auto_parallel_recompute.py 18.3 KB
Newer Older
1
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
2
#
3 4 5
# 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
6
#
7
#     http://www.apache.org/licenses/LICENSE-2.0
8
#
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import copy
import logging

from .pass_base import PassBase, register_pass
from paddle.fluid import core, unique_name
from paddle.fluid import framework as framework
from paddle.fluid.framework import Variable, Operator
from paddle.fluid.backward import _append_grad_suffix_, _get_no_grad_set_name
from paddle.fluid.backward import ProgramStats, _rename_arg_, _find_op_path_
from paddle.distributed.auto_parallel.process_mesh import ProcessMesh
from paddle.distributed.auto_parallel.dist_attribute import OperatorDistributedAttribute
from paddle.distributed.auto_parallel.utils import get_loss_op, set_var_dist_attr, set_dist_op_desc_original_id
from paddle.distributed.auto_parallel.utils import naive_set_dist_op_attr_for_program_by_mesh_and_mapping


class RecomputeState(ProgramStats):
31

32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
    def __init__(self, block, ops):
        super(RecomputeState, self).__init__(block=block, ops=ops)
        self._block = block
        self._ops = ops
        self.var_op_deps = {}

    def build_stats(self):
        for i, op in enumerate(self._ops):
            for name in op.desc.input_arg_names():
                if name in self.var_op_deps:
                    self.var_op_deps[name]["var_as_input_ops"].extend([i])
                else:
                    self.var_op_deps[name] = {}
                    self.var_op_deps[name]["var_as_input_ops"] = [i]
                    self.var_op_deps[name]["var_as_output_ops"] = []

            for name in op.desc.output_arg_names():
                if name in self.var_op_deps:
                    self.var_op_deps[name]["var_as_output_ops"].extend([i])
                else:
                    self.var_op_deps[name] = {}
                    self.var_op_deps[name]["var_as_input_ops"] = []
                    self.var_op_deps[name]["var_as_output_ops"] = [i]

    def get_recompute_segments(self, checkpoints):
        """ get recompute segments from checkpoints """
        segments = []
        start_idx = -1
        pre_segment_end_idx = -1
        while start_idx + 1 < len(checkpoints):
            if start_idx == -1:
                ckpt_name = checkpoints[start_idx + 1]
                if ckpt_name not in self.var_op_deps:
                    start_idx += 1
                    continue
                op_idx_list = self.var_op_deps[ckpt_name]["var_as_output_ops"]
                if op_idx_list:
                    segments.append([0, max(op_idx_list) + 1])
            else:
                flag, min_idx, max_idx = self.is_subgraph(
                    [checkpoints[start_idx]], [checkpoints[start_idx + 1]])
                if flag:
74 75
                    min_idx = self._update_segment_start(
                        min_idx, pre_segment_end_idx)
76 77
                    segments.append([min_idx, max_idx + 1])
                else:
78 79 80
                    logging.info(
                        "Could not recompute op range [{}] - [{}] ".format(
                            min_idx, max_idx + 1))
81 82 83 84
            start_idx += 1

        for i, (idx1, idx2) in enumerate(segments):
            logging.info("recompute segment[{}]".format(i))
85 86 87 88 89 90 91 92
            logging.info("segment start op: [{}]: [{}] [{}]".format(
                self._ops[idx1].desc.type(),
                self._ops[idx1].desc.input_arg_names(),
                self._ops[idx1].desc.output_arg_names()))
            logging.info("segment end op: [{}]: [{}] [{}]".format(
                self._ops[idx2 - 1].desc.type(),
                self._ops[idx2 - 1].desc.input_arg_names(),
                self._ops[idx2 - 1].desc.output_arg_names()))
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

        return segments

    def modify_forward_desc_for_recompute(self, dist_context):
        """
        If program's foward part has 'dropout' op, this function will insert 
        a seed op before it to guarantee that two dropout op have the same outputs.
        """
        op_types = [op.desc.type() for op in self._ops]
        if "dropout" not in op_types:
            return

        op_idx = 0
        while op_idx < len(self._ops):
            cur_op = self._ops[op_idx]
            if "grad" in cur_op.type:
                break
            if cur_op.type != "dropout":
                op_idx += 1
                continue
            if cur_op.input("Seed") is not None and len(cur_op.input("Seed")):
                op_idx += 1
                continue

            cur_op_dist_attr = dist_context.get_op_dist_attr_for_program(cur_op)
            # insert seed op to guarantee that two dropout op have the same outputs
            op_unique_name = unique_name.generate("seed")
            var_unique_name = unique_name.generate_with_ignorable_key(".".join(
                [op_unique_name, 'tmp']))
            seed_var = self._block.create_var(
                name=var_unique_name,
                dtype='int32',
                type=core.VarDesc.VarType.LOD_TENSOR,
                persistable=False,
                stop_gradient=False)

            # set new seed_var's dist_attr
            ref_dims_mapping = [-1]
            ref_process_mesh = cur_op_dist_attr.process_mesh
132 133 134
            seed_var_dist_attr = set_var_dist_attr(dist_context, seed_var,
                                                   ref_dims_mapping,
                                                   ref_process_mesh)
135 136 137 138 139 140 141 142

            seed = 0 if cur_op.attr("fix_seed") is False else int(
                cur_op.attr("seed"))
            seed_op = self._block._insert_op_without_sync(
                index=cur_op.idx,
                type="seed",
                inputs={},
                outputs={"Out": seed_var},
143 144 145 146
                attrs={
                    "seed": seed,
                    "force_cpu": True
                })
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
            # set new seed op's dist_attr
            naive_set_dist_op_attr_for_program_by_mesh_and_mapping(
                seed_op, ref_process_mesh, ref_dims_mapping, dist_context)

            # modify dropout op's desc
            self._ops.insert(op_idx, seed_op)
            cur_op.desc.set_input("Seed", [var_unique_name])
            cur_op.desc.remove_attr("fix_seed")
            cur_op.desc.remove_attr("seed")
            cur_op_dist_attr.set_input_dist_attr(seed_var.name,
                                                 seed_var_dist_attr)
            self._block._sync_with_cpp()
            op_idx += 2


def _find_op_index(block, cur_op):
    for idx in range(block.desc.op_size()):
        if cur_op.desc == block.desc.op(idx):
            return idx
    return -1


def _get_stop_gradients(program, no_grad_set):
    """ get no grad var """
    if no_grad_set is None:
        no_grad_set = set()
    else:
        no_grad_set = _get_no_grad_set_name(no_grad_set)

    no_grad_set_name = set()
    for var in program.list_vars():
        assert isinstance(var, Variable)
        if "@GRAD" in var.name:
            break
        if var.stop_gradient:
            no_grad_set_name.add(_append_grad_suffix_(var.name))
    no_grad_set_name.update(list(map(_append_grad_suffix_, no_grad_set)))
    return no_grad_set_name


def _add_needed_descs_to_block(descs, block, main_block, in_memory_vars,
                               dist_context):
    """
    Get the recomputed ops which will insert the backward part
    """
    if len(descs) == 0:
        return []
    result_descs = []
    op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
    backward = core.op_proto_and_checker_maker.OpRole.Backward
    for desc in descs:
        if isinstance(desc, framework.Operator):
            desc = desc.desc
        if isinstance(desc, tuple):
            desc = desc[0]
        is_needed = False
        for name in desc.output_arg_names():
            if main_block.has_var(name) and main_block.var(name).persistable:
                continue
            if name not in in_memory_vars:
                is_needed = True
        if is_needed:
            new_op_desc = block.desc.append_op()
            new_op_desc.copy_from(desc)
            set_dist_op_desc_original_id(new_op_desc, desc, dist_context)
            new_op_desc._set_attr(op_role_attr_name, backward)
            result_descs.append(new_op_desc)
    return result_descs


@register_pass("auto_parallel_recompute")
class RecomputePass(PassBase):
219

220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
    def __init__(self):
        super(RecomputePass, self).__init__()
        self.set_attr("checkpoints", None)
        self.set_attr("loss", None)
        self.set_attr("dist_context", None)
        self.set_attr("no_grad_set", None)

    def _check_self(self):
        if self.get_attr("dist_context") is None:
            return False
        if self.get_attr("loss") is None:
            return False
        if self.get_attr("checkpoints") is None:
            return False
        return True

    def _check_conflict(self, other_pass):
        return True

    def _apply_single_impl(self, main_programs, startup_programs, context):
        checkpoints = self.get_attr("checkpoints")
        loss = self.get_attr("loss")
        no_grad_set = self.get_attr("no_grad_set")
        self._dist_context = self.get_attr("dist_context")

        main_block = main_programs.global_block()
        no_grad_set_name = _get_stop_gradients(main_programs, no_grad_set)
        # get op_path which is related to loss
        op_path = _find_op_path_(main_block, [loss], [], no_grad_set_name)

        # step 1: build recompute state
        rc_state = RecomputeState(main_block, op_path)
        rc_state.modify_forward_desc_for_recompute(self._dist_context)
        rc_state.build_stats()
        checkpoints = rc_state.sort_checkpoints(checkpoints)
        segments = rc_state.get_recompute_segments(checkpoints)
        if segments == []:
            return

        # step 2: get vars_should_be_hold
        vars_should_be_hold = []
        for segment in segments:
            vars_should_be_hold.extend(
                rc_state.get_out_of_subgraph_vars(segment[0], segment[1]))
        cross_vars = set(vars_should_be_hold) - set(checkpoints)
265 266 267 268
        logging.info(
            "found [{}] vars which cross recompute segment: [{}],"
            "better checkpoints might be set to reduce those vars".format(
                len(cross_vars), cross_vars))
269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315
        vars_should_be_hold.extend(rc_state.get_reserved_vars())
        vars_should_be_hold.extend(rc_state.get_input_nodes())
        vars_should_be_hold = list(set(vars_should_be_hold))
        vars_in_memory = vars_should_be_hold + checkpoints

        # step 3: get recomputed fwd ops desc
        var_name_dict = {}
        ckpt_ops_dict = {}
        buffer_block = main_block.program._create_block()
        for i, segment in enumerate(segments[::-1]):
            fwd_ops = op_path[segment[0]:segment[1]]
            var_suffix = ".subprog_%d" % i
            for op in fwd_ops:
                input_and_output_names = []
                input_and_output_names.extend(op.desc.input_arg_names())
                input_and_output_names.extend(op.desc.output_arg_names())
                cur_op_dist_attr = self._dist_context.get_op_dist_attr_for_program(
                    op)
                assert cur_op_dist_attr is not None
                for name in input_and_output_names:
                    if main_block.var(name).persistable or name in checkpoints:
                        continue
                    if name in vars_should_be_hold:
                        continue
                    if name not in var_name_dict:
                        ref_process_mesh = cur_op_dist_attr.process_mesh
                        if name in op.desc.input_arg_names():
                            ref_dims_mapping = cur_op_dist_attr.get_input_dims_mapping(
                                name)
                        else:
                            ref_dims_mapping = cur_op_dist_attr.get_output_dims_mapping(
                                name)
                        # record recomputed var's old_name and new_name (old_name.subprog_XXX)
                        # create new var with new name
                        var_name_dict[name] = name + var_suffix
                        ref_var = main_block.var(name)
                        rc_var = main_block.create_var(
                            name=var_name_dict[name],
                            shape=ref_var.shape,
                            dtype=ref_var.dtype,
                            type=ref_var.type,
                            persistable=ref_var.persistable,
                            stop_gradient=ref_var.stop_gradient)
                        # set new recomputed var's dist attr
                        set_var_dist_attr(self._dist_context, rc_var,
                                          ref_dims_mapping, ref_process_mesh)
            # get recomputed segment's descs
316 317 318 319
            segment_descs = _add_needed_descs_to_block(fwd_ops, buffer_block,
                                                       main_block,
                                                       vars_in_memory,
                                                       self._dist_context)
320 321 322 323 324
            # rename recomputed ops' input and output var name
            for key in var_name_dict:
                _rename_arg_(segment_descs, key, var_name_dict[key])

            # NOTE: one forward op could be correspond to multiple xxx_grad op.
325
            # When traversing all grad_ops in reverse, need to set a flag to indicate
326 327
            # whether the ckpt and its segment_descs can be used.
            ckpt_op = op_path[segment[1] - 1]
328
            ckpt_ops_dict[ckpt_op.desc.original_id()] = [True, segment_descs]
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351

        # step 4: insert recomputed fwd ops
        ops = main_block.ops
        loss_op = get_loss_op(main_block)
        loss_op_idx = _find_op_index(main_block, loss_op)
        dist_op_context = self._dist_context.dist_op_context
        assert loss_op_idx != -1
        # Traversing all grad_ops in reverse, and if the fwd op corresponding to reverse op is checkpoints,
        # segments ops should be inserted.
        for i in range(len(ops) - 1, loss_op_idx, -1):
            grad_op = ops[i]
            # remove some attrs of dropout_grad op's desc
            if grad_op.type == "dropout_grad":
                grad_op.desc.remove_attr("fix_seed")
                grad_op.desc.remove_attr("seed")
                main_block._sync_with_cpp()

            # rename grad op's var_name which is not in 'vars_in_memory'
            for key in var_name_dict:
                self.reset_op_dist_attr(grad_op, var_name_dict)
                _rename_arg_([grad_op.desc], key, var_name_dict[key])

            # insert recomputed ops
352 353 354
            original_id = grad_op.desc.original_id()
            if original_id in dist_op_context.grad_op_id_to_op_id:
                fwd_op_id = dist_op_context.grad_op_id_to_op_id[original_id]
355 356 357 358 359 360 361 362
                if fwd_op_id in ckpt_ops_dict and ckpt_ops_dict[fwd_op_id][0]:
                    idx = grad_op.idx
                    while idx - 1 >= 0 and ops[idx - 1].type == "sum":
                        idx -= 1
                    segment_descs = ckpt_ops_dict[fwd_op_id][1]
                    for _, op_desc in reversed(list(enumerate(segment_descs))):
                        rc_desc = main_block.desc._insert_op(idx)
                        rc_desc.copy_from(op_desc)
363
                        rc_desc.set_original_id(rc_desc.id())
364 365 366 367
                        rc_op = Operator(main_block, rc_desc)
                        main_block.ops.insert(idx, rc_op)
                        # set recomputed ops' dist attr
                        fwd_op_dist_attr = self._dist_context.get_op_dist_attr_for_program_with_id(
368
                            op_desc.original_id())
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
                        assert fwd_op_dist_attr is not None
                        self.set_op_dist_attr(rc_op, fwd_op_dist_attr,
                                              var_name_dict)

                    ckpt_ops_dict[fwd_op_id][0] = False
                    main_block._sync_with_cpp()

        main_programs._sync_with_cpp()

    def reset_op_dist_attr(self, op, var_name_dict):
        op_dist_attr = self._dist_context.get_op_dist_attr_for_program(op)
        assert op_dist_attr is not None
        for input in op.desc.input_arg_names():
            if input in var_name_dict.keys():
                in_dist_attr = op_dist_attr.get_input_dist_attr(input)
                op_dist_attr.set_input_dist_attr(var_name_dict[input],
                                                 in_dist_attr)
        for output in op.desc.output_arg_names():
            if output in var_name_dict.keys():
                out_dist_attr = op_dist_attr.get_output_dist_attr(output)
                op_dist_attr.set_output_dist_attr(var_name_dict[output],
                                                  out_dist_attr)

    def set_op_dist_attr(self, op, old_dist_attr, var_name_dict):
        new_dist_attr = OperatorDistributedAttribute()
        new_dist_attr.is_recompute = True
        new_dist_attr.impl_idx = old_dist_attr.impl_idx
Z
zhaoyingli 已提交
396
        new_dist_attr.impl_type = old_dist_attr.impl_type
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414
        new_dist_attr.process_mesh = old_dist_attr.process_mesh
        for input in old_dist_attr.inputs_dist_attrs.keys():
            if input in var_name_dict.keys():
                in_dist_attr = old_dist_attr.inputs_dist_attrs[input]
                new_dist_attr.set_input_dist_attr(var_name_dict[input],
                                                  in_dist_attr)
            else:
                in_dist_attr = old_dist_attr.inputs_dist_attrs[input]
                new_dist_attr.set_input_dist_attr(input, in_dist_attr)
        for output in old_dist_attr.outputs_dist_attrs.keys():
            if output in var_name_dict.keys():
                out_dist_attr = old_dist_attr.outputs_dist_attrs[output]
                new_dist_attr.set_output_dist_attr(var_name_dict[output],
                                                   out_dist_attr)
            else:
                out_dist_attr = old_dist_attr.outputs_dist_attrs[output]
                new_dist_attr.set_output_dist_attr(output, out_dist_attr)
        self._dist_context.set_op_dist_attr_for_program(op, new_dist_attr)