reshard.py 41.2 KB
Newer Older
C
caozhou 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
# Copyright (c) 2021 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 copy
from functools import reduce

import paddle
import paddle.fluid.core as core
from paddle.utils import unique_name
from paddle.fluid.layer_helper import LayerHelper
from paddle.fluid.framework import Program, OpProtoHolder
23
from paddle.distributed.fleet.meta_optimizers.common import OpRole
C
caozhou 已提交
24 25
import paddle.fluid.layers.utils as utils
from ..collective import _get_global_env
26 27 28
from .dist_context import DistributedContext
from .dist_attribute import OperatorDistributedAttribute, TensorDistributedAttribute
from .process_group import new_process_group, ProcessGroup, _g_process_group_map
C
caozhou 已提交
29

30 31 32
# NOTE: If op in _g_special_ops, it will not be resharded. 
_g_special_ops = ['check_finite_and_unscale', 'update_loss_scaling']

C
caozhou 已提交
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282

class AllGatherOpDesc:
    """
    Describe the allgather op in the reshard phase.

    Args:
        group (list): Process group.
    """

    def __init__(self, group):
        self._group = group
        self._desc = "all_gather"

    @property
    def group(self):
        return self._group

    @property
    def desc(self):
        return self._desc

    def __repr__(self):
        return f"op: {self._desc}, group: {self._group}."


class SendOpDesc:
    """
    Describe the send op in the reshard phase.

    Args:
        partition_index (list): The index of partition in complete tensor.
        dst (int): The destination process to receive.
    """

    def __init__(self, partition_index, dst):
        self._dst = dst
        self._partition_index = partition_index
        self._desc = "send"

    @property
    def partition_index(self):
        return self._partition_index

    @property
    def dst(self):
        return self._dst

    @property
    def desc(self):
        return self._desc

    def __repr__(self):
        return f"op: {self._desc}, partition_index: {self._partition_index}, dst: {self._dst}."


class RecvOpDesc:
    """
    Describe the recv op in the reshard op.

    Args:
        partition_index (list): The index of partition in complete tensor.
        src (int): The source process to send.
    """

    def __init__(self, partition_index, src):
        self._src = src
        self._partition_index = partition_index
        self._desc = "recv"

    @property
    def partition_index(self):
        return self._partition_index

    @property
    def src(self):
        return self._src

    @property
    def desc(self):
        return self._desc

    def __repr__(self):
        return f"op: {self._desc}, partition_index: {self._partition_index}, src: {self._src}."


class SliceOpDesc:
    """
    Describe the slice op in the reshard phase.

    Args:
        starts (list): It represents starting indices of corresponding axis in ``axes``.
        ends (list):  It represents ending indices of corresponding axis in ``axes``.
        axes (list):  Axes that `starts` and `ends` apply to .
    """

    def __init__(self, starts, ends, axes):
        self._starts = starts
        self._ends = ends
        self._axes = axes
        self._desc = "slice"

    @property
    def starts(self):
        return self._starts

    @property
    def ends(self):
        return self._ends

    @property
    def axes(self):
        return self._axes

    @property
    def desc(self):
        return self._desc

    def __repr__(self):
        return f"op: {self._desc}, starts: {self._starts}, ends: {self._ends}, axes: {self._axes}."


class ConcatOpDesc:
    """
    Describe the concat op in the reshard phase.

    Args:
        partition_index_list (list): A list contains all partition index.
    """

    def __init__(self, partition_index_list):
        self._partition_index_list = partition_index_list
        self._desc = "concat"

    @property
    def partition_index_list(self):
        return self._partition_index_list

    @property
    def desc(self):
        return self._desc

    def __repr__(self):
        return f"op: {self._desc}, partition_index_list: {self._partition_index_list}."


def _compute_partition_shape(complete_shape, dims_mapping, process_shape):
    """Compute the shape of partition."""
    partition_shape = []
    for idx, item in enumerate(complete_shape):
        if dims_mapping[idx] == -1:
            partition_shape.append(item)
        else:
            partition_shape.append(item // process_shape[dims_mapping[idx]])

    return partition_shape


def _compute_process_index(process, process_group, process_shape):
    """Compute the index of process_shape corresponding to the process."""
    relative_process = process_group.index(process)
    process_index = []
    product = reduce(lambda x, y: x * y, process_shape)

    for i in range(len(process_shape)):
        idx = relative_process // (product // process_shape[i])
        product = product // process_shape[i]
        relative_process = relative_process - relative_process // product * product
        process_index.append(idx)

    return process_index


def _compute_partition_index(process, complete_shape, dims_mapping,
                             process_shape, process_group):
    """Compute the partition index in complete tensor."""
    partition_shape = _compute_partition_shape(complete_shape, dims_mapping,
                                               process_shape)
    process_index = _compute_process_index(process, process_group,
                                           process_shape)
    partition_index = []

    for i in range(len(complete_shape)):
        if dims_mapping[i] == -1:
            partition_index.append([0, partition_shape[i]])
        else:
            partition_index.append([
                process_index[dims_mapping[i]] * partition_shape[i],
                (process_index[dims_mapping[i]] + 1) * partition_shape[i]
            ])

    return partition_index


def _compute_concat_info(partition_index_x, partition_index_y):
    """Judge whether two partition can be concatenated and compute concatenated partition index."""
    differ_count = 0
    concat_axis = -1
    first_order = 0
    new_partition = []

    for idx, item in enumerate(partition_index_x):
        if item != partition_index_y[idx]:
            differ_count += 1
            if item[1] == partition_index_y[idx][0] and item[
                    0] < partition_index_y[idx][1]:
                concat_axis = idx
                new_partition.append([item[0], partition_index_y[idx][1]])
            elif item[0] == partition_index_y[idx][1] and item[
                    1] > partition_index_y[idx][0]:
                first_order = 1
                concat_axis = idx
                new_partition.append([partition_index_y[idx][0], item[1]])
        else:
            new_partition.append(item)

    if differ_count == 1:
        return concat_axis, first_order, new_partition
    else:
        return -1, first_order, new_partition


def _concat_partitions(partition_index_list, partition_index):
    """Concat the given partitions without inserting concat op."""
    if not partition_index_list:
        partition_index_list.append(partition_index)
    else:
        i = 0
        has_concat = False
        while i < len(partition_index_list):
            concat_axis, _, new_partition = _compute_concat_info(
                partition_index_list[i], partition_index)
            if concat_axis != -1:
                has_concat = True
                partition_index_list.pop(i)
                _concat_partitions(partition_index_list, new_partition)
                break
            i += 1
        if not has_concat:
            partition_index_list.append(partition_index)


def _is_overlapped(shape_x, shape_y):
    """Judge whether two partitions intersect on the specified dimension."""
    overlapped = False
    if (shape_y[0] <= shape_x[0] < shape_y[1]) or (
            shape_x[0] <= shape_y[0] < shape_x[1]):
        overlapped = True
    return overlapped


283
def _need_reshard(dist_tensor, dist_op, op_input=True):
C
caozhou 已提交
284 285
    """Judge the tensor whether needs to be resharded."""
    is_reshard = False
286 287 288 289 290 291 292
    tensor_dist_attr = dist_tensor.dist_attr
    tensor_name = dist_tensor.serial_tensor.name
    tensor_dims_mapping = tensor_dist_attr.dims_mapping
    tensor_process_mesh = tensor_dist_attr.process_mesh
    op_dist_attr = dist_op.dist_attr
    op_input_dims_mapping = op_dist_attr.get_input_dims_mapping(tensor_name)
    op_process_mesh = op_dist_attr.process_mesh
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317
    if op_input:
        op_input_dims_mapping = op_dist_attr.get_input_dims_mapping(tensor_name)
        op_process_mesh = op_dist_attr.process_mesh
        if all(
                map(lambda x: x is not None, [
                    tensor_dims_mapping, tensor_process_mesh,
                    op_input_dims_mapping, op_process_mesh
                ])):
            if tensor_dims_mapping != op_input_dims_mapping or tensor_process_mesh != op_process_mesh:
                is_reshard = True
    else:
        op_output_dims_mapping = op_dist_attr.get_output_dims_mapping(
            tensor_name)
        op_process_mesh = op_dist_attr.process_mesh
        if all(
                map(lambda x: x is not None, [
                    tensor_dims_mapping, tensor_process_mesh,
                    op_output_dims_mapping, op_process_mesh
                ])):
            if tensor_process_mesh != op_process_mesh:
                is_reshard = True
            if tensor_dims_mapping != op_output_dims_mapping:
                raise ValueError(
                    "It is not supported that tensor dims mapping is different from op output dims mapping."
                )
C
caozhou 已提交
318 319 320 321 322 323 324 325 326 327 328 329 330 331
    return is_reshard


def _compute_complete_shape(slice_shape, process_shape, dims_mapping):
    """compute the complete shape of the slice tensor  with its process mesh and dims mapping"""
    complete_shape = []
    for idx, item in enumerate(slice_shape):
        if dims_mapping[idx] == -1:
            complete_shape.append(item)
        else:
            complete_shape.append(item * process_shape[dims_mapping[idx]])
    return complete_shape


332
def find_op_desc_seq(dist_tensor, dist_op):
C
caozhou 已提交
333 334 335 336
    """
    Find the op description sequence to reshard the source tensor for matching the op requirement.

    Args:
337 338
        dist_tensor (DistributedTensor): A distributed tensor.
        dist_op (DistributedOperator): A distributed operator.
C
caozhou 已提交
339 340 341 342 343

    Returns:
        Dict, the dict represents the required op description sequence corresponding to process, The key of dict is
        process and value is a list containing op description.
    """
344 345 346 347 348 349
    tensor_dist_attr = dist_tensor.dist_attr
    source_tensor = dist_tensor.serial_tensor
    tensor_name = source_tensor.name
    source_dims_mapping = tensor_dist_attr.dims_mapping
    source_process_mesh = tensor_dist_attr.process_mesh
    source_process_group = source_process_mesh.processes
C
caozhou 已提交
350 351
    source_process_shape = source_process_mesh.topology

352 353 354 355
    op_dist_attr = dist_op.dist_attr
    target_process_mesh = op_dist_attr.process_mesh
    target_dims_mapping = op_dist_attr.get_input_dims_mapping(tensor_name)
    target_process_group = target_process_mesh.processes
C
caozhou 已提交
356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651
    target_process_shape = target_process_mesh.topology

    complete_shape = _compute_complete_shape(
        source_tensor.shape, source_process_shape, source_dims_mapping)
    op_desc_seq = {}

    # TODO: if the target process group has the same process with source process group
    if set(target_process_group).intersection(set(
            source_process_group)) and set(target_process_group).difference(
                set(source_process_group)):
        pass

    # in the different process group, it will use send, recv, concat and slice op
    elif target_process_group != source_process_group:
        partition_process_mapping_list = []
        for source_process in source_process_group:
            source_partition_index = _compute_partition_index(source_process, complete_shape, source_dims_mapping, \
                                                              source_process_shape, source_process_group)
            if not partition_process_mapping_list:
                partition_process_mapping_list.append(
                    [source_partition_index, [source_process], [False]])
            else:
                partition_list = list(
                    [item[0] for item in partition_process_mapping_list])
                process_list = list(
                    [item[1] for item in partition_process_mapping_list])
                has_used = list(
                    [item[2] for item in partition_process_mapping_list])
                if partition_list.count(source_partition_index) == 1:
                    index = partition_list.index(source_partition_index)
                    process_list[index].append(source_process)
                    has_used[index].append(False)
                else:
                    partition_process_mapping_list.append(
                        [source_partition_index, [source_process], [False]])

        for target_process in target_process_group:
            has_sent = []
            target_partition_index = _compute_partition_index(
                target_process, complete_shape, target_dims_mapping,
                target_process_shape, target_process_group)
            partition_index_list = []
            all_partition_index_list = []
            for source_process in source_process_group:
                source_partition_index = _compute_partition_index(
                    source_process, complete_shape, source_dims_mapping,
                    source_process_shape, source_process_group)
                to_send_process = None
                if all(_ for _ in list(map(_is_overlapped, source_partition_index, target_partition_index))) \
                        and source_partition_index not in has_sent:
                    idx = list([
                        item[0] for item in partition_process_mapping_list
                    ]).index(source_partition_index)
                    has_used = list(
                        [item[2]
                         for item in partition_process_mapping_list])[idx]
                    process_list = list(
                        [item[1]
                         for item in partition_process_mapping_list])[idx]
                    i = 0
                    while i < len(has_used):
                        if not has_used[i]:
                            to_send_process = process_list[i]
                            has_used[i] = True
                            break
                        i += 1
                    if i == len(has_used):
                        has_used = list(map(lambda x: False, has_used))
                        to_send_process = process_list[0]
                        has_used[0] = True
                    assert to_send_process is not None, "Failed to find the send process."

                    if to_send_process not in op_desc_seq.keys():
                        op_desc_seq[to_send_process] = []
                    if target_process not in op_desc_seq.keys():
                        op_desc_seq[target_process] = []
                    all_partition_index_list.append(source_partition_index)

                    # append send and recv op desc
                    send_op_desc = SendOpDesc(source_partition_index,
                                              target_process)
                    recv_op_desc = RecvOpDesc(source_partition_index,
                                              to_send_process)
                    op_desc_seq[to_send_process].append(send_op_desc)
                    op_desc_seq[target_process].append(recv_op_desc)
                    has_sent.append(source_partition_index)
                    _concat_partitions(partition_index_list,
                                       source_partition_index)

            # append concat op desc
            op_desc_seq[target_process].append(
                ConcatOpDesc(all_partition_index_list))

            # append slice op desc
            slice_starts = []
            slice_ends = []
            slices_axes = []
            concatenated_partition_index = partition_index_list[0]
            for idx, item in enumerate(concatenated_partition_index):
                slice_starts.append(target_partition_index[idx][0] - item[0])
                slice_ends.append(target_partition_index[idx][1] - item[0])
                slices_axes.append(idx)
            op_desc_seq[target_process].append(
                SliceOpDesc(slice_starts, slice_ends, slices_axes))

    # in the same process group, it will use allgahther and slice op
    else:
        partition_index_list = []
        all_partition_index_list = []
        process_index = []
        for source_process in source_process_group:
            source_partition_index = _compute_partition_index(
                source_process, complete_shape, source_dims_mapping,
                source_process_shape, source_process_group)
            if source_partition_index not in partition_index_list:
                partition_index_list.append(source_partition_index)
                process_index.append(
                    [[source_process, ], source_partition_index])
            else:
                process_index[partition_index_list.index(
                    source_partition_index)][0].append(source_process)

        for i in range(len(process_index[0][0])):
            group = []
            for j in range(len(process_index)):
                group.append(process_index[j][0][i])
                if i == 0:
                    all_partition_index_list.append(process_index[j][1])
            for process in group:
                # append slice op desc
                slice_starts = []
                slice_ends = []
                slices_axes = []
                target_partition_index = _compute_partition_index(
                    process, complete_shape, target_dims_mapping,
                    target_process_shape, target_process_group)
                for idx, item in enumerate(target_partition_index):
                    slice_starts.append(item[0])
                    slice_ends.append(item[1])
                    slices_axes.append(idx)

                slice_op_desc = SliceOpDesc(
                    starts=slice_starts, ends=slice_ends, axes=slices_axes)
                op_desc_seq[process] = [AllGatherOpDesc(group=group),
                                        ConcatOpDesc(partition_index_list=all_partition_index_list), slice_op_desc] \
                    if len(group) > 1 else [slice_op_desc]

    return op_desc_seq


def _insert_send_op(block, idx, tensor, dst):
    """Insert send op into block at the given index."""
    op_type = 'send_v2'
    block._insert_op(
        idx,
        type=op_type,
        inputs={'X': [tensor]},
        attrs={
            'ring_id': 0,
            'peer': dst,
            'use_calc_stream': True,
        })


def _insert_recv_op(block, idx, tensor, src):
    """Insert recv op into block at the given index."""
    op_type = 'recv_v2'
    block._insert_op(
        idx,
        type=op_type,
        inputs={'X': [tensor]},
        outputs={'Out': [tensor]},
        attrs={
            'ring_id': 0,
            'peer': src,
            'out_shape': tensor.shape,
            'dtype': tensor.dtype,
            'use_calc_stream': True,
        })


def _insert_concat_op(block, idx, tensors, axis):
    """Insert concat op into block at the given block."""
    inputs = {'X': tensors}
    attrs = {}
    attrs['axis'] = axis
    helper = LayerHelper('concat', **locals())
    with paddle.static.program_guard(block.program):
        out = helper.create_variable_for_type_inference(
            dtype=helper.input_dtype())
    block._insert_op(
        idx, type='concat', inputs=inputs, outputs={'Out': [out]}, attrs=attrs)
    return out


def _insert_slice_op(block, idx, tensor, starts, ends, axes, new_var_name):
    """Insert slice op into block at the given block."""
    inputs = {'Input': tensor}
    infer_flags = list(1 for i in range(len(axes)))
    attrs = {
        "axes": axes,
        "starts": starts,
        "ends": ends,
        "infer_flags": infer_flags
    }
    helper = LayerHelper('slice', **locals())
    out = block.create_var(
        name=new_var_name,
        dtype=tensor.dtype,
        type=core.VarDesc.VarType.LOD_TENSOR)
    block._insert_op(
        idx, type="slice", inputs=inputs, outputs={'Out': [out]}, attrs=attrs)
    return out


def _insert_split_op(block, idx, tensor, num_or_sections):
    """Insert split op into block at the given index."""
    helper = LayerHelper('split', **locals())
    input_shape = tensor.shape
    inputs = {'X': tensor}
    attrs = {'num': num_or_sections, "axis": 0}
    with paddle.static.program_guard(block.program):
        outs = [
            helper.create_variable_for_type_inference(
                dtype=helper.input_dtype()) for i in range(num_or_sections)
        ]
    block._insert_op(
        idx, type="split", inputs=inputs, outputs={'Out': outs}, attrs=attrs)
    return outs


def _insert_allgather_op(block, idx, tensor, ranks):
    """Insert allgather op into block at the given index."""

    def _insert_fill_constant_op(block, idx):
        """Insert fill constant op into block at the given index."""
        helper = LayerHelper("fill_constant", **locals())
        with paddle.static.program_guard(block.program):
            out = helper.create_variable_for_type_inference(dtype="int32")
        inputs = {}
        attrs = {'force_cpu': False}
        attrs['str_value'] = str(int("1"))
        attrs['value'] = int("1")
        attrs['dtype'] = out.dtype
        utils.get_shape_tensor_inputs(
            inputs=inputs, attrs=attrs, shape=[0], op_type='fill_constant')
        block._insert_op(
            idx,
            type='fill_constant',
            inputs=inputs,
            outputs={'Out': [out]},
            attrs=attrs)
        out.stop_gradient = True
        return out

    tensor_list = []
    group = new_process_group(ranks)
    idx_offset = 0

    # instant process group before insert allgather op.
    if not group.is_instantiate():
        # insert fill_constant op
        fill_constant_out = _insert_fill_constant_op(block, idx)
        fill_constant_out.stop_gradient = True

        # insert c_allreduce_sum op
        block._insert_op(
            idx + 1,
            type="c_allreduce_sum",
            inputs={'X': [fill_constant_out]},
            outputs={'Out': [fill_constant_out]},
            attrs={'ring_id': 0,
                   'use_calc_stream': True})

        # insert c_sync_calc_stream op
        block._insert_op(
            idx + 2,
            type="c_sync_calc_stream",
            inputs={'X': [fill_constant_out]},
            outputs={'Out': [fill_constant_out]})
        idx_offset = 3

    # insert c_allgather op
    op_type = 'c_allgather'
    helper = LayerHelper(op_type, **locals())
    with paddle.static.program_guard(block.program):
        allgather_out = helper.create_variable_for_type_inference(
            dtype=tensor.dtype)
    block._insert_op(
        idx + idx_offset,
        type=op_type,
        inputs={'X': [tensor]},
        outputs={'Out': [allgather_out]},
        attrs={
            'ring_id': group.id,
            'use_calc_stream': True,
652
            'nranks': group.nranks
C
caozhou 已提交
653 654 655 656 657
        })
    idx_offset += 1

    # insert split op
    split_out = _insert_split_op(block, idx + idx_offset, allgather_out,
658
                                 group.nranks)
C
caozhou 已提交
659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791
    idx_offset += 1
    tensor_list.extend(split_out)
    return tensor_list, idx_offset


def _concat_partitions_with_op(partition_tensor_list, tensor, partition_index,
                               block, idx):
    """Concat the tensors and insert concat op."""
    if not partition_tensor_list:
        partition_tensor_list.append((tensor, partition_index))
    else:
        i = 0
        has_concat = False
        while i < len(partition_tensor_list):
            concat_axis, first_order, new_partition = _compute_concat_info(
                partition_tensor_list[i][1], partition_index)
            if concat_axis != -1:
                has_concat = True
                _ = _insert_concat_op(block, idx[0], [partition_tensor_list[i][0], tensor], concat_axis) \
                    if first_order == 0 else \
                    _insert_concat_op(block, idx[0], [tensor, partition_tensor_list[i][0]], concat_axis)
                partition_tensor_list.pop(i)
                idx[0] += 1
                _concat_partitions_with_op(partition_tensor_list, _,
                                           new_partition, block, idx)
                break
            i += 1
        if not has_concat:
            partition_tensor_list.append((tensor, partition_index))


HAS_SENT = {}
HAS_RECV = {}
HAS_ALLGATHER = {}


def parse_op_desc(program, rank_id, op_desc_seq, var_name, reshard_op,
                  dist_context):
    """Parse op desc sequence and insert op in the block"""
    global HAS_SENT
    global HAS_RECV
    global HAS_ALLGATHER
    tensor_list = []
    partition_tensor_list = []
    if rank_id not in op_desc_seq.keys():
        return
    op_desc_list = op_desc_seq[rank_id]
    block = program.global_block()
    assert var_name in block.vars.keys(
    ), "The {} cannot be found in the {} program.".format(var_name, rank_id)

    idx = None
    for index, op in list(enumerate(block.ops)):
        if op.desc.id == reshard_op.desc.id:
            idx = index
            break
    assert idx is not None, "The op for reshard cannot be found in the rank {} program.".format(
        rank_id)

    matched_op = block.ops[idx]
    source_tensor = block.vars[var_name]
    for op_desc in op_desc_list:
        if isinstance(op_desc, AllGatherOpDesc):  # noqa: F401
            if var_name not in HAS_ALLGATHER.keys():
                HAS_ALLGATHER[var_name] = []
            if not HAS_ALLGATHER[var_name] or op_desc.group not in list(
                    map(lambda x: x[0], HAS_ALLGATHER[var_name])):
                tensor_list, idx_offset = _insert_allgather_op(
                    block, idx, source_tensor, op_desc.group)
                idx += idx_offset
                tensor_name_list = [var.name for var in tensor_list]
                HAS_ALLGATHER[var_name].append(
                    [op_desc.group, tensor_name_list])
            else:
                for item in HAS_ALLGATHER[var_name]:
                    if op_desc.group == item[0]:
                        tensor_list = [
                            program.global_block().vars[var_name]
                            for var_name in item[1]
                        ]
                        break
            assert tensor_list, "The result of parsing allgather op should not be None."

        elif isinstance(op_desc, SendOpDesc):
            if var_name not in HAS_SENT.keys():
                HAS_SENT[var_name] = []
            if op_desc.dst not in HAS_SENT[var_name]:
                _insert_send_op(block, idx, source_tensor, op_desc.dst)
                idx += 1
                HAS_SENT[var_name].append(op_desc.dst)

        elif isinstance(op_desc, RecvOpDesc):
            if var_name not in HAS_RECV.keys():
                HAS_RECV[var_name] = {}
            if op_desc.src not in HAS_RECV[var_name].keys():
                partition_index = op_desc.partition_index
                shape = []
                for index in partition_index:
                    shape.append(index[1] - index[0])
                recv_tensor = block.create_var(
                    name=unique_name.generate(var_name + "@recv"),
                    shape=shape,
                    dtype=source_tensor.dtype)
                _insert_recv_op(block, idx, recv_tensor, op_desc.src)
                tensor_list.append(recv_tensor)
                idx += 1
                HAS_RECV[var_name][op_desc.src] = recv_tensor
            else:
                tensor_list.append(HAS_RECV[var_name][op_desc.src])

        elif isinstance(op_desc, ConcatOpDesc):
            partition_index_list = op_desc.partition_index_list
            idx_list = [idx]
            for index, tensor in enumerate(tensor_list):
                _concat_partitions_with_op(partition_tensor_list, tensor,
                                           partition_index_list[index], block,
                                           idx_list)
            idx = idx_list[0]

        elif isinstance(op_desc, SliceOpDesc):
            assert len(partition_tensor_list) == 1 or not partition_tensor_list
            to_slice_tensor = partition_tensor_list[0][0] if len(
                partition_tensor_list) == 1 else source_tensor
            new_name = unique_name.generate(var_name + "@RESHARD")
            target_tensor = _insert_slice_op(
                block,
                idx,
                to_slice_tensor,
                starts=op_desc.starts,
                ends=op_desc.ends,
                axes=op_desc.axes,
                new_var_name=new_name)

792 793 794 795
            tensor_attr = TensorDistributedAttribute()
            process_mesh = dist_context.get_op_dist_attr_for_program(
                matched_op).process_mesh
            dims_mapping = dist_context.get_op_dist_attr_for_program(
C
caozhou 已提交
796
                matched_op).get_input_dims_mapping(var_name)
797 798 799 800
            tensor_attr.dims_mapping = dims_mapping
            tensor_attr.process_mesh = process_mesh
            dist_context.set_tensor_dist_attr_for_program(target_tensor,
                                                          tensor_attr)
C
caozhou 已提交
801 802 803 804

            # rename op input name according to new name
            for op in block.ops:
                for name in op.input_arg_names:
805
                    op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
C
caozhou 已提交
806
                    if name == var_name and op_dist_attr is not None:
807
                        op_process_mesh = op_dist_attr.process_mesh
C
caozhou 已提交
808 809
                        op_input_dims_mapping = op_dist_attr.get_input_dims_mapping(
                            var_name)
810
                        if op_process_mesh == process_mesh and op_input_dims_mapping == dims_mapping:
C
caozhou 已提交
811 812 813
                            op.desc._rename_input(name, target_tensor.name)
                            op_dist_attr.set_input_dims_mapping(
                                target_tensor.name, dims_mapping)
814
                            op_dist_attr.set_input_dist_attr(name, None)
C
caozhou 已提交
815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841


def _remove_no_need_ops(auto_parallel_main_prog, dist_context, rank_id):
    """Remove no need ops in the main program"""
    not_remove_op_ref = [
        "create_py_reader", "create_double_buffer_reader", "read"
    ]
    remove_op_idx = []
    block = auto_parallel_main_prog.global_block()
    ops = block.ops
    vars = block.vars
    for idx, op in enumerate(ops):
        # handle read op in the pipeline scene specially, it will be removed in the future.
        if op.type == "read":
            dim_list = []
            for var_name in op.output_arg_names:
                dim_list.extend(vars[var_name].shape)
            for i in range(idx, -1, -1):
                if ops[i].type == "create_py_reader":
                    ops[i]._set_attr("shape_concat", dim_list)
                    break
            continue

        # replace the input and output of c_sync_comm_stream op when in pipeline scene.
        if op.type == "c_sync_comm_stream":
            need_save = []
            for var_name in op.input_arg_names:
842 843 844
                process_mesh = dist_context.get_tensor_dist_attr_for_program(
                    vars[var_name]).process_mesh
                if rank_id in process_mesh.processes:
C
caozhou 已提交
845 846 847 848 849 850 851 852 853 854 855
                    need_save.append(var_name)
            if not need_save:
                remove_op_idx.append(idx)
                continue

            proto = OpProtoHolder.instance().get_op_proto(op.type)
            op.desc.set_input(proto.inputs[0].name, need_save)
            op.desc.set_output(proto.outputs[0].name, need_save)
            continue

        # judge the other op whether should be removed.
856
        op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
C
caozhou 已提交
857
        if op_dist_attr is not None:
858 859
            op_process_mesh = op_dist_attr.process_mesh
            if rank_id not in op_process_mesh.processes and op.type not in not_remove_op_ref:
C
caozhou 已提交
860 861 862 863 864 865
                remove_op_idx.append(idx)

    for idx in remove_op_idx[::-1]:
        block._remove_op(idx)


866
def _remove_no_need_vars(auto_parallel_main_prog, dist_params_grads):
C
caozhou 已提交
867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882
    """Remove no need vars in the main program"""
    remove_vars = set()
    block = auto_parallel_main_prog.global_block()
    ops = block.ops
    vars = block.vars
    need_vars = set()
    for op in ops:
        for var_name in op.input_arg_names:
            if var_name in vars:
                need_vars.add(var_name)
        for var_name in op.output_arg_names:
            if var_name in vars:
                need_vars.add(var_name)
    for var in vars:
        if var not in need_vars:
            remove_vars.add(var)
883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909

    # change dist_params_grads
    param_grad_map = {}
    for op in ops:
        if int(op.attr('op_role')) == int(OpRole.Optimize):
            if "Param" in op.input_names and "Grad" in op.input_names:
                param_name = op.input("Param")[0]
                grad_name = op.input("Grad")[0]
                param_grad_map[param_name] = grad_name

    need_remove_idx = []
    for idx, item in enumerate(dist_params_grads):
        if item[0].name not in param_grad_map.keys():
            need_remove_idx.append(idx)

    for idx in need_remove_idx[::-1]:
        dist_params_grads.pop(idx)

    idx = 0
    while idx < len(dist_params_grads):
        param_name = dist_params_grads[idx][0].name
        grad_name = dist_params_grads[idx][1].name
        if grad_name != param_grad_map[param_name]:
            dist_params_grads[idx] = (vars[param_name],
                                      vars[param_grad_map[param_name]])
        idx += 1

C
caozhou 已提交
910 911 912 913
    for var in remove_vars:
        block._remove_var(var)


914 915
def remove_no_need_in_main(auto_parallel_main_prog, dist_context, rank_id,
                           dist_params_grads):
C
caozhou 已提交
916 917
    """Remove no need vars and ops in the main program."""
    _remove_no_need_ops(auto_parallel_main_prog, dist_context, rank_id)
918
    _remove_no_need_vars(auto_parallel_main_prog, dist_params_grads)
C
caozhou 已提交
919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995


def remove_no_need_in_startup(auto_parallel_main_prog,
                              auto_parallel_startup_prog):
    """Remove no need vars and ops in the startup program."""
    main_input_vars = set()
    main_ops = auto_parallel_main_prog.global_block().ops
    for op in main_ops:
        for var_name in op.input_arg_names:
            main_input_vars.add(var_name)

    startup_block = auto_parallel_startup_prog.global_block()
    startup_output_vars = set()
    startup_ops = startup_block.ops
    for op in startup_ops:
        # skip c_sync_comm_stream op
        if op.type == "c_sync_comm_stream":
            continue
        for var_name in op.output_arg_names:
            startup_output_vars.add(var_name)

    need_vars = set()
    for var_name in startup_output_vars:
        if var_name in main_input_vars:
            need_vars.add(var_name)

    startup_ops = startup_block.ops
    actual_need_vars = set()
    for idx, op in enumerate(startup_ops):
        is_need_op = False
        if op.type == "c_sync_comm_stream":
            continue
        for var_name in op.output_arg_names:
            if var_name in need_vars:
                is_need_op = True
                break
        if is_need_op:
            for var_name in op.output_arg_names:
                actual_need_vars.add(var_name)
            for var_name in op.input_arg_names:
                actual_need_vars.add(var_name)

    remove_vars = set()
    for var_name in startup_block.vars:
        if var_name not in actual_need_vars:
            remove_vars.add(var_name)
    for var in remove_vars:
        startup_block._remove_var(var)

    remove_op_idx = []
    vars = startup_block.vars
    for idx, op in enumerate(startup_block.ops):
        is_no_need_op = False
        if op.type == "c_sync_comm_stream":
            var_names = []
            for var_name in op.input_arg_names:
                if var_name in vars:
                    var_names.append(var_name)
            if not var_names:
                remove_op_idx.append(idx)
            else:
                proto = OpProtoHolder.instance().get_op_proto(op.type)
                op.desc.set_input(proto.inputs[0].name, var_names)
                op.desc.set_output(proto.outputs[0].name, var_names)
            continue

        for var_name in op.output_arg_names:
            if var_name not in vars:
                is_no_need_op = True
                break
        if is_no_need_op:
            remove_op_idx.append(idx)
    for idx in remove_op_idx[::-1]:
        startup_block._remove_op(idx)


def reshard(auto_parallel_main_prog, auto_parallel_startup_prog, rank_id,
996
            dist_context, dist_params_grads):
C
caozhou 已提交
997
    """
998
    Reshard tensor in the program according to its distributed attribute and corresponding op distributed attribute.
C
caozhou 已提交
999 1000 1001 1002 1003

    Args:
        auto_parallel_main_prog (Program): An auto parallel main program.
        auto_parallel_startup_prog (Program): An auto parallel startup program.
        rank_id (int): The process id.
1004
        dist_context (DistributedContext): The distributed context of this rank.
1005
        dist_params_grads (list): The list contains the tuple of param and grad.
C
caozhou 已提交
1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020
    """
    assert isinstance(auto_parallel_main_prog, Program), "The type of auto_parallel_main_prog should be Program, " \
                                         "but got {}.".format(type(auto_parallel_main_prog))
    assert isinstance(auto_parallel_main_prog, Program), "The type of auto_parallel_startup_prog should be Program, " \
                                         "but got {}.".format(type(auto_parallel_startup_prog))
    assert isinstance(rank_id, int), "The type of rank_id should be int, " \
                                         "but got {}.".format(type(rank_id))
    assert isinstance(dist_context, DistributedContext), "The type of dist_context should be DistributedContext, " \
                                         "but got {}.".format(type(dist_context))

    block = auto_parallel_main_prog.global_block()
    idx = 0
    while idx < len(block.ops):
        pre_op_count = len(block.ops)
        op = block.ops[idx]
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031

        def _is_special_op(op):
            global _g_special_ops
            if op.type in _g_special_ops:
                return True
            return False

        if _is_special_op(op):
            idx += 1
            continue

1032 1033
        dist_op = dist_context.get_dist_op_for_program(op)
        if dist_op is not None:
C
caozhou 已提交
1034 1035 1036 1037 1038 1039
            idx_offset = 0
            for var_name in op.input_arg_names:
                # skip lod_tensor_blocking_queue_0
                if var_name == "lod_tensor_blocking_queue_0":
                    continue
                var = block.vars[var_name]
1040 1041 1042 1043
                dist_tensor = dist_context.get_dist_tensor_for_program(var)
                if dist_tensor is not None and _need_reshard(dist_tensor,
                                                             dist_op):
                    reshard_op_desc = find_op_desc_seq(dist_tensor, dist_op)
C
caozhou 已提交
1044 1045 1046 1047 1048 1049 1050 1051 1052
                    parse_op_desc(auto_parallel_main_prog, rank_id,
                                  reshard_op_desc, var_name, op, dist_context)
                    cur_op_count = len(block.ops)
                    idx_offset = idx_offset + cur_op_count - pre_op_count
                    pre_op_count = cur_op_count
            idx = idx + idx_offset + 1
        else:
            idx += 1

1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080
    # insert send and recv op if output process mesh is different from tensor process mesh
    idx = 0
    skip_ops = ["create_py_reader", "create_double_buffer_reader", "read"]
    while idx < len(block.ops):
        pre_op_count = len(block.ops)
        op = block.ops[idx]
        dist_op = dist_context.get_dist_op_for_program(op)
        if dist_op is not None and op.type not in skip_ops:
            for var_name in op.output_arg_names:
                var = block.vars[var_name]
                dist_tensor = dist_context.get_dist_tensor_for_program(var)
                if dist_tensor is not None and _need_reshard(dist_tensor,
                                                             dist_op, False):
                    for index, item in enumerate(
                            dist_op.dist_attr.process_mesh.processes):
                        recv_rank = dist_tensor.dist_attr.process_mesh.processes[
                            index]
                        if rank_id == item:
                            _insert_send_op(block, idx + 1, var, recv_rank)
                        if rank_id == recv_rank:
                            _insert_recv_op(block, idx + 1, var, item)
                    cur_op_count = len(block.ops)
                    idx_offset = idx_offset + cur_op_count - pre_op_count
                    pre_op_count = cur_op_count
            idx = idx + idx_offset + 1
        else:
            idx += 1

C
caozhou 已提交
1081
    # remove no need vars and ops in the main program
1082 1083
    remove_no_need_in_main(auto_parallel_main_prog, dist_context, rank_id,
                           dist_params_grads)
C
caozhou 已提交
1084 1085 1086 1087

    # remove no need vars and ops in the startip program
    remove_no_need_in_startup(auto_parallel_main_prog,
                              auto_parallel_startup_prog)