utils.py 55.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# 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

15
import os
16
import copy
17
import paddle
18
import threading
19
import numpy as np
20 21
import warnings
import logging
22
from functools import reduce
23 24 25

import paddle.fluid.core as core
from paddle.framework.io import _to_LodTensor
26
from paddle.distributed.fleet.meta_optimizers.common import OpRole
27
from paddle.fluid.io import is_parameter, is_belong_to_optimizer
J
JZ-LIANG 已提交
28
from paddle.distributed.auto_parallel.dist_attribute import TensorDistributedAttribute, OperatorDistributedAttribute
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92


def is_valid_list_index(list, index):
    if index >= -len(list) and index < len(list):
        return True
    else:
        return False


def is_dim_shard(mapping):
    if mapping != -1:
        return True
    else:
        return False


def is_dim_replicate(mapping):
    if mapping == -1:
        return True
    else:
        return False


def compute_compatible_dim_mapping(dim_mappings):
    if not dim_mappings:
        return None
    compatible_mapping = dim_mappings[0]
    for mapping in dim_mappings:
        if compatible_mapping == -1:
            compatible_mapping = mapping
        elif mapping == -1:
            continue
        elif compatible_mapping == mapping:
            continue
        else:
            return None
    return compatible_mapping


def compute_compatible_dims_mapping(dims_mapping_list):
    if not dims_mapping_list:
        return None
    length = len(dims_mapping_list[0])
    for dims_mapping in dims_mapping_list:
        assert dims_mapping is not None, \
            "Dims mapping must not be None for compatible computation"
        assert len(dims_mapping) == length, \
            "The length of dims_mapping in list must be same for compatible computation."
    compatible_result = []
    for dim_mappings in zip(*dims_mapping_list):
        compatible_dim_mapping = compute_compatible_dim_mapping(
            list(dim_mappings))
        if compatible_dim_mapping is None:
            return None
        compatible_result.append(compatible_dim_mapping)
    return compatible_result


def compute_compatible_process_mesh(process_mesh_list):
    compatible_process_mesh = None
    if not process_mesh_list:
        return compatible_process_mesh
    for process_mesh in process_mesh_list:
        if process_mesh is not None:
93
            if compatible_process_mesh is None or compatible_process_mesh == process_mesh:
94 95
                compatible_process_mesh = process_mesh
            else:
96
                return None
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
    return compatible_process_mesh


def compute_compatible_and_update_dim_mapping(dims_mapping_list, index_list):
    assert len(dims_mapping_list) == len(index_list)
    changed = False
    dim_mappings = []
    for i in range(len(dims_mapping_list)):
        assert is_valid_list_index(dims_mapping_list[i], index_list[i])
        dim_mappings.append(dims_mapping_list[i][index_list[i]])
    compatible_dim_mapping = compute_compatible_dim_mapping(dim_mappings)
    if compatible_dim_mapping is None:
        return False
    for i in range(len(dims_mapping_list)):
        if compatible_dim_mapping != dims_mapping_list[i][index_list[i]]:
            dims_mapping_list[i][index_list[i]] = compatible_dim_mapping
            changed = True
    return changed


def append_distributed_attr_suffix(name):
    """
    Append auto parallel suffix for distributed attribute name.
    """
    return name + core.kAutoParallelSuffix()


def remove_distributed_attr_suffix(name):
    """
    Remove auto parallel suffix from distributed attribute name.
    """
    return name.strip(core.kAutoParallelSuffix())


def check_distributed_attr_for_program(program, dist_context=None):
132
    from .dist_context import get_default_distributed_context
133 134 135 136 137 138
    if dist_context is None:
        dist_context = get_default_distributed_context()
    assert dist_context.is_initialized_for_program(), \
        "Distributed attributes must be initialized before check."
    for block in program.blocks:
        for tensor in block.vars.values():
139 140
            dist_tensor = dist_context.get_dist_tensor_for_graph(tensor)
            tensor_dist_attr = dist_context.get_tensor_dist_attr_for_program(
141
                tensor)
142
            if (tensor_dist_attr is not None) and (not dist_tensor.is_valid()):
143 144
                return False
        for op in block.ops:
145 146 147
            dist_op = dist_context.get_dist_op_for_graph(tensor)
            op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
            if (op_dist_attr is not None) and (not dist_op.is_valid()):
148 149 150 151
                return False
    return True


152
def print_program_with_dist_attr(program, dist_context=None):
153 154 155 156 157 158
    """
    This function reuses the original program output ability with a distributed context.
    Using lock can avoid multiple threads change the default distributed context simultaneously.
    """
    lock = threading.Lock()
    lock.acquire()
159 160
    from .dist_context import get_default_distributed_context
    from .dist_context import set_default_distributed_context
161 162 163 164 165 166 167 168 169
    if dist_context is None:
        dist_context = get_default_distributed_context()
        print(program)
    else:
        original_default_context = get_default_distributed_context()
        set_default_distributed_context(dist_context)
        print(program)
        set_default_distributed_context(original_default_context)
    lock.release()
170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185


def _get_comm_group(processes, shape, axis, rank):
    """
    Given a rank and the processes mesh the rank belongs to,  
    compute the communication peers of the rank based on the give axis in the mesh.

    Example: 16 processes managed in a 4-Dimensinal mesh with shape of [2, 2, 2, 2].
    the rank communication peers of rank 0 (included) are following:
    in axis 0: [0, 1]
    in axis 1: [0, 2]
    in axis 2: [0, 4]
    in axis 3: [0, 8]
    """

    # NOTE _linear_idx2coordinate assume processes mesh start with 0 and continuous
186 187 188
    # tricks to support processes mesh when it is not start with 0 or continuous
    assert rank in processes, "rank [{}] is NOT in processes group {}".format(
        rank, processes)
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
    rank_relatvie = processes.index(rank)
    coordinate = _linear_idx2coordinate(shape, rank_relatvie)
    coordinates_in_group = [coordinate[:] for i in range(shape[axis])]

    # select comm group
    for i in range(shape[axis]):
        coordinates_in_group[i][axis] = i

    ranks_in_group_relative = [
        _coordinate2linear_idx(shape, coordinate)
        for coordinate in coordinates_in_group
    ]
    ranks_in_group = [processes[idx] for idx in ranks_in_group_relative]

    return sorted(ranks_in_group)


206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
def _get_idx_in_axis(processes, shape, axis, rank):
    """
    Given a rank and the processes mesh the rank belongs to,  
    compute the index of the rank in given axis.

    Example: 27 processes managed in a 3-Dimensinal mesh with shape of [3, 3, 3].
    the index of rank 22 are:
    in axis 0: 1
    in axis 1: 1
    in axis 2: 2
    """

    # NOTE _linear_idx2coordinate assume processes mesh start with 0 and continuous
    #  tricks to support processes mesh when it is not start with 0 or continuous
    rank_relatvie = processes.index(rank)
    coordinate = _linear_idx2coordinate(shape, rank_relatvie)
    return coordinate[axis]


225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
def _coordinate2linear_idx(mesh_shape, coordinate):
    """
    convert a coordinate in multidimensional mesh space into a scala idx in linear space.

    it use Row-major order for dimension conversion. 
    so it has:  [most_significant_dim, ..., least_significant_dim]
    assume: 

        the size of i-th dimension to be:  S[i]
        the index of j-th dimension is: I[j]

    linear_idx of a n dimensional coordinate is: 

        I[n-1] * (S[n-2] * S[n-3] * S[n-4] *     ....    S[0]) +
        I[n-2] * (         S[n-3] * S[n-4] *     ....    S[0]) +       
        I[n-3] * (                  S[n-4] *     ....    S[0]) +  
        ...
        I[1]   * (                                       S[0]) + 
        I[0]

    """
    # NOTE the following function work based on a strong an assumption
247
    # that the processes in mesh are
248
    #    1. starts from 0
249 250
    #    2. continuous
    # it will be wrong if ths above condition doesnot meet,
251
    # e.g. process_mesh = { process_groups = [7, 8, 9,10, 12, 13, 14, 15], mesh = [2, 4]}
252
    # if you want a more general mapping, you should use cartesian product
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312

    assert len(mesh_shape) == len(
        coordinate
    ), "coordinate should have the same size as mesh shape, but got shape: {}, coordinate: {}".format(
        mesh_shape, coordinate)
    for i in range(len(mesh_shape)):
        assert coordinate[
            i] >= 0, "index in dimension [{}] is least than zero. coordinate: {}".format(
                i, coordinate)
        assert coordinate[i] < mesh_shape[
            i], "index beyond extent in dimension [{}]. shape: {}, coordinate: {}".format(
                i, mesh_shape, coordinate)

    base = mesh_shape[-1]
    linear_idx = coordinate[-1]

    # row major order
    for i in range(len(mesh_shape) - 2, -1, -1):
        linear_idx += base * coordinate[i]
        base *= mesh_shape[i]

    return linear_idx


def _linear_idx2coordinate(mesh_shape, linear_idx):
    """
    mapping a linear scala into multidimensional mesh space, return it coordinate in that space.

    it is the inverse function of _coordinate2linear_idx.
    assume: 

        the size of i-th dimension to be:  S[i]
        the index of j-th dimension is: I[j]

    the coordinate given linear_idx is:

        I[0] = linear_idx                                  % S[0]
        I[0] = (linear_idx / S[0])                         % S[1]
        I[0] = (linear_idx / (S[0] * S[1]))                % S[2]
        ....

    """

    assert linear_idx >= 0, "linear index [{}] is least than zero".format(
        linear_idx)
    assert linear_idx < np.prod(
        mesh_shape
    ), "linear index beyond the extent of mesh shape. shape: {}, linear index: {}".format(
        mesh_shape, linear_idx)

    base = 1
    coordinate = [-1] * len(mesh_shape)

    for i in reversed(range(len(mesh_shape))):
        offset = linear_idx / base
        coordinate[i] = int(offset % mesh_shape[i])
        base *= mesh_shape[i]

    # row major order
    return coordinate
313 314


315
def _get_corresponding_rank(dist_context, target_mesh, rank):
316 317 318 319 320 321

    # TODO(JZ-LIANG) a hack method to support varying mesh in Pipeline parallelism case.
    # we assume that all mesh are evenly divide from a parent mesh and should have same size.
    # to revise this in future.

    coordinate = None
322 323
    for mesh in dist_context.process_meshes:
        if rank in mesh.processes and mesh.topology == target_mesh.topology:
324
            coordinate = _linear_idx2coordinate(mesh.topology,
325
                                                mesh.processes.index(rank))
326 327 328 329
            break

    assert coordinate is not None, "could NOT found rank [{}] in any registered mesh".format(
        rank)
330 331
    return target_mesh.processes[_coordinate2linear_idx(mesh.topology,
                                                        coordinate)]
332 333


334 335
def _get_unshard_dist_shape(var, dist_attr):
    var_shape = var.shape
336 337
    mapping = dist_attr.dims_mapping
    mesh = dist_attr.process_mesh.topology
338 339 340 341 342 343 344 345 346 347 348 349 350 351
    assert len(var_shape) == len(
        mapping
    ), "variable shape [{}] and dim_mapping [{}] is NOT match !".format(
        var_shape, mapping)
    new_shape = []
    for idx in range(len(var_shape)):
        if var_shape[idx] == -1 or mapping[idx] == -1:
            new_shape.append(var_shape[idx])
        else:
            new_shape.append(var_shape[idx] * mesh[mapping[idx]])

    return new_shape


352
def make_data_unshard(dist_main_prog, dist_startup_prog, dist_context=None):
353
    from .dist_context import get_default_distributed_context
354 355
    if dist_context is None:
        dist_context = get_default_distributed_context()
356 357 358

    for var in dist_main_prog.list_vars():
        if var.is_data:
359
            tensor_dist_attr = dist_context.get_tensor_dist_attr_for_program(
360 361 362
                var)
            inverse_shape = _get_unshard_dist_shape(var, tensor_dist_attr)
            var.desc.set_shape(inverse_shape)
363
            dim_mapping = tensor_dist_attr.dims_mapping
364
            dim_mapping = [-1] * len(dim_mapping)
365 366
            tensor_dist_attr.dims_mapping = dim_mapping
            dist_context.set_tensor_dist_attr_for_program(var, tensor_dist_attr)
367 368


369 370 371
def _update_addition_info(addition_info):
    """ Update default addition_info with inputs """
    add_info = {"epoch": 0, "batch": 0, "batch_size": 0}
372
    if not addition_info:
373
        return add_info
374
    elif not isinstance(addition_info, dict):
375 376
        raise TypeError("The type of 'addition_info' should be 'dict', "
                        "but got '{}'.".format(str(type(addition_info))))
377
    else:
378 379 380 381 382 383 384 385 386 387 388 389
        for item, value in addition_info.items():
            if item not in ["epoch", "batch", "batch_size"]:
                raise ValueError(
                    "The key of 'addition_info' should be one of the "
                    "['epoch', 'batch', 'batch_size'], but got '{}'."
                    .format(str(item)))
            if not isinstance(value, int):
                raise ValueError(
                    "The value of 'addition_info' should be 'int', "
                    "but got '{}'.".format(str(type(value))))
            add_info[item] = value
        return add_info
390 391 392


def _check_valid_path(file_path):
393
    """ Validity check of input file path """
394 395 396
    if not file_path:
        return file_path
    elif isinstance(file_path, list):
397 398 399 400 401 402 403
        for file in file_path:
            if not isinstance(file, str):
                raise TypeError("The type of file path should be 'str', "
                                "but got '{}'.".format(str(type(file))))
            if not os.path.exists(file):
                raise ValueError("The file path '{}' does not exist."
                                 .format(file))
404 405
        return file_path
    else:
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
        raise TypeError("The type of file path should be 'list', "
                        "but got '{}'.".format(str(type(file_path))))


def _check_param_dict(param_dict):
    if not param_dict:
        raise ValueError("'param_dict' cannot be None.")
    elif not isinstance(param_dict, dict):
        raise TypeError("The type of 'param_dict' should be 'dict', "
                        "but got '{}'.".format(str(type(param_dict))))
    else:
        for name, value in param_dict.items():
            if not isinstance(name, str):
                raise TypeError(
                    "The type of key of 'param_dict' should be 'str', "
                    "but got '{}'.".format(str(type(name))))
            if not isinstance(value, paddle.fluid.LoDTensor):
                raise TypeError(
                    "The type of value of 'param_dict' should be 'LoDTensor', "
                    "but got '{}'.".format(str(type(value))))
        return param_dict


def _check_dist_attr(dist_attr):
    if not dist_attr:
        return dist_attr
    elif not isinstance(dist_attr, dict):
        raise TypeError("The type of 'dist_attr' should be 'dict', "
                        "but got '{}'.".format(str(type(dist_attr))))
    else:
        for name, value in dist_attr.items():
            if not isinstance(name, str):
                raise TypeError(
                    "The type of param name of 'dist_attr' should be 'str', "
                    "but got '{}'.".format(str(type(name))))
            if not isinstance(value, dict):
                raise TypeError(
                    "The type of distributed attribute should be 'dict', "
                    "but got '{}'".format(str(type(value))))
            attr = ['process_shape', 'process_group', 'dims_mapping']
            if list(value.keys()) != attr:
                raise ValueError(
                    "The key of distributed attribute should be "
                    "'['process_shape', 'process_group', 'dims_mapping']', "
                    "but got {}.".format(str(value.keys())))
        return dist_attr
452 453 454 455


def save_distributed_checkpoint(program,
                                checkpoint_path,
456
                                dist_attr_path,
457
                                addition_info=None,
458 459
                                is_integrated=False,
                                dist_context=None):
460 461 462 463 464 465 466
    """ 
    Save model parameter state, optimzer state, distributed attribute and 
    additional information of each rank.

    Args:
        program(Program): The program to be saved.
        checkpoint_path(str): The path of the checkpoint file to be saved.
467 468 469
        dist_attr_path(str): The path of distributed attribute file to be saved.
        addition_info(dict, optional): Additional information, key should be selected in ['epoch', 'batch', 'batch_size'].
            Default values are 0, when 'addition_info' is None. Default: None.
470
        is_integrated(bool, optional): Whether to integrate param before save. Default: False.
471
        dist_context(DistributedContext ,optional): collect related distributed information for program
472 473 474 475 476 477 478

    Returns:
        None

    Examples:
        .. code-block:: python

479 480 481 482
            path = os.path.join("./output", "step_%d" % step)
            os.makedirs(path, exist_ok=True)
            add_info = {'batch': step, "batch_size": global_batch_size}
            save_distributed_checkpoint(program, path, path, add_info)
483
    """
484 485 486 487 488 489 490 491
    from .dist_context import get_default_distributed_context

    assert isinstance(program, paddle.fluid.framework.Program)
    assert isinstance(is_integrated, bool)
    if dist_context is None:
        dist_context = get_default_distributed_context()
    addition_info = _update_addition_info(addition_info)

492
    if not is_integrated:
493 494
        _save_distributed_state_dict(program, addition_info, checkpoint_path)
        _save_distributed_attribute(program, dist_attr_path, dist_context)
495 496 497 498 499 500
    else:
        # TODO: integrate param before save
        raise NotImplementedError(
            "Integrating parameter has not been implemented.")


501
def load_distributed_checkpoint(checkpoint_path, dist_attr_path):
502
    """ 
503
    Load parameter, optimizer, distributed attribute and addition_info.
504 505

    Args:
506 507
        checkpoint_path(list[str]): model parameter file path, must be in order of rank id.
        dist_attr_path(list[str]): distributed attribute file path, must be in order of rank id.
508 509

    Returns:
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
        param_dict(dict): parameters' value of all ranks.
        dist_attr(dict): parameters' distributed attribute.
        addition_info(dict): additional information user saved in last training. 

    Notes:
        The return, 'addition_info', is belonging to the first file of checkpoint_path by default.

    Examples:
        .. code-block:: python

            ckpt_path = ['./model_state_rank0.pdmodel', 
                         './model_state_rank1.pdmodel']
            dist_attr_path = ['./dist_attr_rank0.pdattr', 
                              './dist_attr_rank1.pdattr']
            param_dict, dist_attr, add_info = load_distributed_checkpoint(ckpt_path, dist_attr_path)
    """
    assert _check_valid_path(checkpoint_path), \
        "'checkpoint_path' cannot be None."
    assert _check_valid_path(dist_attr_path), \
        "'dist_attr_path' cannot be None."

    state_dict_info = _load_distributed_state_dict(checkpoint_path)
    dist_attr = _load_distributed_attribute(dist_attr_path)
    param_dict = state_dict_info["model"]
    addition_info = state_dict_info["addition_info"]
    return param_dict, dist_attr, addition_info


def load_checkpoint_into_program(checkpoint_path,
                                 dist_attr_path,
                                 program,
                                 dist_context=None):
    """ 
    Load parameter, optimizer, distributed attribute and addition_info into model.

    Args:
        checkpoint_path(list[str]): model parameter file path, must be in order of rank id.
        dist_attr_path(list[str]): distributed attribute file path, must be in order of rank id.
        program(Program): the program to be updated with checkpoint_path.
        dist_context(DistributedContext ,optional): collect related distributed information for program

    Returns:
        addition_info(dict): user saved in last train.
    
    Notes:
        The return, 'addition_info', is belonging to the first file of checkpoint_path by default.
556 557 558 559 560

    Examples:
        .. code-block:: python

            exe.run(startup_program)
561 562 563 564 565
            ckpt_path = ['./model_state_rank0.pdmodel', 
                         './model_state_rank1.pdmodel']
            dist_attr_path = ['./dist_attr_rank0.pdattr', 
                              './dist_attr_rank1.pdattr']
            load_checkpoint_into_program(ckpt_path, dist_attr_path, main_program)
566
    """
567
    from .dist_context import get_default_distributed_context
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
    assert isinstance(program, paddle.fluid.framework.Program)
    assert _check_valid_path(checkpoint_path), \
        "'checkpoint_path' cannot be None."
    assert _check_valid_path(dist_attr_path), \
        "'dist_attr_path' cannot be None."
    if dist_context is None:
        dist_context = get_default_distributed_context()
    all_state_dict_info = _load_distributed_state_dict(checkpoint_path)
    all_pre_dist_attr = _load_distributed_attribute(dist_attr_path)
    all_cur_dist_attr = get_dist_attr(program, dist_context)
    all_param_dict = all_state_dict_info["model"]
    addition_info = all_state_dict_info["addition_info"]
    sliced_param_dict = merge_and_slice_parameter(
        all_param_dict, all_pre_dist_attr, all_cur_dist_attr)
    load_parameter_into_program(sliced_param_dict, program)

    return addition_info


def load_parameter_into_program(param_dict, program):
    """ 
    Load parameters into program.

    Args:
        param_dict(dict): parameters' name and value.
        program(Program): the program to be updated
    """
596
    assert isinstance(param_dict, dict)
597
    assert program and isinstance(program, paddle.fluid.framework.Program)
598 599
    if not param_dict:
        return
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
    program.set_state_dict(param_dict)


def _save_distributed_attribute(program, dist_attr_path, dist_context):
    """ Save distributed attribute of all parameters """
    # TODO: just save a complete distributed attribute file
    rank_id = paddle.distributed.get_rank()
    dist_attr_name = os.path.join(dist_attr_path,
                                  "dist_attr_rank{}.pdattr".format(rank_id))
    dist_attr_dict = {
        "model": get_dist_attr(program, dist_context),
        "world_size": paddle.distributed.get_world_size()
    }
    paddle.save(dist_attr_dict, dist_attr_name)
    logging.info("Already saved distributed attribute to '{}'.".format(
        dist_attr_path))


def _load_distributed_attribute(dist_attr_path):
    """ Load parameters' distributed attribute from dist_attr_path """
    total_dist_attr = {}
    for dist_attr_file in dist_attr_path:
        dist_attr = paddle.load(dist_attr_file)
        pre_world_size = dist_attr["world_size"]
        assert pre_world_size == len(dist_attr_path), \
            "The number of 'dist_attr_path' must be equal to the last training world size."
        for name, attr in dist_attr["model"].items():
            if name not in total_dist_attr:
                total_dist_attr[name] = attr

    return total_dist_attr


def _save_distributed_state_dict(program, addition_info, checkpoint_path):
    """ Save parameters' state_dict """
    rank = paddle.distributed.get_rank()
    ckpt_file_name = os.path.join(checkpoint_path,
                                  "model_state_rank{}.pdmodel".format(rank))
    state_dict = {
        "model": program.state_dict(),
        "world_size": paddle.distributed.get_world_size(),
        "addition_info": addition_info
    }
    paddle.save(state_dict, ckpt_file_name)
    logging.info("Already saved model to '{}'.".format(checkpoint_path))


def _load_distributed_state_dict(checkpoint_path):
    """ Load parameters' state_dict from checkpoint_path """
    all_state_dict = {}
    for idx, ckpt_file in enumerate(checkpoint_path):
Z
zhaoyingli 已提交
651
        state_dict_info = paddle.load(ckpt_file, return_numpy=True)
652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723
        pre_world_size = state_dict_info["world_size"]
        assert pre_world_size == len(checkpoint_path), \
            "The number of 'checkpoint_path' must be equal to the last training world size."
        if idx == 0:
            addition_info = state_dict_info["addition_info"]
        for name, value in state_dict_info["model"].items():
            if name in all_state_dict:
                all_state_dict[name].append(np.array(value))
            else:
                all_state_dict[name] = [np.array(value)]

    all_state_dict_info = {
        "model": all_state_dict,
        "addition_info": addition_info
    }
    return all_state_dict_info


def get_dist_attr(program, dist_context=None):
    """ 
    Get distributed attribute of current rank.

    Args:
        program(Program): main program for training
    """
    from .dist_context import get_default_distributed_context

    assert isinstance(program, paddle.fluid.framework.Program)
    if dist_context is None:
        dist_context = get_default_distributed_context()
    dist_attr = {}
    for var in program.list_vars():
        if is_parameter(var) or is_belong_to_optimizer(var):
            tensor_dist_attr = dist_context.get_tensor_dist_attr_for_program(
                var)
            process_mesh = tensor_dist_attr.process_mesh
            dims_mapping = tensor_dist_attr.dims_mapping
            dist_attr[var.name] = {
                "process_shape": process_mesh.topology,
                "process_group": process_mesh.processes,
                "dims_mapping": dims_mapping
            }
    return dist_attr


def merge_and_slice_parameter(dist_param_dict, pre_dist_attr, cur_dist_attr):
    """
    Merge parameters with previous dist_attr and slice parameters with current dist_attr

    Arags:
        dist_param_dict(dict): parameters' value of all ranks.
        pre_dist_attr(dict): parameters' dist_attr of last training process.
        cur_dist_attr(dict): parameters' dist_attr of current training process.

    Returns:
        dist_param_dict(dict): parameters' value of current rank.
    """
    assert _check_dist_attr(pre_dist_attr), "'pre_dist_attr' cannot be None."
    assert isinstance(dist_param_dict, dict), \
        "The type of 'dist_param_dict' should be 'dict', but got {}.".format(
            str(type(dist_param_dict)))
    for name, value in dist_param_dict.items():
        if not isinstance(name, str):
            raise TypeError("The key of 'dist_param_dict' is parameter's name, "
                            "and its type should be 'str', but got {}."
                            .format(str(type(name))))
        if not isinstance(value, list) or not all(
                isinstance(v, np.ndarray) for v in value):
            raise TypeError(
                "The value of 'dist_param_dict' is parameter's value of all ranks, "
                "and its type should be 'list(numpy.ndarray)'.")

724 725 726
    if cur_dist_attr is None:
        return {}

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
    param_not_in_pre = []
    param_not_in_cur = []
    logging.info("Start to merge and slice parameters.")
    for var_name in cur_dist_attr.keys():
        if var_name not in pre_dist_attr:
            param_not_in_pre.append(var_name)
            continue

        pre_attr = pre_dist_attr[var_name]
        cur_attr = cur_dist_attr[var_name]
        if pre_attr == cur_attr:
            # skip merge and slice
            rank_id = paddle.distributed.get_rank()
            index = cur_attr["process_group"].index(rank_id)
            param = dist_param_dict[var_name][index]
            dist_param_dict[var_name] = _to_LodTensor(param)
            continue

        pre_param = dist_param_dict[var_name]
        pre_dims_mapping = pre_attr["dims_mapping"]
        cur_dims_mapping = cur_attr["dims_mapping"]
        if len(set(pre_dims_mapping)) > 1 or -1 not in pre_dims_mapping:
            complete_param = _merge_parameter_with_dist_attr(pre_param,
                                                             pre_attr)
            dist_param_dict[var_name] = complete_param
        else:
            complete_param = pre_param[0]
            dist_param_dict[var_name] = _to_LodTensor(complete_param)

        if len(set(cur_dims_mapping)) > 1 or -1 not in cur_dims_mapping:
            sliced_param = _slice_parameter_with_dist_attr(complete_param,
                                                           cur_attr)
            dist_param_dict[var_name] = sliced_param

    for var_name in pre_dist_attr:
        if var_name not in cur_dist_attr:
            param_not_in_cur.append(var_name)
            dist_param_dict.pop(var_name)

    if param_not_in_pre:
        warnings.warn("Parameters '{}' are not found in last training process."
                      .format(str(param_not_in_pre)))
    if param_not_in_cur:
        warnings.warn(
            "Parameters '{}' are not found in current training process."
            .format(str(param_not_in_cur)))

    return dist_param_dict


def _merge_parameter_with_dist_attr(param_list, dist_attr):
    """ Merge parameter with distributed attribute """
    from .reshard import _compute_complete_shape, _compute_partition_index

    dims_mapping = dist_attr["dims_mapping"]
    process_shape = dist_attr["process_shape"]
    process_group = dist_attr["process_group"]
    # get the complete shape of the parameter
    complete_shape = _compute_complete_shape(param_list[0].shape, process_shape,
                                             dims_mapping)
    # merge the parameter with dist_attr
    partition_param_list = []
Z
zhaoyingli 已提交
789
    merged_partiton = []
790 791 792 793
    for process in process_group:
        partition_index = _compute_partition_index(
            process, complete_shape, dims_mapping, process_shape, process_group)
        index = process_group.index(process)
Z
zhaoyingli 已提交
794 795 796 797 798
        if partition_index not in merged_partiton:
            merged_partiton.append(partition_index)
            _merge_parameter(partition_param_list, param_list[index],
                             partition_index, complete_shape)

799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824
    assert len(partition_param_list) == 1 or not partition_param_list, \
        "Fail to merge parameter"
    complete_param = _to_LodTensor(partition_param_list[0][0])
    return complete_param


def _slice_parameter_with_dist_attr(param, dist_attr):
    """ Slice parameter with distributed attribute """
    param = np.array(param) if isinstance(param,
                                          paddle.fluid.LoDTensor) else param
    dims_mapping = dist_attr["dims_mapping"]
    process_shape = dist_attr["process_shape"]
    process_group = dist_attr["process_group"]
    # slice the parameter with dist_attr
    partition_index_list = _get_split_indices(param.shape, dims_mapping,
                                              process_shape, process_group)
    sliced_param_list = _slice_parameter(param, partition_index_list,
                                         len(partition_index_list))
    # get the current parameter's index in sliced_param_list
    rank_id = paddle.distributed.get_rank()
    sliced_param_index = _get_sliced_param_index(
        rank_id, param.shape, dims_mapping, process_shape, process_group)
    sliced_param = _to_LodTensor(sliced_param_list[sliced_param_index])
    return sliced_param


Z
zhaoyingli 已提交
825 826
def _merge_parameter(partition_param_list, param, partition_index,
                     complete_shape):
827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845
    """
    Merge partitial parameters to a complete one.

    Returns:
        None

    Examples:
        .. code-block:: python

            import numpy as np
            partition_param_list = [(np.array([[[1.11, 1.12]]]), [[0,1],[0,1],[0,2]])]
            param = np.array([[[1.13, 1.14]]])
            partition_index = [[0,1],[0,1],[2,4]]

            _merge_parameter(partition_param_list, param, partition_index)
            # partition_param_list: [(np.array([[[1.11, 1.12, 1.13, 1.14]]]), [[0,1],[0,1],[0,4]])]
    """
    from .reshard import _compute_concat_info

Z
zhaoyingli 已提交
846 847 848 849 850 851 852 853 854
    if len(partition_param_list) == 1:
        is_complete_data = True
        for idx, item in enumerate(partition_param_list[0][1]):
            if item[0] != 0 or item[1] != complete_shape[idx]:
                is_complete_data = False
                break
        if is_complete_data:
            return

855 856
    if not partition_param_list:
        partition_param_list.append((param, partition_index))
857
    else:
858 859 860 861 862 863 864 865 866 867 868 869 870
        i = 0
        while i < len(partition_param_list):
            concat_axis, first_order, new_partition = _compute_concat_info(
                partition_param_list[i][1], partition_index)
            if concat_axis != -1:
                if first_order == 0:
                    new_param = np.concatenate(
                        (partition_param_list[i][0], param), axis=concat_axis)
                else:
                    new_param = np.concatenate(
                        (param, partition_param_list[i][0]), axis=concat_axis)

                partition_param_list.pop(i)
Z
zhaoyingli 已提交
871 872
                _merge_parameter(partition_param_list, new_param, new_partition,
                                 complete_shape)
873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952
                break
            i += 1


def _slice_parameter(complete_param, partition_index_list, length):
    """
    Slice a complete parameter.

    Returns:
        sliced_param_list(list): sliced parameters with 'partition_index_list'

    Examples:
        .. code-block:: python

            import numpy as np
            complete_param = np.array([[[1.11, 1.12, 1.13, 1.14, 1.15, 1.16]]])
            rank = 2
            complete_shape = [1, 1, 6]
            dims_mapping = [-1, -1, 0]
            process_shape = [3]
            process_group = [0, 1, 2]

            sliced_param_list = _slice_parameter(complete_param, [[], [], [2, 4]], 3)
            # [array([[[1.11, 1.12]]]), array([[[1.13, 1.14]]]), array([[[1.15, 1.16]]])]
    """
    sliced_param_list = []
    axis = len(complete_param.shape) - length
    sliced_param = np.split(
        complete_param, partition_index_list[axis], axis=axis)
    if length == 1:
        return sliced_param
    for param in sliced_param:
        sliced_param_list.extend(
            _slice_parameter(param, partition_index_list, length - 1))
    return sliced_param_list


def _get_sliced_param_index(rank, complete_shape, dims_mapping, process_shape,
                            process_group):
    """
    Get sliced_param's index of current rank in all sliced parameters list.

    Returns:
        sliced_param_index(int): the index of sliced param in sliced_param_list

    Examples:
        .. code-block:: python

            import numpy as np
            complete_param = np.array([[[1.11, 1.12, 1.13, 1.14, 1.15, 1.16]]])
            rank = 2
            complete_shape = [1, 1, 6]
            dims_mapping = [-1, -1, 0]
            process_shape = [3]
            process_group = [0, 1, 2]

            slice_param = _slice_parameter(complete_param, [[], [], [2, 4]], 3)
            # slice_param: 
            # [array([[[1.11, 1.12]]]), array([[[1.13, 1.14]]]), array([[[1.15, 1.16]]])]

            index = _get_sliced_param_index(rank, complete_shape, dims_mapping
                                            process_shape, process_group)
            # index: 2
    """
    from .reshard import _compute_partition_index

    partition_index = _compute_partition_index(
        rank, complete_shape, dims_mapping, process_shape, process_group)
    sliced_param_index = 0
    for i, shape in enumerate(complete_shape):
        if dims_mapping[i] == -1:
            slice_shape = shape
        else:
            slice_shape = shape // process_shape[dims_mapping[i]]
        if shape == 1:
            index = 0
        else:
            index = (partition_index[i][0] + 1) // slice_shape
        sliced_param_index = sliced_param_index * (shape // slice_shape) + index
    return sliced_param_index
953 954


955 956 957 958 959 960 961
def _get_split_indices(complete_shape, dims_mapping, process_shape,
                       process_group):
    """
    Get split indices of every dimension.

    Returns:
        split_indices_list(list): the split indices of every dimension of the parameter
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
    Examples:
        .. code-block:: python

            import numpy as np
            complete_param = np.array([[[1.11, 1.12, 1.13, 1.14, 1.15, 1.16]]])
            complete_shape = [1, 1, 6]
            dims_mapping = [-1, -1, 0]
            process_shape = [3]
            process_group = [0, 1, 2]

            index = _get_split_indices(complete_shape, dims_mapping, process_shape, process_group)
            # index: [[], [], [2, 4]]
    """
    from .reshard import _compute_partition_index

    split_indices_list = []
    for process in process_group:
        partition_index = _compute_partition_index(
            process, complete_shape, dims_mapping, process_shape, process_group)
        if split_indices_list:
            for dim in range(len(partition_index)):
                split_indices_list[dim].extend(partition_index[dim])
        else:
            split_indices_list = partition_index
    split_indices_list = list(
        map(lambda x, y: list(set(x) - set([y]) - set([0])), split_indices_list,
            complete_shape))
    split_indices_list = [sorted(x) for x in split_indices_list]
    return split_indices_list
Z
zhaoyingli 已提交
992 993 994 995 996 997 998 999 1000


def set_grad_var_shape(program, dist_context):
    from .operators.common import infer_shape
    from paddle.distributed.fleet.meta_optimizers.common import OpRole

    block = program.global_block()
    vars = block.vars
    for op in block.ops:
J
JZ-LIANG 已提交
1001 1002 1003 1004

        if op.type in ["check_finite_and_unscale", "update_loss_scaling"]:
            break

J
JZ-LIANG 已提交
1005
        if op.type in ["sum", "concat"]:
Z
zhaoyingli 已提交
1006 1007 1008 1009 1010 1011
            continue
        if int(op.attr('op_role')) == int(OpRole.Backward):
            op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
            assert op_dist_attr is not None

            for var_name in op.output_arg_names:
1012 1013
                if "@GRAD" not in var_name:
                    continue
Z
zhaoyingli 已提交
1014
                forward_var_name = var_name[:var_name.find("@GRAD")]
J
JZ-LIANG 已提交
1015 1016 1017
                if op.type in [
                        "c_allreduce_sum", "c_identity", "scale", "cast"
                ]:
Z
zhaoyingli 已提交
1018
                    forward_var_name = op.input_arg_names[0]
1019 1020 1021 1022 1023 1024 1025 1026 1027
                elif op.type == "matmul_v2_grad":
                    forward_var_name = None
                    for output_name in op.output_names:
                        if var_name in op.output(output_name):
                            assert "@GRAD" in output_name
                            input_name = output_name[:output_name.find("@GRAD")]
                            assert len(op.input(input_name)) == 1
                            forward_var_name = op.input(input_name)[0]
                    assert forward_var_name is not None
Z
zhaoyingli 已提交
1028 1029 1030 1031

                need_set_shape_list = [
                    "reshape2_grad", "softmax_with_cross_entropy_grad",
                    "transpose2_grad", "softmax_grad", "cross_entropy_grad2",
1032
                    "dropout_grad"
Z
zhaoyingli 已提交
1033 1034 1035
                ]
                forward_list = [
                    "reshape2", "softmax_with_cross_entropy", "transpose2",
1036
                    "softmax", "cross_entropy2", "dropout"
Z
zhaoyingli 已提交
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050
                ]
                if op.type in need_set_shape_list:
                    for forward_op in block.ops:
                        assert int(forward_op.attr('op_role')) != int(
                            OpRole.Backward)
                        idx = need_set_shape_list.index(op.type)
                        forward_op_name = forward_list[idx]
                        if forward_op.type == forward_op_name and forward_var_name in forward_op.input_arg_names:
                            op_dist_attr = dist_context.get_op_dist_attr_for_program(
                                forward_op)
                            break

                forward_input_dist_attr = op_dist_attr.get_input_dist_attr(
                    forward_var_name)
J
JZ-LIANG 已提交
1051

Z
zhaoyingli 已提交
1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063
                assert forward_input_dist_attr is not None, f"{forward_var_name}"
                forward_var = vars[forward_var_name]
                forward_var_dist_attr = dist_context.get_tensor_dist_attr_for_program(
                    forward_var)
                assert forward_var_dist_attr is not None
                grad_var = vars[var_name]
                ref_shape = infer_shape(block, forward_var,
                                        forward_var_dist_attr,
                                        forward_input_dist_attr)

                if list(grad_var.shape) != ref_shape:
                    grad_var.desc.set_shape(ref_shape)
C
caozhou 已提交
1064 1065


1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082
OP_ROLE_KEY = core.op_proto_and_checker_maker.kOpRoleAttrName()
OpRole = core.op_proto_and_checker_maker.OpRole


def is_forward_op(op):
    ref_role1 = int(core.op_proto_and_checker_maker.OpRole.Forward)
    ref_role2 = int(core.op_proto_and_checker_maker.OpRole.Loss)
    op_role = int(op.attr('op_role'))
    return OP_ROLE_KEY in op.attr_names and (op_role == ref_role1 or
                                             op_role == ref_role2)


def is_backward_op(op):
    return OP_ROLE_KEY in op.attr_names and \
            int(op.all_attrs()[OP_ROLE_KEY]) & int(OpRole.Backward)


J
JZ-LIANG 已提交
1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
def is_loss_op(op):
    return OP_ROLE_KEY in op.attr_names and \
        int(op.all_attrs()[OP_ROLE_KEY]) == (int(core.op_proto_and_checker_maker.OpRole.Forward) | int(core.op_proto_and_checker_maker.OpRole.Loss))


def get_loss_op(block):
    loss_ops = []
    for op in block.ops:
        if is_loss_op(op):
            assert len(op.desc.output_arg_names(
            )) == 1, "loss op should only output loss var"
            loss_ops.append(op)

    assert len(loss_ops) == 1, "num of loss op is not equal to one"
    return loss_ops[0]


def set_var_dist_attr(dist_context, var, dims_mapping, process_mesh, **kwargs):
    tensor_dist_attr = TensorDistributedAttribute()
    tensor_dist_attr.dims_mapping = dims_mapping
    # TODO get global mesh group
    tensor_dist_attr.process_mesh = process_mesh
    dist_context.set_tensor_dist_attr_for_program(var, tensor_dist_attr)
    return tensor_dist_attr


def naive_set_dist_op_attr_for_program_by_mesh_and_mapping(new_op, process_mesh,
                                                           ref_mapping, ctx):
    assert process_mesh is not None
    assert ref_mapping is not None

    new_op_dist_attr = OperatorDistributedAttribute()

    for input_varname in new_op.desc.input_arg_names():
        new_op_dist_attr.set_input_dims_mapping(input_varname, ref_mapping)
    for output_varname in new_op.desc.output_arg_names():
        new_op_dist_attr.set_output_dims_mapping(output_varname, ref_mapping)

    new_op_dist_attr.process_mesh = process_mesh
    ctx.set_op_dist_attr_for_program(new_op, new_op_dist_attr)


C
caozhou 已提交
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258
def update_op_dims_mapping_by_default_dist_impl(dist_op):
    changed = False
    op_dist_attr = dist_op.dist_attr
    op_desc = dist_op.serial_op.desc
    # The following statement will be replaced by a more elegent way
    if op_desc.type() == "shape" or op_desc.type() == "slice":
        return False
    output_names = op_desc.output_names()
    xshape_arg_names = []
    if "XShape" in output_names:
        xshape_arg_names = op_desc.output("XShape")
    batch_dim_mappings = []
    for arg_name in op_desc.input_arg_names():
        serial_tensor = dist_op.get_serial_input(arg_name)
        if serial_tensor.is_parameter:
            continue
        dims_mapping = op_dist_attr.get_input_dims_mapping(arg_name)
        if len(dims_mapping) > 1:
            for idx, mapping in enumerate(dims_mapping[1:]):
                assert mapping == -1, \
                    "{} only the batch dimension (0-dim) can be sharded, but the dimension {} is sharded by {} part."\
                        .format(op_desc.type(), idx, mapping)
        batch_dim_mappings.append(dims_mapping[0])
    for arg_name in op_desc.output_arg_names():
        serial_tensor = dist_op.get_serial_output(arg_name)
        if serial_tensor.is_parameter:
            continue
        dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
        if arg_name not in xshape_arg_names:
            if len(dims_mapping) > 1:
                for idx, mapping in enumerate(dims_mapping[1:]):
                    assert mapping == -1, \
                        "{} only the batch dimension (0-dim) can be sharded, but the dimension {} is sharded by {} part."\
                            .format(op_desc.type(), idx, mapping)
            batch_dim_mappings.append(dims_mapping[0])
        else:
            assert dims_mapping[0] == -1, \
                "{} only the batch dimension (1-dim) of XShape can be sharded, but the dimension 0 is sharded by {} part."\
                    .format(op_desc.type(), mapping)
            if len(dims_mapping) > 2:
                for idx, mapping in enumerate(dims_mapping[2:]):
                    assert mapping == -1, \
                        "{} only the batch dimension (1-dim) of XShape can be sharded, but the dimension {} is sharded by {} part."\
                            .format(op_desc.type(), idx, mapping)
            batch_dim_mappings.append(dims_mapping[1])

    compatible_dim_mapping = compute_compatible_dim_mapping(batch_dim_mappings)
    assert compatible_dim_mapping is not None, "There is no compatible dim mapping."
    for arg_name in op_desc.input_arg_names():
        serial_tensor = dist_op.get_serial_input(arg_name)
        if serial_tensor.is_parameter:
            continue
        dims_mapping = op_dist_attr.get_input_dims_mapping(arg_name)
        if compatible_dim_mapping != dims_mapping[0]:
            dims_mapping[0] = compatible_dim_mapping
            changed = True
    for arg_name in op_desc.output_arg_names():
        serial_tensor = dist_op.get_serial_output(arg_name)
        if serial_tensor.is_parameter:
            continue
        dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
        if arg_name not in xshape_arg_names:
            if compatible_dim_mapping != dims_mapping[0]:
                dims_mapping[0] = compatible_dim_mapping
                changed = True
        else:
            if compatible_dim_mapping != dims_mapping[1]:
                dims_mapping[1] = compatible_dim_mapping
                changed = True

    return changed


def update_op_dims_mapping_by_elementwise_like_dist_impl(dist_op):
    changed = False
    op_dist_attr = dist_op.dist_attr
    op_desc = dist_op.serial_op.desc
    input_arg_names = op_desc.input_arg_names()
    input_dims_mapping_dict = {}
    input_dims_mapping_lens = {}
    max_dims_mapping_len = -1
    for arg_name in input_arg_names:
        dims_mapping = op_dist_attr.get_input_dims_mapping(arg_name)
        if max_dims_mapping_len < len(dims_mapping):
            max_dims_mapping_len = len(dims_mapping)
        input_dims_mapping_dict[arg_name] = dims_mapping
        input_dims_mapping_lens[arg_name] = len(dims_mapping)

    dims_mapping_list = []
    for arg_name in input_arg_names:
        if input_dims_mapping_lens[arg_name] < max_dims_mapping_len:
            new_dims_mapping = [-1 for _ in range(max_dims_mapping_len)]
            for i in range(input_dims_mapping_lens[arg_name]):
                new_idx = (max_dims_mapping_len -
                           input_dims_mapping_lens[arg_name]) + i
                new_dims_mapping[new_idx] = input_dims_mapping_dict[arg_name][i]
            dims_mapping_list.append(new_dims_mapping)
        else:
            dims_mapping_list.append(input_dims_mapping_dict[arg_name])
    output_arg_names = op_desc.output_arg_names()
    for arg_name in output_arg_names:
        dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
        assert len(dims_mapping) == max_dims_mapping_len
        dims_mapping_list.append(dims_mapping)

    compatible_dims_mapping = compute_compatible_dims_mapping(dims_mapping_list)
    assert compatible_dims_mapping is not None, "There is no compatible dim mapping."

    for arg_name in input_arg_names:
        if input_dims_mapping_lens[arg_name] < max_dims_mapping_len:
            new_dims_mapping = [
                -1 for _ in range(input_dims_mapping_lens[arg_name])
            ]
            for i in range(input_dims_mapping_lens[arg_name]):
                new_idx = (max_dims_mapping_len -
                           input_dims_mapping_lens[arg_name]) + i
                new_dims_mapping[i] = compatible_dims_mapping[new_idx]
            if new_dims_mapping != input_dims_mapping_dict[arg_name]:
                op_dist_attr.set_input_dims_mapping(arg_name, new_dims_mapping)
                changed = True
        else:
            if compatible_dims_mapping != input_dims_mapping_dict[arg_name]:
                op_dist_attr.set_input_dims_mapping(arg_name,
                                                    compatible_dims_mapping)
                changed = True

    for arg_name in output_arg_names:
        dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
        if compatible_dims_mapping != dims_mapping:
            op_dist_attr.set_output_dims_mapping(arg_name,
                                                 compatible_dims_mapping)
            changed = True

    return changed
1259 1260


1261 1262
def get_all_distributed_main_program(serial_program_info, dist_context,
                                     parallelizer):
1263
    "Get all distributed main programs by dist_context."
1264
    from .dist_context import DistributedOperatorContext, DistributedContext
1265
    cluster = serial_program_info.cluster
1266
    copied_parallelizer = copy.deepcopy(parallelizer)
1267 1268 1269 1270 1271 1272
    all_dist_main_program = []
    ranks = paddle.distributed.get_world_size() if cluster is None else len(
        cluster.get_all_devices("GPU"))
    for rank_id in range(ranks):
        used_dist_context = copy.deepcopy(dist_context)
        used_dist_context._dist_op_context = DistributedOperatorContext()
1273 1274
        _, _, dist_startup_program, dist_main_program, _ = copied_parallelizer._get_dist_program(
            rank_id, used_dist_context)
1275
        # print("dist_main_program: ", dist_main_program)
1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312
        all_dist_main_program.append(dist_main_program)

    return all_dist_main_program


class SerialProgramInfo:
    def __init__(self,
                 train_program,
                 satrtup_program,
                 loss,
                 optimizer,
                 cluster=None):
        self._train_program = train_program
        self._startup_program = satrtup_program
        self._loss = loss
        self._optimizer = optimizer
        self._cluster = cluster

    @property
    def train_program(self):
        return self._train_program

    @property
    def startup_program(self):
        return self._startup_program

    @property
    def loss(self):
        return self._loss

    @property
    def optimizer(self):
        return self._optimizer

    @property
    def cluster(self):
        return self._cluster
1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352


def get_standalone_cost_data(distributed_programs):
    def _compute_runtime(op_cost, op, vars):
        runtime = 0
        try:
            runtime = float(op_cost["op_time"])
        except:
            return runtime
        op_config = op_cost["config"]
        total_static_input_size = 0
        total_actual_input_size = 0
        parsed_info = op_config.split("\n")
        variable = "(Variable)"
        for info in parsed_info:
            variable = "(Variable)" if "(Variable)" in info else "(list<Variable>"
            if variable in info:
                arg_name_lower = info[:info.find(variable) - 1]
                shape_left_boundary = info.find("[")
                shape_right_boundary = info.find("]")
                assert shape_left_boundary > 0 and shape_right_boundary > 0 and shape_right_boundary > shape_left_boundary, "Get shape failed."
                shape = info[shape_left_boundary + 1:
                             shape_right_boundary].split(",")
                shape = list(map(lambda x: int(x.strip()), shape))
                dtype_factor = 1
                total_static_input_size += reduce(lambda x, y: x * y, shape)
                if op.type == "c_embedding":
                    arg_name_lower = "w" if arg_name_lower == "weight" else "ids"
                for arg_name in op.input_names:
                    if arg_name.lower() == arg_name_lower:
                        for var_name in op.input(arg_name):
                            var = vars[var_name]
                            total_actual_input_size += reduce(
                                lambda x, y: x * y, var.shape)
                        break
        assert total_static_input_size > 0 and total_actual_input_size > 0, "Get input size failed."

        actual_runtime = total_actual_input_size / total_static_input_size * runtime
        return actual_runtime

1353 1354
    import paddle.cost_model as cm
    cost_model = cm.CostModel()
1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404
    cost_model.static_cost_data()
    DEFAULT_MULTIPLE = 2
    OP_NAME_MAPPING = {
        "c_embedding": "embedding",
        "matmul_v2": "matmul",
        "transpose2": "transpose",
        "reshape2": "reshape",
        "unsqueeze2": "unsqueeze",
        "reduce_sum": "sum",
        "elementwise_div": "divide"
    }

    standalone_cost_data = []
    not_enum_ops = ["create_py_reader", "create_double_buffer_reader", "read"]
    for distributed_program in distributed_programs:
        cost_data = {}
        vars = distributed_program.global_block().vars
        for op in distributed_program.global_block().ops:
            runtime = 0
            if op.type in not_enum_ops:
                cost_data[op.desc.id()] = runtime
                continue
            dtype = str(vars[op.input_arg_names[0]]
                        .dtype) if op.input_arg_names else "float32"
            if int(op.attr('op_role')) == int(OpRole.Backward):
                if "_grad" in op.type:
                    forward_op_name = op.type[:-5]
                    if forward_op_name in OP_NAME_MAPPING.keys():
                        forward_op_name = OP_NAME_MAPPING[forward_op_name]
                    op_cost = cost_model.get_static_op_time(
                        forward_op_name, forward=False, dtype=dtype)
                    if op_cost:
                        runtime = _compute_runtime(op_cost, op, vars)
                    else:
                        op_cost = cost_model.get_static_op_time(
                            forward_op_name, dtype=dtype)
                        if op_cost:
                            runtime = 2 * _compute_runtime(op_cost, op, vars)
            elif int(op.attr('op_role')) == int(OpRole.Forward):
                op_name = OP_NAME_MAPPING[
                    op.type] if op.type in OP_NAME_MAPPING.keys() else op.type
                op_cost = cost_model.get_static_op_time(op_name)
                if op_cost:
                    runtime = _compute_runtime(op_cost, op, vars)

            cost_data[op.desc.id()] = runtime

        standalone_cost_data.append(cost_data)

    return standalone_cost_data
1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420


def set_dist_op_desc_original_id(dist_op_desc, op_desc, dist_context):
    op_id = op_desc.id()
    op_original_id = op_desc.original_id()
    # First, try to set the original id to the id of the op_desc
    if op_id in dist_context._dist_ops_for_program:
        dist_op_desc.set_original_id(op_id)
        return
    # Second, try to set the original id to the original_id of the op_desc
    elif op_original_id in dist_context._dist_ops_for_program:
        dist_op_desc.set_original_id(op_original_id)
        return
    # Third, print error infomation if we cannot find the original id
    else:
        assert False, "Cannot find the original id in the distributed context"