# 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 jin 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 itertools import os import sys import time import warnings from collections import OrderedDict, namedtuple from contextlib import contextmanager from multiprocessing import Manager # noqa: F401 from multiprocessing import Process # noqa: F401 import numpy as np import paddle from paddle import _legacy_C_ops, framework from paddle.distributed.collective import ( Group, _default_group_name, _get_group_map_by_name, _new_process_group_impl, _set_default_backend, _set_default_store, _set_group_map, _set_group_map_backend, _set_group_map_by_name, _valid_backend_list, ) from paddle.distributed.communication.group import _add_new_group from paddle.distributed.fleet.base.private_helper_function import ( # noqa: F401 wait_server_ready, ) from paddle.distributed.fleet.launch_utils import check_backend # (TODO: GhostScreaming) It will be removed later. from paddle.framework import _set_expected_place from paddle.framework import base as imperative_base from paddle.framework import core, in_dygraph_mode from paddle.nn.layer import layers from paddle.utils import deprecated from . import parallel_helper __all__ = [] ParallelStrategy = core.ParallelStrategy def _build_default_parallel_strategy(): strategy = ParallelStrategy() strategy.nranks = paddle.distributed.ParallelEnv().nranks strategy.local_rank = paddle.distributed.ParallelEnv().local_rank strategy.trainer_endpoints = ( paddle.distributed.ParallelEnv().trainer_endpoints ) strategy.current_endpoint = ( paddle.distributed.ParallelEnv().current_endpoint ) return strategy def _coalesce_tensors(var_groups): coalesced_grads_and_grad_vars = [] for group_id, grad_vars in var_groups.items(): flattened_vars = [] g_var_shapes = [] for g_var in grad_vars: g_var_shapes.append(g_var.shape) flattened_vars.append( paddle.reshape(x=g_var, shape=[np.prod(g_var.shape)]) ) coalesced_grad = paddle.concat(flattened_vars) coalesced_grads_and_grad_vars.append( [coalesced_grad, grad_vars, g_var_shapes] ) return coalesced_grads_and_grad_vars @framework.dygraph_only def _reshape_inplace(x, shape): x_shape = framework._create_tensor(dtype=x.dtype) framework._dygraph_tracer().trace_op( type="reshape2", inputs={'X': x}, outputs={'Out': x, 'XShape': x_shape}, attrs={'shape': shape}, ) @framework.dygraph_only def _split_tensors(coalesced_grads_and_grad_vars): if in_dygraph_mode(): for ( coalesced_grad, origin_grad_vars, grad_shapes, ) in coalesced_grads_and_grad_vars: grad_var_len = [np.prod(g_shape) for g_shape in grad_shapes] attrs = () attrs += ('sections', grad_var_len) attrs += ('axis', 0) _legacy_C_ops.split(coalesced_grad, origin_grad_vars, *attrs) for g_var, g_shape in zip(origin_grad_vars, grad_shapes): g_var.reshape_(shape=g_shape) assert g_var.shape == g_shape @imperative_base.no_grad @framework.dygraph_only def build_groups(vars, group_size): group_idx = 0 memory_counter = 0 var_groups = OrderedDict() dtype = vars[0].dtype for var in vars: bytes = np.prod(var.shape) * core.size_of_dtype(var.dtype) if memory_counter < group_size and dtype == var.dtype: memory_counter += bytes else: memory_counter = bytes dtype = var.dtype group_idx += 1 var_groups.setdefault(group_idx, []).append(var) return _coalesce_tensors(var_groups) @imperative_base.no_grad @framework.dygraph_only def sync_params_buffers( model, comm_group=None, src_rank=0, is_model_parallel=False ): model_vars = [] for _, param in model._obtain_parameters_buffers().items(): if not isinstance(param, core.eager.Tensor): raise TypeError( "The data type of '%s' must be Varbase or eager.Tensor" % param.name ) # is_distributed param not need to sync when in mp mode if isinstance(param, core.eager.Tensor): if is_model_parallel: if hasattr(param, "is_distributed") and param.is_distributed: continue # NOTE(shenliang03): Support situations that do not require synchronization parameters, # such as moe's expert parameters if getattr(param, "no_sync", False): continue if param.type == core.VarDesc.VarType.VOCAB: continue model_vars.append(param.detach()) if len(model_vars) == 0: return # group size is 128M coalesced_vars = build_groups(model_vars, 128 * 1024 * 1024) for coalesced_var, _, _ in coalesced_vars: paddle.distributed.broadcast( coalesced_var, src=src_rank, group=comm_group, sync_op=True ) for coalesced_var, origin_vars, var_shapes in coalesced_vars: var_len = [np.prod(v_shape) for v_shape in var_shapes] paddle.fluid.framework._dygraph_tracer().trace_op( type='split', inputs={'X': coalesced_var}, outputs={'Out': origin_vars}, attrs={'sections': var_len, 'axis': 0}, ) class DataParallel(layers.Layer): """ Run the dygraph module with data parallelism. Currently, DataParallel class only supports to run the dynamic graph with multi-process. Now supports two ways to start training: 1. start by ``paddle.distributed.spawn`` method, for example: ``python demo.py`` (spawn need to be called in ``__main__`` method) 2. start by ``paddle.distributed.launch`` module, for example: ``python -m paddle.distributed.launch --gpus=0,1 demo.py`` . And the content of `demo.py` is the code of examples. Args: layers(Layer): The module that should be executed by data parallel. strategy(ParallelStrategy, optional): (deprecated) The strategy of data parallelism, contains environment configuration related to parallel execution. Default: None. comm_buffer_size(int, optional): It limits the memory size(MB) of one buffer parameters' gradient which is the input of communication calling(e.g NCCLAllReduce). Default: 25. last_comm_buffer_size(float, optional): It limits memory size(MB) of last buffer in communication calling. Making the last communication buffer size small is useful to improve performance. Default: 1. find_unused_parameters(bool, optional): Whether to traverse the entire backward graph from the all tensors in the return value of the wrapped model's forward function. For parameters not involved in loss calculation, their gradients will be marked as ready in advance to prepare reduce. Please note that all forward outputs derived from the wrapped model parameters must participate in the calculation of loss and subsequent gradient calculations. If not, serious error will occur. Note that setting the find_unused_parameters to True will affect computing performance. Therefore, if all parameters are sure to participate in the loss calculation and the autograd graph construction, please set it False. Default: False. Returns: Layer: The data paralleled module. Examples: .. code-block:: python :name: dp-example # required: distributed import paddle import paddle.nn as nn import paddle.optimizer as opt import paddle.distributed as dist class LinearNet(nn.Layer): def __init__(self): super().__init__() self._linear1 = nn.Linear(10, 10) self._linear2 = nn.Linear(10, 1) def forward(self, x): return self._linear2(self._linear1(x)) def train(): # 1. initialize parallel environment dist.init_parallel_env() # 2. create data parallel layer & optimizer layer = LinearNet() dp_layer = paddle.DataParallel(layer) loss_fn = nn.MSELoss() adam = opt.Adam( learning_rate=0.001, parameters=dp_layer.parameters()) # 3. run layer inputs = paddle.randn([10, 10], 'float32') outputs = dp_layer(inputs) labels = paddle.randn([10, 1], 'float32') loss = loss_fn(outputs, labels) loss.backward() adam.step() adam.clear_grad() if __name__ == '__main__': # 1. start by ``paddle.distributed.spawn`` (default) dist.spawn(train, nprocs=2) # 2. start by ``paddle.distributed.launch`` # train() .. note:: ``PyLayer`` is not supported in DataParallel. To solve problems of this kind, it's recommended to skip gradient synchronization among multiple cards by 'no_sync', and manually implement 'all_reduce' before model optimization. There is an example showing specific implemetation processing. Examples: .. code-block:: python :name: dp-pylayer-example # required: distributed import numpy import paddle import paddle.distributed as dist from paddle.autograd import PyLayer from paddle.distributed.fleet.utils.hybrid_parallel_util import fused_allreduce_gradients class cus_tanh(PyLayer): @staticmethod def forward(ctx, x): y = paddle.tanh(x) ctx.save_for_backward(y) return y @staticmethod def backward(ctx, dy): y, = ctx.saved_tensor() grad = dy * (1 - paddle.square(y)) return grad class SimpleNet(paddle.nn.Layer): def __init__(self): super().__init__() self.linear = paddle.nn.Linear(2, 2) def forward(self, inputs): inputs = cus_tanh.apply(inputs) return self.linear(inputs) if __name__ == '__main__': dist.init_parallel_env() model = SimpleNet() model = paddle.DataParallel(model) opt = paddle.optimizer.SGD(learning_rate=0.01, parameters=model.parameters()) for step in range(10): x_data = numpy.random.randn(2, 2).astype(numpy.float32) x = paddle.to_tensor(x_data) x.stop_gradient = False # step 1 : skip gradient synchronization by 'no_sync' with model.no_sync(): y_pred = model(x) loss = y_pred.mean() loss.backward() # step 2 : fuse + allreduce manually before optimization fused_allreduce_gradients(list(model.parameters()), None) opt.step() opt.clear_grad() """ def __init__( self, layers, strategy=None, comm_buffer_size=25, last_comm_buffer_size=1, find_unused_parameters=False, group=None, ): super().__init__(layers.full_name() + "_data_parallel") assert ( in_dygraph_mode() ), "It's not supported to construct DataParallel in static graph mode." self._layers = layers self.find_unused_parameters = find_unused_parameters self.grad_need_sync = True self.group = group self.var_dtype = core.eager.Tensor # NOTE(chenweihang): The ParallelStrategy here is not strictly a strategy. # It just stores some environment variables, which can be constructed by # ParallelEnv. Here it is set as an optional argument. # This parameter is not removed because of compatibility with 1.x writing. if strategy is not None: self._strategy = strategy else: self._strategy = _build_default_parallel_strategy() if self._strategy.nranks > 1: # check the environment assert parallel_helper.__parallel_ctx__clz__ is not None, ( "ParallelContext must be initialized before. You should use init_parallel_env() before" "constructing the DataParallel." ) if in_dygraph_mode(): self.group = ( paddle.distributed.collective._get_default_group() if self.group is None else self.group ) assert isinstance( self.group, paddle.distributed.collective.Group ), "ProcessGroup must be an instance of Group in DataParallel." # sync buffer and params # TODO(liuyuhui) Currently not support xpu. xpu is # still broadcasting parameters when calling layer if not paddle.is_compiled_with_xpu(): sync_params_buffers(self._layers) self.comm_buffer_size = int(comm_buffer_size * 1024 * 1024) # NOTE(shenliang03): We can set environment variables to control # the size of the group, Default: 1MB. The role of this small group is: # when the last group allreduce, the overlap cannot work. Making the # the last group small is useful to improve performance. self.last_comm_buffer_size = int( last_comm_buffer_size * 1024 * 1024 ) self.init_reducer() else: warnings.warn( "The program will return to single-card operation. " "Please check 1, whether you use spawn or fleetrun " "to start the program. 2, Whether it is a multi-card " "program. 3, Is the current environment multi-card." ) def init_reducer(self): layers_param = [] params_set = set() for sublayer in self.sublayers(): for _, param in sublayer.named_parameters(include_sublayers=False): if param is None or param in params_set: continue params_set.add(param) if not isinstance(param, self.var_dtype): raise TypeError( "The data type of '%s' must be '%s'" % (param.name, self.var_dtype) ) if param.trainable: layers_param.append((sublayer, param)) trainable_parameters = list( filter( lambda x: not getattr(x, "no_sync", False), [param for _, param in layers_param], ) ) assert len(trainable_parameters) > 0, ( "This model does not have any parameters to train, and " "does not need to use DataParallel" ) # NOTE(shenliang03): Here we can only use the attributes to judge whether # parameter is sparse(or SelectedRows). The reason is that the sparse message # can't be obtained when bp hasn't happened yet. So if layer supports sparse parameter, # we should add the layer here like "paddle.nn.layer.common.Embedding". def check_layer_sparse(sublayer): if isinstance(sublayer, paddle.nn.layer.common.Embedding): return sublayer._sparse return False is_sparse_gradient = [ check_layer_sparse(sublayer) for sublayer, _ in layers_param ] if in_dygraph_mode(): self.group_indices = core.eager_assign_group_by_size( trainable_parameters, is_sparse_gradient, [self.last_comm_buffer_size, self.comm_buffer_size], ) self._reducer = core.EagerReducer( trainable_parameters, list(reversed(self.group_indices)), is_sparse_gradient, self.group.process_group, [self.last_comm_buffer_size, self.comm_buffer_size], self.find_unused_parameters, ) def _find_varbase(self, obj): var_type = core.eager.Tensor if isinstance(obj, var_type): return [obj] if isinstance(obj, (list, tuple)): return itertools.chain(*map(self._find_varbase, obj)) if isinstance(obj, dict): return itertools.chain(*map(self._find_varbase, obj.values())) return [] @contextmanager def no_sync(self): """ A context manager to stop gradient synchronization. Within no_sync(), gradients of parameters will only be accumulated on model and not synchronized util the first forward-backward out of this context. Examples: .. code-block:: python # required: distributed import paddle import paddle.nn as nn import paddle.distributed as dist class SimpleNet(nn.Layer): def __init__(self): super().__init__() self._linear = nn.Linear(10, 1) def forward(self, x): return self._linear(x) dist.init_parallel_env() model = SimpleNet() dp_model = paddle.DataParallel(model) inputs_1 = paddle.randn([10, 10], 'float32') inputs_2 = paddle.ones([10, 10], 'float32') with dp_model.no_sync(): # gradients will not be synchronized dp_model(inputs_1).backward() # synchronization happens here dp_model(inputs_2).backward() """ tmp_grad_need_sync = self.grad_need_sync self.grad_need_sync = False try: yield finally: self.grad_need_sync = tmp_grad_need_sync def forward(self, *inputs, **kwargs): outputs = self._layers(*inputs, **kwargs) if ( self._strategy.nranks > 1 and framework._dygraph_tracer()._has_grad and self.grad_need_sync ): self._reducer.prepare_for_backward( list(self._find_varbase(outputs)) ) return outputs @deprecated( since="2.0.0", reason="This method does not need to be called anymore." ) def scale_loss(self, loss): """ Deprecated method, now ``scale_loss`` is an empty method, keep this method just for compatibility. """ return loss @deprecated( since="2.0.0", reason="This method does not need to be called anymore." ) def apply_collective_grads(self): """ Deprecated method, now ``apply_collective_grads`` is an empty method, keep this method just for compatibility. """ return def state_dict( self, destination=None, include_sublayers=True, structured_name_prefix="", ): ''' Get all parameters and persistable buffers of current layer and its sub-layers. And set them into a dict Parameters: destination(dict, optional) : If provide, all the parameters and persistable buffers will be set to this dict . Default: None include_sublayers(bool, optional) : If true, also include the parameters and persistable buffers from sublayers. Default: True Retruns: dict: a dict contains all the parameters and persistable buffers. Examples: .. code-block:: python import paddle import paddle.distributed as dist dist.init_parallel_env() emb = paddle.nn.Embedding(10, 10) emb = paddle.DataParallel(emb) state_dict = emb.state_dict() paddle.save(state_dict, "paddle_dy.pdparams") ''' return self._layers.state_dict( destination=destination, include_sublayers=include_sublayers, structured_name_prefix=structured_name_prefix, ) @framework.deprecate_stat_dict def set_state_dict(self, state_dict, use_structured_name=True): ''' Set parameters and persistable buffers from state_dict. All the parameters and buffers will be reset by the tensor in the state_dict Parameters: state_dict(dict) : Dict contains all the parameters and persistable buffers. use_structured_name(bool, optional) : If true, use structured name as key, otherwise, use parameter or buffer name as key. Default: True Returns: None Examples: .. code-block:: python import paddle import paddle.distributed as dist dist.init_parallel_env() emb = paddle.nn.Embedding(10, 10) emb = paddle.DataParallel(emb) state_dict = emb.state_dict() paddle.save(state_dict, "paddle_dy.pdparams") para_state_dict = paddle.load("paddle_dy.pdparams") emb.set_state_dict(para_state_dict) ''' self._layers.set_state_dict( state_dict, use_structured_name=use_structured_name ) # [aliases] Compatible with old method names set_dict = set_state_dict load_dict = set_state_dict # NOTE(chenweihang): Maintain a global parallel env to avoid # initializing ParallelEnv every time and improve performance _global_parallel_env = None class ParallelEnv: """ .. note:: This API is not recommended, if you need to get rank and world_size, it is recommended to use ``paddle.distributed.get_rank()`` and ``paddle.distributed.get_world_size()`` . This class is used to obtain the environment variables required for the parallel execution of ``paddle.nn.Layer`` in dynamic mode. The parallel execution in dynamic mode needs to be started using ``paddle.distributed.launch`` or ``paddle.distributed.spawn`` . Examples: .. code-block:: python import paddle import paddle.distributed as dist def train(): # 1. initialize parallel environment dist.init_parallel_env() # 2. get current ParallelEnv parallel_env = dist.ParallelEnv() print("rank: ", parallel_env.rank) print("world_size: ", parallel_env.world_size) # print result in process 1: # rank: 1 # world_size: 2 # print result in process 2: # rank: 2 # world_size: 2 if __name__ == '__main__': # 1. start by ``paddle.distributed.spawn`` (default) dist.spawn(train, nprocs=2) # 2. start by ``paddle.distributed.launch`` # train() """ def __init__(self): self._rank = int(os.getenv("PADDLE_TRAINER_ID", "0")) self._world_size = int(os.getenv("PADDLE_TRAINERS_NUM", "1")) self._device_type = str(os.getenv("PADDLE_XCCL_BACKEND", "")) # imperative only support one gpu or xpu if self._device_type != "": FLAGS_selected_custom_devices = 'FLAGS_selected_{}s'.format( self._device_type ) selected_custom_devices = os.getenv( FLAGS_selected_custom_devices, "0" ).split(",") self._device_id = int(selected_custom_devices[0]) else: if core.is_compiled_with_cuda(): selected_gpus = os.getenv("FLAGS_selected_gpus", "0").split(",") self._device_id = int(selected_gpus[0]) elif core.is_compiled_with_xpu(): selected_xpus = os.getenv("FLAGS_selected_xpus", "0").split(",") self._device_id = int(selected_xpus[0]) self._trainer_endpoints = os.getenv( "PADDLE_TRAINER_ENDPOINTS", "" ).split(",") self._current_endpoint = os.getenv("PADDLE_CURRENT_ENDPOINT", "") self._nrings = int(os.getenv("FLAGS_nccl_nrings", "1")) assert ( self._nrings > 0 ), "nccl_nrings must be an integer greater than 0." assert ( self._nrings < 9 ), "nccl_nrings should be less than 9, which is enough in most scenarios." @property def rank(self): """ Rank of current trainer. Its value is equal to the value of the environment variable ``PADDLE_TRAINER_ID`` . The default value is 0. Examples: .. code-block:: python # execute this command in terminal: export PADDLE_TRAINER_ID=0 import paddle.distributed as dist env = dist.ParallelEnv() print("The rank is %d" % env.rank) # The rank is 0 """ return self._rank @property def world_size(self): """ The number of trainers (number of processes participating in current job). Its value is equal to the value of the environment variable ``PADDLE_TRAINERS_NUM`` . The default value is 1. Examples: .. code-block:: python # execute this command in terminal: export PADDLE_TRAINERS_NUM=4 import paddle.distributed as dist env = dist.ParallelEnv() print("The world_size is %d" % env.world_size) # The world_size is 4 """ return self._world_size @property def device_id(self): """ The ID of selected GPU card for parallel training. Its value is equal to the value of the environment variable ``FLAGS_selected_gpus`` . The default value is 0. Examples: .. code-block:: python # execute this command in terminal: export FLAGS_selected_gpus=1 import paddle.distributed as dist env = dist.ParallelEnv() print("The device id are %d" % env.device_id) # The device id are 1 """ return self._device_id @property def device_type(self): """ The type of custom device for parallel training. Its value is equal to the value of the environment variable ``PADDLE_XCCL_BACKEND`` . The default value is None. """ return self._device_type @property def current_endpoint(self): """ The endpoint of current trainer, it is in the form of (node IP + port). Its value is equal to the value of the environment variable ``PADDLE_CURRENT_ENDPOINT`` . The default value is "". Examples: .. code-block:: python # execute this command in terminal: export PADDLE_CURRENT_ENDPOINT=127.0.0.1:6170 import paddle.distributed as dist env = dist.ParallelEnv() print("The current endpoint are %s" % env.current_endpoint) # The current endpoint are 127.0.0.1:6170 """ return self._current_endpoint @property def trainer_endpoints(self): """ The endpoints of all trainer nodes in the task, which are used to broadcast the NCCL ID when NCCL2 is initialized. Its value is equal to the value of the environment variable ``PADDLE_TRAINER_ENDPOINTS`` . The default value is "". Examples: .. code-block:: python # execute this command in terminal: export PADDLE_TRAINER_ENDPOINTS=127.0.0.1:6170,127.0.0.1:6171 import paddle.distributed as dist env = dist.ParallelEnv() print("The trainer endpoints are %s" % env.trainer_endpoints) # The trainer endpoints are ['127.0.0.1:6170', '127.0.0.1:6171'] """ return self._trainer_endpoints @property def nrings(self): """ Nrings of current trainer. Its value is equal to the value of the environment variable ``FLAGS_nccl_nrings`` . The default value is 1. Examples: .. code-block:: python # execute this command in terminal: export FLAGS_nccl_nrings=1 import paddle.distributed as dist env = dist.ParallelEnv() print("The nrings is %d" % env.nrings) # the number of ring is 1 """ return self._nrings # [aliases] Compatible with old method names local_rank = rank nranks = world_size dev_id = device_id def _get_global_parallel_env(): global _global_parallel_env if _global_parallel_env is None: _global_parallel_env = ParallelEnv() return _global_parallel_env def _start_kv_server(port, http_server_d, size): from paddle.distributed.fleet.utils.http_server import KVServer http_server = KVServer(int(port), size=size) http_server.start() wait_seconds = 3 while http_server_d.get("running", False) or not http_server.should_stop(): time.sleep(wait_seconds) http_server.stop() def _is_cpuonly(backend): check_backend(backend) if ( backend in ['auto', 'nccl', 'bkcl', 'heter'] and (core.is_compiled_with_cuda() or core.is_compiled_with_xpu()) ) or backend == 'xccl': # passes 'auto' and can use cuda or xpu, use the default logics. so return False return False else: return True def _check_var_exists(var_name): var = os.environ.get(var_name, None) if var is None: raise ValueError( "paddle.distributed initialize error, " "environment variable %s is needed, but not set." % var_name ) def _get_modified_flags(): ret = [] FLAGS = namedtuple('FLAGS', ['name', 'current_value', 'default_value']) global_flags = core.globals() for key in global_flags.keys(): value = global_flags.get(key) default_value = global_flags.get_default(key) if not value == default_value: ret.append(FLAGS(key, value, default_value)) return ret def _print_modified_flags(modified_flags): if len(modified_flags) > 0: sys.stderr.write( "======================= Modified FLAGS detected =======================\n" ) for flag in modified_flags: sys.stderr.write(str(flag)) sys.stderr.write("\n") sys.stderr.write( "=======================================================================\n" ) def init_parallel_env(): """ Initialize parallel training environment in dynamic graph mode. Note: Now initialize both `NCCL` and `GLOO` contexts for communication. Args: backend (string): A string represents the backend used by DataParallel, should be one of 'gloo'(for cpu), 'nccl'(for cuda), 'bkcl'(for xpu), 'auto'(auto detect). The auto detection prefer 'nccl', 'bkcl' than 'gloo'. Returns: None Examples: .. code-block:: python # required: gpu import paddle import paddle.nn as nn import paddle.optimizer as opt import paddle.distributed as dist class LinearNet(nn.Layer): def __init__(self): super().__init__() self._linear1 = nn.Linear(10, 10) self._linear2 = nn.Linear(10, 1) def forward(self, x): return self._linear2(self._linear1(x)) def train(): # 1. initialize parallel environment dist.init_parallel_env() # 2. create data parallel layer & optimizer layer = LinearNet() dp_layer = paddle.DataParallel(layer) loss_fn = nn.MSELoss() adam = opt.Adam( learning_rate=0.001, parameters=dp_layer.parameters()) # 3. run layer inputs = paddle.randn([10, 10], 'float32') outputs = dp_layer(inputs) labels = paddle.randn([10, 1], 'float32') loss = loss_fn(outputs, labels) loss.backward() adam.step() adam.clear_grad() if __name__ == '__main__': dist.spawn(train) """ modified_flags = _get_modified_flags() _print_modified_flags(modified_flags) # 0. get env & check world size global _global_parallel_env # when call init_parallel_env, need update `_global_parallel_env` _global_parallel_env = ParallelEnv() parallel_env = _global_parallel_env # if not parallel, `init_parallel_env` do nothing if parallel_env.world_size < 2: warnings.warn( "Currently not a parallel execution environment, `paddle.distributed.init_parallel_env` will not do anything." ) return # NOTE(xiongkun): support cpu gloo only, add this environment variable to # enable cpu only gloo prarllel training) backend = os.environ.get('PADDLE_DISTRI_BACKEND', 'auto') is_cpu_only = _is_cpuonly(backend) # 1. gpu xpu check, must be gpu or xpu, if not ( is_cpu_only or core.is_compiled_with_cuda() or core.is_compiled_with_xpu() or backend == "xccl" ): raise NotImplementedError( "If you want to use CPU-only version, please use 'gloo' as backend" ) if backend == "xccl": FLAGS_selected_custom_devices = 'FLAGS_selected_{}s'.format( parallel_env.device_type ) _check_var_exists(FLAGS_selected_custom_devices) else: if not is_cpu_only and core.is_compiled_with_cuda(): _check_var_exists("FLAGS_selected_gpus") backend = "nccl" if backend == "auto" else backend elif not is_cpu_only and core.is_compiled_with_xpu(): _check_var_exists('FLAGS_selected_xpus') backend = "bkcl" if backend == "auto" else backend _check_var_exists("PADDLE_TRAINER_ID") _check_var_exists("PADDLE_CURRENT_ENDPOINT") _check_var_exists("PADDLE_TRAINERS_NUM") _check_var_exists("PADDLE_TRAINER_ENDPOINTS") # NOTE(chenweihang): [ why config global place here? ] # the dygraph mode will be set to default mode, # users will not call `dygraph.guard` or `enable_dygraph` # directly, if they want to switch default place, # they need to call a function to change default place, # here just set correctly place to users if backend == "xccl": place = core.CustomPlace( parallel_env.device_type, parallel_env.device_id ) elif is_cpu_only: place = core.CPUPlace() elif core.is_compiled_with_cuda(): place = core.CUDAPlace(parallel_env.device_id) elif core.is_compiled_with_xpu(): place = core.XPUPlace(parallel_env.device_id) _set_expected_place(place) group = None if backend in _valid_backend_list and in_dygraph_mode(): if _default_group_name in _get_group_map_by_name(): return _get_group_map_by_name()[_default_group_name] _set_default_backend(backend) rank = int(os.getenv("PADDLE_TRAINER_ID")) world_size = int(os.getenv("PADDLE_TRAINERS_NUM")) assert rank >= 0 and world_size > rank and world_size > 1, ( "rank must be non-negative and world_size must be the " "maximum rank plus one. Moreover, at least two processes are " "required to create a process group." ) master_addr = os.getenv("MASTER_ADDR", None) master_port = os.getenv("MASTER_PORT", None) endpoints = ( ":".join([master_addr, master_port]) if master_addr and master_port else None ) if endpoints is None: endpoints = os.getenv("PADDLE_MASTER", None) if endpoints is None: endpoints = os.getenv("PADDLE_TRAINER_ENDPOINTS").split(',')[0] assert endpoints, ( "The environment variable 'MASTER_ADDR' and 'MASTER_PORT' " "must be specified, for example 'export MASTER_ADDR=127.0.0.1' " "and 'export MASTER_ADDR=54612'. Or you can start your training" "with paddle.distributed.run module." ) master_addr, master_port = endpoints.split(":") master_port = int(master_port) is_master = rank == 0 stop_check_timeout = int(os.getenv("FLAGS_stop_check_timeout", "900")) default_store = core.TCPStore( master_addr, master_port, is_master, world_size, timeout=stop_check_timeout, ) _set_default_store(default_store) pg = _new_process_group_impl( backend, default_store, rank, world_size, _default_group_name, pg_options=None, ) ranks = list(range(world_size)) group = Group(rank, 0, ranks, pg=pg, name=_default_group_name) _set_group_map_by_name(_default_group_name, group) _set_group_map(0, group) _set_group_map_backend(group, backend) _add_new_group(group) parallel_helper._set_parallel_ctx(True) paddle.distributed.barrier(group=group) return group node_num = {i.split(":")[0] for i in parallel_env.trainer_endpoints} # 3: init gloo context (step 1: httpsever start) init_gloo = int(os.getenv("PADDLE_WITH_GLOO", "0")) if is_cpu_only or init_gloo or backend == "heter": ep_rank_0 = parallel_env.trainer_endpoints[0].split(":") manager = Manager() # glboal dict to store status http_server_d = manager.dict() http_server_d["running"] = False if parallel_env.rank == 0: # The scope for worker used by http server is '_worker' size = {'_worker': parallel_env.world_size} if backend == "heter": size = {'_worker': len(node_num)} http_server = Process( target=_start_kv_server, args=(int(ep_rank_0[1]), http_server_d, size), ) http_server.daemon = True http_server_d["running"] = True http_server.start() # 4. init NCCL ParallelStrategy strategy = ParallelStrategy() if parallel_helper._is_parallel_ctx_initialized(): warnings.warn("The parallel environment has been initialized.") strategy.nranks = parallel_env.world_size strategy.local_rank = parallel_env.rank strategy.trainer_endpoints = parallel_env.trainer_endpoints strategy.current_endpoint = parallel_env.current_endpoint strategy.nrings = parallel_env.nrings # init nccl or bkcl or heter context if is_cpu_only: parallel_helper._set_parallel_ctx( core.GLOOParallelContext(strategy, place) ) elif backend == "heter": parallel_helper._set_parallel_ctx( core.HeterParallelContext(strategy, parallel_env.device_id) ) elif core.is_compiled_with_cuda(): parallel_helper._set_parallel_ctx( core.NCCLParallelContext(strategy, place) ) elif core.is_compiled_with_xpu(): parallel_helper._set_parallel_ctx( core.BKCLParallelContext(strategy, place) ) if backend != "heter": other_endpoints = strategy.trainer_endpoints[:] other_endpoints.remove(strategy.current_endpoint) if not is_cpu_only and strategy.local_rank == 0: wait_server_ready(other_endpoints) parallel_helper._init_parallel_ctx() # 5: init gloo context (step 2: gloo init) # dividing init_gloo into two part beacause nccl and gloo # are separately looking for free ports which sometimes # leads to port-conflict. if (is_cpu_only or backend == "heter") and parallel_env.rank == 0: # compare to init_gloo, we don't need to # init gloo, because we do this in _init_parallel_ctx; http_server_d["running"] = False http_server.join() elif init_gloo: wait_server_ready([parallel_env.trainer_endpoints[0]]) gloo_strategy = core.GlooParallelStrategy() gloo_strategy.rank = parallel_env.rank gloo_strategy.rank_num = parallel_env.world_size gloo_strategy.ip_address = ep_rank_0[0] gloo_strategy.ip_port = int(ep_rank_0[1]) default_init_timeout_seconds = 3600 default_run_timeout_seconds = 9999999 gloo_strategy.init_seconds = default_init_timeout_seconds gloo_strategy.run_seconds = default_run_timeout_seconds gloo = core.GlooParallelContext(gloo_strategy) gloo.init() if parallel_env.rank == 0: http_server_d["running"] = False http_server.join() return group def get_rank(group=None): """ Returns the rank of current trainer in the given group, ranks are consecutive integers in [0, ``world_size``). If none of the group is given, the global group will be used as default. Args: group (Group, optional): The communication group you want to get rank of current trainer, use global group as default if group is None. Returns: (int) The rank of current trainer in the given group. Return -1 if the process is not part of the given group. Warning: Argument ``group`` only supports in dygraph mode. Examples: .. code-block:: python # Execute this script using distributed launch with one card configs. import paddle import paddle.distributed as dist dist.init_parallel_env() print("The rank is %d" % dist.get_rank()) # The rank is 0 """ if in_dygraph_mode() and group: return group.rank assert group is None, "Only support group argument in eager mode." return _get_global_parallel_env().rank def get_world_size(group=None): """ Returns the number of trainers (number of processes participating in current job) in the given group. If none of the group is given, the global group will be used as default. Args: group (Group, optional): The communication group you want to check world size, use global group as default if group is None. Returns: (int) The number of trainers in the given group. Return -1 if the process if not part of the given group. Warning: Argument ``group`` only supports in dygraph mode. Examples: .. code-block:: python # Execute this script using distributed launch with one card configs. import paddle import paddle.distributed as dist dist.init_parallel_env() print("The world_size is %d" % dist.get_world_size()) # The world_size is 1 """ if in_dygraph_mode() and group: return group.world_size assert group is None, "Only support group argument in eager mode." return _get_global_parallel_env().world_size