dist_matmul.py 85.5 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 copy
Z
zhaoyingli 已提交
16
from .common import infer_shape
17
from .common import DistributedOperatorImplContainer
18
from .common import DistributedOperatorImpl
19
from .common import register_distributed_operator_impl_container
20
from .common import register_distributed_operator_impl
J
JZ-LIANG 已提交
21
from .common import set_comm_op_dist_attr_for_program, naive_copy_op_dist_attr_for_program, is_parameter_related
22 23 24 25 26 27
from ..utils import is_dim_shard
from ..utils import is_dim_replicate
from ..utils import is_valid_list_index
from ..utils import compute_compatible_dim_mapping
from ..utils import compute_compatible_dims_mapping
from ..utils import compute_compatible_and_update_dim_mapping
28
from ..utils import set_dist_op_desc_original_id
29
from ..dist_attribute import OperatorDistributedAttribute
30
from paddle.fluid import core, unique_name
J
Jiabin Yang 已提交
31
from paddle.fluid.framework import _non_static_mode
32 33
from paddle.fluid.framework import Program, Parameter, Variable, program_guard
from paddle.fluid.data_feeder import check_variable_and_dtype, check_dtype
34
from paddle.distributed.fleet.meta_optimizers.common import OpRole, OP_ROLE_KEY, OP_ROLE_VAR_KEY
35
from ..process_group import new_process_group
36
from ..utils import _get_comm_group, _get_corresponding_rank
37
from .dist_default import DistributedDefaultImpl0
38 39


40
def copy_op_with_new_input_output(ctx, block, src_op, **kwargs):
41 42
    dist_op_desc = block.desc.append_op()
    dist_op_desc.copy_from(src_op.desc)
43
    set_dist_op_desc_original_id(dist_op_desc, src_op.desc, ctx)
44 45 46 47 48 49 50 51 52 53 54
    for input_name in src_op.desc.input_names():
        assert input_name in kwargs
        dist_op_desc.set_input(input_name, kwargs[input_name])
    for output_name in src_op.desc.output_names():
        assert input_name in kwargs
        dist_op_desc.set_output(output_name, kwargs[output_name])

    block._sync_with_cpp()
    return dist_op_desc


55
def _update_dims_mapping_for_matmul(dist_op):
56
    changed = False
57 58
    op_desc = dist_op.serial_op.desc
    op_dist_attr = dist_op.dist_attr
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
    x_name = op_desc.input('X')[0]
    y_name = op_desc.input('Y')[0]
    out_name = op_desc.output('Out')[0]
    x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
    y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
    out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
    x_dims_mapping_len = len(x_dims_mapping)
    y_dims_mapping_len = len(y_dims_mapping)
    out_dims_mapping_len = len(out_dims_mapping)

    # Add dim mapping to Make sure the length dims_mapping be at least 2
    if x_dims_mapping_len == 1:
        x_dims_mapping.insert(0, -1)
    if y_dims_mapping_len == 1:
        y_dims_mapping.insert(1, -1)

75
    # Deal with dim > 2 and take care of broadcasting
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97
    if out_dims_mapping_len > 2:
        broadcast_x_dims_mapping = []
        broadcast_y_dims_mapping = []
        broadcast_out_dims_mapping = []

        for i in range(out_dims_mapping_len - x_dims_mapping_len):
            broadcast_x_dims_mapping.append(out_dims_mapping[i])
        for i in range(x_dims_mapping_len - 2):
            broadcast_x_dims_mapping.append(x_dims_mapping[i])

        for i in range(out_dims_mapping_len - y_dims_mapping_len):
            broadcast_y_dims_mapping.append(out_dims_mapping[i])
        for i in range(y_dims_mapping_len - 2):
            broadcast_y_dims_mapping.append(y_dims_mapping[i])

        for i in range(out_dims_mapping_len - 2):
            broadcast_out_dims_mapping.append(out_dims_mapping[i])

        compatible_dims_mapping = compute_compatible_dims_mapping([
            broadcast_x_dims_mapping, broadcast_y_dims_mapping,
            broadcast_out_dims_mapping
        ])
98 99
        if compatible_dims_mapping is None:
            return False
100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117

        for i in range(x_dims_mapping_len - 2):
            new_idx = i + (out_dims_mapping_len - x_dims_mapping_len)
            if x_dims_mapping[i] != compatible_dims_mapping[new_idx]:
                x_dims_mapping[i] = compatible_dims_mapping[new_idx]
                changed = True

        for i in range(y_dims_mapping_len - 2):
            new_idx = i + (out_dims_mapping_len - y_dims_mapping_len)
            if y_dims_mapping[i] != compatible_dims_mapping[new_idx]:
                y_dims_mapping[i] = compatible_dims_mapping[new_idx]
                changed = True

        for i in range(out_dims_mapping_len - 2):
            if out_dims_mapping[i] != compatible_dims_mapping[i]:
                out_dims_mapping[i] = compatible_dims_mapping[i]
                changed = True

118
    # The following which uses negative index can be work
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134
    # when len(out_dims_mapping) > 2 and len(out_dims_mapping) <=2
    dim_changed = compute_compatible_and_update_dim_mapping(
        [x_dims_mapping, y_dims_mapping], [-1, -2])
    if dim_changed:
        changed = True

    dim_changed = compute_compatible_and_update_dim_mapping(
        [x_dims_mapping, out_dims_mapping], [-2, -2])
    if dim_changed:
        changed = True

    dim_changed = compute_compatible_and_update_dim_mapping(
        [y_dims_mapping, out_dims_mapping], [-1, -1])
    if dim_changed:
        changed = True

135
    # Remove unnecessary dim mapping to make sure the length of dims_mapping is same as its tensor
136 137 138 139 140 141 142 143 144 145 146 147
    if x_dims_mapping_len == 1:
        x_dims_mapping.pop(0)
    if y_dims_mapping_len == 1:
        y_dims_mapping.pop(1)

    assert len(x_dims_mapping) == x_dims_mapping_len
    assert len(y_dims_mapping) == y_dims_mapping_len
    assert len(out_dims_mapping) == out_dims_mapping_len

    return changed


148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168
def _is_auto_compatible_for_matmul(dist_op):
    op_desc = dist_op.serial_op.desc
    op_dist_attr = dist_op.dist_attr
    x_name = op_desc.input('X')[0]
    y_name = op_desc.input('Y')[0]
    out_name = op_desc.output('Out')[0]
    # Deep copy these dims_mappings for keeping them unchanged.
    x_dims_mapping = copy.deepcopy(op_dist_attr.get_input_dims_mapping(x_name))
    y_dims_mapping = copy.deepcopy(op_dist_attr.get_input_dims_mapping(y_name))
    out_dims_mapping = copy.deepcopy(
        op_dist_attr.get_output_dims_mapping(out_name))
    x_dims_mapping_len = len(x_dims_mapping)
    y_dims_mapping_len = len(y_dims_mapping)
    out_dims_mapping_len = len(out_dims_mapping)

    # Add dim mapping to Make sure the length dims_mapping be at least 2
    if x_dims_mapping_len == 1:
        x_dims_mapping.insert(0, -1)
    if y_dims_mapping_len == 1:
        y_dims_mapping.insert(1, -1)

169 170 171
    # NOTE: Partition is not supported if matmul op has trans.
    if op_desc.type() == "matmul_v2":
        if op_desc.attr('trans_x') or op_desc.attr('trans_y'):
172 173
            if x_dims_mapping[-2:] != [-1, -1
                                       ] or y_dims_mapping[-2:] != [-1, -1]:
174 175 176
                return False
    elif op_desc.type() == "matmul":
        if op_desc.attr('transpose_X') or op_desc.attr('transpose_Y'):
177 178
            if x_dims_mapping[-2:] != [-1, -1
                                       ] or y_dims_mapping[-2:] != [-1, -1]:
179 180
                return False

181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199
    # Deal with dim > 2 and take care of broadcasting
    if out_dims_mapping_len > 2:
        broadcast_x_dims_mapping = []
        broadcast_y_dims_mapping = []
        broadcast_out_dims_mapping = []

        for i in range(out_dims_mapping_len - x_dims_mapping_len):
            broadcast_x_dims_mapping.append(out_dims_mapping[i])
        for i in range(x_dims_mapping_len - 2):
            broadcast_x_dims_mapping.append(x_dims_mapping[i])

        for i in range(out_dims_mapping_len - y_dims_mapping_len):
            broadcast_y_dims_mapping.append(out_dims_mapping[i])
        for i in range(y_dims_mapping_len - 2):
            broadcast_y_dims_mapping.append(y_dims_mapping[i])

        for i in range(out_dims_mapping_len - 2):
            broadcast_out_dims_mapping.append(out_dims_mapping[i])

200 201
        is_same = ((broadcast_x_dims_mapping == broadcast_y_dims_mapping)
                   and (broadcast_x_dims_mapping == broadcast_out_dims_mapping))
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
        if not is_same:
            return False

    # The following which uses negative index can be work
    # when len(out_dims_mapping) > 2 and len(out_dims_mapping) <=2
    is_same = (x_dims_mapping[-1] == y_dims_mapping[-2])
    if not is_same:
        return False

    is_same = (x_dims_mapping[-2] == out_dims_mapping[-2])
    if not is_same:
        return False

    is_same = (y_dims_mapping[-1] == out_dims_mapping[-1])
    if not is_same:
        return False

    return True


222 223 224 225
def _right_operand_parameter_matmul_backward(ctx, *args, **kwargs):

    # by now the backward function only insert the gradient allreduce for dist op itself

226
    dist_op_context = ctx.dist_op_context
227 228 229
    main_block = dist_op_context.work_block
    backward_op = dist_op_context.cur_src_op
    rank_id = dist_op_context.rank_id
230
    dist_attr = ctx.get_op_dist_attr_for_program(backward_op)
231 232 233 234
    assert dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
        str(backward_op))

    # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
235 236
    if rank_id not in dist_attr.process_mesh.processes:
        rank_id = _get_corresponding_rank(ctx, dist_attr.process_mesh, rank_id)
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260

    assert 'Y' in kwargs, "input [{}] is not given".format('Y')
    assert 'X' in kwargs, "input [{}] is not given".format('X')
    assert 'Out@GRAD' in kwargs, "input [{}] is not given".format('Out@GRAD')
    assert 'Y@GRAD' in kwargs, "output [{}] is not given".format('Y@GRAD')
    assert 'X@GRAD' in kwargs, "output [{}] is not given".format('X@GRAD')
    assert len(
        kwargs['Y']
    ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format(
        kwargs['Y'])
    assert len(
        kwargs['X']
    ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format(
        kwargs['X'])
    assert len(
        kwargs['Out@GRAD']
    ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format(
        kwargs['Out'])
    assert len(
        kwargs['Y@GRAD']
    ) == 1, "row_parallel_embedding output Ids take 1 variable but got {}".format(
        kwargs['Y@GRAD'])

    X_var = main_block.var(kwargs['X'][0])
261
    Y_var = main_block._var_recursive(kwargs['Y'][0])
262 263 264
    Out_grad = main_block.var(kwargs['Out@GRAD'][0])
    Y_grad = main_block.var(kwargs['Y@GRAD'][0])

J
JZ-LIANG 已提交
265 266 267
    assert not is_parameter_related(
        X_var.name, main_block
    ), "left operand(X) [{}] of dist matmul should not be parameter".format(
268 269
        X_var.name)

270 271 272
    Y_var_dim_mapping = dist_attr.get_input_dims_mapping(Y_var.name)
    process_mesh_shape = dist_attr.process_mesh.topology
    process_mesh_group = dist_attr.process_mesh.processes
273 274 275 276
    # assert len(
    #     Y_var_dim_mapping
    # ) == 2, "dist matmual only support Y operand with 2 dims now but Y({})'s dim is [{}]".format(
    #     Y_var.name, Y_var_dim_mapping)
277 278 279 280 281 282
    Y_var_partitioned = False
    for dim in Y_var_dim_mapping:
        if dim >= 0 and process_mesh_shape[dim] > 0:
            Y_var_partitioned = True
            break

J
JZ-LIANG 已提交
283
    if is_parameter_related(Y_var.name, main_block) and Y_var_partitioned:
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

        if Y_var_dim_mapping[0] >= 0:
            # row parallel: c_identity + matmul
            assert Y_var_dim_mapping[1] < 0
            parallel_axis = Y_var_dim_mapping[0]

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

            intermediate_var_0 = main_block.create_var(
                name=unique_name.generate_with_ignorable_key(".".join(
                    ["c_identity", 'tmp'])) + "@GRAD",
                dtype=Out_grad.dtype,
                shape=Out_grad.shape,
                type=core.VarDesc.VarType.LOD_TENSOR,
                persistable=False,
                stop_gradient=Out_grad.stop_gradient)

            # copy X_var's dist_attr to intermediate_var_0's dist_attr
            out_grad_dist_attr = dist_attr.get_input_dist_attr(Out_grad.name)
            assert out_grad_dist_attr is not None
            ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                                 out_grad_dist_attr)

310 311 312
            group_ranks = _get_comm_group(process_mesh_group,
                                          process_mesh_shape, parallel_axis,
                                          rank_id)
313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328
            group = new_process_group(group_ranks)
            c_identity_op = main_block.append_op(
                type='c_identity',
                inputs={'X': [Out_grad]},
                outputs={'Out': intermediate_var_0},
                attrs={
                    'ring_id': group.id,
                    'use_calc_stream': True,
                    'use_model_parallel': True,
                    OP_ROLE_KEY: OpRole.Backward,
                })
            check_variable_and_dtype(intermediate_var_0, 'x',
                                     ['float16', 'float32', 'float64'],
                                     'linear')
            check_dtype(intermediate_var_0.dtype, 'dtype',
                        ['float16', 'float32', 'float64'], 'linear')
329 330 331
            set_comm_op_dist_attr_for_program(c_identity_op,
                                              dist_attr.process_mesh,
                                              out_grad_dist_attr, ctx)
332 333 334 335

            new_kwargs = copy.deepcopy(kwargs)
            new_kwargs['Out@GRAD'] = [intermediate_var_0.name]
            matmul_op_desc = copy_op_with_new_input_output(
336
                ctx, main_block, backward_op, **new_kwargs)
337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
        else:
            # col parallel: matmul + allreduce
            assert Y_var_dim_mapping[0] < 0
            parallel_axis = Y_var_dim_mapping[1]
            new_kwargs = copy.deepcopy(kwargs)

            # NOTE (JZ-LIANG) should allow left operand be empty for matmul grad
            has_x_grad = len(kwargs['X@GRAD']) > 0
            if has_x_grad:
                assert len(kwargs['X@GRAD']) == 1
                X_grad = main_block.var(kwargs['X@GRAD'][0])
                intermediate_var_0 = main_block.create_var(
                    name=unique_name.generate_with_ignorable_key(".".join(
                        ["c_identity", 'tmp'])) + "@GRAD",
                    dtype=X_grad.dtype,
                    shape=X_grad.shape,
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    persistable=False,
                    stop_gradient=X_grad.stop_gradient)

                X_grad_dist_attr = dist_attr.get_output_dist_attr(X_grad.name)
                assert X_grad_dist_attr is not None
                ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                                     X_grad_dist_attr)
                new_kwargs['X@GRAD'] = [intermediate_var_0.name]

            matmul_op_desc = copy_op_with_new_input_output(
364
                ctx, main_block, backward_op, **new_kwargs)
365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386

            # NOTE (JZ-LIANG) trick to skip one allreduce if left operand has not grad
            if has_x_grad:
                group_ranks = _get_comm_group(process_mesh_group,
                                              process_mesh_shape, parallel_axis,
                                              rank_id)
                group = new_process_group(group_ranks)
                c_allreduce_sum_op = main_block.append_op(
                    type='c_allreduce_sum',
                    inputs={'X': [intermediate_var_0.name]},
                    outputs={'Out': kwargs['X@GRAD']},
                    attrs={
                        'ring_id': group.id,
                        'use_calc_stream': True,
                        'use_model_parallel': True,
                        OP_ROLE_KEY: OpRole.Backward
                    })
                set_comm_op_dist_attr_for_program(c_allreduce_sum_op,
                                                  dist_attr.process_mesh,
                                                  X_grad_dist_attr, ctx)
    else:
        # replicate
387 388
        matmul_op_desc = copy_op_with_new_input_output(ctx, main_block,
                                                       backward_op, **kwargs)
389 390 391 392 393 394

    main_block._sync_with_cpp()

    # check if need gradient allreduce
    need_gradient_allreduce = False

395
    process_mesh = dist_attr.process_mesh
396 397 398 399 400
    var_dim_mapping = dist_attr.get_input_dims_mapping(X_var.name)
    mesh_shape = process_mesh.topology
    batch_size_axis = var_dim_mapping[0]
    if batch_size_axis > -1 and mesh_shape[batch_size_axis] > 1:
        need_gradient_allreduce = True
401
        group_ranks = _get_comm_group(process_mesh.processes,
402 403 404 405 406
                                      process_mesh.topology, batch_size_axis,
                                      rank_id)
        dp_degree = len(group_ranks)
        dp_group = new_process_group(group_ranks)

J
JZ-LIANG 已提交
407
    if need_gradient_allreduce and is_parameter_related(Y_var.name, main_block):
408
        added_ops = []
409
        Y_Grad_var = main_block.var(kwargs['Y@GRAD'][0])
410 411 412 413 414 415 416 417
        allreduce_op = main_block.append_op(type='c_allreduce_sum',
                                            inputs={'X': [Y_Grad_var]},
                                            outputs={'Out': [Y_Grad_var]},
                                            attrs={
                                                'ring_id': dp_group.id,
                                                'use_calc_stream': True,
                                                OP_ROLE_KEY: OpRole.Backward
                                            })
418 419 420 421 422 423 424 425 426 427 428 429
        added_ops.append(allreduce_op)

        if ctx.gradient_scale:
            scale_op = main_block.append_op(type='scale',
                                            inputs={'X': Y_Grad_var},
                                            outputs={'Out': Y_Grad_var},
                                            attrs={
                                                'scale': 1.0 / dp_degree,
                                                OP_ROLE_KEY: OpRole.Backward
                                            })
            added_ops.append(scale_op)

430 431
        main_block._sync_with_cpp()

432 433 434
        dims_mapping = ctx.get_tensor_dist_attr_for_program(
            Y_Grad_var).dims_mapping
        process_mesh = dist_attr.process_mesh
435
        for op in added_ops:
436 437
            op_attr = OperatorDistributedAttribute()
            op_attr.process_mesh = process_mesh
438 439
            op_attr.set_output_dims_mapping(Y_Grad_var.name, dims_mapping)
            op_attr.set_input_dims_mapping(Y_Grad_var.name, dims_mapping)
440
            ctx.set_op_dist_attr_for_program(op, op_attr)
441 442


443
def _init_param_sync(Weight_var, dist_op_context, startup_block, ctx, rank_id):
444

445 446
    if Weight_var.name in dist_op_context.already_init_sync_vars:
        return
447
    assert startup_block.has_var(Weight_var.name)
448
    dist_op_context.already_init_sync_vars.add(Weight_var.name)
449
    param = startup_block.var(Weight_var.name)
450 451 452
    param_dist_attr = ctx.get_tensor_dist_attr_for_program(param)
    process_mesh = param_dist_attr.process_mesh
    dim_mapping = param_dist_attr.dims_mapping
453 454 455 456 457

    for axis, size in enumerate(process_mesh.topology):
        if size <= 1 or axis in dim_mapping:
            pass
        else:
458
            group_ranks = _get_comm_group(process_mesh.processes,
459 460 461
                                          process_mesh.topology, axis, rank_id)
            sync_group = new_process_group(group_ranks)

462 463 464 465 466 467 468 469 470
            startup_block.append_op(type='c_broadcast',
                                    inputs={'X': param},
                                    outputs={'Out': param},
                                    attrs={
                                        'ring_id': sync_group.id,
                                        'root': 0,
                                        'use_calc_stream': True,
                                        OP_ROLE_KEY: OpRole.Forward
                                    })
471 472 473
    startup_block._sync_with_cpp()


474
class DistributedMatmul(DistributedOperatorImplContainer):
475

476 477
    def __init__(self, op_type):
        super(DistributedMatmul, self).__init__(op_type)
478 479


480
register_distributed_operator_impl_container(DistributedMatmul("matmul"))
481 482 483 484


# ColumnParallel
class DistributedMatmulImpl0(DistributedOperatorImpl):
485

486
    def __init__(self, name):
487
        super(DistributedMatmulImpl0, self).__init__(name)
488
        self._forward_implemented = True
489
        self._backward_implemented = True
490

491 492 493
    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
494 495 496 497 498 499
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
        if is_dim_shard(x_dims_mapping[-1]):
            return False
500 501
        if is_dim_shard(y_dims_mapping[-2]) or is_dim_replicate(
                y_dims_mapping[-1]):
502 503 504 505 506 507
            return False
        for mapping in x_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

508 509 510
    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
511 512 513 514 515 516 517 518 519
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
        if is_dim_replicate(out_dims_mapping[-1]):
            return False
        for mapping in out_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

520
    def is_auto_compatible(self, dist_op):
521 522
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
523
            return False
524
        if not _is_auto_compatible_for_matmul(dist_op):
525 526 527
            return False
        return True

528
    def update_dims_mapping(self, dist_op):
529
        changed = False
530
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
531 532 533 534
        if dim_changed:
            changed = True
        return changed

535 536 537 538 539 540
    @staticmethod
    def forward(ctx, *args, **kwargs):
        """
        kwargs: inputname_mapping & outputname_mapping
        """

541
        dist_op_context = ctx.dist_op_context
542 543 544 545
        main_block = dist_op_context.work_block
        startup_block = dist_op_context.startup_block
        src_op = dist_op_context.cur_src_op
        rank_id = dist_op_context.rank_id
546
        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
547 548 549 550
        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(src_op))

        # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
551 552
        if rank_id not in op_dist_attr.process_mesh.processes:
            rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh,
553 554
                                              rank_id)

555
        # check validation of inputs / outputs
556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
        for input_name in src_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                src_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in src_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                src_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        X_var = main_block.var(kwargs['X'][0])
        Weight_var = main_block.var(kwargs['Y'][0])
        Out_var = main_block.var(kwargs['Out'][0])

        # TODO infer logic comm presentation
        matmul_col_dim_mapping = op_dist_attr.get_input_dims_mapping(
576
            Weight_var.name)[-1]
577 578
        assert matmul_col_dim_mapping >= 0, "col_parallel_matmul's row should be divided by a specific mesh axis, but got [{}]".format(
            matmul_col_dim_mapping)
579 580
        process_mesh_shape = op_dist_attr.process_mesh.topology
        process_mesh_group = op_dist_attr.process_mesh.processes
581 582 583 584 585 586

        parallel_axis = matmul_col_dim_mapping
        group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape,
                                      parallel_axis, rank_id)
        group = new_process_group(group_ranks)

Z
zhaoyingli 已提交
587 588 589 590 591 592 593 594 595 596 597 598 599 600 601
        # infer new var shape with op dist attr
        x_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(X_var)
        assert x_tensor_dist_attr is not None
        identity_var_dist_attr = op_dist_attr.get_input_dist_attr(X_var.name)
        assert identity_var_dist_attr is not None
        ref_shape_x = infer_shape(main_block, X_var, x_tensor_dist_attr,
                                  identity_var_dist_attr)
        # infer out var shape with op dist attr
        out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var)
        assert out_tensor_dist_attr is not None
        out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert out_var_dist_attr is not None
        ref_shape_out = infer_shape(main_block, Out_var, out_tensor_dist_attr,
                                    out_var_dist_attr)

602 603 604 605 606 607 608 609
        intermediate_var_0 = main_block.create_var(
            name=unique_name.generate_with_ignorable_key(".".join(
                ["c_identity", 'tmp'])),
            dtype=X_var.dtype,
            shape=X_var.shape,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=X_var.stop_gradient)
Z
zhaoyingli 已提交
610 611 612
        # set intermediate_var_0's dist_attr with X_var's dist_attr
        ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                             identity_var_dist_attr)
613 614 615 616 617 618 619 620 621 622 623 624 625

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

        c_identity_op = main_block.append_op(
            type='c_identity',
            inputs={'X': [X_var]},
            outputs={'Out': intermediate_var_0},
            attrs={
                'ring_id': group.id,
                'use_calc_stream': True,
                'use_model_parallel': True,
626
                OP_ROLE_KEY: src_op.attr('op_role')
627
            })
Z
zhaoyingli 已提交
628 629
        if intermediate_var_0.shape != ref_shape_x:
            intermediate_var_0.desc.set_shape(ref_shape_x)
630 631 632 633 634 635 636 637 638

        check_variable_and_dtype(intermediate_var_0, 'x',
                                 ['float16', 'float32', 'float64'], 'linear')
        check_dtype(intermediate_var_0.dtype, 'dtype',
                    ['float16', 'float32', 'float64'], 'linear')
        attrs = {
            'transpose_X': False,
            'transpose_Y': False,
            'alpha': 1,
639
            OP_ROLE_KEY: src_op('op_role')
640 641
        }
        inputs = {'X': [intermediate_var_0], 'Y': [Weight_var]}
642 643 644 645
        matmul_op = main_block.append_op(type='matmul',
                                         inputs=inputs,
                                         outputs={'Out': Out_var},
                                         attrs=attrs)
Z
zhaoyingli 已提交
646 647 648 649 650 651 652
        if Out_var.shape != ref_shape_out:
            Out_var.desc.set_shape(ref_shape_out)

        # set dist op's dist_attr with serial op's dist_attr
        # c_identity
        identity_op_dist_attr = OperatorDistributedAttribute()
        identity_op_dist_attr.process_mesh = op_dist_attr.process_mesh
653
        identity_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671
        identity_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        # input
        input_varname = c_identity_op.desc.input_arg_names()[0]
        input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname)
        assert input_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        identity_op_dist_attr.set_input_dist_attr(input_varname,
                                                  input_dist_attr)
        # output
        output_varname = c_identity_op.desc.output_arg_names()[0]
        identity_op_dist_attr.set_output_dist_attr(output_varname,
                                                   input_dist_attr)
        # set op dist attr
        ctx.set_op_dist_attr_for_program(c_identity_op, identity_op_dist_attr)

        # matmul
        matmul_op_dist_attr = OperatorDistributedAttribute()
        matmul_op_dist_attr.process_mesh = op_dist_attr.process_mesh
672
        matmul_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
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
        matmul_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        # input
        for input_varname in matmul_op.desc.input_arg_names():
            if input_varname in src_op.desc.input_arg_names():
                input_dist_attr = op_dist_attr.get_input_dist_attr(
                    input_varname)
                assert input_dist_attr is not None, "dist_attr is {}".format(
                    op_dist_attr)
                matmul_op_dist_attr.set_input_dist_attr(input_varname,
                                                        input_dist_attr)
            else:
                input_var = main_block.var(input_varname)
                tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(
                    input_var)
                matmul_op_dist_attr.set_input_dist_attr(input_varname,
                                                        tensor_dist_attr)
        # output
        output_varname = matmul_op.desc.output_arg_names()[0]
        output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname)
        assert output_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        matmul_op_dist_attr.set_output_dist_attr(output_varname,
                                                 output_dist_attr)
        # set op dist attr
        ctx.set_op_dist_attr_for_program(matmul_op, matmul_op_dist_attr)
698 699

        # init param sync
700
        if Weight_var.is_parameter and not op_dist_attr.is_recompute:
701
            _init_param_sync(Weight_var, dist_op_context, startup_block, ctx,
702 703 704 705 706
                             rank_id)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)
707

708 709 710

# RowParallel
class DistributedMatmulImpl1(DistributedOperatorImpl):
711

712
    def __init__(self, name):
713
        super(DistributedMatmulImpl1, self).__init__(name)
714
        self._forward_implemented = True
715
        self._backward_implemented = True
716

717 718 719
    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
720 721 722 723 724 725
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
        if is_dim_replicate(x_dims_mapping[-1]):
            return False
726 727
        if is_dim_replicate(y_dims_mapping[-2]) or is_dim_shard(
                y_dims_mapping[-1]):
728 729 730 731 732 733 734
            return False
        # Other dimensions must be replicate except the batch dimension
        for mapping in x_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

735 736 737
    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
738 739 740 741 742 743 744 745 746 747
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
        if is_dim_shard(out_dims_mapping[-1]):
            return False
        # Other dimensions must be replicate except the batch dimension
        for mapping in out_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

748
    def is_auto_compatible(self, dist_op):
749 750
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
751
            return False
752

753
        if not _is_auto_compatible_for_matmul(dist_op):
754 755 756 757
            return False

        return True

758
    def update_dims_mapping(self, dist_op):
759
        changed = False
760
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
761 762 763 764
        if dim_changed:
            changed = True
        return changed

765 766 767 768 769 770
    @staticmethod
    def forward(ctx, *args, **kwargs):
        """
        kwargs: inputname_mapping & outputname_mapping
        """

771
        dist_op_context = ctx.dist_op_context
772 773 774 775
        main_block = dist_op_context.work_block
        startup_block = dist_op_context.startup_block
        src_op = dist_op_context.cur_src_op
        rank_id = dist_op_context.rank_id
776
        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
777 778 779 780
        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(src_op))

        # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
781 782
        if rank_id not in op_dist_attr.process_mesh.processes:
            rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh,
783 784
                                              rank_id)

785
        # check validation of inputs / outputs
786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805
        for input_name in src_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                src_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in src_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                src_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        X_var = main_block.var(kwargs['X'][0])
        Weight_var = main_block.var(kwargs['Y'][0])
        Out_var = main_block.var(kwargs['Out'][0])

        # TODO infer logic comm presentation
        matmul_row_dim_mapping = op_dist_attr.get_input_dims_mapping(
806
            Weight_var.name)[-2]
807 808
        assert matmul_row_dim_mapping >= 0, "row_parallel_matmul's row should be divided by a specific mesh axis, but got [{}]".format(
            matmul_row_dim_mapping)
809 810
        process_mesh_shape = op_dist_attr.process_mesh.topology
        process_mesh_group = op_dist_attr.process_mesh.processes
811 812 813 814 815 816 817 818 819 820 821 822 823 824

        parallel_axis = matmul_row_dim_mapping
        group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape,
                                      parallel_axis, rank_id)
        group = new_process_group(group_ranks)

        check_variable_and_dtype(X_var, 'x', ['float16', 'float32', 'float64'],
                                 'linear')
        check_dtype(X_var.dtype, 'dtype', ['float16', 'float32', 'float64'],
                    'linear')
        attrs = {
            'transpose_X': False,
            'transpose_Y': False,
            'alpha': 1,
825
            OP_ROLE_KEY: src_op.attr('op_role')
826 827
        }
        inputs = {'X': X_var, 'Y': Weight_var}
Z
zhaoyingli 已提交
828 829 830 831 832 833 834 835 836

        # infer out var shape with op dist attr
        out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var)
        assert out_tensor_dist_attr is not None
        out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert out_var_dist_attr is not None
        ref_shape = infer_shape(main_block, Out_var, out_tensor_dist_attr,
                                out_var_dist_attr)

837
        intermediate_var_0 = main_block.create_var(
838 839
            name=unique_name.generate_with_ignorable_key(".".join(
                ["c_allreduce_sum", 'tmp'])),
840 841 842 843 844 845 846
            shape=Out_var.shape,
            dtype=Out_var.dtype,
            type=Out_var.type,
            lod_level=Out_var.lod_level,
            persistable=False,
            is_data=False,
            need_check_feed=Out_var.desc.need_check_feed())
Z
zhaoyingli 已提交
847 848 849
        # set intermediate_var_0's dist_attr with Out_var's dist_attr
        ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                             out_var_dist_attr)
850

851 852 853 854
        matmul_op = main_block.append_op(type='matmul',
                                         inputs=inputs,
                                         outputs={'Out': intermediate_var_0},
                                         attrs=attrs)
Z
zhaoyingli 已提交
855 856
        if intermediate_var_0.shape != ref_shape:
            intermediate_var_0.desc.set_shape(ref_shape)
857 858 859 860 861 862 863 864

        c_allreduce_sum_op = main_block.append_op(
            type='c_allreduce_sum',
            inputs={'X': intermediate_var_0},
            outputs={'Out': Out_var},
            attrs={
                'ring_id': group.id,
                'use_calc_stream': True,
865 866
                'use_model_parallel': True,
                OP_ROLE_KEY: src_op.attr('op_role')
867
            })
Z
zhaoyingli 已提交
868 869 870 871 872 873 874
        if Out_var.shape != ref_shape:
            Out_var.desc.set_shape(ref_shape)

        # set dist op's dist_attr with serial op's dist_attr
        # matmul
        matmul_op_dist_attr = OperatorDistributedAttribute()
        matmul_op_dist_attr.process_mesh = op_dist_attr.process_mesh
875
        matmul_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893
        matmul_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in matmul_op.desc.input_arg_names():
            input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname)
            assert input_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            matmul_op_dist_attr.set_input_dist_attr(input_varname,
                                                    input_dist_attr)
        output_varname = matmul_op.desc.output_arg_names()[0]
        output_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert output_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        matmul_op_dist_attr.set_output_dist_attr(output_varname,
                                                 output_dist_attr)
        ctx.set_op_dist_attr_for_program(matmul_op, matmul_op_dist_attr)

        # allreduce
        allreduce_op_dist_attr = OperatorDistributedAttribute()
        allreduce_op_dist_attr.process_mesh = op_dist_attr.process_mesh
894
        allreduce_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
895 896 897 898 899 900 901 902 903 904 905 906 907 908 909
        allreduce_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in c_allreduce_sum_op.desc.input_arg_names():
            input_var = main_block.var(input_varname)
            tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(input_var)
            assert tensor_dist_attr is not None
            allreduce_op_dist_attr.set_input_dist_attr(input_varname,
                                                       tensor_dist_attr)
        for output_varname in c_allreduce_sum_op.desc.output_arg_names():
            output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname)
            assert output_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            allreduce_op_dist_attr.set_output_dist_attr(output_varname,
                                                        output_dist_attr)
        ctx.set_op_dist_attr_for_program(c_allreduce_sum_op,
                                         allreduce_op_dist_attr)
910 911

        # init param sync
912
        if Weight_var.is_parameter and not op_dist_attr.is_recompute:
913
            _init_param_sync(Weight_var, dist_op_context, startup_block, ctx,
914 915 916 917 918
                             rank_id)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)
919

920

921
# ReplicateParallel
922
class DistributedMatmulImpl2(DistributedOperatorImpl):
923

924
    def __init__(self, name):
925
        super(DistributedMatmulImpl2, self).__init__(name)
926

927 928 929
    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
930 931 932 933 934 935 936
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)

        if is_dim_shard(x_dims_mapping[-1]):
            return False
937 938
        if is_valid_list_index(x_dims_mapping, -2) and is_dim_shard(
                x_dims_mapping[-2]):
939 940 941 942
            return False

        if is_dim_shard(y_dims_mapping[-1]):
            return False
943 944
        if is_valid_list_index(y_dims_mapping, -2) and is_dim_shard(
                y_dims_mapping[-2]):
945 946 947 948
            return False

        return True

949 950 951
    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
952 953 954 955 956
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)

        if is_dim_shard(out_dims_mapping[-1]):
            return False
957 958
        if is_valid_list_index(out_dims_mapping, -2) and is_dim_shard(
                out_dims_mapping[-2]):
959 960 961 962
            return False

        return True

963
    def is_auto_compatible(self, dist_op):
964 965
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
966 967
            return False

968
        if not _is_auto_compatible_for_matmul(dist_op):
969 970 971 972
            return False

        return True

973
    def update_dims_mapping(self, dist_op):
974
        changed = False
975
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
976 977 978 979
        if dim_changed:
            changed = True
        return changed

980 981 982 983
    @staticmethod
    def forward(ctx, *args, **kwargs):
        DistributedDefaultImpl0.forward(ctx, *args, **kwargs)

984 985 986 987
    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)

988 989 990 991 992 993 994 995 996

register_distributed_operator_impl("matmul",
                                   DistributedMatmulImpl0("column_parallel"))
register_distributed_operator_impl("matmul",
                                   DistributedMatmulImpl1("row_parallel"))
register_distributed_operator_impl("matmul",
                                   DistributedMatmulImpl2("replicate_parallel"))


997
class DistributedMatmulV2(DistributedOperatorImplContainer):
998

999 1000
    def __init__(self, op_type):
        super(DistributedMatmulV2, self).__init__(op_type)
1001 1002


1003
register_distributed_operator_impl_container(DistributedMatmulV2("matmul_v2"))
1004 1005


1006 1007
# ColumnParallel
class DistributedMatmulV2Impl0(DistributedOperatorImpl):
1008

1009
    def __init__(self, name):
1010
        super(DistributedMatmulV2Impl0, self).__init__(name)
1011
        self._forward_implemented = True
1012
        self._backward_implemented = True
1013

1014 1015 1016
    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
1017 1018 1019 1020 1021 1022
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
        if is_dim_shard(x_dims_mapping[-1]):
            return False
1023 1024
        if is_dim_shard(y_dims_mapping[-2]) or is_dim_replicate(
                y_dims_mapping[-1]):
1025 1026 1027 1028 1029 1030
            return False
        for mapping in x_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

1031 1032 1033
    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
1034 1035 1036 1037 1038 1039 1040 1041 1042
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
        if is_dim_replicate(out_dims_mapping[-1]):
            return False
        for mapping in out_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

1043
    def is_auto_compatible(self, dist_op):
1044 1045
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
1046 1047
            return False

1048
        if not _is_auto_compatible_for_matmul(dist_op):
1049 1050 1051 1052
            return False

        return True

1053
    def update_dims_mapping(self, dist_op):
1054
        changed = False
1055
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
1056 1057 1058 1059
        if dim_changed:
            changed = True
        return changed

1060 1061 1062 1063 1064 1065
    @staticmethod
    def forward(ctx, *args, **kwargs):
        """
        kwargs: inputname_mapping & outputname_mapping
        """

1066
        dist_op_context = ctx.dist_op_context
1067 1068 1069 1070
        main_block = dist_op_context.work_block
        startup_block = dist_op_context.startup_block
        src_op = dist_op_context.cur_src_op
        rank_id = dist_op_context.rank_id
1071
        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
1072 1073 1074 1075
        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(src_op))

        # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
1076 1077
        if rank_id not in op_dist_attr.process_mesh.processes:
            rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh,
1078 1079
                                              rank_id)

1080
        # check validation of inputs / outputs
1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095
        for input_name in src_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                src_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in src_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                src_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        X_var = main_block.var(kwargs['X'][0])
1096
        Weight_var = main_block._var_recursive(kwargs['Y'][0])
1097 1098 1099 1100
        Out_var = main_block.var(kwargs['Out'][0])

        # TODO infer logic comm presentation
        matmul_col_dim_mapping = op_dist_attr.get_input_dims_mapping(
1101
            Weight_var.name)[-1]
1102 1103
        assert matmul_col_dim_mapping >= 0, "col_parallel_matmul's row should be divided by a specific mesh axis, but got [{}]".format(
            matmul_col_dim_mapping)
1104 1105
        process_mesh_shape = op_dist_attr.process_mesh.topology
        process_mesh_group = op_dist_attr.process_mesh.processes
1106 1107 1108 1109 1110 1111

        parallel_axis = matmul_col_dim_mapping
        group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape,
                                      parallel_axis, rank_id)
        group = new_process_group(group_ranks)

Z
zhaoyingli 已提交
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126
        # infer new var shape with op dist attr
        x_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(X_var)
        assert x_tensor_dist_attr is not None
        identity_var_dist_attr = op_dist_attr.get_input_dist_attr(X_var.name)
        assert identity_var_dist_attr is not None
        ref_shape_x = infer_shape(main_block, X_var, x_tensor_dist_attr,
                                  identity_var_dist_attr)
        # infer out var shape with op dist attr
        out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var)
        assert out_tensor_dist_attr is not None
        out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert out_var_dist_attr is not None
        ref_shape_out = infer_shape(main_block, Out_var, out_tensor_dist_attr,
                                    out_var_dist_attr)

1127 1128 1129 1130 1131 1132 1133 1134
        intermediate_var_0 = main_block.create_var(
            name=unique_name.generate_with_ignorable_key(".".join(
                ["c_identity", 'tmp'])),
            dtype=X_var.dtype,
            shape=X_var.shape,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=X_var.stop_gradient)
Z
zhaoyingli 已提交
1135 1136 1137
        # set intermediate_var_0's dist_attr with X_var's dist_attr
        ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                             identity_var_dist_attr)
1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149

        check_variable_and_dtype(
            X_var, 'tensor',
            ['float16', 'float32', 'float64', 'int32', 'int64'], '_c_identity')
        c_identity_op = main_block.append_op(
            type='c_identity',
            inputs={'X': [X_var]},
            outputs={'Out': intermediate_var_0},
            attrs={
                'ring_id': group.id,
                'use_calc_stream': True,
                'use_model_parallel': True,
1150
                OP_ROLE_KEY: src_op.attr('op_role'),
1151
            })
Z
zhaoyingli 已提交
1152 1153
        if intermediate_var_0.shape != ref_shape_x:
            intermediate_var_0.desc.set_shape(ref_shape_x)
1154 1155 1156 1157 1158

        check_variable_and_dtype(intermediate_var_0, 'x',
                                 ['float16', 'float32', 'float64'], 'linear')
        check_dtype(intermediate_var_0.dtype, 'dtype',
                    ['float16', 'float32', 'float64'], 'linear')
1159 1160 1161 1162 1163
        attrs = {
            'trans_x': False,
            'trans_y': False,
            OP_ROLE_KEY: src_op.attr('op_role')
        }
1164
        inputs = {'X': [intermediate_var_0], 'Y': [Weight_var]}
1165 1166 1167 1168
        matmul_v2_op = main_block.append_op(type='matmul_v2',
                                            inputs=inputs,
                                            outputs={'Out': Out_var},
                                            attrs=attrs)
Z
zhaoyingli 已提交
1169 1170 1171 1172 1173 1174 1175
        if Out_var.shape != ref_shape_out:
            Out_var.desc.set_shape(ref_shape_out)

        # set dist op's dist_attr with serial op's dist_attr
        # c_identity
        identity_op_dist_attr = OperatorDistributedAttribute()
        identity_op_dist_attr.process_mesh = op_dist_attr.process_mesh
1176
        identity_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193
        identity_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        # input
        input_varname = c_identity_op.desc.input_arg_names()[0]
        input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname)
        assert input_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        identity_op_dist_attr.set_input_dist_attr(input_varname,
                                                  input_dist_attr)
        # output
        output_varname = c_identity_op.desc.output_arg_names()[0]
        identity_op_dist_attr.set_output_dist_attr(output_varname,
                                                   input_dist_attr)
        ctx.set_op_dist_attr_for_program(c_identity_op, identity_op_dist_attr)

        # matmulv2
        matmulv2_op_dist_attr = OperatorDistributedAttribute()
        matmulv2_op_dist_attr.process_mesh = op_dist_attr.process_mesh
1194
        matmulv2_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
1195 1196 1197 1198 1199 1200 1201
        matmulv2_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in matmul_v2_op.desc.input_arg_names():
            if input_varname in src_op.desc.input_arg_names():
                input_dist_attr = op_dist_attr.get_input_dist_attr(
                    input_varname)
                assert input_dist_attr is not None, "dist_attr is {}".format(
                    op_dist_attr)
1202 1203
                matmulv2_op_dist_attr.set_input_dist_attr(
                    input_varname, input_dist_attr)
Z
zhaoyingli 已提交
1204 1205 1206 1207
            else:
                input_var = main_block.var(input_varname)
                tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(
                    input_var)
1208 1209
                matmulv2_op_dist_attr.set_input_dist_attr(
                    input_varname, tensor_dist_attr)
Z
zhaoyingli 已提交
1210 1211 1212 1213 1214 1215 1216
        for output_varname in matmul_v2_op.desc.output_arg_names():
            output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname)
            assert output_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            matmulv2_op_dist_attr.set_output_dist_attr(output_varname,
                                                       output_dist_attr)
        ctx.set_op_dist_attr_for_program(matmul_v2_op, matmulv2_op_dist_attr)
1217 1218

        # init param sync
1219
        if Weight_var.is_parameter and not op_dist_attr.is_recompute:
1220
            _init_param_sync(Weight_var, dist_op_context, startup_block, ctx,
1221 1222 1223 1224 1225
                             rank_id)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)
1226 1227 1228 1229


# RowParallel
class DistributedMatmulV2Impl1(DistributedOperatorImpl):
1230

1231
    def __init__(self, name):
1232
        super(DistributedMatmulV2Impl1, self).__init__(name)
1233
        self._forward_implemented = True
1234
        self._backward_implemented = True
1235

1236 1237 1238
    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
1239 1240 1241 1242 1243 1244
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
        if is_dim_replicate(x_dims_mapping[-1]):
            return False
1245 1246
        if is_dim_replicate(y_dims_mapping[-2]) or is_dim_shard(
                y_dims_mapping[-1]):
1247 1248 1249 1250 1251 1252 1253
            return False
        # Other dimensions must be replicate except the batch dimension
        for mapping in x_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

1254 1255 1256
    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
1257 1258 1259 1260 1261 1262 1263 1264 1265 1266
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
        if is_dim_shard(out_dims_mapping[-1]):
            return False
        # Other dimensions must be replicate except the batch dimension
        for mapping in out_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

1267
    def is_auto_compatible(self, dist_op):
1268 1269
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
1270 1271
            return False

1272
        if not _is_auto_compatible_for_matmul(dist_op):
1273 1274 1275 1276
            return False

        return True

1277
    def update_dims_mapping(self, dist_op):
1278
        changed = False
1279
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
1280 1281 1282 1283
        if dim_changed:
            changed = True
        return changed

1284 1285 1286 1287 1288 1289
    @staticmethod
    def forward(ctx, *args, **kwargs):
        """
        kwargs: inputname_mapping & outputname_mapping
        """

1290
        dist_op_context = ctx.dist_op_context
1291 1292 1293 1294
        main_block = dist_op_context.work_block
        startup_block = dist_op_context.startup_block
        src_op = dist_op_context.cur_src_op
        rank_id = dist_op_context.rank_id
1295
        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
1296 1297 1298 1299
        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(src_op))

        # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
1300 1301
        if rank_id not in op_dist_attr.process_mesh.processes:
            rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh,
1302 1303
                                              rank_id)

1304
        # check validation of inputs / outputs
1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319
        for input_name in src_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                src_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in src_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                src_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        X_var = main_block.var(kwargs['X'][0])
1320
        Weight_var = main_block._var_recursive(kwargs['Y'][0])
1321 1322 1323 1324
        Out_var = main_block.var(kwargs['Out'][0])

        # TODO infer logic comm presentation
        matmul_row_dim_mapping = op_dist_attr.get_input_dims_mapping(
1325
            Weight_var.name)[-2]
1326 1327
        assert matmul_row_dim_mapping >= 0, "row_parallel_matmul's row should be divided by a specific mesh axis, but got [{}]".format(
            matmul_row_dim_mapping)
1328 1329
        process_mesh_shape = op_dist_attr.process_mesh.topology
        process_mesh_group = op_dist_attr.process_mesh.processes
1330 1331 1332 1333 1334 1335 1336 1337 1338 1339

        parallel_axis = matmul_row_dim_mapping
        group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape,
                                      parallel_axis, rank_id)
        group = new_process_group(group_ranks)

        check_variable_and_dtype(X_var, 'x', ['float16', 'float32', 'float64'],
                                 'linear')
        check_dtype(X_var.dtype, 'dtype', ['float16', 'float32', 'float64'],
                    'linear')
1340 1341 1342 1343 1344
        attrs = {
            'trans_x': False,
            'trans_y': False,
            OP_ROLE_KEY: src_op.attr('op_role')
        }
1345
        inputs = {'X': X_var, 'Y': Weight_var}
Z
zhaoyingli 已提交
1346 1347 1348 1349 1350 1351 1352 1353 1354

        # infer out var shape with op dist attr
        out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var)
        assert out_tensor_dist_attr is not None
        out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert out_var_dist_attr is not None
        ref_shape = infer_shape(main_block, Out_var, out_tensor_dist_attr,
                                out_var_dist_attr)

1355
        intermediate_var_0 = main_block.create_var(
1356 1357
            name=unique_name.generate_with_ignorable_key(".".join(
                ["c_allreduce_sum", 'tmp'])),
1358 1359 1360 1361 1362 1363 1364
            shape=Out_var.shape,
            dtype=Out_var.dtype,
            type=Out_var.type,
            lod_level=Out_var.lod_level,
            persistable=False,
            is_data=False,
            need_check_feed=Out_var.desc.need_check_feed())
Z
zhaoyingli 已提交
1365 1366 1367
        # set intermediate_var_0's dist_attr with Out_var's dist_attr
        ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                             out_var_dist_attr)
1368

1369 1370 1371 1372
        matmul_v2_op = main_block.append_op(type='matmul_v2',
                                            inputs=inputs,
                                            outputs={'Out': intermediate_var_0},
                                            attrs=attrs)
Z
zhaoyingli 已提交
1373 1374
        if intermediate_var_0.shape != ref_shape:
            intermediate_var_0.desc.set_shape(ref_shape)
1375 1376 1377 1378 1379 1380 1381 1382

        c_allreduce_sum_op = main_block.append_op(
            type='c_allreduce_sum',
            inputs={'X': intermediate_var_0},
            outputs={'Out': Out_var},
            attrs={
                'ring_id': group.id,
                'use_calc_stream': True,
1383 1384
                'use_model_parallel': True,
                OP_ROLE_KEY: src_op.attr('op_role')
1385
            })
Z
zhaoyingli 已提交
1386 1387 1388 1389 1390 1391 1392
        if Out_var.shape != ref_shape:
            Out_var.desc.set_shape(ref_shape)

        # set dist op's dist_attr with serial op's dist_attr
        # matmulv2
        matmulv2_op_dist_attr = OperatorDistributedAttribute()
        matmulv2_op_dist_attr.process_mesh = op_dist_attr.process_mesh
1393
        matmulv2_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411
        matmulv2_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in matmul_v2_op.desc.input_arg_names():
            input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname)
            assert input_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            matmulv2_op_dist_attr.set_input_dist_attr(input_varname,
                                                      input_dist_attr)
        output_varname = matmul_v2_op.desc.output_arg_names()[0]
        output_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert output_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        matmulv2_op_dist_attr.set_output_dist_attr(output_varname,
                                                   output_dist_attr)
        ctx.set_op_dist_attr_for_program(matmul_v2_op, matmulv2_op_dist_attr)

        # allreduce
        allreduce_op_dist_attr = OperatorDistributedAttribute()
        allreduce_op_dist_attr.process_mesh = op_dist_attr.process_mesh
1412
        allreduce_op_dist_attr.impl_type = op_dist_attr.impl_type
Z
zhaoyingli 已提交
1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427
        allreduce_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in c_allreduce_sum_op.desc.input_arg_names():
            input_var = main_block.var(input_varname)
            tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(input_var)
            assert tensor_dist_attr is not None
            allreduce_op_dist_attr.set_input_dist_attr(input_varname,
                                                       tensor_dist_attr)
        for output_varname in c_allreduce_sum_op.desc.output_arg_names():
            output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname)
            assert output_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            allreduce_op_dist_attr.set_output_dist_attr(output_varname,
                                                        output_dist_attr)
        ctx.set_op_dist_attr_for_program(c_allreduce_sum_op,
                                         allreduce_op_dist_attr)
1428 1429

        # init param sync
1430
        if Weight_var.is_parameter and not op_dist_attr.is_recompute:
1431
            _init_param_sync(Weight_var, dist_op_context, startup_block, ctx,
1432 1433 1434 1435 1436
                             rank_id)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)
1437 1438


1439
# ReplicateParallel
1440
class DistributedMatmulV2Impl2(DistributedOperatorImpl):
1441

1442
    def __init__(self, name):
1443
        super(DistributedMatmulV2Impl2, self).__init__(name)
1444

1445 1446 1447
    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
1448 1449 1450 1451 1452 1453 1454
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)

        if is_dim_shard(x_dims_mapping[-1]):
            return False
1455 1456
        if is_valid_list_index(x_dims_mapping, -2) and is_dim_shard(
                x_dims_mapping[-2]):
1457 1458 1459 1460
            return False

        if is_dim_shard(y_dims_mapping[-1]):
            return False
1461 1462
        if is_valid_list_index(y_dims_mapping, -2) and is_dim_shard(
                y_dims_mapping[-2]):
1463 1464 1465
            return False
        return True

1466 1467 1468 1469 1470
    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
1471 1472 1473 1474 1475
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)

        if is_dim_shard(out_dims_mapping[-1]):
            return False
1476 1477
        if is_valid_list_index(out_dims_mapping, -2) and is_dim_shard(
                out_dims_mapping[-2]):
1478 1479 1480 1481
            return False

        return True

1482
    def is_auto_compatible(self, dist_op):
1483 1484
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
1485 1486
            return False

1487
        if not _is_auto_compatible_for_matmul(dist_op):
1488 1489 1490 1491
            return False

        return True

1492
    def update_dims_mapping(self, dist_op):
1493
        changed = False
1494
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
1495 1496 1497 1498
        if dim_changed:
            changed = True
        return changed

1499 1500 1501 1502
    @staticmethod
    def forward(ctx, *args, **kwargs):
        DistributedDefaultImpl0.forward(ctx, *args, **kwargs)

1503 1504 1505 1506
    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)

1507

1508 1509 1510 1511
register_distributed_operator_impl("matmul_v2",
                                   DistributedMatmulV2Impl0("column_parallel"))
register_distributed_operator_impl("matmul_v2",
                                   DistributedMatmulV2Impl1("row_parallel"))
1512
register_distributed_operator_impl(
1513
    "matmul_v2", DistributedMatmulV2Impl2("replicate_parallel"))
1514 1515 1516


class DistributedMul(DistributedOperatorImplContainer):
1517

1518 1519 1520 1521 1522 1523 1524 1525 1526
    def __init__(self, op_type):
        super(DistributedMul, self).__init__(op_type)


register_distributed_operator_impl_container(DistributedMul("mul"))


# ColumnParallel
class DistributedMulImpl0(DistributedOperatorImpl):
1527

1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541
    def __init__(self, name):
        super(DistributedMulImpl0, self).__init__(name)
        self._forward_implemented = True
        self._backward_implemented = True

    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
        if is_dim_shard(x_dims_mapping[-1]):
            return False
1542 1543
        if is_dim_shard(y_dims_mapping[-2]) or is_dim_replicate(
                y_dims_mapping[-1]):
1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668
            return False
        for mapping in x_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
        if is_dim_replicate(out_dims_mapping[-1]):
            return False
        for mapping in out_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

    def is_auto_compatible(self, dist_op):
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
            return False

        if not _is_auto_compatible_for_matmul(dist_op):
            return False

        return True

    def update_dims_mapping(self, dist_op):
        changed = False
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
        if dim_changed:
            changed = True
        return changed

    @staticmethod
    def forward(ctx, *args, **kwargs):
        """
        kwargs: inputname_mapping & outputname_mapping
        """

        dist_op_context = ctx.dist_op_context
        main_block = dist_op_context.work_block
        startup_block = dist_op_context.startup_block
        src_op = dist_op_context.cur_src_op
        rank_id = dist_op_context.rank_id
        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(src_op))

        # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
        if rank_id not in op_dist_attr.process_mesh.processes:
            rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh,
                                              rank_id)

        # check validation of inputs / outputs
        for input_name in src_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                src_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in src_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                src_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        X_var = main_block.var(kwargs['X'][0])
        Weight_var = main_block._var_recursive(kwargs['Y'][0])
        Out_var = main_block.var(kwargs['Out'][0])

        # TODO infer logic comm presentation
        matmul_col_dim_mapping = op_dist_attr.get_input_dims_mapping(
            Weight_var.name)[-1]
        assert matmul_col_dim_mapping >= 0, "col_parallel_matmul's row should be divided by a specific mesh axis, but got [{}]".format(
            matmul_col_dim_mapping)
        process_mesh_shape = op_dist_attr.process_mesh.topology
        process_mesh_group = op_dist_attr.process_mesh.processes

        parallel_axis = matmul_col_dim_mapping
        group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape,
                                      parallel_axis, rank_id)
        group = new_process_group(group_ranks)

        # infer new var shape with op dist attr
        x_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(X_var)
        assert x_tensor_dist_attr is not None
        identity_var_dist_attr = op_dist_attr.get_input_dist_attr(X_var.name)
        assert identity_var_dist_attr is not None
        ref_shape_x = infer_shape(main_block, X_var, x_tensor_dist_attr,
                                  identity_var_dist_attr)
        # infer out var shape with op dist attr
        out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var)
        assert out_tensor_dist_attr is not None
        out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert out_var_dist_attr is not None
        ref_shape_out = infer_shape(main_block, Out_var, out_tensor_dist_attr,
                                    out_var_dist_attr)

        intermediate_var_0 = main_block.create_var(
            name=unique_name.generate_with_ignorable_key(".".join(
                ["c_identity", 'tmp'])),
            dtype=X_var.dtype,
            shape=X_var.shape,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=X_var.stop_gradient)
        # set intermediate_var_0's dist_attr with X_var's dist_attr
        ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                             identity_var_dist_attr)

        check_variable_and_dtype(
            X_var, 'tensor',
            ['float16', 'float32', 'float64', 'int32', 'int64'], '_c_identity')
        c_identity_op = main_block.append_op(
            type='c_identity',
            inputs={'X': [X_var]},
            outputs={'Out': intermediate_var_0},
            attrs={
                'ring_id': group.id,
                'use_calc_stream': True,
                'use_model_parallel': True,
1669
                OP_ROLE_KEY: src_op.attr('op_role')
1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680
            })
        if intermediate_var_0.shape != ref_shape_x:
            intermediate_var_0.desc.set_shape(ref_shape_x)

        check_variable_and_dtype(intermediate_var_0, 'x',
                                 ['float16', 'float32', 'float64'], 'linear')
        check_dtype(intermediate_var_0.dtype, 'dtype',
                    ['float16', 'float32', 'float64'], 'linear')
        # attrs = {'trans_x': False, 'trans_y': False}
        attrs = {
            "x_num_col_dims": src_op.desc.attr("x_num_col_dims"),
1681 1682
            "y_num_col_dims": src_op.desc.attr("y_num_col_dims"),
            OP_ROLE_KEY: src_op.attr('op_role')
1683 1684
        }
        inputs = {'X': [intermediate_var_0], 'Y': [Weight_var]}
1685 1686 1687 1688
        mul_op = main_block.append_op(type='mul',
                                      inputs=inputs,
                                      outputs={'Out': Out_var},
                                      attrs=attrs)
1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721
        if Out_var.shape != ref_shape_out:
            Out_var.desc.set_shape(ref_shape_out)

        # set dist op's dist_attr with serial op's dist_attr
        # c_identity
        identity_op_dist_attr = OperatorDistributedAttribute()
        identity_op_dist_attr.process_mesh = op_dist_attr.process_mesh
        identity_op_dist_attr.impl_type = op_dist_attr.impl_type
        identity_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        # input
        input_varname = c_identity_op.desc.input_arg_names()[0]
        input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname)
        assert input_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        identity_op_dist_attr.set_input_dist_attr(input_varname,
                                                  input_dist_attr)
        # output
        output_varname = c_identity_op.desc.output_arg_names()[0]
        identity_op_dist_attr.set_output_dist_attr(output_varname,
                                                   input_dist_attr)
        ctx.set_op_dist_attr_for_program(c_identity_op, identity_op_dist_attr)

        # matmulv2
        matmulv2_op_dist_attr = OperatorDistributedAttribute()
        matmulv2_op_dist_attr.process_mesh = op_dist_attr.process_mesh
        matmulv2_op_dist_attr.impl_type = op_dist_attr.impl_type
        matmulv2_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in mul_op.desc.input_arg_names():
            if input_varname in src_op.desc.input_arg_names():
                input_dist_attr = op_dist_attr.get_input_dist_attr(
                    input_varname)
                assert input_dist_attr is not None, "dist_attr is {}".format(
                    op_dist_attr)
1722 1723
                matmulv2_op_dist_attr.set_input_dist_attr(
                    input_varname, input_dist_attr)
1724 1725 1726 1727
            else:
                input_var = main_block.var(input_varname)
                tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(
                    input_var)
1728 1729
                matmulv2_op_dist_attr.set_input_dist_attr(
                    input_varname, tensor_dist_attr)
1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749
        for output_varname in mul_op.desc.output_arg_names():
            output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname)
            assert output_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            matmulv2_op_dist_attr.set_output_dist_attr(output_varname,
                                                       output_dist_attr)
        ctx.set_op_dist_attr_for_program(mul_op, matmulv2_op_dist_attr)

        # init param sync
        if Weight_var.is_parameter and not op_dist_attr.is_recompute:
            _init_param_sync(Weight_var, dist_op_context, startup_block, ctx,
                             rank_id)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)


# RowParallel
class DistributedMulImpl1(DistributedOperatorImpl):
1750

1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764
    def __init__(self, name):
        super(DistributedMulImpl1, self).__init__(name)
        self._forward_implemented = True
        self._backward_implemented = True

    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)
        if is_dim_replicate(x_dims_mapping[-1]):
            return False
1765 1766
        if is_dim_replicate(y_dims_mapping[-2]) or is_dim_shard(
                y_dims_mapping[-1]):
1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862
            return False
        # Other dimensions must be replicate except the batch dimension
        for mapping in x_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
        if is_dim_shard(out_dims_mapping[-1]):
            return False
        # Other dimensions must be replicate except the batch dimension
        for mapping in out_dims_mapping[1:-1]:
            if is_dim_shard(mapping):
                return False
        return True

    def is_auto_compatible(self, dist_op):
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
            return False

        if not _is_auto_compatible_for_matmul(dist_op):
            return False

        return True

    def update_dims_mapping(self, dist_op):
        changed = False
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
        if dim_changed:
            changed = True
        return changed

    @staticmethod
    def forward(ctx, *args, **kwargs):
        """
        kwargs: inputname_mapping & outputname_mapping
        """

        dist_op_context = ctx.dist_op_context
        main_block = dist_op_context.work_block
        startup_block = dist_op_context.startup_block
        src_op = dist_op_context.cur_src_op
        rank_id = dist_op_context.rank_id
        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(src_op))

        # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
        if rank_id not in op_dist_attr.process_mesh.processes:
            rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh,
                                              rank_id)

        # check validation of inputs / outputs
        for input_name in src_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                src_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in src_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                src_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        X_var = main_block.var(kwargs['X'][0])
        Weight_var = main_block._var_recursive(kwargs['Y'][0])
        Out_var = main_block.var(kwargs['Out'][0])

        # TODO infer logic comm presentation
        matmul_row_dim_mapping = op_dist_attr.get_input_dims_mapping(
            Weight_var.name)[-2]
        assert matmul_row_dim_mapping >= 0, "row_parallel_matmul's row should be divided by a specific mesh axis, but got [{}]".format(
            matmul_row_dim_mapping)
        process_mesh_shape = op_dist_attr.process_mesh.topology
        process_mesh_group = op_dist_attr.process_mesh.processes

        parallel_axis = matmul_row_dim_mapping
        group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape,
                                      parallel_axis, rank_id)
        group = new_process_group(group_ranks)

        check_variable_and_dtype(X_var, 'x', ['float16', 'float32', 'float64'],
                                 'linear')
        check_dtype(X_var.dtype, 'dtype', ['float16', 'float32', 'float64'],
                    'linear')
        # attrs = {'trans_x': False, 'trans_y': False}
        attrs = {
            "x_num_col_dims": src_op.desc.attr("x_num_col_dims"),
1863 1864
            "y_num_col_dims": src_op.desc.attr("y_num_col_dims"),
            OP_ROLE_KEY: src_op.attr('op_role')
1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876
        }
        inputs = {'X': X_var, 'Y': Weight_var}

        # infer out var shape with op dist attr
        out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var)
        assert out_tensor_dist_attr is not None
        out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert out_var_dist_attr is not None
        ref_shape = infer_shape(main_block, Out_var, out_tensor_dist_attr,
                                out_var_dist_attr)

        intermediate_var_0 = main_block.create_var(
1877 1878
            name=unique_name.generate_with_ignorable_key(".".join(
                ["c_allreduce_sum", 'tmp'])),
1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889
            shape=Out_var.shape,
            dtype=Out_var.dtype,
            type=Out_var.type,
            lod_level=Out_var.lod_level,
            persistable=False,
            is_data=False,
            need_check_feed=Out_var.desc.need_check_feed())
        # set intermediate_var_0's dist_attr with Out_var's dist_attr
        ctx.set_tensor_dist_attr_for_program(intermediate_var_0,
                                             out_var_dist_attr)

1890 1891 1892 1893
        mul_op = main_block.append_op(type='mul',
                                      inputs=inputs,
                                      outputs={'Out': intermediate_var_0},
                                      attrs=attrs)
1894 1895 1896 1897 1898 1899 1900 1901 1902 1903
        if intermediate_var_0.shape != ref_shape:
            intermediate_var_0.desc.set_shape(ref_shape)

        c_allreduce_sum_op = main_block.append_op(
            type='c_allreduce_sum',
            inputs={'X': intermediate_var_0},
            outputs={'Out': Out_var},
            attrs={
                'ring_id': group.id,
                'use_calc_stream': True,
1904 1905
                'use_model_parallel': True,
                OP_ROLE_KEY: src_op.attr('op_role')
1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961
            })
        if Out_var.shape != ref_shape:
            Out_var.desc.set_shape(ref_shape)

        # set dist op's dist_attr with serial op's dist_attr
        # matmulv2
        matmulv2_op_dist_attr = OperatorDistributedAttribute()
        matmulv2_op_dist_attr.process_mesh = op_dist_attr.process_mesh
        matmulv2_op_dist_attr.impl_type = op_dist_attr.impl_type
        matmulv2_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in mul_op.desc.input_arg_names():
            input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname)
            assert input_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            matmulv2_op_dist_attr.set_input_dist_attr(input_varname,
                                                      input_dist_attr)
        output_varname = mul_op.desc.output_arg_names()[0]
        output_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name)
        assert output_dist_attr is not None, "dist_attr is {}".format(
            op_dist_attr)
        matmulv2_op_dist_attr.set_output_dist_attr(output_varname,
                                                   output_dist_attr)
        ctx.set_op_dist_attr_for_program(mul_op, matmulv2_op_dist_attr)

        # allreduce
        allreduce_op_dist_attr = OperatorDistributedAttribute()
        allreduce_op_dist_attr.process_mesh = op_dist_attr.process_mesh
        allreduce_op_dist_attr.impl_type = op_dist_attr.impl_type
        allreduce_op_dist_attr.impl_idx = op_dist_attr.impl_idx
        for input_varname in c_allreduce_sum_op.desc.input_arg_names():
            input_var = main_block.var(input_varname)
            tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(input_var)
            assert tensor_dist_attr is not None
            allreduce_op_dist_attr.set_input_dist_attr(input_varname,
                                                       tensor_dist_attr)
        for output_varname in c_allreduce_sum_op.desc.output_arg_names():
            output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname)
            assert output_dist_attr is not None, "dist_attr is {}".format(
                op_dist_attr)
            allreduce_op_dist_attr.set_output_dist_attr(output_varname,
                                                        output_dist_attr)
        ctx.set_op_dist_attr_for_program(c_allreduce_sum_op,
                                         allreduce_op_dist_attr)

        # init param sync
        if Weight_var.is_parameter and not op_dist_attr.is_recompute:
            _init_param_sync(Weight_var, dist_op_context, startup_block, ctx,
                             rank_id)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)


# ReplicateParallel
class DistributedMulImpl2(DistributedOperatorImpl):
1962

1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975
    def __init__(self, name):
        super(DistributedMulImpl2, self).__init__(name)

    def is_input_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        x_name = op_desc.input('X')[0]
        y_name = op_desc.input('Y')[0]
        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
        y_dims_mapping = op_dist_attr.get_input_dims_mapping(y_name)

        if is_dim_shard(x_dims_mapping[-1]):
            return False
1976 1977
        if is_valid_list_index(x_dims_mapping, -2) and is_dim_shard(
                x_dims_mapping[-2]):
1978 1979 1980
            return False
        if is_dim_shard(y_dims_mapping[-1]):
            return False
1981 1982
        if is_valid_list_index(y_dims_mapping, -2) and is_dim_shard(
                y_dims_mapping[-2]):
1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
            return False
        return True

    def is_output_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        out_name = op_desc.output('Out')[0]
        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)

        if is_dim_shard(out_dims_mapping[-1]):
            return False
1996 1997
        if is_valid_list_index(out_dims_mapping, -2) and is_dim_shard(
                out_dims_mapping[-2]):
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032
            return False

        return True

    def is_auto_compatible(self, dist_op):
        if (not self.is_input_compatible(dist_op)) or \
            (not self.is_output_compatible(dist_op)):
            return False

        if not _is_auto_compatible_for_matmul(dist_op):
            return False

        return True

    def update_dims_mapping(self, dist_op):
        changed = False
        dim_changed = _update_dims_mapping_for_matmul(dist_op)
        if dim_changed:
            changed = True
        return changed

    @staticmethod
    def forward(ctx, *args, **kwargs):
        DistributedDefaultImpl0.forward(ctx, *args, **kwargs)

    @staticmethod
    def backward(ctx, *args, **kwargs):
        _right_operand_parameter_matmul_backward(ctx, *args, **kwargs)


register_distributed_operator_impl("mul",
                                   DistributedMulImpl0("column_parallel"))
register_distributed_operator_impl("mul", DistributedMulImpl1("row_parallel"))
register_distributed_operator_impl("mul",
                                   DistributedMulImpl2("replicate_parallel"))