
| 94 | + raise ValueError( |
| 95 | + f"Array sharding mesh must match, but got {mesh=}, {sharding.mesh=}" |
| 96 | + ) |
| 97 | + if sharding._logical_device_ids is not None: # pylint: disable=protected-access |
| 98 | + raise ValueError( |
| 99 | + "Array sharding's _logical_device_ids must be None, but got" |
| 100 | + f" {sharding._logical_device_ids=}" # pylint: disable=protected-access |
| 101 | + ) |
| 102 | + sharded_dim = -1 |
| 103 | + for dim_idx, dim_spec in enumerate(sharding.spec): |
| 104 | + flat_dim_spec, _ = jax.tree.flatten(dim_spec) |
| 105 | + if mesh_axis in flat_dim_spec: |
| 106 | + sharded_dim = dim_idx |
| 107 | + break |
| 108 | + sharded_dim_idxs.append(sharded_dim) |
| 109 | + |
| 110 | + # Transform mesh_axis_indices_or_sections into a list of axis boundaries, |
| 111 | + # with the last entry being the size of the mesh_axis. |
| 112 | + if mesh_axis_indices_or_sections is None: |
| 113 | + # If mesh_axis_indices_or_sections is None, the arrays will be divided |
| 114 | + # along the mesh_axis. |
| 115 | + mesh_axis_indices_or_sections = mesh.axis_sizes[mesh_axis_idx] |
| 116 | + if isinstance(mesh_axis_indices_or_sections, int): |
| 117 | + # Expand the mesh_axis_indices_or_sections to a list indicating the |
| 118 | + # boundaries of mesh axis. |
| 119 | + if mesh.axis_sizes[mesh_axis_idx] % mesh_axis_indices_or_sections != 0: |
| 120 | + raise ValueError( |
| 121 | + "The size of the `mesh_axis` must be divisible by" |
| 122 | + " `mesh_axis_indices_or_sections`. Got" |
| 123 | + f" {mesh.axis_sizes[mesh_axis_idx]} and" |
| 124 | + f" {mesh_axis_indices_or_sections=}" |
| 125 | + ) |
| 126 | + axis_size = mesh.axis_sizes[mesh_axis_idx] // mesh_axis_indices_or_sections |
| 127 | + mesh_axis_sections = list( |
| 128 | + range(axis_size, mesh.axis_sizes[mesh_axis_idx] + 1, axis_size) |
| 129 | + ) |
| 130 | + else: |
| 131 | + mesh_axis_sections = mesh_axis_indices_or_sections |
| 132 | + for i, boundary in enumerate(mesh_axis_sections): |
| 133 | + if boundary <= 0 or boundary >= mesh.axis_sizes[mesh_axis_idx]: |
| 134 | + raise ValueError( |
| 135 | + "Mesh axis sections values must be in range (0," |
| 136 | + f" axis_size={mesh.axis_sizes[mesh_axis_idx]}) to avoid an empty" |
| 137 | + f" section, but got {mesh_axis_sections=}." |
| 138 | + ) |
| 139 | + if i > 0 and mesh_axis_sections[i] <= mesh_axis_sections[i - 1]: |
| 140 | + raise ValueError( |
| 141 | + "Mesh axis sections must be monotonically increasing, but got" |
| 142 | + f" {mesh_axis_sections=}." |
| 143 | + ) |
| 144 | + mesh_axis_sections += [mesh.axis_sizes[mesh_axis_idx]] |
| 145 | + |
| 146 | + submeshes = [] |
| 147 | + axis_boundary_start = 0 |
| 148 | + slices = [slice(None)] * len(mesh.axis_sizes) |
| 149 | + for axis_boundary_end in mesh_axis_sections: |
| 150 | + slices[mesh_axis_idx] = slice(axis_boundary_start, axis_boundary_end) |
| 151 | + submeshes.append( |
| 152 | + jax.sharding.Mesh(mesh.devices[tuple(slices)], mesh.axis_names) |
| 153 | + ) |
| 154 | + axis_boundary_start = axis_boundary_end |
| 155 | + |
| 156 | + submeshes_tuple = tuple(submeshes) |
| 157 | + submesh_shardings = [ |
| 158 | + _get_per_mesh_shardings( |
| 159 | + submeshes_tuple, x.sharding.spec, x.sharding.memory_kind |
| 160 | + ) |
| 161 | + for x in flat_arrays |
| 162 | + ] |
| 163 | + |
| 164 | + flat_split_arrays = pw_jax.jaxlib_pathways._split_by_mesh_axis( # pylint: disable=protected-access |
| 165 | + arrays=flat_arrays, |
| 166 | + sharded_dim_idxs=sharded_dim_idxs, |
| 167 | + mesh_axis_sizes=mesh.axis_sizes, |
| 168 | + mesh_axis_idx=mesh_axis_idx, |
| 169 | + mesh_axis_sections=mesh_axis_sections, |
| 170 | + submesh_shardings=submesh_shardings, |
| 171 | + donate=donate, |
| 172 | + ) |
| 173 | + |
| 174 | + # Convert the flat arrays to a list of a PyTree per submesh. |
| 175 | + outer_treedef = jax.tree.structure(["*"] * len(flat_split_arrays)) |
| 176 | + inner_treedef = jax.tree.structure(["*"] * len(submeshes)) |
| 177 | + return [ |
| 178 | + jax.tree.unflatten(treedef, flat_submesh_arrays) |
| 179 | + for flat_submesh_arrays in jax.tree.transpose( |
| 180 | + outer_treedef, inner_treedef, flat_split_arrays |
| 181 | + ) |
| 182 | + ] |
0 commit comments