collective.py 98.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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 numpy as np
import os
17 18
import pickle
import io
19 20
import datetime
import time
21
from ..fluid.layer_helper import LayerHelper
22
from ..fluid.framework import Variable
23
from ..fluid.framework import in_dygraph_mode
24
from ..fluid.framework import OpProtoHolder
J
Jiabin Yang 已提交
25
from ..fluid.framework import _non_static_mode
26
from ..fluid.framework import _in_legacy_dygraph
27
from ..fluid.framework import convert_np_dtype_to_dtype_
J
Jiangxinz 已提交
28
from ..fluid.framework import _varbase_creator
29 30 31 32
from ..fluid.data_feeder import convert_dtype
from ..fluid.data_feeder import check_variable_and_dtype
from ..fluid.data_feeder import check_type
from ..fluid.data_feeder import check_dtype
33 34
from ..fluid.layers.tensor import fill_constant
from ..fluid.layers import utils
B
Baibaifan 已提交
35
from ..fluid.dygraph import layers
36 37 38 39
from ..fluid.dygraph.parallel import prepare_context
import paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
40
from paddle import _C_ops, _legacy_C_ops
J
Jiangxinz 已提交
41
import paddle.fluid.dygraph_utils as dygraph_utils
42
import contextlib
43

44
__all__ = []
45 46 47


class ReduceOp:
L
lilong12 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
    """
    Specify the type of operation used for element-wise reductions.
    It should be one of the following values:

        ReduceOp.SUM

        ReduceOp.MAX

        ReduceOp.MIN

        ReduceOp.PROD

    Examples:
        .. code-block:: python

63
            # required: distributed
L
lilong12 已提交
64
            import paddle
65
            import paddle.distributed as dist
L
lilong12 已提交
66

67 68 69
            dist.init_parallel_env()
            if dist.get_rank() == 0:
                data = paddle.to_tensor([[4, 5, 6], [4, 5, 6]])
L
lilong12 已提交
70
            else:
71 72 73 74
                data = paddle.to_tensor([[1, 2, 3], [1, 2, 3]])
            dist.all_reduce(data, op=dist.ReduceOp.SUM)
            print(data)
            # [[5, 7, 9], [5, 7, 9]] (2 GPUs)
L
lilong12 已提交
75
    """
76 77 78 79
    SUM = 0
    MAX = 1
    MIN = 2
    PROD = 3
80
    AVG = 4
81 82


K
kuizhiqing 已提交
83 84 85 86
class Group():
    """
    The abstract representation of group.
    """
87

88
    def __init__(self, rank, rank_num, id=0, ranks=[], pg=None, name=None):
89 90
        self.rank = rank
        self.nranks = rank_num
K
kuizhiqing 已提交
91 92
        self.id = id
        self.ranks = ranks
93 94
        self.pg = pg
        self.name = name
K
kuizhiqing 已提交
95 96 97 98 99 100 101 102 103 104 105 106 107 108

    def is_member(self):
        if self.rank < 0:
            return False
        if self.nranks < 2:
            return False
        return True

    def get_group_rank(self, rank):
        if self.is_member() and rank in self.ranks:
            return self.ranks.index(rank)
        else:
            return -1

109 110 111 112
    @property
    def process_group(self):
        return self.pg

L
LiYuRio 已提交
113 114 115 116
    @property
    def world_size(self):
        return self.nranks if self.rank >= 0 else -1

117 118 119 120
    def __repr__(self):
        debug_str = "rank: {}, nranks: {}, id: {}, ranks: ".format(
            self.rank, self.nranks, self.id)
        debug_str += ", ".join(map(str, self.ranks))
121 122
        debug_str += "; name: "
        debug_str += self.name if self.name else "None"
123 124
        return debug_str

K
kuizhiqing 已提交
125 126 127 128 129 130 131 132 133 134 135 136 137 138

_global_env = None


def _get_global_env():
    global _global_env
    if not _global_env:
        _global_env = paddle.distributed.ParallelEnv()
    return _global_env


# group map : the map of all group, 0 for GlobalGroup
# Dict[int, Group]
_group_map = {}
139
_global_env_gid = 0
K
kuizhiqing 已提交
140

141 142 143 144
# group map by name : the map of all groups from their names
# Dict[name, Group]
_group_map_by_name = {}

145 146 147 148
# backend map by group : the map of all backend from their groups
# Dict[group, backend]
_group_map_backend = {}

149 150 151
# Name of the default group for init_parallel_env
_default_group_name = "_default_pg"

152
_valid_backend_list = ['nccl', 'gloo', 'hccl', 'heter', 'xccl']
153 154
_default_store = None  # the default tcp store
_default_backend = None
155 156
_default_timeout = datetime.timedelta(seconds=1800)
_start_ring_id = 0
157

K
kuizhiqing 已提交
158

L
lilong12 已提交
159 160 161 162 163 164 165 166 167 168
def _set_default_backend(backend):
    global _default_backend
    _default_backend = backend


def _set_default_store(store):
    global _default_store
    _default_store = store


K
kuizhiqing 已提交
169 170
def _get_group_map():
    global _group_map
171
    if _global_env_gid not in _group_map:
K
kuizhiqing 已提交
172
        genv = _get_global_env()
173 174 175
        _group_map[_global_env_gid] = Group(genv.rank,
                                            genv.world_size,
                                            ranks=list(range(genv.world_size)))
K
kuizhiqing 已提交
176 177 178 179
    return _group_map


def _get_global_group():
180
    return _get_group_map()[_global_env_gid]
K
kuizhiqing 已提交
181 182


183 184 185 186 187 188
def _get_group_map_by_name():
    global _group_map_by_name
    return _group_map_by_name


def _get_default_group():
L
lilong12 已提交
189
    global _group_map_by_name
190 191
    assert is_initialized(), ("Call paddle.distributed.init_parallel_env first "
                              "to initialize the distributed environment.")
192 193 194
    return _get_group_map_by_name()[_default_group_name]


L
lilong12 已提交
195 196 197 198 199 200 201 202 203 204 205 206
def _set_group_map(gid, group):
    global _group_map
    assert gid not in _group_map
    _group_map[gid] = group


def _set_group_map_by_name(name, group):
    global _group_map_by_name
    assert name not in _group_map_by_name
    _group_map_by_name[name] = group


207 208 209 210 211 212
def _set_group_map_backend(group, backend):
    global _group_map_backend
    assert group not in _group_map_backend
    _group_map_backend[group] = backend


K
kuizhiqing 已提交
213
def _new_ring_id():
214 215 216 217 218 219 220
    # NOTE(liyurui): For compatible reason, auto parallel and eager mode relay on previous syntax.
    if in_dygraph_mode():
        global _start_ring_id
        _start_ring_id += 1
        return _start_ring_id + max(_get_global_env().nrings, 9)
    else:
        return len(_get_group_map()) + max(_get_global_env().nrings, 9)
K
kuizhiqing 已提交
221 222


223 224 225 226 227 228 229 230 231 232 233 234 235
def _get_reduce_op(reduce_op, func_name):
    if reduce_op == ReduceOp.SUM:
        return core.ReduceOp.SUM
    elif reduce_op == ReduceOp.MAX:
        return core.ReduceOp.MAX
    elif reduce_op == ReduceOp.MIN:
        return core.ReduceOp.MIN
    elif reduce_op == ReduceOp.PROD:
        return core.ReduceOp.PRODUCT
    else:
        raise ValueError("Unknown reduce_op type for {}.".format(func_name))


K
kuizhiqing 已提交
236 237 238 239 240 241
def get_group(id=0):
    """

    Get group instance by group id.

    Args:
K
kuizhiqing 已提交
242
        id (int): the group id. Default value is 0.
K
kuizhiqing 已提交
243 244 245 246 247 248 249 250 251 252 253 254 255 256

    Returns:
        Group: the group instance.

    Examples:
        .. code-block:: python

            ...
            gid = paddle.distributed.new_group([2,4,6])
            paddle.distributed.get_group(gid.id)

    """

    gm = _get_group_map()
J
Jiangxinz 已提交
257
    return gm[id] if id in gm else None
K
kuizhiqing 已提交
258 259


260 261 262 263 264 265
def _new_process_group_impl(backend,
                            store,
                            rank,
                            world_size,
                            group_name,
                            pg_options,
L
lilong12 已提交
266 267 268
                            group_id=0,
                            src_rank=None,
                            dst_rank=None):
269
    pg = None
270
    genv = _get_global_env()
L
lilong12 已提交
271 272 273 274
    if backend != 'heter':
        assert src_rank is None and dst_rank is None, (
            "src_rank and dst_rank "
            "can only be set for heter backend.")
L
lilong12 已提交
275
    assert backend in _valid_backend_list, "Unsupported backend: %s." % backend
276
    if backend == "gloo":
277 278
        place = core.CPUPlace()
        pg = core.ProcessGroupGloo(store, rank, world_size, place, group_id)
279
    elif backend == "nccl":
280 281
        place = core.CUDAPlace(genv.device_id)
        pg = core.ProcessGroupNCCL(store, rank, world_size, place, group_id)
282
    elif backend == "hccl":
283 284
        place = core.NPUPlace(genv.device_id)
        pg = core.ProcessGroupHCCL(store, rank, world_size, place, group_id)
285 286 287
    elif backend == "xccl":
        place = core.CustomPlace(genv.device_type, genv.device_id)
        pg = core.ProcessGroupCustom(store, rank, world_size, place, group_id)
288
    elif backend == "heter":
289 290 291 292 293
        place = None
        if core.is_compiled_with_cuda():
            place = core.CUDAPlace(genv.device_id)
        elif core.is_compiled_with_npu():
            place = core.NPUPlace(genv.device_id)
294 295 296 297 298 299 300 301 302 303 304 305 306
        cluster_id = int(os.getenv("CLUSTER_ID", "-1"))
        assert cluster_id >= 0, "please set the CLUSTER_ID variable."
        cluster_size = os.getenv("CLUSTER_SIZE", None)
        assert cluster_size, "please set the CLUSTER_SIZE variable."
        cluster_size = cluster_size.split(",")
        cluster_size = [int(s) for s in cluster_size]
        switch_ep = os.getenv("CLUSTER_SWITCH", None)
        assert switch_ep, "please set the CLUSTER_SWITCH variable."
        cluster_size_cumsum = np.cumsum(cluster_size)
        cluster_offset = 0 if cluster_id == 0 else cluster_size_cumsum[
            cluster_id - 1]
        global_rank = cluster_offset + rank
        global_world_size = cluster_size_cumsum[-1]
307
        global_rank, global_world_size = _get_global_config(backend, rank)
308 309 310 311 312 313 314 315 316 317 318 319 320
        pg = core.ProcessGroupHeter(store,
                                    rank=global_rank,
                                    world_size=global_world_size,
                                    place=place,
                                    gid=group_id,
                                    local_rank=rank,
                                    local_size=world_size,
                                    gloo_rank=cluster_id,
                                    gloo_size=len(cluster_size),
                                    with_switch=True,
                                    switch_endpoint=switch_ep,
                                    src_rank=src_rank,
                                    dst_rank=dst_rank)
321 322 323 324

    return pg


S
ShenLiang 已提交
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
def barrier(group=None):
    """

    Barrier among all participators in the group.

    Args:
        group (Group): The group instance return by new_group or None for global default group.

    Returns:
        None.

    Examples:
        .. code-block:: python

            import paddle
            from paddle.distributed import init_parallel_env

            paddle.set_device('gpu:%d'%paddle.distributed.ParallelEnv().dev_id)
            init_parallel_env()
            paddle.distributed.barrier()
    """
    if group is not None and not group.is_member():
        return

L
lilong12 已提交
349
    if in_dygraph_mode():
350 351 352 353 354
        group = _get_default_group() if group is None else group
        task = group.process_group.barrier()
        task.wait()
        return

S
ShenLiang 已提交
355 356 357
    ring_id = 0 if group is None else group.id

    temp = fill_constant([1], dtype="int32", value="1")
J
Jiabin Yang 已提交
358
    if _non_static_mode():
359
        return _legacy_C_ops.barrier(temp, temp, 'ring_id', ring_id)
W
wanghuancoder 已提交
360 361 362

    op_type = 'barrier'

S
ShenLiang 已提交
363 364 365
    if not isinstance(ring_id, int):
        raise ValueError("The type of 'group' for barrier must be int.")
    helper = LayerHelper(op_type, **locals())
366 367 368 369
    helper.append_op(type=op_type,
                     inputs={'X': [temp]},
                     outputs={'Out': [temp]},
                     attrs={'ring_id': ring_id})
S
ShenLiang 已提交
370 371


L
lilong12 已提交
372 373 374 375 376 377 378
# _custom_gid provides a way for users to
# set the group id, which is usually useful
# to be compatible with the static mode.
_custom_gid = None


def _set_custom_gid(gid):
379
    global _custom_gid
L
lilong12 已提交
380 381 382
    _custom_gid = gid


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
def _barrier_by_tcp_store(group_name, store, timeout):
    global_rank = paddle.distributed.get_rank()
    global_world_size = paddle.distributed.get_world_size()

    if global_world_size < 2:
        return

    barrier_prefix = "Barrier/" + group_name + "/"
    is_master = (global_rank == 0)

    def _check_keys_ready(wait_keys):
        start_time = time.time()
        while len(wait_keys) > 0:
            time.sleep(0.1)
            elapse_time = time.time() - start_time
            if datetime.timedelta(seconds=elapse_time) > timeout:
                raise RuntimeError(
                    "Timeout while initializing process group {}."
                    "Keys {} are not ready sinck rank {} is waiting them."
                    "Two reason may cause this error:\n 1. The create process group api should be called by all ranks.\n"
                    " 2. Try to increase the waiting time.\n".format(
                        group_name, wait_keys, global_rank))
            wait_keys = list(
                filter(lambda key: int(store.get(key)) != 1, wait_keys))

    # all the workers set their exiting key and exit
    # the master will wait for all workers' exiting key, ensure to exit in the end
    if is_master:
        wait_keys = [
            barrier_prefix + str(rank) for rank in range(1, global_world_size)
        ]
        _check_keys_ready(wait_keys)
    else:
        store.add(barrier_prefix + str(global_rank), 1)


def new_group(ranks=None, backend=None, timeout=_default_timeout):
K
kuizhiqing 已提交
420 421
    """

K
kuizhiqing 已提交
422
    Creates a new distributed communication group.
K
kuizhiqing 已提交
423 424

    Args:
K
kuizhiqing 已提交
425
        ranks (list): The global ranks of group members.
K
kuizhiqing 已提交
426
        backend (str): The backend used to create group, only nccl is supported now.
427
        timeout (datetime.timedelta, optional): The waiting timeout for store relevant options, default is 30 minutes.
K
kuizhiqing 已提交
428 429

    Returns:
K
kuizhiqing 已提交
430
        Group: The group instance.
K
kuizhiqing 已提交
431 432 433 434 435 436 437

    Examples:
        .. code-block:: python

            import paddle

            paddle.distributed.init_parallel_env()
K
kuizhiqing 已提交
438 439 440
            tindata = paddle.randn(shape=[2, 3])
            gp = paddle.distributed.new_group([2,4,6])
            paddle.distributed.all_reduce(tindata, group=gp, use_calc_stream=False)
K
kuizhiqing 已提交
441 442

    """
443
    global _custom_gid
444
    global _group_map
L
lilong12 已提交
445
    if in_dygraph_mode():
446
        global _default_group_name
L
lilong12 已提交
447
        gid = _custom_gid if _custom_gid else _new_ring_id()
448
        group_name = _default_group_name + str(gid)
L
lilong12 已提交
449
        if backend != 'heter' and (ranks is None or len(ranks) > 1):
450 451 452 453 454 455 456 457 458
            global_group = _get_default_group()
            global_rank = global_group.rank
            global_ranks = global_group.ranks
            backend = _default_backend if backend is None else backend
            if ranks is None:
                ranks = global_ranks
            assert len(ranks) <= len(global_ranks), (
                "Size of new group must be less than or "
                "equal to that of the default global group.")
459 460
        size = len(ranks)
        ranks = sorted(ranks)
L
lilong12 已提交
461 462 463 464
        if backend == 'heter' or (size > 1 and global_rank in ranks):
            rank = 0 if backend == 'heter' else ranks.index(global_rank)
            src_rank = ranks[0] if backend == 'heter' else None
            dst_rank = ranks[1] if backend == 'heter' else None
465 466 467 468 469 470 471 472 473
            pg = _new_process_group_impl(backend,
                                         _default_store,
                                         rank,
                                         size,
                                         group_name,
                                         pg_options=None,
                                         group_id=gid,
                                         src_rank=src_rank,
                                         dst_rank=dst_rank)
474 475 476 477 478 479
        else:
            rank = -1
            pg = None
        group = Group(rank, size, id=gid, ranks=ranks, pg=pg, name=group_name)
        _group_map_by_name[group_name] = group
        _group_map[gid] = group
480
        _group_map_backend[group] = backend
481

482
        # TODO(shenliang03): This is a temporary solution to solve the problem of
483
        # hang caused by tcp
484
        paddle.distributed.barrier(group=group)
485 486 487 488 489
        # NOTE(liyurui): All processors should hang and wait using tcp store, in case master exit before sub-group is created.
        if backend != 'heter':
            _barrier_by_tcp_store(group_name, _default_store, timeout)
        else:
            print("Warning: store barrier is not supported for heter backend.")
490
        return group
K
kuizhiqing 已提交
491 492 493 494 495 496 497 498 499 500 501 502 503 504

    if not backend:
        backend = 'nccl'
    assert backend == 'nccl', ("backend other than nccl is not supported yet")

    genv = _get_global_env()
    global_rank = genv.rank

    ring_id = _new_ring_id()

    if global_rank not in ranks:
        gp = Group(-1, -1, ring_id, ranks)
        _group_map[ring_id] = gp
    else:
505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
        ranks = sorted(ranks)
        group_rank = ranks.index(global_rank)
        group_size = len(ranks)
        gp = Group(group_rank, group_size, ring_id, ranks)
        _group_map[ring_id] = gp

        if group_size >= 2:
            strategy = core.ParallelStrategy()
            strategy.nranks = group_size
            strategy.local_rank = group_rank
            strategy.trainer_endpoints = [
                genv.trainer_endpoints[i] for i in ranks
            ]
            strategy.current_endpoint = genv.current_endpoint
            strategy.nrings = 1

            if core.is_compiled_with_cuda():
                place = core.CUDAPlace(genv.device_id)
                core.NCCLParallelContext(strategy,
                                         place).init_with_ring_id(ring_id)
525 526 527 528
            elif core.is_compiled_with_npu():
                place = core.NPUPlace(genv.device_id)
                core.HCCLParallelContext(strategy,
                                         place).init_with_ring_id(ring_id)
529 530 531 532
            elif core.is_compiled_with_mlu():
                place = core.MLUPlace(genv.device_id)
                core.CNCLParallelContext(strategy,
                                         place).init_with_ring_id(ring_id)
533 534 535 536
            elif core.is_compiled_with_xpu():
                place = core.XPUPlace(genv.device_id)
                core.BKCLParallelContext(strategy,
                                         place).init_with_ring_id(ring_id)
537 538 539 540 541
            else:
                assert False, ("no cuda device found")
        else:
            return gp

542
    # TODO(shenliang03): This is a temporary solution to solve the problem of
543
    # hang caused by cross-creation of new_group
544
    tmp = paddle.to_tensor(
J
Jiabin Yang 已提交
545
        [1], dtype="int32") if _non_static_mode() else fill_constant(
546
            [0], dtype="int32", value="1")
547 548
    paddle.distributed.all_reduce(tmp, use_calc_stream=True)
    paddle.distributed.wait(tmp)
K
kuizhiqing 已提交
549 550
    return gp

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
def is_initialized():
    """

    Check whether the distributed environment has been initialized

    Returns (bool): `True` if distributed environment has been initialized, otherwise `False`.

    Examples:
        .. code-block:: python

            # required: distributed
            import paddle

            print(paddle.distributed.is_initialized())
            # False

            paddle.distributed.init_parallel_env()
            print(paddle.distributed.is_initialized())
            # True

    """
    global _group_map_by_name
    return _default_group_name in _group_map_by_name


def destroy_process_group(group=None):
    """
    Destroy a given group for communication

    Args:
        group (ProcessGroup, optional): The group to be destroyed. All of process groups, including 
                                        the default group, will be destroyed and the distributed 
                                        environment will be deinitialized.
    
    Returns : None

    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
593
            import paddle.distributed as dist
594

595 596
            dist.init_parallel_env()
            group = dist.new_group([0, 1])
597

598 599
            dist.destroy_process_group(group)
            print(dist.is_initialized())
600
            # True
601 602
            dist.destroy_process_group()
            print(dist.is_initialized())
603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621
            # False

    """
    global _group_map
    global _group_map_by_name

    pg = _get_default_group() if group is None else group
    assert _group_map.get(pg.id, None) is not None, "Invalid group."

    if group is None:
        _group_map.clear()
        _group_map_by_name.clear()
        _group_map_backend.clear()
    else:
        del _group_map[pg.id]
        del _group_map_by_name[pg.name]
        del _group_map_backend[pg]


K
kuizhiqing 已提交
622 623 624 625 626 627 628 629
def wait(tensor, group=None, use_calc_stream=True):
    """

    wait to sync stream for group.

    Args:
        tensor (Tensor): The Tensor used before sync.
        group (Group): The Group instance to perform sync.
K
kuizhiqing 已提交
630 631
        use_calc_stream (bool): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
K
kuizhiqing 已提交
632 633 634 635 636 637 638 639 640 641

    Returns:
        None.

    Examples:
        .. code-block:: python

            import paddle

            paddle.distributed.init_parallel_env()
K
kuizhiqing 已提交
642
            tindata = paddle.randn(shape=[2, 3])
K
kuizhiqing 已提交
643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660
            paddle.distributed.all_reduce(tindata, use_calc_stream=True)
            paddle.distributed.wait(tindata)

    """

    if group is not None and not group.is_member():
        return

    ring_id = 0 if group is None else group.id

    if use_calc_stream:
        _sync_calc_stream(tensor)
    else:
        _sync_comm_stream(tensor, ring_id)


def _sync_calc_stream(tensor):

J
Jiabin Yang 已提交
661
    if _non_static_mode():
662
        return _legacy_C_ops.c_sync_calc_stream(tensor, tensor)
K
kuizhiqing 已提交
663 664 665 666 667 668 669

    op_type = 'c_sync_calc_stream'

    helper = LayerHelper(op_type, **locals())
    helper.append_op(
        type=op_type,
        inputs={'X': [tensor]},
670 671
        outputs={'Out': [tensor]},
    )
672

673

K
kuizhiqing 已提交
674
def _sync_comm_stream(tensor, ring_id=0):
675

J
Jiabin Yang 已提交
676
    if _non_static_mode():
677 678
        return _legacy_C_ops.c_sync_comm_stream([tensor], [tensor], 'ring_id',
                                                ring_id)
679

K
kuizhiqing 已提交
680
    op_type = 'c_sync_comm_stream'
681

K
kuizhiqing 已提交
682 683 684 685 686
    helper = LayerHelper(op_type, **locals())
    helper.append_op(
        type=op_type,
        inputs={'X': [tensor]},
        outputs={'Out': [tensor]},
687 688
        attrs={'ring_id': ring_id},
    )
K
kuizhiqing 已提交
689 690 691


def broadcast(tensor, src, group=None, use_calc_stream=True):
692 693 694
    """

    Broadcast a tensor from the source to all others.
695 696
    As shown below, one process is started with a GPU and GPU0 owns data 0. Through broadcast operator,
    data 0 will be sent to all GPUs from GPU0.
697 698 699 700 701

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/broadcast.png
        :width: 800
        :alt: broadcast
        :align: center
702 703

    Args:
704 705
        tensor (Tensor): The Tensor to send if current rank is the source, or the Tensor to receive otherwise. Its data type
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
706
        src (int): The source rank.
K
kuizhiqing 已提交
707
        group (Group): The group instance return by new_group or None for global default group.
K
kuizhiqing 已提交
708 709
        use_calc_stream (bool): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
710 711 712 713 714 715 716

    Returns:
        None.

    Examples:
        .. code-block:: python

717
            # required: distributed
718
            import paddle
719
            import paddle.distributed as dist
720

721 722 723
            dist.init_parallel_env()
            if dist.get_rank() == 0:
                data = paddle.to_tensor([[4, 5, 6], [4, 5, 6]])
724
            else:
725 726 727 728
                data = paddle.to_tensor([[1, 2, 3], [1, 2, 3]])
            dist.broadcast(data, src=1)
            print(data)
            # [[1, 2, 3], [1, 2, 3]] (2 GPUs)
729
    """
K
kuizhiqing 已提交
730 731 732 733 734 735 736

    if group is not None and not group.is_member():
        return

    if not isinstance(src, int):
        raise ValueError("src should be int.")

L
lilong12 已提交
737
    if in_dygraph_mode():
738 739 740 741 742 743 744 745 746 747 748
        group = _get_default_group() if group is None else group
        gsrc = group.get_group_rank(src)
        assert gsrc >= 0, ("src rank out of group, need global rank")
        task = group.process_group.broadcast(tensor, gsrc)
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task

    ring_id = ring_id = 0 if group is None else group.id
K
kuizhiqing 已提交
749
    gsrc = src if group is None else group.get_group_rank(src)
K
kuizhiqing 已提交
750
    assert gsrc >= 0, ("src rank out of group, need global rank")
K
kuizhiqing 已提交
751

J
Jiabin Yang 已提交
752
    if _non_static_mode():
753 754 755
        return _legacy_C_ops.c_broadcast(tensor, tensor, 'root', gsrc,
                                         'use_calc_stream', use_calc_stream,
                                         'ring_id', ring_id)
756 757

    op_type = 'c_broadcast'
758 759 760 761
    check_variable_and_dtype(tensor, 'tensor', [
        'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8',
        'bool'
    ], 'broadcast')
762 763

    helper = LayerHelper(op_type, **locals())
764 765 766 767 768 769 770 771
    helper.append_op(type=op_type,
                     inputs={'X': [tensor]},
                     outputs={'Out': [tensor]},
                     attrs={
                         'root': gsrc,
                         'use_calc_stream': use_calc_stream,
                         'ring_id': ring_id,
                     })
772 773


K
kuizhiqing 已提交
774
def all_reduce(tensor, op=ReduceOp.SUM, group=None, use_calc_stream=True):
775 776 777
    """

    Reduce a tensor over all ranks so that all get the result.
778 779
    As shown below, one process is started with a GPU and the data of this process is represented
    by its group rank. The reduce operator is sum. Through all_reduce operator, 
780 781 782 783 784 785
    each GPU will have the sum of the data from all GPUs.

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/allreduce.png
        :width: 800
        :alt: all_reduce
        :align: center
786 787 788

    Args:
        tensor (Tensor): The input Tensor. It also works as the output Tensor. Its data type
789 790
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
        op (ReduceOp.SUM|ReduceOp.MAX|ReduceOp.MIN|ReduceOp.PROD): Optional. The operation used. Default value is ReduceOp.SUM.
K
kuizhiqing 已提交
791
        group (Group): The group instance return by new_group or None for global default group.
K
kuizhiqing 已提交
792 793
        use_calc_stream (bool): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
794 795 796 797 798 799 800

    Returns:
        None.

    Examples:
        .. code-block:: python

801
            # required: distributed
802
            import paddle
803
            import paddle.distributed as dist
804

805 806
            dist.init_parallel_env()
            if dist.get_rank() == 0:
807
                data = paddle.to_tensor([[4, 5, 6], [4, 5, 6]])
808
            else:
809
                data = paddle.to_tensor([[1, 2, 3], [1, 2, 3]])
810 811 812
            dist.all_reduce(data)
            print(data)
            # [[5, 7, 9], [5, 7, 9]] (2 GPUs)
813
    """
K
kuizhiqing 已提交
814 815 816
    if group is not None and not group.is_member():
        return

L
lilong12 已提交
817
    if in_dygraph_mode():
818
        op_type = _get_reduce_op(op, "all_reduce")
819 820 821 822 823 824 825 826
        group = _get_default_group() if group is None else group
        task = group.process_group.allreduce(tensor, op_type)
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task

K
kuizhiqing 已提交
827
    ring_id = 0 if group is None else group.id
J
Jiabin Yang 已提交
828
    if _non_static_mode():
829
        if op == ReduceOp.SUM:
830 831 832
            return _legacy_C_ops.c_allreduce_sum_(tensor, 'use_calc_stream',
                                                  use_calc_stream, 'ring_id',
                                                  ring_id)
833
        elif op == ReduceOp.MAX:
834 835 836
            return _legacy_C_ops.c_allreduce_max_(tensor, 'use_calc_stream',
                                                  use_calc_stream, 'ring_id',
                                                  ring_id)
837
        elif op == ReduceOp.MIN:
838 839 840
            return _legacy_C_ops.c_allreduce_min_(tensor, 'use_calc_stream',
                                                  use_calc_stream, 'ring_id',
                                                  ring_id)
841
        elif op == ReduceOp.PROD:
842 843 844
            return _legacy_C_ops.c_allreduce_prod_(tensor, 'use_calc_stream',
                                                   use_calc_stream, 'ring_id',
                                                   ring_id)
845 846 847
        else:
            raise ValueError("Unknown parameter: {}.".format(op))

848 849 850 851
    check_variable_and_dtype(tensor, 'tensor', [
        'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8',
        'bool'
    ], 'all_reduce')
852 853 854 855 856 857 858 859
    if op == ReduceOp.SUM:
        op_type = 'c_allreduce_sum'
    elif op == ReduceOp.MAX:
        op_type = 'c_allreduce_max'
    elif op == ReduceOp.MIN:
        op_type = 'c_allreduce_min'
    elif op == ReduceOp.PROD:
        op_type = 'c_allreduce_prod'
K
kuizhiqing 已提交
860 861
    if not isinstance(ring_id, int):
        raise ValueError("The type of 'ring_id' for all_reduce should be int.")
862
    helper = LayerHelper(op_type, **locals())
863 864 865 866 867 868 869
    helper.append_op(type=op_type,
                     inputs={'X': [tensor]},
                     outputs={'Out': [tensor]},
                     attrs={
                         'ring_id': ring_id,
                         'use_calc_stream': use_calc_stream
                     })
870 871


K
kuizhiqing 已提交
872
def reduce(tensor, dst, op=ReduceOp.SUM, group=None, use_calc_stream=True):
873 874
    """

875 876
    Reduce a tensor to the destination from all others. As shown below, one process is started with a GPU and the data of this process is represented
    by its group rank. The destination of the reduce operator is GPU0 and the process is sum. Through reduce operator,
877 878 879 880 881 882
    the GPU0 will owns the sum of all data from all GPUs.

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/reduce.png
        :width: 800
        :alt: reduce
        :align: center
883 884 885

    Args:
        tensor (Tensor): The output Tensor for the destination and the input Tensor otherwise. Its data type
886
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
887
        dst (int): The destination rank id.
888
        op (ReduceOp.SUM|ReduceOp.MAX|ReduceOp.MIN|ReduceOp.PROD): Optional. The operation used. Default value is ReduceOp.SUM.
K
kuizhiqing 已提交
889
        group (Group): The group instance return by new_group or None for global default group.
K
kuizhiqing 已提交
890 891
        use_calc_stream (bool): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
892 893 894 895 896 897 898

    Returns:
        None.

    Examples:
        .. code-block:: python

899
            # required: distributed
900
            import paddle
901
            import paddle.distributed as dist
902

903 904 905
            dist.init_parallel_env()
            if dist.get_rank() == 0:
                data = paddle.to_tensor([[4, 5, 6], [4, 5, 6]])
906
            else:
907 908 909 910 911
                data = paddle.to_tensor([[1, 2, 3], [1, 2, 3]])
            dist.reduce(data, dst=0)
            print(data)
            # [[5, 7, 9], [5, 7, 9]] (2 GPUs, out for rank 0)
            # [[1, 2, 3], [1, 2, 3]] (2 GPUs, out for rank 1)
912
    """
K
kuizhiqing 已提交
913 914 915
    if group is not None and not group.is_member():
        return

L
lilong12 已提交
916
    if in_dygraph_mode():
917
        op_type = _get_reduce_op(op, "reduce")
918 919 920 921 922 923 924 925 926
        group = _get_default_group() if group is None else group
        gdst = group.get_group_rank(dst)
        assert gdst >= 0, ("dst rank out of group, need global rank")
        task = group.process_group.reduce(tensor, gdst, op_type)
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task
K
kuizhiqing 已提交
927 928 929

    ring_id = 0 if group is None else group.id
    gdst = dst if group is None else group.get_group_rank(dst)
K
kuizhiqing 已提交
930
    assert gdst >= 0, ("dst rank out of group, need global rank")
K
kuizhiqing 已提交
931

J
Jiabin Yang 已提交
932
    if _non_static_mode():
933
        if op == ReduceOp.SUM:
934 935 936
            return _legacy_C_ops.c_reduce_sum(tensor, tensor, 'use_calc_stream',
                                              use_calc_stream, 'ring_id',
                                              ring_id, 'root_id', gdst)
937
        elif op == ReduceOp.MAX:
938 939 940
            return _legacy_C_ops.c_reduce_max(tensor, tensor, 'use_calc_stream',
                                              use_calc_stream, 'ring_id',
                                              ring_id, 'root_id', gdst)
941
        elif op == ReduceOp.MIN:
942 943 944
            return _legacy_C_ops.c_reduce_min(tensor, tensor, 'use_calc_stream',
                                              use_calc_stream, 'ring_id',
                                              ring_id, 'root_id', gdst)
945
        elif op == ReduceOp.PROD:
946 947 948 949
            return _legacy_C_ops.c_reduce_prod(tensor, tensor,
                                               'use_calc_stream',
                                               use_calc_stream, 'ring_id',
                                               ring_id, 'root_id', gdst)
950 951 952 953
        else:
            raise ValueError("Unknown parameter: {}.".format(op))

    op_type = 'c_reduce'
954 955 956 957
    check_variable_and_dtype(tensor, 'tensor', [
        'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8',
        'bool'
    ], 'reduce')
958 959 960 961 962 963 964 965 966 967 968

    if op == ReduceOp.SUM:
        op_type = 'c_reduce_sum'
    elif op == ReduceOp.MAX:
        op_type = 'c_reduce_max'
    elif op == ReduceOp.MIN:
        op_type = 'c_reduce_min'
    elif op == ReduceOp.PROD:
        op_type = 'c_reduce_prod'

    helper = LayerHelper(op_type, **locals())
969 970 971 972 973 974 975 976
    helper.append_op(type=op_type,
                     inputs={'X': [tensor]},
                     outputs={'Out': [tensor]},
                     attrs={
                         'ring_id': ring_id,
                         'use_calc_stream': use_calc_stream,
                         'root_id': gdst,
                     })
977 978


K
kuizhiqing 已提交
979
def all_gather(tensor_list, tensor, group=None, use_calc_stream=True):
980 981
    """

982
    Gather tensors from all participators and all get the result. As shown
983 984
    below, one process is started with a GPU and the data of this process is represented
    by its group rank. Through the all_gather operator, each GPU will have data
985 986 987 988 989 990
    from all GPUs.

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/allgather.png
        :width: 800
        :alt: all_gather
        :align: center
991 992 993

    Args:
        tensor_list (list): A list of output Tensors. Every element in the list must be a Tensor whose data type
994
            should be float16, float32, float64, int32, int64, int8, uint8, bool, complex64 or complex128.
995
        tensor (Tensor): The Tensor to send. Its data type
996
            should be float16, float32, float64, int32, int64, int8, uint8, bool, complex64 or complex128.
K
kuizhiqing 已提交
997
        group (Group): The group instance return by new_group or None for global default group.
K
kuizhiqing 已提交
998 999
        use_calc_stream (bool): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
1000 1001 1002 1003 1004 1005 1006

    Returns:
        None.

    Examples:
        .. code-block:: python

1007
            # required: distributed
1008
            import paddle
1009
            import paddle.distributed as dist
1010

1011
            dist.init_parallel_env()
1012
            tensor_list = []
1013 1014
            if dist.get_rank() == 0:
                data = paddle.to_tensor([[4, 5, 6], [4, 5, 6]])
1015
            else:
1016 1017 1018 1019
                data = paddle.to_tensor([[1, 2, 3], [1, 2, 3]])
            dist.all_gather(tensor_list, data)
            print(tensor_list)
            # [[[4, 5, 6], [4, 5, 6]], [[1, 2, 3], [1, 2, 3]]] (2 GPUs)
1020
    """
K
kuizhiqing 已提交
1021 1022 1023
    if group is not None and not group.is_member():
        return

1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034
    def convert_to_complex(list_of_tensor):
        list_of_complex = []
        for tensor in list_of_tensor:
            list_of_complex.append(paddle.as_complex(tensor))
        return list_of_complex

    is_input_complex = (tensor.dtype == paddle.complex64
                        or tensor.dtype == paddle.complex128)
    if is_input_complex:
        tensor = paddle.as_real(tensor)

L
lilong12 已提交
1035
    if in_dygraph_mode():
1036
        group = _get_default_group() if group is None else group
1037 1038 1039 1040 1041 1042
        if len(tensor_list) == 0:
            tensor_shape = list(tensor.shape)
            tensor_shape[0] *= group.nranks
            out = paddle.empty(tensor_shape, tensor.dtype)
        else:
            out = paddle.concat(tensor_list, axis=0)
1043 1044 1045
        task = group.process_group.all_gather(tensor, out)
        task.wait()
        tensor_list.clear()
1046 1047 1048 1049 1050
        list_of_tensor = paddle.split(out, group.nranks, 0)
        if is_input_complex:
            tensor_list.extend(convert_to_complex(list_of_tensor))
        else:
            tensor_list.extend(list_of_tensor)
1051 1052
        return

K
kuizhiqing 已提交
1053 1054 1055
    ring_id = 0 if group is None else group.id
    nranks = _get_global_group().nranks if group is None else group.nranks

J
Jiabin Yang 已提交
1056
    if _non_static_mode():
1057 1058 1059
        out = _legacy_C_ops.c_allgather(tensor, 'use_calc_stream',
                                        use_calc_stream, 'ring_id', ring_id,
                                        'nranks', nranks)
1060
    else:
1061 1062 1063
        op_type = 'c_allgather'
        helper = LayerHelper(op_type, **locals())
        out = helper.create_variable_for_type_inference(dtype=tensor.dtype)
1064 1065 1066 1067
        if not isinstance(tensor_list, list):
            raise ValueError("The type of 'tensor_list' for all_gather "
                             "should be list.")
        for elem in tensor_list:
1068 1069 1070 1071 1072 1073 1074 1075
            check_variable_and_dtype(elem, 'tensor_list', [
                'float16', 'float32', 'float64', 'int32', 'int64', 'bool',
                'int8', 'uint8', 'complex64', 'complex128'
            ], 'all_gather')
        check_variable_and_dtype(tensor, 'tensor', [
            'float16', 'float32', 'float64', 'int32', 'int64', 'bool', 'int8',
            'uint8', 'complex64', 'complex128'
        ], 'all_gather')
1076 1077 1078 1079 1080 1081 1082 1083
        helper.append_op(type=op_type,
                         inputs={'X': [tensor]},
                         outputs={'Out': [out]},
                         attrs={
                             'ring_id': ring_id,
                             'use_calc_stream': use_calc_stream,
                             'nranks': nranks
                         })
1084

1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097
    list_of_tensor = paddle.split(out, nranks, 0)
    if is_input_complex:
        tensor_list.extend(convert_to_complex(list_of_tensor))
    else:
        tensor_list.extend(list_of_tensor)


def _convert_object_to_tensor(obj):
    _pickler = pickle.Pickler
    f = io.BytesIO()
    _pickler(f).dump(obj)
    data = np.frombuffer(f.getvalue(), dtype=np.uint8)
    tensor = paddle.to_tensor(data)
1098
    return tensor, tensor.numel()
1099 1100


1101
def _convert_tensor_to_object(tensor, len_of_tensor):
1102
    _unpickler = pickle.Unpickler
1103
    return _unpickler(io.BytesIO(tensor.numpy()[:len_of_tensor])).load()
1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130


def all_gather_object(object_list, obj, group=None):
    """

    Gather picklable objects from all participators and all get the result. Similiar to all_gather(), but python object can be passed in.

    Args:
        object_list (list): A list of output object. The datatype of every element in the list is same as the input obj.
        obj (Any): The picklable object to send.
        group (Group): The group instance return by new_group or None for global default group.

    Returns:
        None.

    Warning:
        This API only supports the dygraph mode.

    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
            object_list = []
1131
            if dist.get_rank() == 0:
1132 1133 1134
                obj = {"foo": [1, 2, 3]}
            else:
                obj = {"bar": [4, 5, 6]}
1135 1136 1137
            dist.all_gather_object(object_list, obj)
            print(object_list)
            # [{'foo': [1, 2, 3]}, {'bar': [4, 5, 6]}] (2 GPUs)
1138 1139 1140 1141
    """
    assert in_dygraph_mode(
    ), "all_gather_object doesn't support static graph mode."

1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154
    tensor, len_of_tensor = _convert_object_to_tensor(obj)

    # gather len_of_tensor from all ranks
    list_len_of_tensor = []
    all_gather(list_len_of_tensor, len_of_tensor, group)
    # get the max length from list
    max_len_of_tensor = int(max(list_len_of_tensor).item())
    # resize the input tensor to max length avoid hang in all gather
    # Note(liyurui): Maybe we should support various length all_gather?
    # Now this operation is efficient for we don't support resize in python.
    numpy_data = tensor.numpy()
    numpy_data = np.resize(numpy_data, [max_len_of_tensor])
    input_tensor = paddle.to_tensor(numpy_data)
1155 1156

    tensor_list = []
1157 1158 1159 1160
    all_gather(tensor_list, input_tensor, group)
    for i, tensor in enumerate(tensor_list):
        object_list.append(
            _convert_tensor_to_object(tensor, list_len_of_tensor[i]))
1161 1162


K
kuizhiqing 已提交
1163
def scatter(tensor, tensor_list=None, src=0, group=None, use_calc_stream=True):
1164 1165
    """

1166
    Scatter a tensor to all participators. As shown below, one process is started with a GPU and the source of the scatter
1167 1168 1169 1170 1171 1172
    is GPU0. Through scatter operator, the data in GPU0 will be sent to all GPUs averagely.

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/scatter.png
        :width: 800
        :alt: scatter
        :align: center
1173 1174 1175

    Args:
        tensor (Tensor): The output Tensor. Its data type
1176
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
1177
        tensor_list (list|tuple): A list/tuple of Tensors to scatter. Every element in the list must be a Tensor whose data type
1178
            should be float16, float32, float64, int32, int64, int8, uint8 or bool. Default value is None.
K
kuizhiqing 已提交
1179
        src (int): The source rank id. Default value is 0.
K
kuizhiqing 已提交
1180
        group (Group): The group instance return by new_group or None for global default group.
K
kuizhiqing 已提交
1181 1182
        use_calc_stream (bool): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
1183 1184 1185 1186 1187 1188 1189

    Returns:
        None.

    Examples:
        .. code-block:: python

1190
            # required: distributed
1191
            import paddle
1192
            import paddle.distributed as dist
1193

1194 1195 1196 1197 1198
            dist.init_parallel_env()
            if dist.get_rank() == 0:
                data1 = paddle.to_tensor([7, 8, 9])
                data2 = paddle.to_tensor([10, 11, 12])
                dist.scatter(data1, src=1)
1199
            else:
1200 1201 1202 1203 1204 1205
                data1 = paddle.to_tensor([1, 2, 3])
                data2 = paddle.to_tensor([4, 5, 6])
                dist.scatter(data1, tensor_list=[data1, data2], src=1)
            print(data1, data2)
            # [1, 2, 3] [10, 11, 12] (2 GPUs, out for rank 0)
            # [4, 5, 6] [4, 5, 6] (2 GPUs, out for rank 1)
1206
    """
K
kuizhiqing 已提交
1207 1208 1209 1210 1211 1212
    if group is not None and not group.is_member():
        return

    if not isinstance(src, int):
        raise ValueError("src should be int.")

L
lilong12 已提交
1213
    if in_dygraph_mode():
1214 1215 1216 1217 1218 1219 1220 1221 1222
        group = _get_default_group() if group is None else group
        gsrc = group.get_group_rank(src)
        rank = group.rank
        nranks = group.nranks
    else:
        ring_id = 0 if group is None else group.id
        gsrc = src if group is None else group.get_group_rank(src)
        rank = _get_global_group().rank if group is None else group.rank
        nranks = _get_global_group().nranks if group is None else group.nranks
K
kuizhiqing 已提交
1223
    assert gsrc >= 0, ("src rank out of group, need global rank")
K
kuizhiqing 已提交
1224 1225

    if rank != gsrc:
1226 1227 1228 1229
        tensor_list = []
        for _ in range(nranks):
            tensor_list.append(tensor)
    temp = paddle.concat(tensor_list, axis=0)
L
lilong12 已提交
1230
    if in_dygraph_mode():
1231 1232 1233 1234 1235 1236 1237
        task = group.process_group.scatter(temp, tensor, gsrc)
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task

L
lilong12 已提交
1238
    if _non_static_mode():
1239 1240 1241
        return _legacy_C_ops.c_scatter(temp, tensor, 'use_calc_stream',
                                       use_calc_stream, 'ring_id', ring_id,
                                       'nranks', nranks, 'root', gsrc)
W
wanghuancoder 已提交
1242
    op_type = 'c_scatter'
1243 1244 1245 1246
    check_variable_and_dtype(tensor, 'tensor', [
        'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8',
        'bool'
    ], 'scatter')
1247
    helper = LayerHelper(op_type, **locals())
1248 1249 1250 1251 1252 1253 1254 1255 1256
    helper.append_op(type=op_type,
                     inputs={'X': [temp]},
                     outputs={'Out': [tensor]},
                     attrs={
                         'ring_id': ring_id,
                         'root': gsrc,
                         'use_calc_stream': use_calc_stream,
                         'nranks': nranks,
                     })
1257 1258


1259
def _c_identity(tensor, group=None):
L
lilong12 已提交
1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270
    """
    Return a copy of the tensor, mainly used with model parallel.

    Args:
        tensor (Tensor): The input Tensor. Its data type
            should be float16, float32, float64, int32 or int64.
        group (int): The id of the process group to work on.

    Returns:
        Tensor.
    """
1271 1272 1273 1274
    if group is not None and not group.is_member():
        return
    ring_id = 0 if group is None else group.id

J
Jiabin Yang 已提交
1275
    if _non_static_mode():
1276 1277 1278
        return _legacy_C_ops.c_identity(tensor, 'use_calc_stream', True,
                                        'ring_id', ring_id,
                                        'use_model_parallel', True)
L
lilong12 已提交
1279 1280 1281
    op_type = 'c_identity'
    helper = LayerHelper(op_type, **locals())
    out = helper.create_variable_for_type_inference(dtype=tensor.dtype)
1282

L
lilong12 已提交
1283 1284 1285
    check_variable_and_dtype(
        tensor, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'],
        '_c_identity')
1286

1287 1288 1289 1290 1291 1292 1293 1294
    helper.append_op(type=op_type,
                     inputs={'X': tensor},
                     outputs={'Out': out},
                     attrs={
                         'ring_id': ring_id,
                         'use_calc_stream': True,
                         'use_model_parallel': True,
                     })
L
lilong12 已提交
1295 1296 1297
    return out


1298
def _c_concat(tensor, group=None):
1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
    """
    Return allgather of the tensor, mainly used with model parallel.

    Args:
        tensor (Tensor): The input Tensor. Its data type
            should be float16, float32, float64, int32 or int64.
        group (int): The id of the process group to work on.

    Returns:
        Tensor.
    """
    if group is not None and not group.is_member():
        return
1312 1313
    group = _get_default_group() if group is None else group
    ring_id = group.id
1314

1315
    global_rank = _get_global_env().rank
1316 1317
    rank = group.rank
    nranks = group.nranks
1318

J
Jiabin Yang 已提交
1319
    if _non_static_mode():
1320 1321 1322 1323
        return _legacy_C_ops.c_concat(tensor, 'ring_id', ring_id,
                                      'use_calc_stream', True, 'rank', rank,
                                      'nranks', nranks, 'use_model_parallel',
                                      True)
1324 1325 1326 1327 1328 1329 1330 1331 1332

    op_type = 'c_concat'
    helper = LayerHelper(op_type, **locals())
    out = helper.create_variable_for_type_inference(dtype=tensor.dtype)

    check_variable_and_dtype(
        tensor, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'],
        '_c_concat')

1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
    helper.append_op(type=op_type,
                     inputs={'X': tensor},
                     outputs={'Out': out},
                     attrs={
                         'ring_id': ring_id,
                         'use_calc_stream': True,
                         'use_model_parallel': True,
                         'nranks': nranks,
                         'rank': rank
                     })
1343 1344 1345
    return out


1346
def _c_split(tensor, group=None):
L
lilong12 已提交
1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358
    """
    Split tensor evenly among all members, mainly used with model parallel.

    Args:
        tensor (Tensor): The input Tensor. Its data type
            should be float16, float32, float64, int32 or int64.
        rank (int): The rank of the current process.
        group (int): The id of the process group to work on.

    Returns:
        Tensor.
    """
1359 1360 1361 1362
    if group is not None and not group.is_member():
        return
    ring_id = 0 if group is None else group.id

1363 1364 1365 1366
    global_rank = _get_global_env().rank
    rank = global_rank if group is None else group.get_group_rank(global_rank)
    nranks = _get_global_env().world_size if group is None else group.nranks

J
Jiabin Yang 已提交
1367
    if _non_static_mode():
1368 1369 1370
        return _legacy_C_ops.c_split(tensor, 'use_calc_stream', True, 'ring_id',
                                     ring_id, 'rank', rank, 'nranks', nranks,
                                     'use_model_parallel', True)
1371

L
lilong12 已提交
1372 1373 1374
    op_type = 'c_split'
    helper = LayerHelper(op_type, **locals())
    out = helper.create_variable_for_type_inference(dtype=tensor.dtype)
1375

L
lilong12 已提交
1376 1377 1378
    check_variable_and_dtype(
        tensor, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'],
        '_c_split')
1379

1380 1381 1382 1383 1384 1385 1386 1387 1388 1389
    helper.append_op(type=op_type,
                     inputs={'X': tensor},
                     outputs={'Out': out},
                     attrs={
                         'ring_id': ring_id,
                         'use_calc_stream': True,
                         'rank': rank,
                         'nranks': nranks,
                         'use_model_parallel': True,
                     })
L
lilong12 已提交
1390 1391 1392
    return out


1393 1394 1395 1396 1397
def _mp_allreduce(tensor,
                  op=ReduceOp.SUM,
                  group=None,
                  use_calc_stream=True,
                  use_model_parallel=True):
1398
    """[it is same as allreduce above, but it supports model parallel. And it support inplace startegy]
1399 1400 1401 1402
    """
    if group is not None and not group.is_member():
        return

1403
    if in_dygraph_mode():
1404
        group = _get_default_group() if group is None else group
1405 1406
        assert op == ReduceOp.SUM, "Unknown parameter: {}.".format(op)

1407
        from paddle.autograd import PyLayer
1408

1409
        class mp_allreduce_eager(PyLayer):
1410

1411
            @staticmethod
1412
            def forward(ctx, tensor, group, use_calc_stream,
1413
                        use_model_parallel):
1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424
                ctx.ring_id = group.id

                if use_calc_stream:
                    op_type = _get_reduce_op(op, "_mp_allreduce")
                    group.process_group.allreduce_on_calc_stream(
                        tensor, op_type)
                    return tensor
                else:
                    return _legacy_C_ops.c_allreduce_sum_(
                        tensor, 'use_calc_stream', use_calc_stream, 'ring_id',
                        ring_id, "use_model_parallel", use_model_parallel)
1425 1426 1427

            @staticmethod
            def backward(ctx, dy):
1428 1429 1430
                return _legacy_C_ops.c_identity(dy, 'use_calc_stream', True,
                                                'ring_id', ctx.ring_id,
                                                'use_model_parallel', True)
1431

1432
        return mp_allreduce_eager.apply(tensor, group, use_calc_stream,
1433 1434
                                        use_model_parallel)

1435 1436
    ring_id = 0 if group is None else group.id
    if _in_legacy_dygraph():
1437
        if op == ReduceOp.SUM:
1438 1439 1440 1441
            return _legacy_C_ops.c_allreduce_sum_(tensor, 'use_calc_stream',
                                                  use_calc_stream, 'ring_id',
                                                  ring_id, "use_model_parallel",
                                                  use_model_parallel)
1442 1443
        else:
            raise ValueError("Unknown parameter: {}.".format(op))
1444 1445 1446 1447 1448 1449 1450 1451 1452

    op_type = 'c_allreduce_sum'
    helper = LayerHelper(op_type, **locals())
    out = helper.create_variable_for_type_inference(dtype=tensor.dtype)

    check_variable_and_dtype(
        tensor, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'],
        op_type)

1453 1454 1455 1456 1457 1458 1459 1460
    helper.append_op(type=op_type,
                     inputs={'X': tensor},
                     outputs={'Out': out},
                     attrs={
                         'ring_id': ring_id,
                         'use_calc_stream': use_calc_stream,
                         'use_model_parallel': use_model_parallel,
                     })
1461
    return out
1462 1463


1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477
def _c_lookup_table(table, index, start_index=0, name=None):
    """
    Lookup table according to index.

    Args:
        table (Tensor): The input Tensor. Its data type
            should be float16, float32, float64.
        index (Tensor): The index to lookup table.
        start_index (int): The initial index for table range.
        name (string): The name of the api

    Returns:
        Tensor.
    """
J
Jiabin Yang 已提交
1478
    if _non_static_mode():
1479 1480
        return _legacy_C_ops.c_embedding(table, index, "start_index",
                                         start_index)
1481

1482 1483 1484 1485 1486
    op_type = 'c_embedding'
    helper = LayerHelper(op_type, **locals())
    dtype = helper.input_dtype(input_param_name='table')
    check_variable_and_dtype(index, 'input', ['int32', 'int64'], op_type)
    tmp = helper.create_variable_for_type_inference(dtype)
1487 1488 1489 1490 1491 1492 1493
    helper.append_op(type='c_embedding',
                     inputs={
                         'Ids': index,
                         'W': table
                     },
                     outputs={'Out': tmp},
                     attrs={"start_index": start_index})
1494 1495
    return tmp

1496

B
Baibaifan 已提交
1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511
class _Linear(layers.Layer):
    """
    Linear
    """

    def __init__(self,
                 in_features,
                 out_features,
                 weight_attr=None,
                 bias_attr=None,
                 name=None):
        super(_Linear, self).__init__()
        self._dtype = self._helper.get_default_dtype()
        self._weight_attr = weight_attr
        self._bias_attr = bias_attr
1512 1513 1514 1515 1516 1517 1518 1519
        self.weight = self.create_parameter(shape=[in_features, out_features],
                                            attr=self._weight_attr,
                                            dtype=self._dtype,
                                            is_bias=False)
        self.bias = self.create_parameter(shape=[out_features],
                                          attr=self._bias_attr,
                                          dtype=self._dtype,
                                          is_bias=True)
B
Baibaifan 已提交
1520 1521 1522
        self.name = name

    def forward(self, input):
1523 1524 1525 1526
        out = _linear(x=input,
                      weight=self.weight,
                      bias=self.bias,
                      name=self.name)
B
Baibaifan 已提交
1527 1528 1529 1530 1531 1532 1533 1534
        return out

    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'in_features={}, out_features={}, dtype={}{}'.format(
            self.weight.shape[0], self.weight.shape[1], self._dtype, name_str)


1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554
def _c_softmax_with_cross_entropy(logits,
                                  label,
                                  group=None,
                                  return_softmax=False):
    if group is not None and not group.is_member():
        return
    ring_id = 0 if group is None else group.id
    global_rank = _get_global_env().rank
    rank = global_rank if group is None else group.get_group_rank(global_rank)
    nranks = _get_global_env().world_size if group is None else group.nranks

    input_dims = len(list(logits.shape))
    label_dims = len(list(label.shape))
    if input_dims - 1 != label_dims and input_dims != label_dims:
        raise ValueError(
            'Expected nput_dims - 1 = label_dims or input_dims == label_dims\
             (got nput_dims{}, label_dims{})'.format(input_dims, label_dims))
    if input_dims - 1 == label_dims:
        label = paddle.unsqueeze(label, axis=-1)

J
Jiabin Yang 已提交
1555
    if _non_static_mode():
1556
        softmax, loss = _legacy_C_ops.c_softmax_with_cross_entropy(
1557 1558 1559 1560 1561 1562
            logits, label, 'ring_id', ring_id, 'rank', rank, 'nranks', nranks)
        if not return_softmax:
            return loss
        else:
            return loss, softmax

W
WangXi 已提交
1563 1564 1565 1566 1567 1568 1569 1570
    attrs = {
        'ring_id': ring_id,
        'rank': rank,
        'nranks': nranks,
    }
    helper = LayerHelper('c_softmax_with_cross_entropy', **locals())
    softmax = helper.create_variable_for_type_inference(dtype=logits.dtype)
    loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
1571 1572 1573 1574 1575 1576 1577 1578 1579 1580
    helper.append_op(type='c_softmax_with_cross_entropy',
                     inputs={
                         'Logits': logits,
                         'Label': label
                     },
                     outputs={
                         'Softmax': softmax,
                         'Loss': loss
                     },
                     attrs=attrs)
W
WangXi 已提交
1581 1582 1583 1584 1585 1586

    if return_softmax:
        return loss, softmax

    return loss

1587

B
Baibaifan 已提交
1588 1589 1590 1591
def _linear(x, weight, bias=None, name=None):
    """
    Fuction Linear
    """
J
Jiabin Yang 已提交
1592
    if _non_static_mode():
B
Baibaifan 已提交
1593
        pre_bias = _varbase_creator(dtype=x.dtype)
1594 1595
        _legacy_C_ops.matmul(x, weight, pre_bias, 'transpose_X', False,
                             'transpose_Y', False, "alpha", 1)
1596 1597 1598
        return dygraph_utils._append_bias_in_dygraph(pre_bias,
                                                     bias,
                                                     axis=len(x.shape) - 1)
B
Baibaifan 已提交
1599 1600 1601
    else:
        helper = LayerHelper('linear', **locals())
        dtype = x.dtype
B
Baibaifan 已提交
1602 1603
        assert len(
            x.shape) < 4, "X latitude is not supported greater than 3 now."
B
Baibaifan 已提交
1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615

        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                 'linear')
        check_dtype(dtype, 'dtype', ['float16', 'float32', 'float64'], 'linear')

        inputs = {'X': [x], 'Y': [weight]}
        attrs = {
            'transpose_X': False,
            'transpose_Y': False,
            'alpha': 1,
        }
        tmp = helper.create_variable_for_type_inference(dtype)
1616 1617 1618 1619
        helper.append_op(type='matmul_v2',
                         inputs=inputs,
                         outputs={'Out': tmp},
                         attrs=attrs)
B
Baibaifan 已提交
1620 1621
        if bias is not None:
            res = helper.create_variable_for_type_inference(dtype)
1622 1623 1624 1625 1626 1627 1628
            helper.append_op(type='elementwise_add',
                             inputs={
                                 'X': [tmp],
                                 'Y': [bias]
                             },
                             outputs={'Out': [res]},
                             attrs={'axis': len(x.shape) - 1})
B
Baibaifan 已提交
1629 1630 1631 1632 1633
        else:
            res = tmp
        return res


1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646
def _set_var_distributed(var):
    if var is None:
        return

    var.is_distributed = True

    # NOTE: use current_block and find_var_recursive to support while_loop
    startup_block = paddle.static.default_startup_program().current_block()
    main_block = paddle.static.default_main_program().current_block()
    startup_block._find_var_recursive(var.name).is_distributed = True
    main_block._find_var_recursive(var.name).is_distributed = True


L
lilong12 已提交
1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657
def _parallel_linear(x,
                     num_rows,
                     num_cols,
                     axis,
                     param_attr,
                     bias_attr,
                     gather_out,
                     inner_rank,
                     nranks,
                     split_tensor,
                     name,
1658
                     group=None):
1659 1660
    """
    Parallel Linear
1661 1662 1663

    axis the dimension of the parameter of linear layer. 
    axis = 0: the row dimension
1664
    axis = 1: the col dimension
1665
    
1666
    """
1667 1668 1669 1670
    if group is not None and not group.is_member():
        return
    ring_id = 0 if group is None else group.id

L
lilong12 已提交
1671 1672
    if axis == 0:
        if split_tensor:
1673
            x = _c_split(x, group=group)
1674
    else:
L
lilong12 已提交
1675 1676
        x = _c_identity(x, group=group)

1677 1678 1679 1680 1681
    linear = paddle.nn.Linear(num_rows,
                              num_cols,
                              weight_attr=param_attr,
                              bias_attr=bias_attr,
                              name=name)
1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693

    # NOTE: npu linear function use matmul_v2 but linear use matmul
    linear_function = _linear if core.is_compiled_with_npu()\
        else paddle.nn.functional.linear
    linear_out = linear_function(
        x,
        linear.weight,
        # NOTE(wangxi): row split, bias need add after allreduce
        None if axis == 0 else linear.bias,
        linear.name)

    _set_var_distributed(linear.weight)
1694 1695 1696 1697
    # set is_distributed for splited bias
    # if a linear layer is splited by row, each rank would hold a complete bias and they should be the same in each rank.
    # if a linear layer is splited by col, the bias would also be split into each rank as its weight
    if axis == 1 and linear._bias_attr != False:
1698
        _set_var_distributed(linear.bias)
L
lilong12 已提交
1699 1700 1701 1702 1703

    if not gather_out: return linear_out

    out_shape = list(linear_out.shape)
    out_shape[0] *= 1 if axis == 0 else nranks
1704
    main_block = paddle.static.default_main_program().current_block()
L
lilong12 已提交
1705 1706 1707 1708 1709 1710 1711 1712 1713
    out = main_block.create_var(
        shape=out_shape,
        dtype=linear_out.dtype,
        type=linear_out.type,
        lod_level=linear_out.lod_level,
        persistable=False,
        is_data=False,
        need_check_feed=linear_out.desc.need_check_feed())
    if axis == 0:
1714 1715 1716 1717 1718 1719 1720 1721
        main_block.append_op(type='c_allreduce_sum',
                             inputs={'X': linear_out},
                             outputs={'Out': out},
                             attrs={
                                 'ring_id': ring_id,
                                 'use_calc_stream': True,
                                 'use_model_parallel': True
                             })
1722 1723
        if linear.bias is not None:
            out = out + linear.bias
L
lilong12 已提交
1724
    else:
1725 1726 1727 1728 1729 1730 1731 1732 1733 1734
        main_block.append_op(type='c_concat',
                             inputs={'X': linear_out},
                             outputs={'Out': out},
                             attrs={
                                 'rank': inner_rank,
                                 'ring_id': ring_id,
                                 'nranks': nranks,
                                 'use_calc_stream': True,
                                 'use_model_parallel': True
                             })
L
lilong12 已提交
1735
    return out
1736 1737


L
lilong12 已提交
1738 1739 1740 1741 1742 1743 1744
def _parallel_embedding(x,
                        per_part_embeddings,
                        origin_size,
                        param_attr,
                        inner_rank,
                        num_partitions,
                        name,
1745
                        group=None):
1746 1747 1748
    """
    Parallel Embedding
    """
1749 1750 1751 1752
    if group is not None and not group.is_member():
        return
    ring_id = 0 if group is None else group.id

1753 1754 1755 1756 1757 1758 1759 1760 1761
    helper = LayerHelper("_parallel_embedding", **locals())

    per_part_size = per_part_embeddings
    rank = inner_rank

    vocab_start_index = rank * per_part_size
    dtype = helper.get_default_dtype()
    size = [per_part_size, origin_size[1]]

1762 1763 1764 1765
    weight = helper.create_parameter(attr=param_attr,
                                     shape=size,
                                     dtype=dtype,
                                     is_bias=False)
1766 1767

    if num_partitions == 1:
1768 1769 1770 1771 1772
        return paddle.nn.functional.embedding(x,
                                              weight=weight,
                                              padding_idx=None,
                                              sparse=False,
                                              name=name)
1773

1774 1775
    startup_block = paddle.static.default_startup_program().global_block()
    main_block = paddle.static.default_main_program().global_block()
1776 1777 1778 1779 1780
    startup_block.vars[weight.name].is_distributed = True
    main_block.vars[weight.name].is_distributed = True

    output_parallel = paddle.distributed.collective._c_lookup_table(
        weight, x, start_index=vocab_start_index, name=name)
1781 1782 1783 1784
    out = paddle.distributed.collective._mp_allreduce(output_parallel,
                                                      group=group,
                                                      use_calc_stream=True,
                                                      use_model_parallel=True)
L
lilong12 已提交
1785
    return out
1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808


def split(x,
          size,
          operation,
          axis=0,
          num_partitions=1,
          gather_out=True,
          weight_attr=None,
          bias_attr=None,
          name=None):
    """

    Split the weight of the specified operation into multiple devices
    and do the computation in parallel.

    Now the following three cases are supported.

    Case 1: Parallel Embedding
        The weight of the embedding operation is a NxM matrix with N rows and M columns.
        With parallel embedding, the weight is split into num_partitions partitions, each
        of which is a matrix with (N/num_partitions + 1) rows and M column where the last
        row as the padding idx.
K
kuizhiqing 已提交
1809

1810 1811 1812 1813 1814 1815 1816 1817 1818
        Suppose we split the NxM weight into two partitons on device_0 and device_1
        respectively. Then, one each device, the final weight has (N/2 + 1) rows with the
        index range from 0 to N/2. On device_0, all values in the input within [0, N/2 -1]
        keep unchanged and all other values are changed to N/2 which is the padding index and
        are mapped to all zeros after embedding. In the same way, on device_1, the value V in the
        input within [N/2, N-1] will be changed to (V - N/2), and all other values are changed
        to N/2 and are mapped to all zeros after embedding. Finally, the results on the two
        devices are sum-reduced.

1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833
        The Embedding put on single card is as shown below:

        .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/split_embedding_single.png
            :width: 800
            :height: 350
            :alt: single_embedding
            :align: center

        Parallel Embedding is shown as below:

        .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/split_embedding_split.png
            :width: 800
            :alt: split_embedding
            :align: center

1834 1835 1836 1837 1838
    Case 2: Row Parallel Linear
        The weight of the linear operation is a NxM matrix with N rows and M columns.
        With row parallel linear, the weight is split into num_partitions partitions, each
        of which is a matrix with N/num_partitions rows and M column.

1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856
        The linear layer put on single card is shown as below, the input variable is represented by X,
        the weight matrix is represented by W and the output vaiable is O. The linear layer on single card is 
        simple matrix multiplication operation, O = X * W.

        .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/split_single.png
            :width: 800
            :alt: single_linear
            :align: center

        Row Parallel Linear is shown as below. As the name suggests, Row Parallel Linear splits the weight matrix W into
        [[W_row1], [W_row2]] along the row. And accordingly the input is splitted along the column into [X_col1, X_col2] and multiply their
        respective weight matrices. Finally apply AllReduce on the output from each card to get the final output.

        .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/split_row.png
            :width: 800
            :alt: split_row
            :align: center

1857 1858 1859 1860 1861
    Case 3: Column Parallel Linear
        The weight of the linear operation is a NxM matrix with N rows and M columns.
        With column parallel linear, the weight is split into num_paratitions partitions, each
        of which is a matrix with N rows and M/num_partitions column.

1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878
        The linear layer put on single card has been illustrated on case 2 and Column Parallel Linear
        is shown as below. The Column Parallel Linear splits the weight matrix W into [W_col1, W_col2] along the column and 
        these splitted matrices respectively multiply the input. Finally apply AllGather on the output from each card to get the final output. 

        .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/split_col.png
            :width: 800
            :alt: split_col
            :align: center
    
    As observed, the column parallel linear and row parallel linear can be combined to skip one ALLGATHER communication
    operator. Furthermore the Attention and MLP can be combined to imporve the performance as shown below.

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/split_col_row.png
            :width: 800
            :alt: split_col_row
            :align: center

1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898
    Args:
        x (Tensor): Input tensor. It's data type should be float16, float32, float64, int32 or int64.
        size (list|tuple): A list or tuple with two elements indicating the shape of the weight.
        operation (str): The name of the operation. The supported operations are 'linear' and 'embedding'.
        axis (int, Optional): Indicate along which axis to split the weight. Default: 0.
        num_partitions (int, Optional): How many parts the weight is partitioned. Default: 1.
        gather_out (bool, Optional): Whether to gather the output after computation. By default, the output
            on each partitions will be gathered after computation. Default: True.
        weight_attr (ParamAttr, Optional): The parameter attribute for the learnable
            weights(Parameter) of the specified operation. Default: None.
        bias_attr (ParamAttr, Optional): The parameter attribute for the bias
            of the specified operation. Default: None.
        name (str, Optional): The default value is None. Normally there is no need for user to set this
            property. Default: None. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor.

    Examples:
        .. code-block:: python
1899

1900
            # required: distributed
1901
            import paddle
1902
            import paddle.distributed.fleet as fleet
1903

1904
            paddle.enable_static()
1905
            paddle.set_device('gpu:%d'%paddle.distributed.ParallelEnv().dev_id)
1906
            fleet.init(is_collective=True)
1907
            data = paddle.randint(0, 8, shape=[10,4])
1908
            emb_out = paddle.distributed.split(
1909 1910 1911 1912
                data,
                (8, 8),
                operation="embedding",
                num_partitions=2)
1913

1914
    """
1915 1916 1917 1918
    assert isinstance(
        size,
        (list, tuple)), ("The type of size for "
                         "paddle.distributed.split must be list or tuple.")
1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930
    assert len(size) == 2, ("Number of elements in size of "
                            "paddle.distributed.split must be two.")
    assert isinstance(operation, str), ("The type of operation for "
                                        "paddle.distributed.split must be str.")
    supported_operations = [
        'linear',
        'embedding',
    ]
    assert operation in supported_operations, (
        "The operation for "
        "paddle.distributed.split must be one of {}.".format(
            supported_operations))
J
Jiabin Yang 已提交
1931
    if _non_static_mode():
L
lilong12 已提交
1932 1933 1934 1935
        raise ValueError(
            "paddle.distributed.split cannot be used in dynamic "
            "graph mode, plese use ParallelEmbedding, ParallelRowLinear, "
            "ParallelColumnLinear instead.")
1936
    else:
1937
        from .fleet import fleet
1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948
        assert fleet._role_maker, ("To use paddle.distributed.split, "
                                   "you must call fleet.init() firstly.")
        rank = fleet.worker_index()
        nranks = fleet.worker_num()

    # rank within a model parallel group
    inner_rank = rank % num_partitions

    if operation == "embedding":
        assert axis == 0, ("We only support to split the weight of embedding "
                           "along the first axis now.")
1949 1950 1951
        assert size[0] % num_partitions == 0, \
            "The length of the vocabulary must be divisible by num_partitions " \
            "but received vocabulary={} num_partitions={}".format(size[0], num_partitions)
1952

1953
        per_part_size = size[0] // num_partitions
1954 1955 1956 1957 1958 1959 1960 1961
        emb_out = _parallel_embedding(x,
                                      per_part_size,
                                      size,
                                      weight_attr,
                                      inner_rank,
                                      num_partitions,
                                      name,
                                      group=None)
B
Baibaifan 已提交
1962
        return emb_out
1963
    else:
L
lilong12 已提交
1964
        should_split = False
1965 1966 1967
        if axis == 0:
            assert size[0] % num_partitions == 0, (
                "Number of rows of the weight for linear ({}) must be"
1968 1969
                " divisible by num_partitions ({})".format(
                    size[0], num_partitions))
1970 1971
            per_part_size = size[0] // num_partitions
            linear_size = (per_part_size, size[1])
L
lilong12 已提交
1972
            if x.shape[-1] == size[0]: should_split = True
1973 1974 1975 1976

        elif axis == 1:
            assert size[1] % num_partitions == 0, (
                "Number of column of the weight for linear ({}) must be"
1977 1978
                " divisible by num_partitions ({})".format(
                    size[1], num_partitions))
1979 1980 1981 1982 1983 1984
            per_part_size = size[1] // num_partitions
            linear_size = (size[0], per_part_size)
        else:
            raise ValueError("The value of axis must be 0 or 1, but the value "
                             "given is {}.".format(axis))

1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996
        linear_out = _parallel_linear(x,
                                      linear_size[0],
                                      linear_size[1],
                                      axis,
                                      weight_attr,
                                      bias_attr,
                                      gather_out,
                                      inner_rank,
                                      num_partitions,
                                      should_split,
                                      name=name,
                                      group=None)
1997
        return linear_out
L
lilong12 已提交
1998 1999


L
lilong12 已提交
2000 2001
def alltoall(in_tensor_list, out_tensor_list, group=None, use_calc_stream=True):
    """
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
    Scatter tensors in in_tensor_list to all participators averagely and gather the result tensors in out_tensor_list.
    As shown below, the in_tensor_list in GPU0 includes 0_0 and 0_1, and GPU1 includes 1_0 and 1_1.
    Through alltoall operator, the 0_0 in GPU0 will be sent to GPU0 and 0_1 to GPU1, 1_0 in GPU1 sent to GPU0 and 1_1 to GPU1.
    Finally the out_tensor_list in GPU0 includes 0_0 and 1_0, and GPU1 includes 0_1 and 1_1.

    .. image:: https://githubraw.cdn.bcebos.com/PaddlePaddle/docs/develop/docs/api/paddle/distributed/img/alltoall.png
        :width: 800
        :alt: alltoall
        :align: center

L
lilong12 已提交
2012 2013
    Args:
        in_tensor_list (list): A list of input Tensors. Every element in the list must be a Tensor whose data type
2014
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
2015
        out_tensor_list (list): A list of output Tensors. The data type of its elements should be the same as the
L
lilong12 已提交
2016 2017
            data type of the input Tensors.
        group (Group, optional): The group instance return by new_group or None for global default group. Default: None.
2018
        use_calc_stream (bool, optional): Whether to use calculation stream (True) or communication stream. Default: True.
2019
    
L
lilong12 已提交
2020 2021
    Returns:
        None.
2022
    
L
lilong12 已提交
2023 2024
    Examples:
        .. code-block:: python
2025

L
lilong12 已提交
2026 2027
            # required: distributed
            import paddle
2028 2029 2030
            import paddle.distributed as dist

            dist.init_parallel_env()
L
lilong12 已提交
2031
            out_tensor_list = []
2032 2033 2034
            if dist.get_rank() == 0:
                data1 = paddle.to_tensor([[1, 2, 3], [4, 5, 6]])
                data2 = paddle.to_tensor([[7, 8, 9], [10, 11, 12]])
L
lilong12 已提交
2035
            else:
2036 2037 2038 2039 2040 2041
                data1 = paddle.to_tensor([[13, 14, 15], [16, 17, 18]])
                data2 = paddle.to_tensor([[19, 20, 21], [22, 23, 24]])
            dist.alltoall([data1, data2], out_tensor_list)
            print(out_tensor_list)
            # [[[1, 2, 3], [4, 5, 6]], [[13, 14, 15], [16, 17, 18]]] (2 GPUs, out for rank 0)
            # [[[7, 8, 9], [10, 11, 12]], [[19, 20, 21], [22, 23, 24]]] (2 GPUs, out for rank 1)
L
lilong12 已提交
2042 2043 2044 2045
    """
    if group is not None and not group.is_member():
        return

L
lilong12 已提交
2046
    if in_dygraph_mode():
2047
        group = _get_default_group() if group is None else group
2048 2049
        backend = _group_map_backend[group]
        assert backend != 'gloo', ("backend gloo is not supported yet")
2050 2051 2052
    else:
        ring_id = 0 if group is None else group.id

L
lilong12 已提交
2053
    temp = paddle.concat(in_tensor_list, axis=0)
李季 已提交
2054
    nranks = len(in_tensor_list)
L
lilong12 已提交
2055
    if in_dygraph_mode():
2056 2057 2058 2059 2060 2061
        if len(out_tensor_list) == 0:
            tensor_shape = list(in_tensor_list[0].shape)
            tensor_shape[0] *= nranks
            out = paddle.empty(tensor_shape, in_tensor_list[0].dtype)
        else:
            out = paddle.concat(out_tensor_list, axis=0)
2062 2063 2064 2065 2066 2067
        task = group.process_group.alltoall(temp, out)
        task.wait()
        out_tensor_list.clear()
        out_tensor_list.extend(paddle.split(out, nranks, 0))
        return

J
Jiabin Yang 已提交
2068
    if _non_static_mode():
2069 2070
        out = _legacy_C_ops.alltoall(temp, 'use_calc_stream', use_calc_stream,
                                     'ring_id', ring_id)
L
lilong12 已提交
2071
    else:
W
wanghuancoder 已提交
2072 2073 2074 2075 2076
        op_type = 'alltoall'
        helper = LayerHelper(op_type, **locals())
        out = helper.create_variable_for_type_inference(
            dtype=in_tensor_list[0].dtype)

L
lilong12 已提交
2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090
        if not isinstance(in_tensor_list, list):
            raise ValueError("The type of 'in_tensor_list' for all_to_all "
                             "should be list.")
        for elem in in_tensor_list:
            check_variable_and_dtype(
                elem, 'in_tensor_list',
                ['float16', 'float32', 'float64', 'int32', 'int64'],
                'all_to_all')
        if not isinstance(out_tensor_list, list):
            raise ValueError("The type of 'out_tensor_list' for all_to_all "
                             "should be list.")
        if len(out_tensor_list) != 0:
            raise ValueError("The 'out_tensor_list' for all_to_all "
                             "must be an empty list.")
2091 2092 2093 2094 2095 2096 2097
        helper.append_op(type=op_type,
                         inputs={'X': [temp]},
                         outputs={'Out': [out]},
                         attrs={
                             'ring_id': ring_id,
                             'use_calc_stream': use_calc_stream,
                         })
L
lilong12 已提交
2098 2099 2100
    out_tensor_list.extend(paddle.split(out, nranks, 0))


2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113
def alltoall_single(in_tensor,
                    out_tensor,
                    in_split_sizes=None,
                    out_split_sizes=None,
                    group=None,
                    use_calc_stream=True):
    """
    Scatter a single input tensor to all participators and gather the received tensors in out_tensor.

    .. note::
        ``alltoall_single`` is only supported in eager mode.

    Args:
2114
        in_tensor (Tensor): Input tensor. The data type should be float16, float32, float64, int32, int64, int8, uint8 or bool.
2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136
        out_tensor (Tensor): Output Tensor. The data type should be the same as the data type of the input Tensor.
        in_split_sizes (list[int], optional): Split sizes of ``in_tensor`` for dim[0]. If not given, dim[0] of ``in_tensor`` 
            must be divisible by group size and ``in_tensor`` will be scattered averagely to all participators. Default: None.
        out_split_sizes (list[int], optional): Split sizes of ``out_tensor`` for dim[0]. If not given, dim[0] of ``out_tensor`` 
            must be divisible by group size and ``out_tensor`` will be gathered averagely from all participators. Default: None.
        group (Group, optional): The group instance return by ``new_group`` or None for global default group. Default: None.
        use_calc_stream (bool, optional): Whether to use calculation stream (True) or communication stream. Default: True.
    
    Returns:
        None, if ``use_calc_stream`` is set to ``True``; ``Task`` of ``group``, if ``use_calc_stream`` is set to ``False``.
    
    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
            rank = dist.get_rank()
            size = dist.get_world_size()

2137 2138 2139 2140
            # case 1 (2 GPUs)
            data = paddle.arange(2, dtype='int64') + rank * 2
            # data for rank 0: [0, 1]
            # data for rank 1: [2, 3]
2141
            output = paddle.empty([2], dtype='int64')
2142 2143
            dist.alltoall_single(data, output)
            print(output)
2144 2145 2146
            # output for rank 0: [0, 2]
            # output for rank 1: [1, 3]

2147
            # case 2 (2 GPUs)
2148
            in_split_sizes = [i + 1 for i in range(size)]
2149 2150
            # in_split_sizes for rank 0: [1, 2]
            # in_split_sizes for rank 1: [1, 2]
2151
            out_split_sizes = [rank + 1 for i in range(size)]
2152 2153 2154 2155 2156
            # out_split_sizes for rank 0: [1, 1]
            # out_split_sizes for rank 1: [2, 2]
            data = paddle.ones([sum(in_split_sizes), size], dtype='float32') * rank
            # data for rank 0: [[0., 0.], [0., 0.], [0., 0.]]
            # data for rank 1: [[1., 1.], [1., 1.], [1., 1.]]
2157 2158
            output = paddle.empty([(rank + 1) * size, size], dtype='float32')
            group = dist.new_group([0, 1])
2159
            task = dist.alltoall_single(data,
2160 2161 2162 2163 2164 2165
                                        output,
                                        in_split_sizes,
                                        out_split_sizes,
                                        use_calc_stream=False,
                                        group=group)
            task.wait()
2166
            print(output)
2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177
            # output for rank 0: [[0., 0.], [1., 1.]]
            # output for rank 1: [[0., 0.], [0., 0.], [1., 1.], [1., 1.]]

    """
    if group is not None and not group.is_member():
        return

    assert in_dygraph_mode(), "Only suppport alltoall_single in eager mode."
    # _check_single_tensor

    group = _get_default_group() if group is None else group
2178 2179 2180
    backend = _group_map_backend[group]
    assert backend != 'gloo', ("backend gloo is not supported yet")

2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192
    in_split_sizes = [] if in_split_sizes is None else in_split_sizes
    out_split_sizes = [] if out_split_sizes is None else out_split_sizes

    task = group.process_group.alltoall_single(in_tensor, out_tensor,
                                               in_split_sizes, out_split_sizes)
    if use_calc_stream:
        task.wait()
        return
    else:
        return task


S
ShenLiang 已提交
2193 2194 2195 2196
def _get_group_rank(global_rank, group=None):
    return global_rank if group is None else group.get_group_rank(global_rank)


L
lilong12 已提交
2197 2198 2199 2200 2201 2202
def send(tensor, dst=0, group=None, use_calc_stream=True):
    """
    Send a tensor to the receiver.

    Args:
        tensor (Tensor): The Tensor to send. Its data type
2203
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
L
lilong12 已提交
2204
        dst (int): The destination rank id.
L
lilong12 已提交
2205 2206
        group (Group, optional): The group instance return by new_group or None for global default group. Default: None.
        use_calc_stream (bool, optional): Whether to use calculate stream or communication stream. Default: True.
2207
    
L
lilong12 已提交
2208 2209 2210 2211 2212
    Returns:
        None.

    Examples:
        .. code-block:: python
2213

L
lilong12 已提交
2214
            # required: distributed
L
lilong12 已提交
2215
            import paddle
2216
            import paddle.distributed as dist
2217

2218 2219
            dist.init_parallel_env()
            if dist.get_rank() == 0:
L
lilong12 已提交
2220
                data = paddle.to_tensor([7, 8, 9])
2221
                dist.send(data, dst=1)
L
lilong12 已提交
2222
            else:
2223 2224 2225 2226
                data = paddle.to_tensor([1, 2, 3])
                dist.recv(data, src=0)
            print(data)
            # [7, 8, 9] (2 GPUs)
L
lilong12 已提交
2227 2228 2229
    """
    if group is not None and not group.is_member():
        return
S
ShenLiang 已提交
2230
    dst = _get_group_rank(dst, group)
L
lilong12 已提交
2231
    if in_dygraph_mode():
2232
        group = _get_default_group() if group is None else group
2233 2234
        backend = _group_map_backend[group]
        assert backend != 'gloo', ("backend gloo is not supported yet")
S
ShenLiang 已提交
2235
        task = group.process_group.send(tensor, dst)
2236 2237 2238 2239 2240 2241
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task

L
lilong12 已提交
2242 2243
    ring_id = 0 if group is None else group.id

J
Jiabin Yang 已提交
2244
    if _non_static_mode():
2245 2246
        return _legacy_C_ops.send_v2(tensor, 'use_calc_stream', use_calc_stream,
                                     'ring_id', ring_id, 'peer', dst)
W
wanghuancoder 已提交
2247
    op_type = 'send_v2'
L
lilong12 已提交
2248 2249 2250 2251 2252
    check_variable_and_dtype(
        tensor, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'],
        'send')

    helper = LayerHelper(op_type, **locals())
2253 2254 2255 2256 2257 2258 2259
    helper.append_op(type=op_type,
                     inputs={'X': [tensor]},
                     attrs={
                         'ring_id': ring_id,
                         'peer': dst,
                         'use_calc_stream': use_calc_stream,
                     })
L
lilong12 已提交
2260 2261 2262 2263 2264 2265 2266 2267


def recv(tensor, src=0, group=None, use_calc_stream=True):
    """
    Receive a tensor to the sender.

    Args:
        tensor (Tensor): The Tensor to receive. Its data type
2268
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
L
lilong12 已提交
2269
        src (int): The source rank id.
L
lilong12 已提交
2270 2271
        group (Group, optional): The group instance return by new_group or None for global default group. Default: None.
        use_calc_stream (bool, optional): Whether to use calculate stream or communication stream. Default: True.
2272
    
L
lilong12 已提交
2273 2274 2275 2276 2277
    Returns:
        None.

    Examples:
        .. code-block:: python
2278

L
lilong12 已提交
2279
            # required: distributed
L
lilong12 已提交
2280
            import paddle
2281
            import paddle.distributed as dist
2282

2283 2284
            dist.init_parallel_env()
            if dist.get_rank() == 0:
L
lilong12 已提交
2285
                data = paddle.to_tensor([7, 8, 9])
2286
                dist.send(data, dst=1)
L
lilong12 已提交
2287
            else:
2288 2289 2290 2291
                data = paddle.to_tensor([1, 2, 3])
                dist.recv(data, src=0)
            print(data)
            # [7, 8, 9] (2 GPUs)
L
lilong12 已提交
2292 2293 2294
    """
    if group is not None and not group.is_member():
        return
2295

S
ShenLiang 已提交
2296
    src = _get_group_rank(src, group)
L
lilong12 已提交
2297
    if in_dygraph_mode():
2298
        group = _get_default_group() if group is None else group
2299 2300
        backend = _group_map_backend[group]
        assert backend != 'gloo', ("backend gloo is not supported yet")
S
ShenLiang 已提交
2301
        task = group.process_group.recv(tensor, src)
2302 2303 2304 2305 2306 2307
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task

L
lilong12 已提交
2308 2309
    ring_id = 0 if group is None else group.id

J
Jiabin Yang 已提交
2310
    if _non_static_mode():
2311 2312 2313
        return _legacy_C_ops.recv_v2(tensor, 'use_calc_stream', use_calc_stream,
                                     'ring_id', ring_id, 'peer', src, 'dtype',
                                     tensor.dtype, 'out_shape', tensor.shape)
W
wanghuancoder 已提交
2314
    op_type = 'recv_v2'
L
lilong12 已提交
2315 2316 2317 2318
    check_variable_and_dtype(
        tensor, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'],
        'recv')
    helper = LayerHelper(op_type, **locals())
2319 2320 2321 2322 2323 2324 2325 2326 2327
    helper.append_op(type=op_type,
                     outputs={'Out': [tensor]},
                     attrs={
                         'ring_id': ring_id,
                         'peer': src,
                         'out_shape': tensor.shape,
                         'dtype': tensor.dtype,
                         'use_calc_stream': use_calc_stream,
                     })
2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349


def _check_single_tensor(tensor, tensor_name):
    if not isinstance(tensor, (core.eager.Tensor, paddle.Tensor)):
        raise RuntimeError("Invalid function argument. Expected parameter {}"
                           "to be of type paddle.Tensor, but it's {}".format(
                               tensor_name, type(tensor)))


def _check_tensor_list(tensor_list, tensor_name):
    if not isinstance(tensor_list, list) or \
        not all(isinstance(t, (core.eager.Tensor, paddle.Tensor)) for t in tensor_list):
        raise RuntimeError("Invalid function argument. Expected parameter {}"
                           "to be of type paddle.Tensor".format(tensor_name))


def isend(tensor, dst, group=None):
    """
    Sends a tensor asynchronously

    Args:
        tensor (Tensor): The Tensor to send. Its data type
2350
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367
        dst (int): The destination rank.
        group (Group, optional): The group instance return by new_group or None for global default group. Default: None.
    
    Returns:
        A distributed task object.

    Warning:    
        This API only supports the dygraph mode.

    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
2368
            if dist.get_rank() == 0:
2369
                data = paddle.to_tensor([7, 8, 9])
2370
                task = dist.isend(data, dst=1)
2371 2372
            else:
                data = paddle.to_tensor([1, 2, 3])
2373
                task = dist.irecv(data, src=0)
2374 2375
            task.wait()
            print(data)
2376
            # [7, 8, 9] (2 GPUs)
2377 2378 2379 2380 2381 2382 2383 2384

    """
    _check_single_tensor(tensor, "tensor")
    if group is not None and not group.is_member():
        return

    if in_dygraph_mode():
        group = _get_default_group() if group is None else group
2385 2386
        backend = _group_map_backend[group]
        assert backend != 'gloo', ("backend gloo is not supported yet")
2387 2388 2389 2390
        group_dst_rank = group.get_group_rank(dst)
        assert group_dst_rank >= 0, ("dst rank out of group, need global rank")
        return group.process_group.send(tensor, group_dst_rank)
    else:
2391
        raise RuntimeError("Only support eager dygraph mode.")
2392 2393 2394 2395 2396 2397 2398 2399


def irecv(tensor, src=None, group=None):
    """
    Receive a tensor to the sender.

    Args:
        tensor (Tensor): The Tensor to receive. Its data type
2400
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
2401 2402 2403 2404
        src (int): The source rank id.
        group (Group, optional): The group instance return by new_group or None for global default group. Default: None.

    Returns:
2405
        A distributed task object.
2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417

    Warning:    
        This API only supports the dygraph mode.

    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
2418
            if dist.get_rank() == 0:
2419
                data = paddle.to_tensor([7, 8, 9])
2420
                task = dist.isend(data, dst=1)
2421 2422
            else:
                data = paddle.to_tensor([1, 2, 3])
2423
                task = dist.irecv(data, src=0)
2424 2425
            task.wait()
            print(data)
2426
            # [7, 8, 9] (2 GPUs)
2427 2428 2429 2430 2431 2432 2433
    """
    _check_single_tensor(tensor, "tensor")
    if group is not None and not group.is_member():
        return

    if in_dygraph_mode():
        group = _get_default_group() if group is None else group
2434 2435
        backend = _group_map_backend[group]
        assert backend != 'gloo', ("backend gloo is not supported yet")
2436 2437 2438 2439
        group_src_rank = group.get_group_rank(src)
        assert group_src_rank >= 0, ("src rank out of group, need global rank")
        return group.process_group.recv(tensor, group_src_rank)
    else:
2440
        raise RuntimeError("Only support eager dygraph mode.")
2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582


class P2POp(object):
    """
    A class that makes point-to-point operations for "batch_isend_irecv".

    This class creates the type of P2P operation, communication buffer, peer rank,
    Group. Instances of this class will be passed to
    ``paddle.distributed.batch_isend_irecv`` for point-to-point communication.

    Args:
        op (callable): A function to send data to or receive data from a peer process.
            The type of ``op`` is either ``paddle.distributed.isend`` or ``paddle.distributed.irecv``.
        tensor (Tensor): Tensor to send or receive.
        peer (int): The destination or source rank.
        group (Group, optional): The group instance return by new_group or None for global 
            default group. Default: None.

    """

    def __init__(self, op, tensor, peer, group=None):
        if op not in [isend, irecv]:
            raise RuntimeError("Invalid ``op`` function. Expected ``op`` "
                               "to be of type ``paddle.distributed.isend`` or "
                               "``paddle.distributed.irecv``.")
        _check_single_tensor(tensor, "tensor")

        self.op = op
        self.tensor = tensor
        self.peer = peer
        self.group = _get_default_group() if group is None else group


@contextlib.contextmanager
def _with_batch_p2p_guard(backend):
    if backend == "nccl":
        core.ProcessGroupNCCL.group_start()
    try:
        yield
    finally:
        if backend == "nccl":
            core.ProcessGroupNCCL.group_end()


def _check_p2p_op_list(p2p_op_list):
    """
    Helper to check that the ``p2p_op_list`` is a list of P2POp instances and
    all ops use the same backend.
    """
    if not isinstance(p2p_op_list, list) or not all(
            isinstance(p2p_op, P2POp) for p2p_op in p2p_op_list):
        raise RuntimeError("Invalid ``p2p_op_list``. Each op is expected to "
                           "to be of type ``paddle.distributed.P2POp``.")

    backend = _group_map_backend[p2p_op_list[0].group]
    if not all(backend == _group_map_backend[p2p_op.group]
               for p2p_op in p2p_op_list):
        raise RuntimeError("All groups need to use the same backend.")


def batch_isend_irecv(p2p_op_list):
    """
    Send or Receive a batch of tensors asynchronously and return a list of requests.

    Process each of the point-to-point operations in ``p2p_op_list`` and return the 
    corresponding tasks. NCCL are currently supported.

    Args:
        p2p_op_list: A list of point-to-point operations(type of each operator is
            ``paddle.distributed.P2POp``). The order of the isend/irecv in the list
            matters and it needs to match with corresponding isend/irecv on the
            remote end.

    Returns:
        A list of distributed tasks returned by calling the corresponding
        op in the op_list. 

    Warning:    
        This API only supports the dygraph mode.

    Examples:
        .. code-block:: python

            # required: distributed

            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
            rank = dist.get_rank()
            world_size = dist.get_world_size()

            send_t = paddle.arange(2) + rank
            # paddle.tensor([0, 1])  # Rank-0
            # paddle.tensor([1, 2])  # Rank-1

            recv_t = paddle.empty(shape=[2], dtype=send_t.dtype)

            send_op = dist.P2POp(dist.isend, send_t, (rank + 1) % world_size)
            recv_op = dist.P2POp(dist.irecv, recv_t, (rank - 1 + world_size) % world_size)

            tasks = dist.batch_isend_irecv([send_op, recv_op])

            for task in tasks:
                task.wait()
            
            print(recv_t)
            # paddle.tensor([1, 2])     # Rank-0
            # paddle.tensor([0, 1])     # Rank-1
    """
    _check_p2p_op_list(p2p_op_list)
    group = p2p_op_list[0].group
    if group is not None and not group.is_member():
        return

    if in_dygraph_mode():
        group = _get_default_group() if group is None else group
        backend = _group_map_backend[group]
        tasks = []
        with _with_batch_p2p_guard(backend):
            for p2p_op in p2p_op_list:
                op = p2p_op.op
                tensor = p2p_op.tensor
                peer = p2p_op.peer
                comm_group = p2p_op.group
                task = op(tensor, peer, comm_group)
                if task is not None:
                    tasks.append(task)
        return tasks
    else:
        raise RuntimeError("Don't support static graph mode currently.")


def reduce_scatter(tensor,
                   tensor_list,
                   op=ReduceOp.SUM,
                   group=None,
                   use_calc_stream=True):
    """
    Reduces, then scatters a list of tensors to all processes in a group

    Args:
2583 2584 2585
        tensor (Tensor): Output tensor. Its data type should be float16, float32, float64, int32, int64, int8, uint8 or bool.
        tensor_list (list[Tensor]): List of tensors to reduce and scatter. Every element in the list must be a Tensor whose data type
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
2586
        op (ReduceOp.SUM|ReduceOp.MAX|ReduceOp.MIN|ReduceOp.PROD): Optional. The operation used. Default: ReduceOp.SUM.
2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606
        group (Group, optional): The group instance return by new_group or None for global 
            default group. Default: None.
        use_calc_stream (bool, optional): Whether this op should be an async op.

    Returns:
        Async task handle, if use_calc_stream is set to False.
        None, if use_calc_stream or if not part of the group.
    
    Warning:    
        This API only supports the dygraph mode.


    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
2607 2608 2609
            if dist.get_rank() == 0:
                data1 = paddle.to_tensor([0, 1])
                data2 = paddle.to_tensor([2, 3])
2610
            else:
2611 2612 2613 2614 2615 2616
                data1 = paddle.to_tensor([4, 5])
                data2 = paddle.to_tensor([6, 7])
            dist.reduce_scatter(data1, [data1, data2])
            print(data1)
            # [4, 6] (2 GPUs, out for rank 0)
            # [8, 10] (2 GPUs, out for rank 1)
2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627

    """
    _check_single_tensor(tensor, "tensor")
    _check_tensor_list(tensor_list, "tensor_list")

    if group is not None and not group.is_member():
        return

    if in_dygraph_mode():
        op_type = _get_reduce_op(op, "reduce_scatter")
        group = _get_default_group() if group is None else group
2628 2629
        backend = _group_map_backend[group]
        assert backend != 'gloo', ("backend gloo is not supported yet")
2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650

        temp = paddle.concat(tensor_list, axis=0)
        task = group.process_group._reduce_scatter_base(tensor, temp, op_type)
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task
    else:
        raise RuntimeError("Don't support static graph mode currently.")


def _reduce_scatter_base(output,
                         input,
                         op=ReduceOp.SUM,
                         group=None,
                         use_calc_stream=True):
    """
    Reduces, then scatters a flattened tensor to all processes in a group.

    Args:
2651 2652 2653
        output (Tensor): Output tensor. Its data type should be float16, float32, float64, int32, int64, int8, uint8 or bool.
        input (Tensor): Input tensor that is of size output tensor size times world size. Its data type 
            should be float16, float32, float64, int32, int64, int8, uint8 or bool.
2654
        op (ReduceOp.SUM|ReduceOp.MAX|ReduceOp.MIN|ReduceOp.PROD): Optional. The operation used. Default: ReduceOp.SUM.
2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671
        group (ProcessGroup, optional): The process group to work on. If None,
            the default process group will be used.
        use_calc_stream (bool, optional): Wether to use calculation stream (True) or communication stream (False).
            Default to True.
    Returns:
        Async task handle, if use_calc_stream is set to False.
        None, if use_calc_stream or if not part of the group.

    Examples:
        .. code-block:: python

            # required: distributed
            import paddle
            import paddle.distributed as dist

            dist.init_parallel_env()
            rank = dist.get_rank()
2672 2673 2674 2675 2676
            data = paddle.arange(4) + rank
            # [0, 1, 2, 3] (2 GPUs, for rank 0)
            # [1, 2, 3, 4] (2 GPUs, for rank 1)
            output = paddle.empty(shape=[2], dtype=data.dtype)
            dist.collective._reduce_scatter_base(output, data)
2677
            print(output)
2678 2679
            # [1, 3] (2 GPUs, out for rank 0)
            # [5, 7] (2 GPUs, out for rank 1)
2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698

    """
    _check_single_tensor(output, "output")
    _check_single_tensor(input, "input")

    if group is not None and not group.is_member():
        return

    if in_dygraph_mode():
        op_type = _get_reduce_op(op, "_reduce_scatter_base")
        group = _get_default_group() if group is None else group
        task = group.process_group._reduce_scatter_base(output, input, op_type)
        if use_calc_stream:
            task.wait()
            return None
        else:
            return task
    else:
        raise RuntimeError("Don't support static graph mode currently.")