dist_default.py 18.9 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
from .common import DistributedOperatorImplContainer
16
from .common import DistributedOperatorImpl
17
from .common import register_distributed_operator_impl_container
J
JZ-LIANG 已提交
18
from .common import register_distributed_operator_impl, is_parameter_related
19 20 21 22 23 24
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
25
from ..utils import set_dist_op_desc_original_id
26
from ..dist_attribute import OperatorDistributedAttribute
27 28 29 30 31
from paddle.fluid import core, unique_name
from paddle.fluid.framework import in_dygraph_mode
from paddle.fluid.framework import Program, Parameter, Variable, program_guard
from paddle.fluid.data_feeder import check_variable_and_dtype, check_dtype
from paddle.distributed.fleet.meta_optimizers.common import OpRole, OP_ROLE_KEY, OP_ROLE_VAR_KEY
32
from ..process_group import new_process_group
33 34
from ..utils import _get_comm_group, _get_corresponding_rank

35 36
__op_not_need_param_init__ = ["while", "cond"]

37

38
class DistributedDefault(DistributedOperatorImplContainer):
39 40
    def __init__(self, op_type):
        super(DistributedDefault, self).__init__(op_type)
41 42


43
register_distributed_operator_impl_container(DistributedDefault("default"))
44 45


46
# Replicated Default
47 48
class DistributedDefaultImpl0(DistributedOperatorImpl):
    def __init__(self, name):
49
        super(DistributedDefaultImpl0, self).__init__(name)
50 51 52
        self._forward_implemented = True
        self._backward_implemented = True

53
    def is_input_compatible(self, dist_op):
54 55 56 57 58 59 60 61 62 63 64 65
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        for arg_name in op_desc.input_arg_names():
            serial_tensor = dist_op.get_serial_input(arg_name)
            if serial_tensor.is_parameter:
                continue
            dims_mapping = op_dist_attr.get_input_dims_mapping(arg_name)
            if len(dims_mapping) > 1:
                for mapping in dims_mapping[1:]:
                    if mapping != -1:
                        return False
        return True
66

67
    def is_output_compatible(self, dist_op):
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        output_names = op_desc.output_names()
        xshape_arg_names = []
        if "XShape" in output_names:
            xshape_arg_names = op_desc.output("XShape")
        for arg_name in op_desc.output_arg_names():
            serial_tensor = dist_op.get_serial_output(arg_name)
            if serial_tensor.is_parameter:
                continue
            dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
            if arg_name not in xshape_arg_names:
                if len(dims_mapping) > 1:
                    for mapping in dims_mapping[1:]:
                        if mapping != -1:
                            return False
            else:
                if dims_mapping[0] != -1:
                    return False
                if len(dims_mapping) > 2:
                    for mapping in dims_mapping[2:]:
                        if mapping != -1:
                            return False
        return True

    def is_auto_compatible(self, dist_op):
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        batch_dim_mappings = []
        # Check input compatibility
        for arg_name in op_desc.input_arg_names():
            serial_tensor = dist_op.get_serial_input(arg_name)
            if serial_tensor.is_parameter:
                continue
            dims_mapping = op_dist_attr.get_input_dims_mapping(arg_name)
            if len(dims_mapping) > 1:
                for mapping in dims_mapping[1:]:
                    if mapping != -1:
                        return False
            batch_dim_mappings.append(dims_mapping[0])

        # Check output compatibility
        output_names = op_desc.output_names()
        xshape_arg_names = []
        if "XShape" in output_names:
            xshape_arg_names = op_desc.output("XShape")
        for arg_name in op_desc.output_arg_names():
            serial_tensor = dist_op.get_serial_output(arg_name)
            if serial_tensor.is_parameter:
                continue
            dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
            if arg_name not in xshape_arg_names:
                if len(dims_mapping) > 1:
                    for mapping in dims_mapping[1:]:
                        if mapping != -1:
                            return False
                batch_dim_mappings.append(dims_mapping[0])
            else:
                if dims_mapping[0] != -1:
                    return False
                if len(dims_mapping) > 2:
                    for mapping in dims_mapping[2:]:
                        if mapping != -1:
                            return False
                batch_dim_mappings.append(dims_mapping[1])

        # Check batch dim mapping compatibility
        if not all(batch_dim_mappings[0] == dim_mapping
                   for dim_mapping in batch_dim_mappings):
            return False

        return True
140

141
    def update_dims_mapping(self, dist_op):
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194
        changed = False
        op_desc = dist_op.serial_op.desc
        op_dist_attr = dist_op.dist_attr
        # The following statement will be replaced by a more elegent way
        if op_desc.type() == "shape" or op_desc.type() == "slice":
            return False
        output_names = op_desc.output_names()
        xshape_arg_names = []
        if "XShape" in output_names:
            xshape_arg_names = op_desc.output("XShape")
        batch_dim_mappings = []
        for arg_name in op_desc.input_arg_names():
            serial_tensor = dist_op.get_serial_input(arg_name)
            if serial_tensor.is_parameter:
                continue
            dims_mapping = op_dist_attr.get_input_dims_mapping(arg_name)
            batch_dim_mappings.append(dims_mapping[0])
        for arg_name in op_desc.output_arg_names():
            serial_tensor = dist_op.get_serial_output(arg_name)
            if serial_tensor.is_parameter:
                continue
            dims_mapping = op_dist_attr.get_output_dims_mapping(arg_name)
            if arg_name not in xshape_arg_names:
                batch_dim_mappings.append(dims_mapping[0])
            else:
                batch_dim_mappings.append(dims_mapping[1])

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

        return changed
195 196 197 198

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

199
        dist_op_context = ctx.dist_op_context
200 201 202 203
        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
204

205
        # check validation of inputs / outputs
206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222
        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)

        # replicate op in dist program
        dist_op_desc = main_block.desc.append_op()
        dist_op_desc.copy_from(src_op.desc)
223
        set_dist_op_desc_original_id(dist_op_desc, src_op.desc, ctx)
224 225 226 227 228 229 230 231
        for input_name in src_op.desc.input_names():
            dist_op_desc.set_input(input_name, kwargs[input_name])
        for output_name in src_op.desc.output_names():
            dist_op_desc.set_output(output_name, kwargs[output_name])

        main_block._sync_with_cpp()

        # param initialization sync
232 233 234
        if src_op.type in __op_not_need_param_init__:
            return

235 236 237
        for varname in dist_op_desc.input_arg_names():
            if startup_block.has_var(varname) and startup_block.var(
                    varname
238 239
            ).is_parameter and varname not in dist_op_context.already_init_sync_vars:
                dist_op_context.already_init_sync_vars.add(varname)
240
                param = startup_block.var(varname)
241 242 243
                param_dist_attr = ctx.get_tensor_dist_attr_for_program(param)
                process_mesh = param_dist_attr.process_mesh
                dims_mapping = param_dist_attr.dims_mapping
244 245

                # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
246 247 248
                if rank_id not in process_mesh.processes:
                    rank_id = _get_corresponding_rank(ctx, process_mesh,
                                                      rank_id)
249

250
                # NOTE all not splited axis should be presented in mesh
251 252 253 254
                for axis, size in enumerate(process_mesh.topology):
                    if size <= 1 or axis in dims_mapping:
                        pass
                    else:
255 256 257
                        group_ranks = _get_comm_group(process_mesh.processes,
                                                      process_mesh.topology,
                                                      axis, rank_id)
258 259 260 261 262 263 264 265 266 267 268 269 270 271
                        sync_group = new_process_group(group_ranks)

                        new_op = 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
                            })

                        # set distributed attribute
272 273
                        op_attr = OperatorDistributedAttribute()
                        op_attr.process_mesh = process_mesh
274 275 276
                        op_attr.set_output_dims_mapping(param.name,
                                                        dims_mapping)
                        op_attr.set_input_dims_mapping(param.name, dims_mapping)
277
                        ctx.set_op_dist_attr_for_program(new_op, op_attr)
278 279 280 281 282 283 284

                startup_block._sync_with_cpp()

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

        # by now the backward function only insert the gradient allreduce for dist op itself
285
        dist_op_context = ctx.dist_op_context
286 287
        main_block = dist_op_context.work_block
        backward_op = dist_op_context.cur_src_op
288
        dist_attr = ctx.get_op_dist_attr_for_program(backward_op)
289 290
        assert dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
            str(backward_op))
291
        rank_id = dist_op_context.rank_id
292

293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310
        # check validation of inputs / outputs
        for input_name in backward_op.desc.input_names():
            assert input_name in kwargs, "input [{}] is not given".format(
                input_name)
            assert len(kwargs[input_name]) == len(
                backward_op.desc.input(input_name)
            ), "number of tensor for input [{}] is not match".format(input_name)
        for output_name in backward_op.desc.output_names():
            assert output_name in kwargs, "input [{}] is not given".format(
                output_name)
            assert len(kwargs[output_name]) == len(
                backward_op.desc.output(output_name)
            ), "number of tensor for input [{}] is not match".format(
                output_name)

        # replicate op in dist program
        dist_op_desc = main_block.desc.append_op()
        dist_op_desc.copy_from(backward_op.desc)
311 312
        # Refer to the related dist op
        set_dist_op_desc_original_id(dist_op_desc, backward_op.desc, ctx)
313 314 315 316 317 318 319
        for input_name in backward_op.desc.input_names():
            dist_op_desc.set_input(input_name, kwargs[input_name])
        for output_name in backward_op.desc.output_names():
            dist_op_desc.set_output(output_name, kwargs[output_name])

        main_block._sync_with_cpp()

320
        # check if need gradient allreduce
321
        # if there is a non-gradient & non-parameter input and its batch dimension is splited,
322 323 324 325
        # we need insert gradient allreduce for the gradient of parameter in its output
        need_gradient_allreduce = False
        for input_name in backward_op.desc.input_names():
            for varname in backward_op.desc.input(input_name):
J
JZ-LIANG 已提交
326 327
                if "@GRAD" not in varname and not is_parameter_related(
                        varname, main_block):
328 329

                    # NOTE input var's dim_mapping of backward op should be the same with input var instead of corresponding varname of forward op
330
                    process_mesh = dist_attr.process_mesh
331 332 333
                    var_dim_mapping = dist_attr.get_input_dims_mapping(varname)

                    # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism
334 335 336
                    if rank_id not in process_mesh.processes:
                        rank_id = _get_corresponding_rank(ctx, process_mesh,
                                                          rank_id)
337 338 339 340 341

                    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
342 343 344
                        group_ranks = _get_comm_group(process_mesh.processes,
                                                      process_mesh.topology,
                                                      batch_size_axis, rank_id)
345 346 347 348 349 350 351 352
                        dp_degree = len(group_ranks)
                        dp_group = new_process_group(group_ranks)
                        break

        if need_gradient_allreduce:
            allreduce_vars = []
            for input_name in backward_op.desc.input_names():
                for varname in backward_op.desc.input(input_name):
J
JZ-LIANG 已提交
353 354
                    if "@GRAD" not in varname and is_parameter_related(
                            varname, main_block):
355 356 357 358 359 360
                        # NOTE: When amp and recompute pass are effective at the same time,
                        # if a parameter is casted and recomputed, the 'parameter@GARD' can not
                        # be found in the grad_op's output.
                        if "subprog_" in varname:
                            varname = varname[:varname.index(".subprog_")]

361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399
                        assert len(
                            backward_op.desc.input(input_name)
                        ) == 1, "parameter input to grad op should be length 1, but got [{}]".format(
                            backward_op.desc.input(input_name))

                        assert varname + "@GRAD" in backward_op.desc.output_arg_names(
                        ), "parameter's grad [{}] not found in the grad op's output".format(
                            varname + "@GRAD")
                        assert len(
                            backward_op.desc.output(input_name + "@GRAD")
                        ) == 1, "parameter grad of grad op should be length 1, but got [{}]".format(
                            backward_op.desc.output(input_name + "@GRAD"))
                        allreduce_vars.append(
                            backward_op.desc.output(input_name + "@GRAD")[0])

            if len(allreduce_vars) > 0:

                for varname in allreduce_vars:

                    grad_var = main_block.var(varname)
                    allreduce_op = main_block.append_op(
                        type='c_allreduce_sum',
                        inputs={'X': [grad_var]},
                        outputs={'Out': [grad_var]},
                        attrs={
                            'ring_id': dp_group.id,
                            'use_calc_stream': True,
                            OP_ROLE_KEY: OpRole.Backward
                        })

                    scale_op = main_block.append_op(
                        type='scale',
                        inputs={'X': grad_var},
                        outputs={'Out': grad_var},
                        attrs={
                            'scale': 1.0 / dp_degree,
                            OP_ROLE_KEY: OpRole.Backward
                        })

400 401 402
                    dims_mapping = ctx.get_tensor_dist_attr_for_program(
                        grad_var).dims_mapping
                    process_mesh = dist_attr.process_mesh
403
                    for op in [allreduce_op, scale_op]:
404 405
                        op_attr = OperatorDistributedAttribute()
                        op_attr.process_mesh = process_mesh
406 407 408 409
                        op_attr.set_output_dims_mapping(grad_var.name,
                                                        dims_mapping)
                        op_attr.set_input_dims_mapping(grad_var.name,
                                                       dims_mapping)
410
                        ctx.set_op_dist_attr_for_program(op, op_attr)
411 412 413 414 415 416

                main_block._sync_with_cpp()


register_distributed_operator_impl(
    "default", DistributedDefaultImpl0("replicate_parallel"))