# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import sys import json import shlex import copy import pathlib import subprocess import logging import pickle import time import paddle import paddle.fluid.core as core from paddle.fluid import program_guard from paddle.fluid.backward import append_backward from paddle.distributed.utils.log_utils import get_logger from paddle.distributed.passes import new_pass, PassContext from .dist_context import DistributedContext from .dist_context import set_default_distributed_context from .completion import Completer from .partitioner import Partitioner from .process_group import get_all_process_groups from .process_group import get_process_group from .process_group import get_world_process_group from .process_group import _g_process_group_map, ProcessGroup from .utils import make_data_unshard from .utils import set_grad_var_shape from .utils import SerialProgramInfo from .reshard import Resharder from .cluster import Cluster from .mapper import mapping from .dist_op import DistributedOperator from .dist_tensor import DistributedTensor from .planner import Planner _logger = get_logger(logging.INFO) class AutoParallelizer: """ AutoParallelizer is the main controller class to do the auto parallel process. And the auto parallel process will be triggered in the wrapped parallelize function. To facilitate the auto parallelization, it will contain information about program, cluster and the related context. In this basic version, the program information will be retrevied from Fleet object, and the cluster information can be retrevied in the new created Cluster object, and the context information can be retrevied in the new created DistributedContext. """ def __init__(self, fleet): self._fleet = fleet self._optimizer = self._fleet.user_defined_optimizer self._dist_strategy = self._fleet._user_defined_strategy self._dist_context = DistributedContext() self._cluster = None self._cluster_topo_path = os.getenv("PADDLE_CLUSTER_TOPO_PATH", None) if self._cluster_topo_path is not None: self._cluster = Cluster() self._cluster.build_from_file(self._cluster_topo_path) # Prepare information for auto mapping self._rank_mapping_path = os.getenv("PADDLE_RANK_MAPPING_PATH", None) enable_auto_mapping_env = os.getenv("PADDLE_ENABLE_AUTO_MAPPING", None) if enable_auto_mapping_env is None: self._enable_auto_mapping = False else: self._enable_auto_mapping = True self._pass_context = PassContext() self._need_rank_mapping = os.getenv("PADDLE_NEED_RANK_MAPPING") self._need_rank_mapping = ( True if self._need_rank_mapping and self._need_rank_mapping.lower() == 'true' else False ) # self._pass_context = None def _remove_distributed_attrs(self, main_program): suffix = core.kAutoParallelSuffix() # distributed attributes for variable have been removed # in previous process. for block in main_program.blocks: for op in block.ops: for attr_name in op.attr_names: if suffix in attr_name: op._remove_attr(attr_name) def _apply_pre_optimization_passes( self, main_program, startup_program, loss, params_grads, no_grad_set ): # apply amp pass if self._dist_strategy.amp: config = copy.deepcopy(self._dist_strategy.amp_configs) config["dist_context"] = self._dist_context config["params_grads"] = params_grads config["loss"] = loss if config["use_pure_fp16"]: config["base_opt"] = self._optimizer auto_parallel_fp16_pass = new_pass("auto_parallel_fp16", config) auto_parallel_fp16_pass.apply( [main_program], [startup_program], self._pass_context ) else: auto_parallel_amp_pass = new_pass("auto_parallel_amp", config) auto_parallel_amp_pass.apply( [main_program], [startup_program], self._pass_context ) # apply recompute pass if self._dist_strategy.recompute: config = copy.deepcopy(self._dist_strategy.recompute_configs) config["dist_context"] = self._dist_context config["no_grad_set"] = copy.deepcopy(no_grad_set) config["loss"] = loss auto_parallel_recompute_pass = new_pass( "auto_parallel_recompute", config ) auto_parallel_recompute_pass.apply( [main_program], [startup_program], self._pass_context ) def _generate_backward( self, main_program, startup_program, loss, parameter_list, no_grad_set, callbacks, ): with program_guard(main_program, startup_program): params_grads = append_backward( loss, parameter_list, no_grad_set, callbacks, distop_context=self._dist_context.dist_op_context, ) self._completer = Completer(self._dist_context) self._completer.complete_backward_annotation(main_program) self._dist_context.block_state.parse_backward_blocks(main_program) return params_grads def _apply_optimize(self, main_program, startup_program, params_grads): optimizer = copy.deepcopy(self._optimizer) with program_guard(main_program, startup_program): optimize_ops = optimizer.apply_gradients(params_grads) self._dist_context._serial_optimizer = optimizer # update completion self._completer = Completer(self._dist_context) self._completer.complete_update_annotation(main_program) return optimize_ops def _apply_post_optimization_passes( self, main_program, startup_program, rank, params_grads ): if self._dist_strategy.sharding: config = copy.deepcopy(self._dist_strategy.sharding_configs) config["dist_context"] = self._dist_context config["params_grads"] = params_grads config["global_rank"] = rank auto_parallel_sharding_pass = new_pass( "auto_parallel_sharding", config ) auto_parallel_sharding_pass.apply( [main_program], [startup_program], self._pass_context ) params_grads = self._pass_context.get_attr("params_grads") config = copy.deepcopy(self._dist_strategy.sharding_configs) config["dist_context"] = self._dist_context config["params_grads"] = params_grads config["rank_id"] = rank auto_parallel_clip_pass = new_pass("auto_parallel_grad_clip", config) auto_parallel_clip_pass.apply( [main_program], [startup_program], self._pass_context ) if self._dist_strategy.gradient_merge: config = copy.deepcopy(self._dist_strategy.gradient_merge_configs) config["dist_context"] = self._dist_context config["params_grads"] = params_grads auto_parallel_gradient_merge_pass = new_pass( "auto_parallel_gradient_merge_pass", config ) auto_parallel_gradient_merge_pass.apply( [main_program], [startup_program], self._pass_context ) def _get_dist_program(self, rank, dist_context=None, relaunch_phase=False): completed_main_program = None serial_main_program = self._main_program.clone() serial_startup_program = self._startup_program.clone() serial_loss = serial_main_program.global_block().var(self._loss.name) # generating serial if dist_context is None: # Annotation completion self._dist_context = DistributedContext() _logger.info("Start annotation dist attr.") self._completer = Completer(self._dist_context) completed_main_program = ( self._completer.complete_forward_annotation(serial_main_program) ) else: completed_main_program = serial_main_program self._dist_context = copy.deepcopy(dist_context) # parse forward sub block self._dist_context.block_state.parse_forward_blocks(serial_main_program) # serial backward pass params_grads = self._generate_backward( completed_main_program, serial_startup_program, serial_loss, self._parameter_list, self._no_grad_set, self._callbacks, ) # serial forward pass self._apply_pre_optimization_passes( completed_main_program, serial_startup_program, serial_loss, params_grads, self._no_grad_set, ) # Logical partition partitioner = Partitioner(self._dist_context, rank) ( dist_main_prog, dist_startup_prog, dist_params_grads, ) = partitioner.partition( completed_main_program, serial_startup_program, params_grads ) # TODO refactor the placement of optimizer # generate optimize program dist_optimize_ops = self._apply_optimize( dist_main_prog, dist_startup_prog, dist_params_grads ) set_grad_var_shape(dist_main_prog, self._dist_context) make_data_unshard(dist_main_prog, dist_startup_prog, self._dist_context) resharder = Resharder( dist_main_prog, dist_startup_prog, rank, self._dist_context, dist_params_grads, ) resharder.reshard() self._apply_post_optimization_passes( dist_main_prog, dist_startup_prog, rank, dist_params_grads ) g_process_group_map = None if not relaunch_phase: g_process_group_map = copy.deepcopy(_g_process_group_map) _g_process_group_map.clear() _g_process_group_map[0] = ProcessGroup(1000, []) for process_mesh in self._dist_context._process_meshes: _g_process_group_map[0].add_ranks(process_mesh.processes) return ( dist_optimize_ops, dist_params_grads, dist_startup_prog, dist_main_prog, g_process_group_map, ) def parallelize( self, loss, startup_program, parameter_list=None, no_grad_set=None, callbacks=None, ): assert startup_program is not None self._loss = loss self._startup_program = startup_program self._main_program = loss.block.program self._parameter_list = parameter_list self._no_grad_set = no_grad_set self._callbacks = callbacks if self._enable_auto_mapping and self._need_rank_mapping: # Do the mapping pass before parallelization assert ( self._cluster is not None ), "The cluster must not be none when using auto mapping." dist_programs = {} world_process_group = get_world_process_group() dist_context = None # auto search if self._dist_strategy.auto_search: logging.info("Start searching dist attr.") serial_program_info = SerialProgramInfo( self._main_program, self._startup_program, self._loss, self._optimizer, self._cluster, ) planner = Planner( serial_program_info, self, algorithm_config={"name": "mcmc", "max_search_times": 5}, ) dist_context, _ = planner.search() logging.info("End searching dist attr.") # serialize the dist context by planner if dist_context is not None: logging.info("Start serialize searched dist attr") cwd = pathlib.Path().resolve() searched_dist_context_path = os.path.join( cwd, f"searched_dist_context_{time.time()}.pkl" ) saved_dist_context = {} ops_dist_attr = {} tensors_dist_attr = {} for key, dist_op in dist_context._dist_ops_for_program.items(): ops_dist_attr[key] = dist_op.dist_attr for ( key, dist_tensor, ) in dist_context._dist_tensors_for_program.items(): tensors_dist_attr[key] = dist_tensor.dist_attr saved_dist_context["ops_dist_attr"] = ops_dist_attr saved_dist_context["tensors_dist_attr"] = tensors_dist_attr saved_dist_context[ "process_meshes" ] = dist_context._process_meshes with open( searched_dist_context_path, "wb" ) as dist_context_file: pickle.dump(saved_dist_context, dist_context_file) os.environ[ 'PADDLE_SEARCHED_DIST_CONTEXT_PATH' ] = searched_dist_context_path logging.info( f"End serialize searched dist attr to {searched_dist_context_path}" ) for rank in world_process_group.ranks: ( dist_optimize_ops, dist_params_grads, dist_startup_prog, dist_main_prog, g_process_group_map, ) = self._get_dist_program(rank, dist_context) dist_programs[rank] = [dist_main_prog, g_process_group_map] # Do the mapping between the distributed program graph and the cluster graph rank_mapping_dict = mapping(dist_programs, self._cluster) rank_mapping = list(rank_mapping_dict.values()) # Relaunch the training by using the rank mapping file with open(self._rank_mapping_path, "w") as rank_mapping_file: json.dump(rank_mapping, rank_mapping_file) enable_elastic = os.getenv("PADDLE_ENABLE_ELASTIC") enable_elastic = ( True if enable_elastic and enable_elastic.lower() == 'true' else False ) if enable_elastic: print("Auto mapping finished, now do elastic re-launch") sys.exit( paddle.distributed.fleet.elastic.manager.ELASTIC_AUTO_PARALLEL_EXIT_CODE ) original_cmd_args = os.getenv("PADDLE_ORIGINAL_CMD_ARGS") rank_mapping_args = " ".join( ["--rank_mapping_path", self._rank_mapping_path] ) if os.environ.get("WITH_COVERAGE", "OFF") == "ON": coverage_args = ["-m", "coverage", "run", "--branch", "-p"] else: coverage_args = [] new_cmd_args = ( "-m paddle.distributed.fleet.launch" + " " + rank_mapping_args + " " + original_cmd_args ) new_cmd = ( [sys.executable, "-u"] + coverage_args + shlex.split(new_cmd_args) ) new_process = subprocess.Popen(new_cmd) new_process.wait() assert ( new_process.returncode == 0 ), "Launch failed with rank mapping" print("Successfully do the second launch for auto mapping!") sys.exit(0) else: # Parallelization after the mapping pass rank = paddle.distributed.get_rank() dist_context = None searched_dist_context_path = os.getenv( "PADDLE_SEARCHED_DIST_CONTEXT_PATH", None ) if searched_dist_context_path is not None: with open( searched_dist_context_path, "rb" ) as dist_context_file: saved_dist_context = pickle.load(dist_context_file) dist_context = DistributedContext() for op in self._main_program.global_block().ops: dist_attr = saved_dist_context["ops_dist_attr"][ op.desc.id() ] dist_op = DistributedOperator(op, dist_attr) dist_context.add_dist_op_for_program(dist_op) vars = self._main_program.global_block().vars for var in vars.values(): dist_attr = saved_dist_context["tensors_dist_attr"][ var.desc.id() ] dist_tensor = DistributedTensor(var, dist_attr) dist_context.add_dist_tensor_for_program(dist_tensor) dist_context._process_meshes = saved_dist_context[ "process_meshes" ] else: if self._dist_strategy.auto_search: serial_program_info = SerialProgramInfo( self._main_program, self._startup_program, self._loss, self._optimizer, cluster=self._cluster, ) planner = Planner( serial_program_info, self, algorithm_config={ "name": "mcmc", "max_search_times": 5, }, ) dist_context, _ = planner.search() # rebuild g_process_group if dist_context is not None: pg0 = get_process_group(0) for process_mesh in dist_context._process_meshes: pg0.add_ranks(process_mesh.processes) ( dist_optimize_ops, dist_params_grads, dist_startup_prog, dist_main_prog, _, ) = self._get_dist_program(rank, dist_context, relaunch_phase=True) # NOTE: This is a trick to fix hang in pipeline mode when dist context is searched by planner if self._dist_strategy.auto_search: is_pipeline = False for op in dist_main_prog.global_block().ops: if op.type == "send_v2" or op.type == "recv_v2": is_pipeline = True break if is_pipeline: with paddle.static.program_guard(dist_main_prog): paddle.distributed.barrier(get_process_group(0)) # Traverse different rank programs and traverse each op of them, # instantiate communication by process_mapping. all_process_groups = get_all_process_groups() for process_group in all_process_groups: if len(_g_process_group_map) > 0: tmp = paddle.to_tensor([1], dtype="int32") paddle.distributed.all_reduce( tmp, sync_op=True, group=_g_process_group_map[0] ) paddle.device.cuda.synchronize() if rank not in process_group.ranks: continue process_group.instantiate() # Copy distributed info to the default context set_default_distributed_context(self._dist_context) # The last step: remove all distributed attributes to be compatible # with inference. self._remove_distributed_attrs(dist_main_prog) return ( dist_optimize_ops, dist_params_grads, dist_startup_prog, dist_main_prog, ) def __deepcopy__(self, memo): cls = self.__class__ result = cls.__new__(cls) memo[id(self)] = result for k, v in self.__dict__.items(): if ( k == "_main_program" or k == "_startup_program" or k == "_dist_context" or k == "_fleet" or k == "_loss" ): setattr(result, k, v) else: setattr(result, k, copy.deepcopy(v, memo)) return result