# Copyright (c) 2020 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. from __future__ import print_function import copy import warnings import paddle import os from types import MethodType import numpy as np from paddle.fluid.framework import dygraph_only, _global_flags from paddle.fluid import compiler from .role_maker import UserDefinedRoleMaker, PaddleCloudRoleMaker, RoleMakerBase from .strategy_compiler import StrategyCompiler from .distributed_strategy import DistributedStrategy from .meta_optimizer_factory import MetaOptimizerFactory from .runtime_factory import RuntimeFactory from paddle.fluid.wrapped_decorator import wrap_decorator from paddle.fluid.dygraph import parallel_helper from paddle.fluid.ir import apply_build_strategy from . import topology as tp from .topology import ParallelMode from ..meta_parallel import TensorParallel, model_parallel_random_seed from ..meta_parallel import PipelineParallel, ShardingParallel from ..meta_optimizers import HybridParallelOptimizer, HeterParallelOptimizer from paddle import _C_ops from paddle.fluid import core from paddle.fluid.dygraph import to_variable from paddle.distributed.fleet.utils.recompute import LegacyRecomputeFunction from paddle.fluid.dygraph.varbase_patch_methods import _grad_scalar __all__ = [] _grad_scalar = None class _RecomputeModelWrapper(paddle.nn.Layer): def __init__(self, model, segments=2, preserve_rng_state=True): super(_RecomputeModelWrapper, self).__init__() assert isinstance(model, paddle.nn.Sequential), ( "The model passed to RecomputeModelWrapper must be of type " "paddle.nn.Sequential.") self._model = model self._segments = segments self._preserve_rng_state = preserve_rng_state self._layers = list(model.children()) self._segment_size = len(self._layers) // segments def _run_func(self, begin, end): def do_run(input): for i in range(begin, end): input = self._layers[i](input) return input return do_run def _checkpoint(self, func, *args, **kwargs): return LegacyRecomputeFunction.apply(func, self._preserve_rng_state, *args) def forward(self, input): end = 0 for begin in range(0, self._segment_size * (self._segments - 1), self._segment_size): end = begin + self._segment_size input = self._checkpoint(self._run_func(begin, end), input) return self._run_func(end, len(self._layers))(input) def apply_ir_passes(main_program, startup_program, config): build_strategy = config._user_defined_strategy.build_strategy._copy() if not _global_flags()['FLAGS_apply_pass_to_program']: return build_strategy pipeline_opt = getattr(main_program, "_pipeline_opt", {}) if pipeline_opt: main_program = pipeline_opt["section_program"] startup_program = startup_program._pipeline_opt["startup_program"] pass_attrs = {"use_cuda": config._is_collective} fuse_all_reduce = config._user_defined_strategy.fuse_all_reduce_ops if fuse_all_reduce and build_strategy.fuse_all_optimizer_ops: # FIXME(zjl): currently, fuse_all_optimizer_ops # have conflict with fuse_all_reduce_ops because # RawProgramOptimizer also inserts coalesce_tensor # into program. These two procedures may conflict # in which vars are to be fused. warnings.warn( 'Currently, the fuse_all_optimizer_ops pass has conflict with fuse_all_reduce_ops pass. Disable the fuse_all_optimizer_ops pass temporarily.' ) build_strategy.fuse_all_optimizer_ops = False return apply_build_strategy(main_program, startup_program, build_strategy, pass_attrs) def _inited_runtime_handler_(func): def __impl__(*args, **kwargs): cls = args[0] if cls._runtime_handle is None: raise ValueError("Fleet can not find suitable runtime handler") return func(*args, **kwargs) return __impl__ def _is_non_distributed_check_(func): def __impl__(*args, **kwargs): cls = args[0] if cls._role_maker is not None and cls._role_maker._is_non_distributed( ) is True: warnings.warn( "%s() function doesn't work when use non_distributed fleet." % (func.__name__)) return return func(*args, **kwargs) return __impl__ inited_runtime_handler = wrap_decorator(_inited_runtime_handler_) is_non_distributed_check = wrap_decorator(_is_non_distributed_check_) class Fleet(object): """ Unified API for distributed training of PaddlePaddle Please reference the https://github.com/PaddlePaddle/FleetX for details Returns: Fleet: A Fleet instance Example for collective training: .. code-block:: python import paddle paddle.enable_static() import paddle.distributed.fleet as fleet fleet.init(is_collective=True) strategy = fleet.DistributedStrategy() optimizer = paddle.optimizer.SGD(learning_rate=0.001) optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy) # do distributed training Example for parameter server training: .. code-block:: python import paddle paddle.enable_static() import paddle.distributed.fleet as fleet strategy = fleet.DistributedStrategy() fleet.init(strategy=strategy) optimizer = paddle.optimizer.SGD(learning_rate=0.001) optimizer = fleet.distributed_optimizer(optimizer) if fleet.is_first_worker(): print("this is first worker") print("current node index: {}".format(fleet.worker_index())) print("total number of worker num: {}".format(fleet.worker_num())) if fleet.is_worker(): print("this is worker") print("worker endpoints: {}".format(fleet.worker_endpoints(to_string=True))) print("server num: {}".format(fleet.server_num())) print("server endpoints: {}".format(fleet.server_endpoints(to_string=True))) if fleet.is_server(): print("this is server") fleet.stop_worker() """ def __init__(self): self._role_maker = None self.strategy_compiler = None self._is_collective = False self._runtime_handle = None self._util = None self._context = {} def init(self, role_maker=None, is_collective=False, strategy=None): """ Initialize role_maker in Fleet. This function is responsible for the distributed architecture what you want to run your code behind. Args: role_maker (RoleMakerBase, optional): A ``RoleMakerBase`` containing the configuration of environment variables related to distributed training.If you did not initialize the rolemaker by yourself, it will be automatically initialized to PaddleRoleMaker. The default value is None. is_collective (Boolean, optional): A ``Boolean`` variable determines whether the program runs on the CPU or GPU. False means set distributed training using CPU, and True means GPU.The default value is False.The default value is False. strategy (DistributedStrategy): Extra properties for distributed training. For details, please refer to paddle.distributed.fleet.DistributedStrategy. Default: None. Returns: None Examples1: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() Examples2: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init(is_collective=True) Examples3: .. code-block:: python import paddle.distributed.fleet as fleet role = fleet.PaddleCloudRoleMaker() fleet.init(role) Examples4: .. code-block:: python import paddle.distributed.fleet as fleet strategy = fleet.DistributedStrategy() fleet.init(strategy=strategy) """ if strategy is None: strategy = DistributedStrategy() self._user_defined_strategy = copy.deepcopy(strategy) if role_maker is None: if isinstance(is_collective, bool): self._is_collective = is_collective self._role_maker = PaddleCloudRoleMaker( is_collective=self._is_collective) else: raise ValueError( "`is_collective` should be instance of `bool`, but got {}". format(type(is_collective))) else: if isinstance(role_maker, RoleMakerBase): self._role_maker = role_maker self._is_collective = role_maker._is_collective else: raise ValueError( "`role_maker` should be subclass of `RoleMakerBase`, but got {}" .format(type(role_maker))) self._role_maker._generate_role() import paddle.distributed.fleet as fleet fleet.util._set_role_maker(self._role_maker) self.strategy_compiler = StrategyCompiler() if self._role_maker._is_non_distributed() and self._is_collective: if paddle.fluid.core.is_compiled_with_cuda(): gpus_num = paddle.fluid.core.get_cuda_device_count() if gpus_num != 1: raise ValueError( "CUDA_VISIBLE_DEVICES shoule be set only 1 card if you use `python` to launch fleet program." ) if paddle.fluid.framework._non_static_mode(): if self.worker_num() == 1: # if worker_num is 1, should construct default topology & hcg self._topology = tp.CommunicateTopology() self._hcg = tp.HybridCommunicateGroup(self._topology) return if parallel_helper._is_parallel_ctx_initialized(): warnings.warn( "The dygraph parallel environment has been initialized.") else: # FLAGS_nccl_nrings is used for dynamic graph multi-stream communication if "FLAGS_nccl_nrings" in os.environ: warnings.warn( "You have set the environment variable FLAGS_nccl_nrings " "outside the program, so the nccl_comm_num in " "DistributedStrategy will not take effect here.") else: os.environ["FLAGS_nccl_nrings"] = str( self._user_defined_strategy.nccl_comm_num) paddle.distributed.init_parallel_env() # hybrid parallel not support for npu/xpu if self._user_defined_strategy.heter_ccl_mode == False: # init hybrid parallel environment in dygraph if tp._HYBRID_PARALLEL_GROUP is None: self._init_hybrid_parallel_env() else: warnings.warn( "The dygraph hybrid parallel environment has been initialized." ) elif self._is_collective: use_sharding = self._user_defined_strategy.sharding # global group global_rank = self.worker_index() global_world_size = self.worker_num() # NOTE(wangxi): see sharding_optimizer global_ring_id = 3 if use_sharding else 0 global_ranks = list(range(global_world_size)) if tp._HYBRID_PARALLEL_GROUP is None: tp._CommunicateGroup() cg = tp._HYBRID_PARALLEL_GROUP self._hcg = cg cg.set_comm_group('global', global_rank, global_world_size, global_ring_id, global_ranks) use_tensor_parallel = self._user_defined_strategy.tensor_parallel use_mp = use_sharding or use_tensor_parallel # hybrid group if use_mp is False: return mp_degree_sharding = 1 mp_degree_tensor_parallel = 1 if use_sharding: sharding_configs = self._user_defined_strategy.sharding_configs mp_degree_sharding = int(sharding_configs['mp_degree']) if use_tensor_parallel: tensor_parallel_configs = self._user_defined_strategy.tensor_parallel_configs mp_degree_tensor_parallel = int( tensor_parallel_configs['tensor_parallel_degree']) if use_sharding and use_tensor_parallel: assert mp_degree_sharding == mp_degree_tensor_parallel mp_degree = mp_degree_sharding if use_sharding else mp_degree_tensor_parallel if mp_degree > 1: assert global_world_size % mp_degree == 0 # NOTE(wangxi): mp_ring_id sync with sharding_optimizer.py _build_groups mp_ring_id = 0 mp_rank = global_rank % mp_degree mp_group_id = global_rank // mp_degree mp_group_ranks = [ idx for idx in global_ranks if idx // mp_degree == mp_group_id ] cg.set_comm_group('model', mp_rank, mp_degree, mp_ring_id, mp_group_ranks) def _init_hybrid_parallel_env(self): """initialize the hybrid environment """ self.hybrid_configs = self._user_defined_strategy.hybrid_configs self.dp_degree = self.hybrid_configs["dp_degree"] self.mp_degree = self.hybrid_configs["mp_degree"] self.pp_degree = self.hybrid_configs["pp_degree"] self.sharding_degree = self.hybrid_configs["sharding_degree"] assert self.mp_degree >= 0, "mp_degree should be greater or equal to 0" assert self.pp_degree >= 0, "pp_degree should be greater or equal to 0" assert self.sharding_degree >= 0, "sharding_degree should be greater or equal to 0" self.mp_degree = max(self.mp_degree, 1) self.pp_degree = max(self.pp_degree, 1) if self.dp_degree < 0: nranks = paddle.distributed.get_world_size() self.dp_degree = nranks // (self.mp_degree * self.pp_degree) self.dp_degree = max(self.dp_degree, 1) self._topology = tp.CommunicateTopology( hybrid_group_names=["data", "pipe", "sharding", "model"], dims=[ self.dp_degree, self.pp_degree, self.sharding_degree, self.mp_degree ]) self._hcg = tp.HybridCommunicateGroup(self._topology) if self.mp_degree > 1: tensor_parallel_configs = self._user_defined_strategy.tensor_parallel_configs tensor_init_seed = tensor_parallel_configs["tensor_init_seed"] if tensor_init_seed == -1: model_parallel_random_seed() else: model_parallel_random_seed(tensor_init_seed) def get_hybrid_communicate_group(self): assert self._hcg is not None return self._hcg def get_hybrid_parallel_topology(self): assert self._topology is not None return self._topology def is_first_worker(self): """ Check whether the node is the first instance of worker. Returns: bool: True if this is the first node of worker, False if not. Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.is_first_worker() """ return self._role_maker._is_first_worker() def worker_index(self): """ Get current worker index. Returns: int: node id Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.worker_index() """ return self._role_maker._worker_index() def worker_num(self): """ Get current total worker number. Returns: int: worker numbers Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.worker_num() """ return self._role_maker._worker_num() def node_num(self): return self._role_maker._get_node_num() def local_rank(self): return self._role_maker._get_local_rank() def local_device_ids(self): return self._role_maker._get_local_device_ids() def world_device_ids(self): return self._role_maker._get_world_device_ids() def is_worker(self): """ Check whether the node is an instance of worker. Returns: bool: True if this is a node of worker, False if not. Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.is_worker() """ return self._role_maker._is_worker() def is_coordinator(self): return self._role_maker._is_coordinator() def worker_endpoints(self, to_string=False): """ Get current worker endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"]. Returns: list/string: server endpoints Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.worker_endpoints() """ if to_string: return ",".join(self._role_maker._get_trainer_endpoints()) else: return self._role_maker._get_trainer_endpoints() def server_num(self): """ Get current total worker number. Returns: int: server number Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.server_num() """ return len(self._role_maker._get_pserver_endpoints()) def server_index(self): """ Get current server index. Returns: int: node id Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.server_index() """ return self._role_maker._server_index() def server_endpoints(self, to_string=False): """ Get current server endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"]. Returns: list/string: server endpoints Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.server_endpoints() """ if to_string: return ",".join(self._role_maker._get_pserver_endpoints()) else: return self._role_maker._get_pserver_endpoints() def is_server(self): """ Check whether the node is an instance of server. Returns: bool: True if this is a node of server, False if not. Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() fleet.is_server() """ return self._role_maker._is_server() def barrier_worker(self): """ barrier all workers Returns: None """ self._role_maker._barrier("worker") @is_non_distributed_check @inited_runtime_handler def init_worker(self, scopes=None): """ initialize `Communicator` for parameter server training. Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.init_worker() """ self._runtime_handle._init_worker(scopes) @is_non_distributed_check @inited_runtime_handler def init_coordinator(self, scopes=None): """ initialize coordinator node """ self._runtime_handle._init_coordinator(scopes) def make_fl_strategy(self): self._runtime_handle._make_fl_strategy() @is_non_distributed_check @inited_runtime_handler def get_fl_client(self): """ get worker(training node) ptr """ return self._runtime_handle._worker @is_non_distributed_check @inited_runtime_handler def init_server(self, *args, **kwargs): """ init_server executor to initialize startup program, if the `args` is not empty, it will run load_persistables for increment training. Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.init_server() """ self._runtime_handle._init_server(*args, **kwargs) @is_non_distributed_check @inited_runtime_handler def load_model(self, path, mode): """ load fleet model from path Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.load_model("path", mode=0) """ self._runtime_handle._load_persistables(path, mode) @is_non_distributed_check @inited_runtime_handler def load_one_table(self, table_id, path, mode): """ load fleet one table from path Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.load_one_table(0, "path", mode=0) """ self._runtime_handle._load_one_table(table_id, path, mode) @is_non_distributed_check @inited_runtime_handler def load_inference_model(self, path, mode): """ load fleet inference model from path Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.load_inference_model("path", mode=1) """ self._runtime_handle._load_inference_model(path, mode) @is_non_distributed_check @inited_runtime_handler def run_server(self): """ run server will run pserver main program with executor. Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) if fleet.is_server(): fleet.init_server() """ self._runtime_handle._run_server() @is_non_distributed_check @inited_runtime_handler def stop_worker(self): """ stop `Communicator` and give training complete notice to parameter server. Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.init_server() """ self._runtime_handle._stop_worker() @is_non_distributed_check @inited_runtime_handler def save(self, dirname, feed=[], fetch=[], **configs): inference = True if not feed and not fetch: inference = False place = paddle.CPUPlace() executor = paddle.static.Executor(place) if inference: feeded_var_names = [] fetch_var_names = [] for var in feed: if isinstance(var, str): feeded_var_names.append(var) elif isinstance(var, paddle.static.Variable): feeded_var_names.append(var.name) else: raise ValueError("feed must be [str|Variable]") for var in fetch: if isinstance(var, str): fetch_var_names.append(var) elif isinstance(var, paddle.static.Variable): fetch_var_names.append(var.name) else: raise ValueError("feed must be [str|Variable]") fetch_vars = [ paddle.static.default_main_program().global_block().var(name) for name in fetch_var_names ] self._runtime_handle._save_inference_model(executor, dirname, feeded_var_names, fetch_vars, None, True, 0) else: increment_mode = 0 if "mode" in configs: increment_mode = int(configs["mode"]) self._runtime_handle._save_persistables(executor, dirname, main_program=None, mode=increment_mode) @is_non_distributed_check @inited_runtime_handler def save_inference_model(self, executor, dirname, feeded_var_names, target_vars, main_program=None, export_for_deployment=True, mode=0): """ save inference model for inference. Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.init_server() """ # warnings.warn( # "'save_inference_model' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead." # ) self._runtime_handle._save_inference_model(executor, dirname, feeded_var_names, target_vars, main_program, export_for_deployment, mode) @is_non_distributed_check @inited_runtime_handler def save_persistables(self, executor, dirname, main_program=None, mode=0): """ saves all persistable tensors from :code:`main_program` to the folder :code:`dirname`. You can refer to The :code:`dirname` is used to specify the folder where persistable tensors are going to be saved. If you would like to save tensors in separate files, set :code:`filename` None. Args: executor(Executor): The executor to run for saving persistable tensors. You can refer to :ref:`api_guide_executor_en` for more details. dirname(str, optional): The saving directory path. When you need to save the parameter to the memory, set it to None. main_program(Program, optional): The program whose persistbale tensors will be saved. Default: None. Returns: None Examples: .. code-block:: text import paddle paddle.enable_static() import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) exe = paddle.static.Executor(paddle.CPUPlace()) fleet.save_persistables(exe, "dirname", paddle.static.default_main_program()) """ # warnings.warn( # "'save_persistables' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead." # ) self._runtime_handle._save_persistables(executor, dirname, main_program, mode) @is_non_distributed_check @inited_runtime_handler def save_cache_model(self, dirname, **configs): return self._runtime_handle._save_cache_model(dirname, **configs) @is_non_distributed_check @inited_runtime_handler def check_save_pre_patch_done(self): return self._runtime_handle._check_save_pre_patch_done() @is_non_distributed_check @inited_runtime_handler def save_one_table(self, table_id, path, mode): """ save fleet one table from path Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() # build net # fleet.distributed_optimizer(...) fleet.save_one_table(0, "path", mode=0) """ self._runtime_handle._save_one_table(table_id, path, mode) @is_non_distributed_check @inited_runtime_handler def save_dense_params(self, executor, dirname, scope, program, var_names=None): """ save fleet one table from path Returns: None Examples: .. code-block:: python import paddle.distributed.fleet as fleet fleet.init() import paddle place = paddle.fluid.CPUPlace() exe = paddle.fluid.Executor(place) # build net # fleet.distributed_optimizer(...) fleet.save_dense_params(exe, "path", scope=paddle.static.global_scope(), program=paddle.static.default_main_program()) """ self._runtime_handle._save_dense_params(executor, dirname, scope, program, var_names) def shrink(self, threshold=None): self._runtime_handle._shrink(threshold) def distributed_optimizer(self, optimizer, strategy=None): """ Optimizer for distributed training. For the distributed training, this method would rebuild a new instance of DistributedOptimizer. Which has basic Optimizer function and special features for distributed training. Args: optimizer(Optimizer): The executor to run for init server. strategy(DistributedStrategy): Extra properties for distributed optimizer. It is recommended to use DistributedStrategy in fleet.init(). The strategy here is for compatibility. If the strategy in fleet.distributed_optimizer() is not None, then it will overwrite the DistributedStrategy in fleet.init(), which will take effect in distributed training. Returns: Fleet: instance of fleet. Examples: .. code-block:: python import paddle import paddle.distributed.fleet as fleet fleet.init(is_collective=True) strategy = fleet.DistributedStrategy() optimizer = paddle.optimizer.SGD(learning_rate=0.001) optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy) """ self.user_defined_optimizer = optimizer if strategy is not None: if self._is_collective: warnings.warn( "It is recommended to use DistributedStrategy " "in fleet.init(). The strategy here is only for compatibility. " "If the strategy in fleet.distributed_optimizer() is " "not None, then it will overwrite the DistributedStrategy in fleet.init(), " "which will take effect in distributed training.") self._user_defined_strategy = copy.deepcopy(strategy) self._context = {} if paddle.fluid.framework._non_static_mode(): if self.worker_num() > 1: if self._user_defined_strategy.heter_ccl_mode == False: return HybridParallelOptimizer(optimizer, self._hcg, self._user_defined_strategy) else: return HeterParallelOptimizer(optimizer, self._user_defined_strategy) else: return optimizer return self @dygraph_only def distributed_model(self, model): """ Return distributed data parallel model (Only work in dygraph mode) Args: model (Layer): the user-defind model which inherits Layer. Returns: distributed data parallel model which inherits Layer. Examples: .. code-block:: python import paddle import paddle.nn as nn from paddle.distributed import fleet class LinearNet(nn.Layer): def __init__(self): super(LinearNet, self).__init__() self._linear1 = nn.Linear(10, 10) self._linear2 = nn.Linear(10, 1) def forward(self, x): return self._linear2(self._linear1(x)) # 1. initialize fleet environment fleet.init(is_collective=True) # 2. create layer & optimizer layer = LinearNet() loss_fn = nn.MSELoss() adam = paddle.optimizer.Adam( learning_rate=0.001, parameters=layer.parameters()) # 3. get data_parallel model using fleet adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) # 4. run layer inputs = paddle.randn([10, 10], 'float32') outputs = dp_layer(inputs) labels = paddle.randn([10, 1], 'float32') loss = loss_fn(outputs, labels) print("loss:", loss.numpy()) loss.backward() adam.step() adam.clear_grad() """ assert model is not None, "model should not be None" if self.worker_num() <= 1: return model amp_enable = False recompute_enable = False strategy = self._user_defined_strategy if strategy.amp == True: amp_enable = True amp_level = "O2" if strategy.amp_configs['use_pure_fp16'] else "O1" if amp_level.upper() == "O2": model = paddle.amp.decorate(models=model, optimizers=None, level="O2", master_weight=None, save_dtype=None) init_loss_scaling = strategy.amp_configs['init_loss_scaling'] incr_ratio = strategy.amp_configs['incr_ratio'] decr_ratio = strategy.amp_configs['decr_ratio'] incr_every_n_steps = strategy.amp_configs['incr_every_n_steps'] decr_every_n_nan_or_inf = strategy.amp_configs[ 'decr_every_n_nan_or_inf'] use_dynamic_loss_scaling = strategy.amp_configs[ 'use_dynamic_loss_scaling'] global _grad_scalar _grad_scalar = paddle.amp.GradScaler( init_loss_scaling=init_loss_scaling, incr_ratio=incr_ratio, decr_ratio=decr_ratio, incr_every_n_steps=incr_every_n_steps, decr_every_n_nan_or_inf=decr_every_n_nan_or_inf, use_dynamic_loss_scaling=use_dynamic_loss_scaling) if strategy.recompute == True: recompute_enable = True model = _RecomputeModelWrapper(model) if self._user_defined_strategy.heter_ccl_mode == True: distributed_model = paddle.DataParallel( model, comm_buffer_size=self._user_defined_strategy. fuse_grad_size_in_MB, last_comm_buffer_size=self._user_defined_strategy. last_comm_group_size_MB, find_unused_parameters=self._user_defined_strategy. find_unused_parameters) return distributed_model if self._hcg.get_parallel_mode() == ParallelMode.SHARDING_PARALLEL: model = ShardingParallel(model, self._hcg, strategy=self._user_defined_strategy) elif self._hcg.get_parallel_mode() == ParallelMode.DATA_PARALLEL: # NOTE (JZ-LIANG) init parameters broadcast within sharding group # normally it should be done inside DataParallel if self.sharding_degree > 1: from paddle.distributed.fleet.utils.hybrid_parallel_util import broadcast_mp_parameters, broadcast_sharding_parameters assert self.sharding_degree == self._hcg.get_sharding_parallel_world_size( ) broadcast_sharding_parameters(model, self._hcg) model = paddle.DataParallel( model, comm_buffer_size=self._user_defined_strategy. fuse_grad_size_in_MB, last_comm_buffer_size=self._user_defined_strategy. last_comm_group_size_MB, find_unused_parameters=self._user_defined_strategy. find_unused_parameters) elif self._hcg.get_parallel_mode() == ParallelMode.TENSOR_PARALLEL: model = TensorParallel(model, self._hcg, strategy=self._user_defined_strategy) elif self._hcg.get_parallel_mode() == ParallelMode.PIPELINE_PARALLEL: model = PipelineParallel(model, self._hcg, strategy=self._user_defined_strategy) return model @dygraph_only def state_dict(self): """ Get state dict information from optimizer. (Only work in dygraph mode) Returns: state_dict(dict) : dict contains all the Tensor used by optimizer Examples: .. code-block:: python import numpy as np import paddle from paddle.distributed import fleet fleet.init(is_collective=True) value = np.arange(26).reshape(2, 13).astype("float32") a = paddle.to_tensor(value) layer = paddle.nn.Linear(13, 5) adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters()) adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) state_dict = adam.state_dict() """ # imitate target optimizer retrieval return self.user_defined_optimizer.state_dict() @dygraph_only def set_state_dict(self, state_dict): """ Load optimizer state dict. (Only work in dygraph mode) Args: state_dict(dict) : Dict contains all the Tensor needed by optimizer Returns: None Examples: .. code-block:: python import numpy as np import paddle from paddle.distributed import fleet fleet.init(is_collective=True) value = np.arange(26).reshape(2, 13).astype("float32") a = paddle.to_tensor(value) layer = paddle.nn.Linear(13, 5) adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters()) adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) state_dict = adam.state_dict() paddle.save(state_dict, "paddle_dy") para_state_dict = paddle.load("paddle_dy") adam.set_state_dict(para_state_dict) """ # imitate target optimizer retrieval return self.user_defined_optimizer.set_state_dict(state_dict) @dygraph_only def set_lr(self, value): """ Set the value of the learning rate manually in the optimizer. (Only work in dygraph mode) Args: value (float|Tensor): the value of learning rate Returns: None Examples: .. code-block:: python import numpy as np import paddle from paddle.distributed import fleet fleet.init(is_collective=True) value = np.arange(26).reshape(2, 13).astype("float32") a = paddle.to_tensor(value) layer = paddle.nn.Linear(13, 5) adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters()) adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) lr_list = [0.2, 0.3, 0.4, 0.5, 0.6] for i in range(5): adam.set_lr(lr_list[i]) lr = adam.get_lr() print("current lr is {}".format(lr)) # Print: # current lr is 0.2 # current lr is 0.3 # current lr is 0.4 # current lr is 0.5 # current lr is 0.6 """ # imitate target optimizer retrieval return self.user_defined_optimizer.set_lr(value) @dygraph_only def get_lr(self): """ Get current step learning rate. (Only work in dygraph mode) Returns: float: The learning rate of the current step. Examples: .. code-block:: python import numpy as np import paddle from paddle.distributed import fleet fleet.init(is_collective=True) value = np.arange(26).reshape(2, 13).astype("float32") a = paddle.to_tensor(value) layer = paddle.nn.Linear(13, 5) adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters()) adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) lr = adam.get_lr() print(lr) # 0.01 """ # imitate target optimizer retrieval return self.user_defined_optimizer.get_lr() @dygraph_only def step(self): """ Execute the optimizer once. (Only work in dygraph mode) Returns: None Examples: .. code-block:: python import paddle import paddle.nn as nn from paddle.distributed import fleet class LinearNet(nn.Layer): def __init__(self): super(LinearNet, self).__init__() self._linear1 = nn.Linear(10, 10) self._linear2 = nn.Linear(10, 1) def forward(self, x): return self._linear2(self._linear1(x)) # 1. initialize fleet environment fleet.init(is_collective=True) # 2. create layer & optimizer layer = LinearNet() loss_fn = nn.MSELoss() adam = paddle.optimizer.Adam( learning_rate=0.001, parameters=layer.parameters()) # 3. get data_parallel model using fleet adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) # 4. run layer inputs = paddle.randn([10, 10], 'float32') outputs = dp_layer(inputs) labels = paddle.randn([10, 1], 'float32') loss = loss_fn(outputs, labels) print("loss:", loss.numpy()) loss.backward() adam.step() adam.clear_grad() """ # imitate target optimizer retrieval return self.user_defined_optimizer.step() @dygraph_only def clear_grad(self): """ Clear the gradients of all optimized parameters for model. (Only work in dygraph mode) Returns: None Examples: .. code-block:: python import paddle import paddle.nn as nn from paddle.distributed import fleet class LinearNet(nn.Layer): def __init__(self): super(LinearNet, self).__init__() self._linear1 = nn.Linear(10, 10) self._linear2 = nn.Linear(10, 1) def forward(self, x): return self._linear2(self._linear1(x)) # 1. initialize fleet environment fleet.init(is_collective=True) # 2. create layer & optimizer layer = LinearNet() loss_fn = nn.MSELoss() adam = paddle.optimizer.Adam( learning_rate=0.001, parameters=layer.parameters()) # 3. get data_parallel model using fleet adam = fleet.distributed_optimizer(adam) dp_layer = fleet.distributed_model(layer) # 4. run layer inputs = paddle.randn([10, 10], 'float32') outputs = dp_layer(inputs) labels = paddle.randn([10, 1], 'float32') loss = loss_fn(outputs, labels) print("loss:", loss.numpy()) loss.backward() adam.step() adam.clear_grad() """ # imitate target optimizer retrieval return self.user_defined_optimizer.clear_grad() def _get_amp_optimizer(self): # imitate target optimizer retrieval amp_optimizer = None for optimizer in self.strategy_compiler._get_applied_meta_optimizer(): if hasattr(optimizer, 'amp_init'): amp_optimizer = optimizer break if amp_optimizer is None: if hasattr(self.user_defined_optimizer, 'amp_init'): amp_optimizer = self.user_defined_optimizer assert amp_optimizer is not None, \ "amp_init can only be used when the amp(auto mixed precision) strategy is turned on." return amp_optimizer def get_loss_scaling(self): """Return the real-time loss scaling factor. """ amp_optimizer = self._get_amp_optimizer() return amp_optimizer.get_loss_scaling() def amp_init(self, place, scope=None, test_program=None, use_fp16_test=False): """ Init the amp training, such as cast fp32 parameters to fp16 type. Args: place(CUDAPlace): place is used to initialize fp16 parameters with fp32 values. scope(Scope): The scope is used to find fp32 parameters. test_program(Program): The program is used for testing. use_fp16_test(bool): Whether to use fp16 testing. Examples: .. code-block:: python import numpy as np import paddle import paddle.nn.functional as F paddle.enable_static() def run_example_code(): place = paddle.CUDAPlace(0) exe = paddle.static.Executor(place) data = paddle.static.data(name='X', shape=[None, 1, 28, 28], dtype='float32') conv2d = paddle.static.nn.conv2d(input=data, num_filters=6, filter_size=3) # 1) Use fp16_guard to control the range of fp16 kernels used. with paddle.static.amp.fp16_guard(): bn = paddle.static.nn.batch_norm(input=conv2d, act="relu") pool = F.max_pool2d(bn, kernel_size=2, stride=2) hidden = paddle.static.nn.fc(pool, size=10) loss = paddle.mean(hidden) # 2) Create the optimizer and set `multi_precision` to True. # Setting `multi_precision` to True can avoid the poor accuracy # or the slow convergence in a way. optimizer = paddle.optimizer.Momentum(learning_rate=0.01, multi_precision=True) # 3) These ops in `custom_black_list` will keep in the float32 computation type. amp_list = paddle.static.amp.CustomOpLists( custom_black_list=['pool2d']) # 4) The entry of Paddle AMP. # Enable pure fp16 training by setting `use_pure_fp16` to True. optimizer = paddle.static.amp.decorate( optimizer, amp_list, init_loss_scaling=128.0, use_dynamic_loss_scaling=True, use_pure_fp16=True) # If you don't use the default_startup_program(), you sholud pass # your defined `startup_program` into `minimize`. optimizer.minimize(loss) exe.run(paddle.static.default_startup_program()) # 5) Use `amp_init` after FP32 parameters initialization(such as `exe.run(startup_program)`). # If you want to perform the testing process, you should pass `test_program` into `amp_init`. optimizer.amp_init(place, scope=paddle.static.global_scope()) if paddle.is_compiled_with_cuda() and len(paddle.static.cuda_places()) > 0: run_example_code() """ amp_optimizer = self._get_amp_optimizer() return amp_optimizer.amp_init(place, scope, test_program, use_fp16_test) def _final_strategy(self): if "valid_strategy" not in self._context: print( "WARNING: You may need to call minimize function before this function is called" ) return {} else: return self._context["valid_strategy"] def _get_applied_meta_list(self): if "applied_meta_list" not in self._context: print( "WARNING: You may need to call minimize function before _get_applied_meta_list called" ) return [] else: return self._context["applied_meta_list"] def _get_applied_graph_list(self): if "applied_graph_list" not in self._context: print( "WARNING: You may need to call minimize function before _get_applied_graph_list called" ) return [] else: return self._context["applied_graph_list"] def minimize(self, loss, startup_program=None, parameter_list=None, no_grad_set=None): """ Add distributed operations to minimize ``loss`` by updating ``parameter_list``. Args: loss (Tensor): A ``Tensor`` containing the value to minimize. startup_program (Program, optional): :ref:`api_fluid_Program` for initializing parameters in ``parameter_list``. The default value is None, at this time :ref:`api_fluid_default_startup_program` will be used. parameter_list (Iterable, optional): Iterable of ``Tensor`` or ``Tensor.name`` to update to minimize ``loss``. The default value is None, at this time all parameters will be updated. no_grad_set (set, optional): Set of ``Tensor`` or ``Tensor.name`` that don't need to be updated. The default value is None. Returns: tuple: tuple (optimize_ops, params_grads), A list of operators appended by minimize and a list of (param, grad) tensor pairs, param is ``Parameter``, grad is the gradient value corresponding to the parameter. The returned tuple can be passed to ``fetch_list`` in ``Executor.run()`` to indicate program pruning. If so, the program will be pruned by ``feed`` and ``fetch_list`` before run, see details in ``Executor``. Examples: .. code-block:: python import paddle paddle.enable_static() import paddle.distributed.fleet as fleet import paddle.nn.functional as F hid_dim = 10 label_dim = 2 input_x = paddle.static.data(name='x', shape=[None, 13], dtype='float32') input_y = paddle.static.data(name='y', shape=[None, 1], dtype='int64') fc_1 = paddle.static.nn.fc(x=input_x, size=hid_dim, activation='tanh') fc_2 = paddle.static.nn.fc(x=fc_1, size=hid_dim, activation='tanh') prediction = paddle.static.nn.fc(x=[fc_2], size=label_dim, activation='softmax') cost = F.cross_entropy(input=prediction, label=input_y) avg_cost = paddle.mean(x=cost) fleet.init(is_collective=True) strategy = fleet.DistributedStrategy() optimizer = paddle.optimizer.SGD(learning_rate=0.001) optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy) optimizer.minimize(avg_cost) # for more examples, please reference https://github.com/PaddlePaddle/FleetX """ if not isinstance(loss, list): return self._minimize_impl(loss, startup_program, parameter_list, no_grad_set) else: if paddle.fluid.framework._non_static_mode( ) or self._role_maker._is_non_distributed() or self._is_collective: raise ValueError("loss can be list only in PS mode") return self._minimize_losses_impl(loss, startup_program, parameter_list, no_grad_set) def _minimize_impl(self, loss, startup_program=None, parameter_list=None, no_grad_set=None): context = {} context["user_defined_strategy"] = copy.deepcopy( self._user_defined_strategy) if paddle.fluid.framework._non_static_mode(): # imitate target optimizer retrieval target_opt = self.user_defined_optimizer self._context = context return target_opt.minimize(loss) # cache original feed forward program self.origin_main_program = loss.block.program # add distributed attr if not hasattr(self.origin_main_program, "distributed_info_"): setattr(self.origin_main_program, "distributed_info_", dict()) self.origin_main_program.distributed_info_[ "dp_degree"] = self._user_defined_strategy.sharding_configs[ "dp_degree"] self.origin_main_program.distributed_info_[ "mp_degree"] = self._user_defined_strategy.sharding_configs[ "mp_degree"] self.origin_main_program.distributed_info_[ "pp_degree"] = self._user_defined_strategy.sharding_configs[ "pp_degree"] self.origin_main_program.distributed_info_[ "sharding_degree"] = self._user_defined_strategy.sharding_configs[ "sharding_degree"] context["origin_main_program"] = self.origin_main_program context["origin_main_programs"] = [self.origin_main_program] context["loss"] = loss if startup_program == None: self.origin_startup_program = \ paddle.static.default_startup_program().clone(for_test=False) startup_program = paddle.static.default_startup_program() else: self.origin_startup_program = \ startup_program.clone(for_test=False) context["origin_startup_program"] = startup_program context["origin_startup_programs"] = [startup_program] context["role_maker"] = self._role_maker # Use the auto-parallel's routines instead if self._user_defined_strategy.semi_auto or self._user_defined_strategy.auto_search: from ...auto_parallel.parallelizer import AutoParallelizer auto_parallelizer = AutoParallelizer(self) optimize_ops, params_grads, dist_startup_prog, dist_main_prog = auto_parallelizer.parallelize( loss, startup_program, parameter_list, no_grad_set) return optimize_ops, params_grads, dist_startup_prog, dist_main_prog # compile time distributed_optimizer_list = \ MetaOptimizerFactory()._get_valid_meta_optimizers( self.user_defined_optimizer) context["user_defined_strategy"] = copy.deepcopy( self._user_defined_strategy) copy_user_defined_strategy = copy.deepcopy(self._user_defined_strategy) # trigger the auto-parallel in very strict condition # strategy = DistributedStrategy() # strategy.auto = True # optimizer = paddle.optimizer.SGD(learning_rate=0.1) # optimizer = fleet.distributed_optimizer(optimizer, strategy) if copy_user_defined_strategy._is_strict_auto(): # turn on all the strategy for each optimizer for opt in distributed_optimizer_list: opt._enable_strategy(copy_user_defined_strategy, context) valid_optimizer_list = [] valid_graph_optimizer_list = [] can_not_apply_optimizer_list = [] # recall meta optimizers for ranking for opt in distributed_optimizer_list: opt._set_basic_info(loss, self._role_maker, self.user_defined_optimizer, copy_user_defined_strategy) if opt._can_apply() and not opt._is_graph_out(): valid_optimizer_list.append(opt) elif opt._can_apply() and opt._is_graph_out(): valid_graph_optimizer_list.append(opt) else: can_not_apply_optimizer_list.append(opt) # combine recalled meta optimizers to be a valid meta optimizer meta_optimizer, graph_optimizer = \ self.strategy_compiler.generate_optimizer( loss, self._role_maker, self.user_defined_optimizer, copy_user_defined_strategy, valid_optimizer_list, valid_graph_optimizer_list) valid_strategy = self.strategy_compiler._get_valid_strategy( copy_user_defined_strategy, can_not_apply_optimizer_list) context["valid_strategy"] = copy.deepcopy(valid_strategy) # print("valid_strategy:", context["valid_strategy"]) # print("user_defined_strategy:", context["user_defined_strategy"]) applied_meta_list = self.strategy_compiler._get_applied_meta_list() applied_graph_list = self.strategy_compiler._get_applied_graph_list() context['applied_meta_list'] = applied_meta_list context['applied_graph_list'] = applied_graph_list self._context = context self.valid_strategy = valid_strategy self.valid_strategy._enable_env() optimize_ops = [] params_grads = [] if self._role_maker._is_non_distributed() and not self._is_collective: if self._runtime_handle is None: self._runtime_handle = RuntimeFactory()._create_runtime(context) compiled_program = compiler.CompiledProgram( self.origin_main_program).with_data_parallel( loss_name=loss.name, share_vars_from=None) loss.block.program._graph = compiled_program return self.user_defined_optimizer.minimize(loss, startup_program, parameter_list, no_grad_set=no_grad_set) if meta_optimizer: # print("before minimize program id:", id(loss.block.program)) optimize_ops, params_grads = meta_optimizer.minimize( loss, startup_program, parameter_list, no_grad_set=no_grad_set) # print("after minimize program id:", id(loss.block.program)) default_program = paddle.static.default_main_program() # print("default program id:", id(default_program)) if id(default_program) != id(loss.block.program): paddle.fluid.framework.switch_main_program(loss.block.program) # print("default program id after switch:", id(default_program)) else: optimize_ops, params_grads = self.user_defined_optimizer.minimize( loss, startup_program, parameter_list, no_grad_set=no_grad_set) context["program_optimize_ops"] = optimize_ops context["program_params_grads"] = params_grads if graph_optimizer: # print("before graph minimize program id:", id(loss.block.program)) optimize_ops, params_grads = graph_optimizer.minimize( loss, startup_program, parameter_list, no_grad_set=no_grad_set) # since we do not encourage users to use graph operations # if a graph optimizer takes effect, mostly # optimizers_ops and params_grads are None # i.e. users can not modify current computation graph anymore context["graph_optimize_ops"] = optimize_ops context["graph_optimize_grads"] = params_grads else: apply_ir_passes(loss.block.program, startup_program, self) if not self._role_maker._is_heter_parameter_server_mode: program = paddle.static.default_main_program() opt_info = {} if program._fleet_opt is None else program._fleet_opt opt_info["mpi_size"] = self.worker_num() opt_info["mpi_rank"] = self.worker_index() for k, v in self._user_defined_strategy.trainer_desc_configs.items( ): if v or k not in opt_info: opt_info[k] = v program._fleet_opt = opt_info if self._runtime_handle is None: self._runtime_handle = RuntimeFactory()._create_runtime(context) import paddle.distributed.fleet as fleet fleet.util._set_strategy(context["valid_strategy"]) return optimize_ops, params_grads def _minimize_losses_impl(self, losses, startup_programs=None, parameter_list=None, no_grad_set=None): context = {} # cache original feed forward program self.origin_main_program = losses[0].block.program context["origin_main_program"] = self.origin_main_program context["origin_main_programs"] = [] for loss in losses: context["origin_main_programs"].append(loss.block.program) context["loss"] = losses if startup_programs is None: if len(losses) == 1: startup_programs = [paddle.static.default_startup_program()] else: raise ValueError( "startup_program can't be None when loss is list.") self.origin_startup_program = startup_programs[0].clone(for_test=False) context["origin_startup_program"] = startup_programs[0] context["origin_startup_programs"] = [] for program in startup_programs: context["origin_startup_programs"].append(program) context["role_maker"] = self._role_maker context["user_defined_strategy"] = copy.deepcopy( self._user_defined_strategy) context["valid_strategy"] = copy.deepcopy(self._user_defined_strategy) self._context = context self.valid_strategy = context["valid_strategy"] self.valid_strategy._enable_env() optimize_ops = [] params_grads = [] from ..meta_optimizers import ParameterServerOptimizer ps_optimizer = ParameterServerOptimizer(self.user_defined_optimizer) ps_optimizer._set_basic_info(losses, self._role_maker, self.user_defined_optimizer, self._user_defined_strategy) optimize_ops, params_grads = ps_optimizer.minimize_losses_impl( losses, startup_programs, parameter_list, no_grad_set=no_grad_set) # default_program = paddle.static.default_main_program() # if id(default_program) != id(losses[0].block.program): # paddle.fluid.framework.switch_main_program(losses[0].block.program) context["program_optimize_ops"] = optimize_ops context["program_params_grads"] = params_grads for loss in losses: program = loss.block.program opt_info = {} if program._fleet_opt is None else program._fleet_opt opt_info["mpi_size"] = self.worker_num() opt_info["mpi_rank"] = self.worker_index() for k, v in self._user_defined_strategy.trainer_desc_configs.items( ): if v or k not in opt_info: opt_info[k] = v program._fleet_opt = opt_info # print("fleet base opt info:", id(program), program._fleet_opt) if self._runtime_handle is None: self._runtime_handle = RuntimeFactory()._create_runtime(context) import paddle.distributed.fleet as fleet fleet.util._set_strategy(context["valid_strategy"]) return optimize_ops, params_grads @dygraph_only def distributed_scaler(self, scaler): def unscale_method(self, optimizer): if not self._enable: return if getattr(optimizer, '_param_groups', None) and isinstance( optimizer._param_groups[0], dict): param_grads = [] param_grads_fp16 = [] param_grads_fp32 = [] for group in optimizer._param_groups: for param in group['params']: if param._grad_ivar() is not None: param_grads.append(param._grad_ivar()) if param._grad_ivar( ).dtype == core.VarDesc.VarType.FP16: param_grads_fp16.append(param._grad_ivar()) else: param_grads_fp32.append(param._grad_ivar()) else: param_grads = [ param._grad_ivar() for param in optimizer._parameter_list if param._grad_ivar() is not None ] param_grads_fp16 = [ param._grad_ivar() for param in optimizer._parameter_list if (param._grad_ivar() is not None) and ( param._grad_ivar().dtype == core.VarDesc.VarType.FP16) ] param_grads_fp32 = [ param._grad_ivar() for param in optimizer._parameter_list if (param._grad_ivar() is not None) and ( param._grad_ivar().dtype == core.VarDesc.VarType.FP32) ] temp_found_inf_fp16 = to_variable(np.array([0]).astype(np.bool_)) temp_found_inf_fp32 = to_variable(np.array([0]).astype(np.bool_)) if len(param_grads_fp16): _C_ops.check_finite_and_unscale(param_grads_fp16, self._scale, param_grads_fp16, temp_found_inf_fp16) if len(param_grads_fp32): _C_ops.check_finite_and_unscale(param_grads_fp32, self._scale, param_grads_fp32, temp_found_inf_fp32) self._found_inf = 1 if temp_found_inf_fp16 or temp_found_inf_fp32 else 0 is_found_inf = paddle.to_tensor([self._found_inf], dtype="int32") # TODO(shenliang03) Since dp allreduce in the optimizer is # after the gradscaler, check_finite needs to synchronize global # information. In the future, we should use check_group to speed. paddle.distributed.all_reduce(is_found_inf, op=paddle.distributed.ReduceOp.MAX, group=None) self._found_inf = is_found_inf.numpy()[0] # Only data_parallel doesn't need to modify scaler if self._hcg.get_parallel_mode() is not ParallelMode.DATA_PARALLEL: scaler._unscale = MethodType(unscale_method, scaler) return scaler