File size: 40,499 Bytes
4008bf9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 |
# Copyright 2022 The T5X 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.
"""Utilities for partitioning."""
import abc
import collections
import dataclasses
import typing
from typing import Any, Callable, Optional, Sequence, Tuple, Union
import cached_property
import jax
import numpy as np
from absl import logging
from flax import traverse_util
from flax.linen import partitioning as flax_partitioning
from jax import numpy as jnp
from jax import random
from jax.experimental import multihost_utils
from jax.experimental.mesh_utils import create_hybrid_device_mesh
from jax.experimental.pjit import pjit as jax_pjit
from jax.sharding import Mesh, PartitionSpec
JaxDevice = Any
TpuMesh = Tuple[int, int, int, int] # (x, y, z, num_cores).
OtherMesh = Tuple[int, int]
HardwareMesh = Union[TpuMesh, OtherMesh]
PyTreeDef = type(jax.tree_util.tree_structure(None))
TrainState = Any
LogicalAxisRules = Sequence[Tuple[str, Optional[str]]]
if typing.TYPE_CHECKING: # See b/163639353
cached_property = property # pylint: disable=invalid-name
else:
cached_property = cached_property.cached_property
class AxisNames(tuple):
"""Tuple of strings specifying name for each axis.
We create a separate class for this so JAX's pytree utilities can distinguish
it from a tuple that should be treated as a pytree, instead treating it as a
leaf.
"""
def __new__(cls, *names):
return tuple.__new__(AxisNames, names)
def __repr__(self):
return "AxisNames%s" % tuple.__repr__(self)
# pjit wrappers for cpu fallback.
# ----------------------------------------------------------------------------
# TODO(levskaya): This function is now no different than jax_pjit, but callers
# currently depend on `backend` argument
def pjit(
fun: Callable, # pylint: disable=g-bare-generic
in_axis_resources,
out_axis_resources,
static_argnums: Union[int, Sequence[int]] = (),
donate_argnums: Union[int, Sequence[int]] = (),
backend: Optional[str] = None,
):
"""Wrapper for pjit."""
del backend
return jax_pjit(
fun,
in_axis_resources,
out_axis_resources,
static_argnums=static_argnums,
donate_argnums=donate_argnums,
)
# pjit wrappers for cpu fallback.
# -----------------------------------------------------------------------------
# TODO(levskaya): upstream this fallback behavior to jax pjit.
def pjit_with_cpu_fallback(
fun: Callable, # pylint: disable=g-bare-generic
in_axis_resources,
out_axis_resources,
static_argnums: Union[int, Sequence[int]] = (),
donate_argnums: Union[int, Sequence[int]] = (),
backend: Optional[str] = None,
):
"""Wrapper for pjit that calls normal jit on cpu."""
if jax.devices(backend)[0].platform == "cpu":
return jax.jit(fun, static_argnums=static_argnums, donate_argnums=donate_argnums)
else:
return jax_pjit(
fun,
in_axis_resources,
out_axis_resources,
static_argnums=static_argnums,
donate_argnums=donate_argnums,
)
def with_sharding_constraint(x, axis_resources):
"""Wrapper for pjit with_sharding_constraint, no-op on cpu or outside pjit."""
if jax.devices()[0].platform == "cpu" or not global_mesh_defined():
return x
else:
return jax.experimental.pjit.with_sharding_constraint(x, axis_resources)
# pjit Mesh creation functions.
# -----------------------------------------------------------------------------
def bounds_from_last_device(last_device: JaxDevice) -> HardwareMesh:
"""Get the bound from the given last device."""
# Must be passed the device at the highest-coordinate corner of the
# relevant mesh, which is a requirement we know is satisfied by the last
# device in jax.devices().
if hasattr(last_device, "coords"):
x, y, z = last_device.coords
return x + 1, y + 1, z + 1, last_device.core_on_chip + 1
else:
# On non-TPU platforms, the "mesh" is hosts x devices per host in order
# to take advantage of faster within-host interconnect.
return jax.host_count(), jax.local_device_count()
def get_coords(device: JaxDevice) -> HardwareMesh:
"""Returns the coordinates of the given device."""
if hasattr(device, "coords"):
return (*device.coords, device.core_on_chip)
return (device.process_index, device.id % jax.local_device_count())
def global_mesh_defined():
"""Checks if global xmap/pjit mesh resource environment is defined."""
maps_env = jax.experimental.maps.thread_resources.env
return maps_env.physical_mesh.devices.shape != () # pylint: disable=g-explicit-bool-comparison
def get_mesh(
model_parallel_submesh: HardwareMesh,
input_devices: Sequence[JaxDevice] = (),
input_local_devices: Sequence[JaxDevice] = (),
tile_by_host_if_needed: bool = True,
backend: Optional[str] = None,
) -> Mesh:
"""Construct an xmap/pjit Mesh for the given model-parallel submesh.
The resulting mesh has two resource axes: 'model', with the provided submesh
shape, and 'data', which covers the rest of the mesh.
Args:
model_parallel_submesh: a HardwareMesh spec, namely (x,y,z,core) on TPU for
a single model-parallel replica's "tile" in the physical device mesh. The
first three elements (`x`, `y`, and `z`) should be factors of the pod
slice; e.g., if you are using df_4x8, then `x` should be a factor of 4
(one of 1, 2, 4), `y` should be a factor of 8 (one of 1, 2, 4, 8), and `z`
must be 1, because TPU v3 slices are only 2D. `z` can be >1 for TPU v4
(and maybe later TPUs) that allow 3D slices. `core` is the number of cores
to use from each TPU node. As communication is usually fastest inside the
same node, if you need a tile of more than 1 core, then
you should first increase `core`: e.g., for TPU v3, (1,1,1,2) is better
than (2,1,1,1). To pick a good spec, try a few possible values until you
get high TPU utilization.
input_devices: the devices to use, will use jax.devices() if this is not
set.
input_local_devices: the local devices to use, will use jax.local_devices()
if this is not set.
tile_by_host_if_needed: JAX currently requires that the parts of any sharded
array that are located on one host's local devices form a single
contiguous slice. A best effort will be made to achieve this without
"tiling" the device assignment over hosts (which can reduce XLA collective
performance). If this flag is True, then the device assignment will be
tiled over hosts if necessary to satisfy this constraint and create a
buildable mesh; if false, mesh construction will fail instead.
backend: get devices from the pinned backend, if specified. This is
useful for explicitly specifying the devices other than relying on
jax_platform_name.
Returns:
A xmap / pjit Mesh containing the virtual device mesh with data, model axes.
"""
input_devices = input_devices or jax.devices(backend)
input_local_devices = input_local_devices or jax.local_devices(0, backend)
# Sort input_devices based on coords, as backends might not return devices
# in order.
last_device = sorted(input_devices, key=get_coords)[-1]
last_input_local_devices = sorted(input_local_devices, key=get_coords)[-1]
logging.info(
"last device coords : %r\nlast local device coords: %r",
get_coords(last_device),
get_coords(last_input_local_devices),
)
global_hardware_mesh = bounds_from_last_device(last_device)
mesh_ndim = len(global_hardware_mesh)
local_hardware_mesh = bounds_from_last_device(last_input_local_devices)
mesh_err = (
f"each dimension of the model parallel submesh {model_parallel_submesh} "
"must be a factor of the corresponding dimension of the global device "
f"mesh {global_hardware_mesh}"
)
assert not any(g % m for g, m in zip(global_hardware_mesh, model_parallel_submesh)), mesh_err
assert not any(g % l for g, l in zip(global_hardware_mesh, local_hardware_mesh))
devices = np.empty(global_hardware_mesh, dtype=object)
for device in input_devices:
device_coords = get_coords(device)
devices[device_coords] = device
tile_by_host = tile_by_host_if_needed
if len(global_hardware_mesh) == 4:
# enable contiguous local chunks without host tiling by making Z major
global_hardware_mesh = typing.cast(Tuple[int, int, int, int], global_hardware_mesh)
model_parallel_submesh = typing.cast(Tuple[int, int, int, int], model_parallel_submesh)
gx, gy, gz, gc = global_hardware_mesh
mx, my, mz, mc = model_parallel_submesh
if (mx == gx > 1 and my == mz == 1) or (mx == 1 and my == gy > 1 and mz == gz > 1):
logging.info("ensuring YZ plane has a Z-major device order")
# YZ should be ZY
assert mc == gc, (mc, gc)
global_hardware_mesh = gx, gz, gy, gc
model_parallel_submesh = mx, mz, my, mc
devices = devices.swapaxes(1, 2)
tile_by_host = False
if (my == gy > 1 and mx == mz == 1) or (my == 1 and mx == gx > 1 and mz == gz > 1):
logging.info("ensuring XZ plane has a Z-major device order")
# XZ should be ZX
assert mc == gc, (mc, gc)
global_hardware_mesh = gz, gy, gx, gc
model_parallel_submesh = mz, my, mx, mc
devices = devices.swapaxes(0, 2)
tile_by_host = False
if tile_by_host:
logging.warning(
"Tiling device assignment mesh by hosts, which may lead to "
"reduced XLA collective performance. To avoid this, modify "
"the model parallel submesh or run with more tasks per host."
)
tile_err = (
"to tile the mesh by hosts, each dimension of the model parallel "
"submesh must be either a factor or a multiple of the corresponding "
"dimension of the per-host submesh"
)
def dh_dd_mh_md(g: int, m: int, l: int) -> Tuple[int, int, int, int]:
"""Split a global mesh dimension into four tiling components.
Args:
g: global mesh bounds dimension size
m: model-parallel submesh bounds dimension size
l: local submesh bounds dimension size
Returns:
The resulting tuple divides the dimension into the hosts component of
the data-parallel submesh, the devices component of the data-parallel
submesh, the hosts component of the model-parallel submesh, and the
devices component of the model-parallel submesh.
"""
d = g // m
if m >= l:
assert not m % l, tile_err
return (d, 1, m // l, l)
else:
assert not l % m, tile_err
return (d // (l // m), l // m, 1, m)
# e.g. [(x_data_hosts, x_data_devs, x_model_hosts, x_model_devs), ...]
dh_dd_mh_md_tups = map(
dh_dd_mh_md,
global_hardware_mesh,
model_parallel_submesh,
local_hardware_mesh,
)
# reshape to e.g. (x_dh, x_dd, x_mh, x_md, y_dh, ...)
devices = devices.reshape(*(s for t in dh_dd_mh_md_tups for s in t)) # pylint: disable=g-complex-comprehension
# TODO(jekbradbury): reorder local subgroups for ring locality
# Transpose to [data_host], [data_device], [model_host], [model_device]
# block ordering e.g. (x_dh, y_dh, ..., x_dd, y_dd, ...)
devices = devices.transpose(
*(4 * i for i in range(mesh_ndim)),
*(4 * i + 1 for i in range(mesh_ndim)),
*(4 * i + 2 for i in range(mesh_ndim)),
*(4 * i + 3 for i in range(mesh_ndim)),
)
else:
# e.g. [(x_data, x_model), (y_data, y_model), ...]
model_data_tups = [(g // m, m) for g, m in zip(global_hardware_mesh, model_parallel_submesh)]
# reshape to e.g. (x_data, x_model, y_data, y_model...)
devices = devices.reshape(*(s for t in model_data_tups for s in t)) # pylint: disable=g-complex-comprehension
# TODO(jekbradbury): reorder small subgroups for ring locality
# transpose to e.g. (x_data, y_data, ..., x_model, ...)
devices = devices.transpose(*(2 * i for i in range(mesh_ndim)), *(2 * i + 1 for i in range(mesh_ndim)))
# reshape to (data, model)
devices = devices.reshape(-1, np.prod(model_parallel_submesh))
global_mesh = Mesh(devices, ["data", "model"])
logging.info("global_mesh axis_names: %s", global_mesh.axis_names)
logging.info("global_mesh devices: %s", global_mesh.devices)
logging.info("global_mesh devices shape: %s", global_mesh.devices.shape)
return global_mesh
def get_cpu_mesh() -> Mesh:
"""Trivial mesh for CPU Testing."""
devices = np.empty((jax.host_count(), jax.local_device_count()), dtype=object)
for device in jax.devices():
devices[device.process_index, device.id % jax.local_device_count()] = device
return Mesh(devices, ["data", "model"])
def get_gpu_mesh(num_partitions: int) -> Mesh:
"""Mesh for GPUs that preferentially places 'model' on NVLink."""
nvlink_size = jax.local_device_count()
dcn_size = jax.process_count()
nvlink_mp = min(num_partitions, nvlink_size)
nvlink_dp, extra1 = divmod(nvlink_size, nvlink_mp)
dcn_mp, extra2 = divmod(num_partitions, nvlink_mp)
assert not (
extra1 or extra2
), "number of partitions on GPU must be a factor or multiple of the number of local devices"
dcn_dp = dcn_size // dcn_mp
devices = create_hybrid_device_mesh(
mesh_shape=[nvlink_dp, nvlink_mp],
dcn_mesh_shape=[dcn_dp, dcn_mp],
process_is_granule=True,
)
global_mesh = Mesh(devices, ["data", "model"])
logging.info("global_mesh axis_names: %s", global_mesh.axis_names)
logging.info("global_mesh devices: %s", global_mesh.devices)
return global_mesh
def default_mesh(
num_partitions: int,
model_parallel_submesh: Optional[HardwareMesh] = None,
backend: Optional[str] = None,
) -> Mesh:
"""Attempt to return a default mesh for simple cases.
Args:
num_partitions: number of partitions to use, will be ignored if
model_parallel_submesh is provided.
model_parallel_submesh: 4-tuple that specifies the x,y,z,c submesh to use as
the model-parallel device tile.
backend: get devices from the pinned backend, if specified. This is useful
for explicitly specifying the devices other than relying on
jax_platform_name.
Returns:
xmap/pjit 2D Mesh with 'data', 'model' mesh axes.
"""
last_device = jax.devices(backend)[-1]
platform = last_device.platform
device_kind = last_device.device_kind
bounds = bounds_from_last_device(last_device)
if model_parallel_submesh:
return get_mesh(model_parallel_submesh, backend=backend)
if platform == "cpu":
return get_cpu_mesh()
elif platform == "gpu":
return get_gpu_mesh(num_partitions)
mps = None
if device_kind in ("TPU v2", "TPU v3"):
if num_partitions == 1:
mps = (1, 1, 1, 1)
elif num_partitions == 2:
mps = (1, 1, 1, 2)
elif num_partitions == 4:
mps = (2, 1, 1, 2)
elif num_partitions == 8:
mps = (2, 2, 1, 2)
elif num_partitions == 16:
mps = (4, 2, 1, 2)
# assume the use of megacore on TPU v4
elif (device_kind == "TPU v4" or device_kind == "TPU v4 lite") and bounds[3] == 1:
if num_partitions == 1:
mps = (1, 1, 1, 1)
elif num_partitions == 2:
mps = (1, 2, 1, 1)
elif num_partitions == 4:
if bounds[0] >= 4:
mps = (4, 1, 1, 1)
else:
mps = (2, 2, 1, 1)
elif num_partitions == 8:
if bounds[2] >= 8:
mps = (1, 1, 8, 1)
else:
mps = (4, 2, 1, 1)
elif num_partitions == 16:
if bounds[2] >= 16:
mps = (1, 1, 16, 1)
elif bounds[0] >= 8:
mps = (8, 2, 1, 1)
elif bounds[0] >= 4:
mps = (4, 4, 1, 1)
else:
mps = (2, 2, 4, 1)
if mps is None:
raise ValueError(
"No default mesh for this configuration: specify " "config.model_parallel_submesh explicitly."
)
return get_mesh(mps, backend=backend)
# Data chunking helper.
# -----------------------------------------------------------------------------
@dataclasses.dataclass
class LocalChunkInfo:
# The logical slice of an array located on this host's local devices.
slice: Tuple[slice, ...]
# A unique index for this host/local chunk among chunks with the same slice.
replica_id: int
class LocalChunker:
"""Utility class to aid chunking of sharded arrays in multihost settings."""
def __init__(self, global_mesh: Mesh):
self.global_mesh = global_mesh
local_mesh = global_mesh.local_mesh
first_local_device = local_mesh.devices.reshape(-1)[0]
host_location = collections.OrderedDict(
zip(
global_mesh.shape.keys(),
list(zip(*np.nonzero(global_mesh.devices == first_local_device)))[0],
)
)
self.num_chunks = collections.OrderedDict()
self.chunk_ids = collections.OrderedDict()
self.mesh_axes = list(global_mesh.shape.keys())
for mesh_axis in self.mesh_axes:
num_devices_per_chunk = local_mesh.shape[mesh_axis]
self.num_chunks[mesh_axis] = global_mesh.shape[mesh_axis] // num_devices_per_chunk
self.chunk_ids[mesh_axis] = host_location[mesh_axis] // num_devices_per_chunk
def get_local_chunk_info(
self, global_shape: Tuple[int, ...], mesh_axes: Sequence[Optional[str]]
) -> LocalChunkInfo:
"""Get the local chunk info for a given array shape and sharded axes.
Args:
global_shape: the global, unsharded shape of the array to chunk.
mesh_axes: a sequence of names (or None) of equal rank to `global_shape`
that specifies which mesh dimensions the array is sharded along.
Returns:
LocalChunkInfo containing the logical slices of the array found on this
host's local devices, as well as the replica index for this chunk among
chunks with the same slice. The latter is used to determine which
host should write this chunk during checkpointing.
"""
local_slice = [slice(None) for dim in global_shape]
sharded_mesh_axes = set()
for i, (mesh_axis, size) in enumerate(zip(mesh_axes, global_shape)):
if not mesh_axis:
continue
sharded_mesh_axes.add(mesh_axis)
if not isinstance(mesh_axis, str):
raise NotImplementedError("TODO(jekbradbury)")
chunk_id = self.chunk_ids[mesh_axis]
chunk_size = size // self.num_chunks[mesh_axis]
local_slice[i] = slice(chunk_id * chunk_size, (chunk_id + 1) * chunk_size)
replicated_mesh_axes = [mesh_axis for mesh_axis in self.mesh_axes if mesh_axis not in sharded_mesh_axes]
replica_id = 0
for mesh_axis in replicated_mesh_axes:
chunk_id = self.chunk_ids[mesh_axis]
replica_id = replica_id * self.num_chunks[mesh_axis] + chunk_id
return LocalChunkInfo(tuple(local_slice), replica_id)
def standard_logical_axis_rules(
activation_partitioning_dims: int = 1,
parameter_partitioning_dims: int = 1,
additional_rules: Optional[LogicalAxisRules] = None,
) -> LogicalAxisRules:
"""Default sharding rules for T5X model in terms of logical axis names.
Args:
activation_partitioning_dims: enables 2-D activation sharding when set to 2.
parameter_partitioning_dims: enables 2-D parameter sharding when set to 2.
additional_rules: additional rules (a sequence of tuples) that will be
appended to the standard rules.
Returns:
Sequence of logical axis rules
"""
logging.info(
"`activation_partitioning_dims` = %d, `parameter_partitioning_dims` = %d",
activation_partitioning_dims,
parameter_partitioning_dims,
)
if activation_partitioning_dims == 1 and parameter_partitioning_dims == 1:
rules = [
("batch", "data"),
("vocab", "model"),
("embed", None),
("mlp", "model"),
("heads", "model"),
("kv", None),
("joined_kv", "model"), # joined heads+kv dim in 2D attn param layouts
]
elif activation_partitioning_dims == 2 and parameter_partitioning_dims == 1:
rules = [
("batch", "data"),
("vocab", "model"),
("mlp", "model"),
("heads", "model"),
("kv", None),
("joined_kv", "model"),
("embed", "model"),
]
elif activation_partitioning_dims == 1 and parameter_partitioning_dims == 2:
rules = [
("batch", "data"),
("vocab", "model"),
("mlp", "model"),
("heads", "model"),
("kv", None),
("joined_kv", "model"),
("embed", "data"),
]
elif activation_partitioning_dims == 2 and parameter_partitioning_dims == 2:
rules = [
("batch", "data"),
("vocab", "model"),
("mlp", "model"),
("heads", "model"),
("kv", None),
("joined_kv", "model"),
("embed", "model"),
("embed", "data"),
]
else:
raise ValueError(
f"`activation_partitioning_dims` = {activation_partitioning_dims} "
f"`parameter_partitioning_dims` = {parameter_partitioning_dims} "
"is not supported."
)
# Add the common rules for the replicated logical axes names.
replicated_rules = [
("relpos_buckets", None),
("abspos_buckets", None),
("length", None),
("layers", None),
("stack", None),
("mlp_activations", None),
]
rules.extend(replicated_rules)
if additional_rules:
rules.extend(additional_rules)
return rules
# NB: This needs to be top-level for the jax compilation cache.
def _id_fn(x, ix):
"""Identity function for copying parameters to the devices, sharded."""
# A pure identity such as `lambda x, *: x` can get optimized away, so we
# include a random.split as a cheap function that cannot be optimized away.
y = random.split(random.PRNGKey(jnp.array(ix, dtype=jnp.uint32)))
return x, y
@dataclasses.dataclass
class DataLayout:
"""Represents data layout for the partitioned model."""
batch_size: int
shard_id: int
num_shards: int
is_first_host_in_replica_set: bool
PartitionedCallable = Callable[..., Any]
CompiledPartitionedCallable = Callable[..., Any]
class BasePartitioner(metaclass=abc.ABCMeta):
"""Interface for partitioning computations across hardware devices."""
def __init__(
self,
num_partitions: Optional[int] = None,
model_parallel_submesh: Optional[HardwareMesh] = None,
params_on_devices: bool = True,
backend: Optional[str] = None,
):
"""Configures the partitioner.
Args:
num_partitions: the number of partitions to use. Ignored if
`model_parallel_submesh` is provided.
model_parallel_submesh: 4-tuple that specifies the x,y,z,c submesh to use
as the model-parallel device tile. This submesh is used for the larger
of the two parameter dimensions, and, if 2-D activation sharding is
enabled, for the model dimension of activations. The rest of the mesh is
used for data parallelism and, if 2-D parameter sharding is enabled, the
other parameter dimension.
params_on_devices: whether to keep the params on devices, if False -
params stay in the host memory. Note that some partitioners might ignore
this setting, for example if they don't support storing all params on
device memory.
backend: get devices from the pinned backend, if specified. This is useful
for explicitly specifying the devices other than relying on
jax_platform_name.
"""
if not num_partitions and not model_parallel_submesh:
raise ValueError("At least one of `num_partitions` or " "`model_parallel_submesh` must be set.")
if model_parallel_submesh is not None and len(model_parallel_submesh) != 4:
logging.error(
(
"`model_parallel_submesh` must be either None or a 4-tuple. Got"
" `model_parallel_submesh`=%s. A ValueError will be raised"
" beginning March 1, 2022."
),
model_parallel_submesh,
)
if bool(num_partitions) and bool(model_parallel_submesh):
logging.error(
(
"At most one of `num_partitions` or `model_parallel_submesh` can be"
" set. Got `num_partitions=%s` and `model_parallel_submesh`=%s. A"
" ValueError will be raised beginning March 21, 2022."
),
num_partitions,
model_parallel_submesh,
)
self._num_partitions = num_partitions
self._model_parallel_submesh = model_parallel_submesh
self._params_on_devices = params_on_devices
self._data_axis = "data"
self._backend = backend
@property
def mesh(self) -> Mesh:
raise NotImplementedError
@property
def data_partition_spec(self) -> PartitionSpec:
return PartitionSpec(self._data_axis)
def get_data_layout(self, batch_size: Optional[int] = None, host_index: Optional[int] = None) -> DataLayout:
"""Returns filled `DataLayout` based on the partitioned model layout.
Args:
batch_size: if set, indicates the requested batch size. The exception will
be raised if this batch size is not compatible with the layout. If not
set, the batch size is inferred from the layout.
host_index: indicates the host index to use for the calculations, if not
set - use JAX-provided one. Should be in [0, num_hosts) interval and the
order should match the order of corresponding CPU devices in
`jax.devices()`.
Returns:
Filled `DataLayout` structure.
"""
if host_index is not None:
raise NotImplementedError("Explicit host_index is not yet implemented.")
if self._data_axis is None:
return DataLayout(
batch_size=batch_size,
shard_id=0,
num_shards=1,
is_first_host_in_replica_set=(jax.process_index() == 0),
)
mesh_size = self._local_chunker.global_mesh.shape[self._data_axis]
batch_size = batch_size or mesh_size
if batch_size % mesh_size:
raise ValueError(
f"Batch size ({batch_size}) must be divisible by corresponding " f"mesh size ({mesh_size})."
)
num_shards = self._local_chunker.num_chunks[self._data_axis]
if batch_size % num_shards:
raise ValueError(f"Batch size ({batch_size}) must be divisible by number of " f"replicas ({num_shards}).")
replica_id = self._local_chunker.get_local_chunk_info((batch_size,), [self._data_axis]).replica_id
return DataLayout(
batch_size=int(batch_size),
shard_id=int(self._local_chunker.chunk_ids[self._data_axis]),
num_shards=int(num_shards),
is_first_host_in_replica_set=(replica_id == 0),
)
def get_local_chunk_info(
self, global_shape: Tuple[int, ...], mesh_axes: Sequence[Optional[str]]
) -> LocalChunkInfo:
"""Returns the local chunk info for a given array shape and sharded axes."""
return self._local_chunker.get_local_chunk_info(global_shape, mesh_axes)
@property
def params_on_devices(self):
return self._params_on_devices
def move_params_to_devices(self, train_state: TrainState, train_state_axes: TrainState) -> TrainState:
"""Moves the optimizer parameters to devices."""
p_id_fn = self.partition(
_id_fn,
in_axis_resources=(train_state_axes, None),
out_axis_resources=(train_state_axes, None),
donate_argnums=(0,),
)
if jax.config.jax_array and jax.process_count() > 1:
train_state = multihost_utils.host_local_array_to_global_array(train_state, self.mesh, train_state_axes)
train_state, _ = p_id_fn(train_state, jnp.ones((), dtype=jnp.uint32))
return train_state
@property
@abc.abstractmethod
def _local_chunker(self):
"""Returns the chunker that matches the parameters of this partitioner."""
raise NotImplementedError
def get_logical_axes(self, train_state: TrainState) -> TrainState:
"""Returns a copy of TrainState with Optional[AxisNames] as leaves."""
# By default, return None for the logical axes.
return train_state.restore_state(jax.tree_map(lambda x: None, train_state.state_dict()))
def get_mesh_axes(self, train_state: TrainState) -> TrainState:
"""Returns a copy of TrainState with Optional[PartitionSpecs] as leaves."""
raise NotImplementedError
@abc.abstractmethod
def partition(
self,
fn: Callable, # pylint: disable=g-bare-generic
in_axis_resources,
out_axis_resources,
static_argnums: Union[int, Sequence[int]] = (),
donate_argnums: Union[int, Sequence[int]] = (),
) -> PartitionedCallable:
"""Partitions the computation using partitioner-specific implementation.
Args:
fn: the function to partition.
in_axis_resources: Pytree of structure matching that of arguments to `fn`,
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 valid resource assignment specifications are:
`None`: in which case the value will be replicated on all devices
`PartitionSpec`: a tuple of length at most equal to the rank of the
partitioned value. Each element can be a `None`, a mesh axis or a
tuple of mesh axes, and specifies the set of resources assigned to
partition the value's dimension matching its position in the spec.
out_axis_resources: Like `in_axis_resources`, but specifies resource
assignment for function outputs.
static_argnums: an optional int or collection of ints that specify which
positional arguments to treat as static (compile-time constant) in the
partitioned function.
donate_argnums: an optional int or collection of ints that specify which
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.
Returns:
A partitioned version of the input function.
"""
raise NotImplementedError
@abc.abstractmethod
def compile(self, partitioned_fn: PartitionedCallable, *args) -> CompiledPartitionedCallable:
"""Compiles and returns the partitioned function, or the original.
Args:
partitioned_fn: The partitioned function.
*args: Sample arguments to the partitioned function matching the input
shapes that will be passed to the compiled function.
Returns:
The compiled function, or the original if this partitioner does not
support compilation.
"""
raise NotImplementedError
class PjittedFnWithContext(PartitionedCallable):
"""Wraps pjitted function to apply the appropriate contexts."""
def __init__(
self,
pjitted_fn,
partition_mesh: Mesh,
logical_axis_rules: flax_partitioning.LogicalRules = (),
):
self._pjitted_fn = pjitted_fn
self._mesh = partition_mesh
self._logical_axis_rules = logical_axis_rules
def __call__(self, *args):
with Mesh(self._mesh.devices, self._mesh.axis_names), flax_partitioning.axis_rules(self._logical_axis_rules):
return self._pjitted_fn(*args)
def lower(self, *args):
with Mesh(self._mesh.devices, self._mesh.axis_names), flax_partitioning.axis_rules(self._logical_axis_rules):
return self._pjitted_fn.lower(*args)
class BasePjitPartitioner(BasePartitioner):
"""Partitioner that uses T5X version of jax.pjit."""
@cached_property
def _local_chunker(self) -> LocalChunker:
return LocalChunker(self.mesh)
@cached_property
def mesh(self) -> Mesh:
return default_mesh(self._num_partitions, self._model_parallel_submesh, self._backend)
def partition(
self,
fn: Callable, # pylint: disable=g-bare-generic
in_axis_resources,
out_axis_resources,
static_argnums: Union[int, Sequence[int]] = (),
donate_argnums: Union[int, Sequence[int]] = (),
) -> PjittedFnWithContext:
pjitted = pjit(
fn,
in_axis_resources=in_axis_resources,
out_axis_resources=out_axis_resources,
static_argnums=static_argnums,
donate_argnums=donate_argnums,
backend=self._backend,
)
return PjittedFnWithContext(pjitted, self.mesh)
def compile(self, partitioned_fn: PjittedFnWithContext, *args) -> CompiledPartitionedCallable:
return partitioned_fn.lower(*args).compile()
class PjitPartitioner(BasePjitPartitioner):
"""Partitioner that uses named axes and jax.pjit."""
def __init__(
self,
num_partitions: Optional[int] = None,
model_parallel_submesh: Optional[HardwareMesh] = None,
params_on_devices: bool = True,
backend: Optional[str] = None,
logical_axis_rules: Optional[LogicalAxisRules] = None,
use_cpu_pjit: Optional[bool] = False,
):
"""PjitPartitioner constructor.
See https://github.com/google-research/text-to-text-transfer-transformer/blob/main/README.mdx/usage/partitioning for details.
Args:
num_partitions: an integer that specifies the size of the model parallel
submesh to be automatically selected for the current topology. See
`model_parallel_submesh` for details on how this submesh is used.
Mutually exlusive with `model_parallel_submesh`.
model_parallel_submesh: is a 4-tuple that specifies the `(x, y, z, c)`
submesh model-parallel device tile, an axis of accelerator parallelism
orthogonal to data parallelism. Array axes in a model's parameters or
activations can be sharded over this submesh using axis rules (see
`logical_axis_rules`) that map them to 'model'. The effective number of
model sub-partitions is equal to `np.prod(model_parallel_submesh)` and
must evenly divide the total number of devices (i.e.,
`jax.device_count() % np.prod(model_parallel_submesh) == 0`). The rest
of the TPU mesh is the data parallel submesh, providing
`jax.device_count() // np.prod(model_parallel_submesh)` partitions. It
is used for data (batch) parallelism and to shard other array axes that
are mapped to 'data'. This argument is mutually exclusive with
`num_partitions`.
params_on_devices: whether to keep the params on devices, if False -
params stay in the host memory. Note that some partitioners might ignore
this setting, for example if they don't support storing all params on
device memory.
backend: get devices from the pinned backend, if specified. This is
useful for explicitly specifying the devices other than relying on
jax_platform_name.
logical_axis_rules: a priority-ordered sequence of KV tuples that maps
logical axis names to either `None` (not sharded), 'model' (to shard
across the model-parallel submesh), or 'data' (to shard across the
data-parallel submesh).
use_cpu_pjit: enables wrapper function for pjit which just jits the
function if using CPU backend.
"""
super().__init__(
num_partitions=num_partitions,
model_parallel_submesh=model_parallel_submesh,
params_on_devices=params_on_devices,
backend=backend,
)
if logical_axis_rules is None:
logical_axis_rules = standard_logical_axis_rules()
self._logical_axis_rules = tuple(logical_axis_rules)
(self._data_axis,) = flax_partitioning.logical_to_mesh_axes(["batch"], logical_axis_rules)
self._use_cpu_pjit = use_cpu_pjit
def partition(
self,
fn: Callable, # pylint: disable=g-bare-generic
in_axis_resources,
out_axis_resources,
static_argnums: Union[int, Sequence[int]] = (),
donate_argnums: Union[int, Sequence[int]] = (),
) -> PjittedFnWithContext:
"""Partitions the function using jax.pjit."""
if self._use_cpu_pjit:
pjit_fn = pjit_with_cpu_fallback
else:
pjit_fn = pjit
pjitted = pjit_fn(
fn,
in_axis_resources=in_axis_resources,
out_axis_resources=out_axis_resources,
static_argnums=static_argnums,
donate_argnums=donate_argnums,
backend=self._backend,
)
return PjittedFnWithContext(pjitted, self.mesh, self._logical_axis_rules)
@property
def logical_axis_rules(self):
"""Returns the logical axis rules."""
return self._logical_axis_rules
def get_logical_axes(self, train_state: TrainState) -> TrainState:
"""Returns a copy of TrainState with Optional[AxisNames] as leaves."""
return train_state.as_logical_axes()
def get_mesh_axes(self, train_state: TrainState) -> TrainState:
"""Returns a copy of TrainState with Optional[PartitionSpecs] as leaves."""
logical_axes = self.get_logical_axes(train_state)
def _logical_to_mesh_axes(param_name, logical_axes):
if logical_axes is None:
return None
elif logical_axes is traverse_util.empty_node:
return traverse_util.empty_node
try:
return flax_partitioning.logical_to_mesh_axes(logical_axes, self._logical_axis_rules)
except ValueError as e:
raise ValueError(f"Failed to map logical axes for {param_name}") from e
flat_logical_axes = traverse_util.flatten_dict(logical_axes.state_dict(), keep_empty_nodes=True, sep="/")
flat_mesh_axes = {k: _logical_to_mesh_axes(k, v) for k, v in flat_logical_axes.items()}
return logical_axes.restore_state(traverse_util.unflatten_dict(flat_mesh_axes, sep="/"))
|