import itertools import collections.abc import contextlib import hashlib import io import logging import os import pickle import time import warnings from collections import namedtuple from datetime import timedelta from typing import Any, Callable, Dict, Optional, Tuple, Union, List import torch from torch._C._distributed_c10d import ( AllreduceCoalescedOptions, AllreduceOptions, AllToAllOptions, _DistributedBackendOptions, BarrierOptions, BroadcastOptions, GatherOptions, PrefixStore, ProcessGroup, ReduceOp, ReduceOptions, ReduceScatterOptions, ScatterOptions, Store, DebugLevel, get_debug_level, Work ) from .constants import default_pg_timeout from .c10d_logger import _exception_logger, _time_logger from .rendezvous import register_rendezvous_handler, rendezvous # noqa: F401 __all__ = [ 'Backend', 'BackendConfig', 'GroupMember', 'P2POp', 'all_gather', 'all_gather_coalesced', 'all_gather_multigpu', 'all_gather_object', 'all_reduce', 'all_reduce_coalesced', 'all_reduce_multigpu', 'all_to_all', 'all_to_all_single', 'barrier', 'batch_isend_irecv', 'broadcast', 'broadcast_multigpu', 'broadcast_object_list', 'destroy_process_group', 'gather', 'gather_object', 'get_backend_config', 'get_backend', 'get_rank', 'get_world_size', 'group', 'init_process_group', 'irecv', 'is_gloo_available', 'is_initialized', 'is_mpi_available', 'is_backend_available', 'is_nccl_available', 'is_torchelastic_launched', 'is_ucc_available', 'isend', 'monitored_barrier', 'new_group', 'new_subgroups', 'new_subgroups_by_enumeration', 'recv', 'reduce', 'reduce_multigpu', 'reduce_scatter', 'reduce_scatter_multigpu', 'scatter', 'scatter_object_list', 'send', 'supports_complex', 'AllreduceCoalescedOptions', 'AllreduceOptions', 'AllToAllOptions', 'BarrierOptions', 'BroadcastOptions', 'GatherOptions', 'PrefixStore', 'ProcessGroup', 'ReduceOp', 'ReduceOptions', 'ReduceScatterOptions', 'ScatterOptions', 'Store', 'DebugLevel', 'get_debug_level', 'Work', 'default_pg_timeout', 'get_group_rank', 'get_global_rank', 'get_process_group_ranks', 'reduce_op', 'all_gather_into_tensor', 'reduce_scatter_tensor', ] _MPI_AVAILABLE = True _NCCL_AVAILABLE = True _GLOO_AVAILABLE = True _UCC_AVAILABLE = True _pickler = pickle.Pickler _unpickler = pickle.Unpickler # Change __module__ of all imported types from torch._C._distributed_c10d that are public def _export_c_types(): _public_types_to_change_module = [ AllreduceCoalescedOptions, AllreduceOptions, AllToAllOptions, BarrierOptions, BroadcastOptions, GatherOptions, PrefixStore, ProcessGroup, ReduceOp, ReduceOptions, ReduceScatterOptions, ScatterOptions, Store, DebugLevel, get_debug_level, Work ] for type in _public_types_to_change_module: type.__module__ = "torch.distributed.distributed_c10d" _export_c_types() try: from torch._C._distributed_c10d import ProcessGroupMPI ProcessGroupMPI.__module__ = "torch.distributed.distributed_c10d" __all__ += ["ProcessGroupMPI"] except ImportError: _MPI_AVAILABLE = False try: from torch._C._distributed_c10d import ProcessGroupNCCL ProcessGroupNCCL.__module__ = "torch.distributed.distributed_c10d" __all__ += ["ProcessGroupNCCL"] except ImportError: _NCCL_AVAILABLE = False try: from torch._C._distributed_c10d import ProcessGroupGloo from torch._C._distributed_c10d import _ProcessGroupWrapper ProcessGroupGloo.__module__ = "torch.distributed.distributed_c10d" __all__ += ["ProcessGroupGloo"] except ImportError: _GLOO_AVAILABLE = False try: from torch._C._distributed_c10d import ProcessGroupUCC ProcessGroupUCC.__module__ = "torch.distributed.distributed_c10d" __all__ += ["ProcessGroupUCC"] except ImportError: _UCC_AVAILABLE = False logger = logging.getLogger(__name__) PG_WRAPPER_STORE_PREFIX = "pg_wrapper" # Some reduce ops are not supported by complex numbers and will result in an error. # We currently provide complex support to the distributed API by viewing # complex tensors as real (torch.view_as_real), meaning that calling # these unsupported ops will return garbage values rather than error out. # (e.g. max(2+3i, 3+2i) = 3+3i) # We'd like calls to unsupported ops to error out accordingly, # rather than returning garbage values. def supports_complex(reduceOp: ReduceOp) -> bool: denyList = [ ReduceOp.MAX, ReduceOp.MIN, ReduceOp.PRODUCT, ReduceOp.BAND, ReduceOp.BOR, ReduceOp.BXOR, ] return reduceOp not in denyList class Backend: """ An enum-like class of available backends: GLOO, NCCL, UCC, MPI, and other registered backends. The values of this class are lowercase strings, e.g., ``"gloo"``. They can be accessed as attributes, e.g., ``Backend.NCCL``. This class can be directly called to parse the string, e.g., ``Backend(backend_str)`` will check if ``backend_str`` is valid, and return the parsed lowercase string if so. It also accepts uppercase strings, e.g., ``Backend("GLOO")`` returns ``"gloo"``. .. note:: The entry ``Backend.UNDEFINED`` is present but only used as initial value of some fields. Users should neither use it directly nor assume its existence. """ UNDEFINED = "undefined" GLOO = "gloo" NCCL = "nccl" UCC = "ucc" MPI = "mpi" _BackendPlugin = namedtuple("_BackendPlugin", ["creator_fn", "extended_api"]) _plugins: Dict[str, _BackendPlugin] = {} backend_list = [UNDEFINED, GLOO, NCCL, UCC, MPI] default_device_backend_map: Dict[str, str] = { 'cpu' : GLOO, 'cuda' : NCCL, } backend_capability: Dict[str, List[str]] = { GLOO : ["cpu", "cuda"], NCCL : ["cuda"], UCC : ["cpu", "cuda"], MPI : ["cpu"], } backend_type_map: Dict[str, ProcessGroup.BackendType] = { UNDEFINED: ProcessGroup.BackendType.UNDEFINED, GLOO : ProcessGroup.BackendType.GLOO, NCCL: ProcessGroup.BackendType.NCCL, UCC: ProcessGroup.BackendType.UCC, } def __new__(cls, name: str): if not isinstance(name, str): raise ValueError(f"Backend name must be a string, but got: {name}") value = getattr(Backend, name.upper(), Backend.UNDEFINED) if value == Backend.UNDEFINED: value = name.lower() return value @classmethod def register_backend(cls, name, func, extended_api=False, devices: Optional[Union[str, List[str]]] = None): """ Registers a new backend with the given name and instantiating function. This class method is used by 3rd party ``ProcessGroup`` extension to register new backends. Args: name (str): Backend name of the ``ProcessGroup`` extension. It should match the one in ``init_process_group()``. func (function): Function handler that instantiates the backend. The function should be implemented in the backend extension and takes four arguments, including ``store``, ``rank``, ``world_size``, and ``timeout``. extended_api (bool, optional): Whether the backend supports extended argument structure. Default: ``False``. If set to ``True``, the backend will get an instance of ``c10d::DistributedBackendOptions``, and a process group options object as defined by the backend implementation. device (str or list of str, optional): device type this backend supports, e.g. "cpu", "cuda", etc. If `None`, assuming both "cpu" and "cuda" .. note:: This support of 3rd party backend is experimental and subject to change. """ # Allow UCC plugin if Pytorch is not built with native support. # TODO: remove this exception once UCC plugin is fully deprecated. if (name != Backend.UCC or (name == Backend.UCC and is_ucc_available())): assert not hasattr(Backend, name.upper()), ( f"{name.upper()} c10d backend already exist" ) assert name.upper() not in Backend._plugins, ( f"{name.upper()} c10d backend creator function already exist" ) setattr(Backend, name.upper(), name.lower()) Backend.backend_list.append(name.lower()) if devices is not None: for device in devices: if device != 'cpu' and device != 'cuda': Backend.default_device_backend_map[device] = name.lower() Backend.backend_type_map[name.lower()] = ProcessGroup.BackendType.CUSTOM # Update device capability matrix in Backend class if devices is None: # This is more of a backward support for groups like `threaded`: # assume default devices "cpu" and "cuda", but warn warnings.warn( f"Device capability of {name} unspecified, assuming `cpu` and " "`cuda`. Please specify it via the `devices` argument of " "`register_backend`." ) Backend.backend_capability[name.lower()] = ["cpu", "cuda"] elif isinstance(devices, str): # Single device string specified. Simply convert to list. Backend.backend_capability[name.lower()] = [devices] else: Backend.backend_capability[name.lower()] = devices Backend._plugins[name.upper()] = Backend._BackendPlugin(func, extended_api) class BackendConfig: def __init__(self, backend: Union[str, Backend]): self.device_backend_map: Dict[torch.device, Backend] = {} if backend == Backend.UNDEFINED: # default config when backend is not specified # supported since PyTorch 2.0 for device in Backend.default_device_backend_map: if is_backend_available(Backend.default_device_backend_map[device]): self.device_backend_map[device] = Backend.default_device_backend_map[device] elif backend.lower() in Backend.backend_list: # Cases for when backend is a single string (without device types) # e.g. "nccl", "gloo", "ucc", "mpi" supported_devices = Backend.backend_capability[backend.lower()] backend_val = Backend(backend) self.device_backend_map = { device : backend_val for device in supported_devices } elif ":" in backend.lower(): # Backend specified in "device:backend" format # make sure the backend string is in the correct format # "{device_type1}:{backend1},{device_type2}:{backend2}" # e.g. "cpu:gloo,cuda:nccl" backend_str_error_message = f"""The custom backend string argument is invalid: {backend}. Custom backend string is an experimental feature where the backend string must be in the format: ":,:...". e.g. 'cpu:gloo,cuda:nccl'""" # parse the backend string and populate the device_backend_map for device_backend_pair_str in backend.lower().split(","): device_backend_pair = device_backend_pair_str.split(":") if len(device_backend_pair) != 2: raise ValueError(f"Invalid device:backend pairing: \ {device_backend_pair_str}. {backend_str_error_message}") device, backend = device_backend_pair if device in self.device_backend_map: raise ValueError(f"Duplicate device type {device} \ in backend string: {backend}. {backend_str_error_message}") self.device_backend_map[device] = Backend(backend) else: # User specified a single backend name whose device capability is # unknown, assuming it can support the default devices of PyTorch # (cpu and cuda) warnings.warn( f"Device capability of {backend} unknown, assuming `cpu` and " "`cuda`. You can specify it in `device:backend` format in " "`init_process_group` call." ) backend_val = Backend(backend) self.device_backend_map = { "cpu" : backend_val, "cuda" : backend_val, "xpu" : backend_val, } logger.info( f"Using backend config: {self.device_backend_map}" # noqa: G004 ) def __repr__(self): # string with all the device:backend pairs separated by commas return ",".join(f"{device}:{backend}" for device, backend in self.device_backend_map.items()) def get_device_backend_map(self): return self.device_backend_map class _reduce_op: r""" Deprecated enum-like class for reduction operations: ``SUM``, ``PRODUCT``, ``MIN``, and ``MAX``. :class:`~torch.distributed.ReduceOp` is recommended to use instead. """ def __init__(self): # __members__ is a dict storing key-value pairs for enum classes for k, v in ReduceOp.RedOpType.__members__.items(): setattr(self, k, v) self.__members__ = ReduceOp.RedOpType.__members__ def __getattribute__(self, key): warnings.warn( "torch.distributed.reduce_op is deprecated, please use " "torch.distributed.ReduceOp instead" ) return object.__getattribute__(self, key) reduce_op = _reduce_op() class P2POp: """ A class to build point-to-point operations for ``batch_isend_irecv``. This class builds the type of P2P operation, communication buffer, peer rank, Process Group, and tag. Instances of this class will be passed to ``batch_isend_irecv`` for point-to-point communications. Args: op (Callable): A function to send data to or receive data from a peer process. The type of ``op`` is either ``torch.distributed.isend`` or ``torch.distributed.irecv``. tensor (Tensor): Tensor to send or receive. peer (int): Destination or source rank. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. tag (int, optional): Tag to match send with recv. """ def __init__(self, op: Callable, tensor: torch.Tensor, peer: int, group: Optional[ProcessGroup] = None, tag: int = 0): self.op = op self.tensor = tensor self.peer = peer self.group = group self.tag = tag def __new__(cls, op: Callable, tensor: torch.Tensor, peer: int, group: Optional[ProcessGroup] = None, tag: int = 0): _check_op(op) _check_single_tensor(tensor, "tensor") return object.__new__(cls) class _CollOp: """ A class to capture collective operations. Args: op (Callable): A collective function, e.g. ``torch.distributed.all_reduce``. tensor (Tensor): Tensor to operate on. dst_tensor (Tensor, optional): Provided when source and destinaton tensors are not the same. redop (ReduceOp, optional): reduce operation. root (int, optional): root of broadcast or reduce. """ def __init__(self, op: Callable, tensor: torch.Tensor, dst_tensor: Optional[torch.Tensor] = None, redop: Optional[ReduceOp] = None, root: Optional[int] = None): self.op = op self.tensor = tensor self.dst_tensor = dst_tensor self.redop = redop self.root = root # DO NOT USE THESE FIELDS DIRECTLY. # Use them through the _world object to make sure the _world override mechanism _pg_map: Dict[ProcessGroup, Tuple[str, Optional[Store]]] = {} _pg_names: Dict[ProcessGroup, str] = {} _pg_group_ranks: Dict[ProcessGroup, Dict[int, int]] = {} # For a pg, it is a map from ProcessGroup to BackendConfig _pg_backend_config: Dict[ProcessGroup, str] = {} _group_count = 0 _tags_to_pg: Dict[str, List[ProcessGroup]] = {} _pg_to_tag: Dict[ProcessGroup, str] = {} class _World: """ Container class for c10d process group state. This is used during registration and lookup of PG state. .. warning:: This is an experimental API intended to expose the inner workings of c10d and is subject to change.. """ def __init__(self): self._default_pg = None self._pg_coalesce_state: Dict[ProcessGroup, List[Union[_CollOp, P2POp]]] = {} self._pg_default_device: Dict[ProcessGroup, torch.device] = {} @property def default_pg(self): """ The default ProcessGroup includes all ranks of the cluster. This is used by c10d APIs when a ProcessGroup is needed but None is provided. """ return self._default_pg @default_pg.setter def default_pg(self, value): self._default_pg = value @property def pg_map(self) -> Dict[ProcessGroup, Tuple[str, Optional[Store]]]: """ Cached process groups For NCCL and GLOO pg, it is a map from ProcessGroup to (Backend, Store) For MPI pg, it is a map from ProcessGroup to (Backend, None) TODO don't expose the map, expose fine grained ops """ global _pg_map return _pg_map @property def pg_names(self) -> Dict[ProcessGroup, str]: """ Process group's names, map from ProcessGroup to str. TODO don't expose the map, expose fine grained ops """ global _pg_names return _pg_names @property def pg_group_ranks(self) -> Dict[ProcessGroup, Dict[int, int]]: """ Process group's global rank to local rank mapping TODO don't expose the map, expose fine grained ops """ global _pg_group_ranks return _pg_group_ranks @property def pg_backend_config(self) -> Dict[ProcessGroup, str]: """ Process group's backend config TODO don't expose the map, expose fine grained ops """ global _pg_backend_config return _pg_backend_config @property def group_count(self) -> int: """ Process group count for default naming. TODO don't expose group_count, use something else instead """ global _group_count return _group_count @group_count.setter def group_count(self, value): """ Count is used when computing the name of ProcessGroups when using global synchronization. """ global _group_count _group_count = value @property def tags_to_pg(self) -> Dict[str, List[ProcessGroup]]: global _tags_to_pg return _tags_to_pg @property def pg_to_tag(self) -> Dict[ProcessGroup, str]: global _pg_to_tag return _pg_to_tag @property def pg_coalesce_state(self) -> Dict[ProcessGroup, List[Union[_CollOp, P2POp]]]: return self._pg_coalesce_state @property def pg_default_device(self) -> Dict[ProcessGroup, torch.device]: return self._pg_default_device @property def pg_config_info(self) -> List[Dict[str, Union[int, str]]]: """ Returns a list of dict with process groups and backends with their unique IDs and configurations (types and ranks). """ config_info = [] for pg, backend in self.pg_map.items(): # backend is a tuple with the first element being the backend type ("nccl", etc.) backend_type = Backend.backend_type_map[backend[0]] config_info.append( { "pg_id": pg._id(), "backend_id": pg._backend_id(backend_type), "backend_config": self.pg_backend_config[pg], "ranks": self.pg_group_ranks[pg], } ) return config_info _world = _World() """Holds the singleton instance of ``_World`` used by c10. Experimental extension point to override it""" class _WorldMeta(type): """ Meta class of ``group`` and ``GroupMember`` so they can have the class property ``WORLD``. """ # Points to the default PG once initialized. @property def WORLD(cls) -> Optional[ProcessGroup]: return _world.default_pg @WORLD.setter def WORLD(cls, pg: Optional[ProcessGroup]): _world.default_pg = pg class group(metaclass=_WorldMeta): pass class GroupMember(metaclass=_WorldMeta): NON_GROUP_MEMBER = -100 # Default process group state _default_pg_init_method = None STORE_BASED_BARRIER_PREFIX = "store_based_barrier_key" def _get_pg_default_device(group: Optional[ProcessGroup] = None): """ Returns the device to use with ``group`` for control flow usage (object collectives, barrier). There are selection rules: 1. If user specifies exactly one backend in ``init_process_group`` call: use that backend 2. Else if user specifies multiple "device:backend" pairs in init_process_group: If "cpu" is among those pairs, use "cpu" (because the object is in cpu memory); Otherwise, use the first backend (sort of a random pick). Args: group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Returns: torch.device: The device to use with ``group``. """ group = group or _get_default_group() if group in _world.pg_default_device: # Previously searched and cached; just return return _world.pg_default_device[group] if not isinstance(group, ProcessGroup): # Provide backward compatibility to cases where `group` passed in is # actually a Backend (like `ProcessGroupGloo`) rather than a # `ProcessGroup` in PT 2.0 sense warnings.warn( f"You are using a Backend {type(group)} as a ProcessGroup. " "This usage is deprecated since PyTorch 2.0. Please use a public API " "of PyTorch Distributed instead." ) # Most users create Gloo with private API for object collectives _world.pg_default_device[group] = torch.device("cpu") return _world.pg_default_device[group] """ ``group._device_types`` is a property pybind that returns the devices ("cpu", "cuda", etc) supported by ``group``. Can be multiple if the ``group`` supports multiple devices. """ devices = group._device_types if len(devices) == 1: # User fixed exactly one backend in `init_process_group` _world.pg_default_device[group] = devices[0] elif len(devices) == 0: # No backend has been registered with this PG (maybe because no # collective has been run?) We pick cpu as the default and hopefully # this would lazily init Gloo or other available cpu backend. _world.pg_default_device[group] = torch.device("cpu") elif torch.device("cpu") in devices: # There are multiple backends in this PG and cpu is among them. # cpu is preferred as the object is in cpu memory. No need for device # copy. _world.pg_default_device[group] = torch.device("cpu") else: # No cpu in the backend list. Randomly pick the first backend _world.pg_default_device[group] = devices[0] logger.info( f"Using device {_world.pg_default_device[group]} for object " # noqa: G004 "collectives." ) return _world.pg_default_device[group] @_time_logger def _store_based_barrier(rank, store, group_name, rendezvous_count, timeout, logging_interval=timedelta(seconds=10)): """ Barrier based on store which is used for synchronizing processes after ``init_process_group`` or ``new_group``. Intended to be used only with those two methods and is not a generic alternative to ``barrier()``. """ store_key = f"{STORE_BASED_BARRIER_PREFIX}:{group_name}" store.add(store_key, 1) logger.info("Added key: %s to store for rank: %s", store_key, rank) # Now wait for all workers to check in with the store. world_size = rendezvous_count worker_count = store.add(store_key, 0) last_worker_key = f"{store_key}:last_worker" if worker_count == world_size: store.set(last_worker_key, "1") start = time.time() while True: try: # This will throw an exception after the logging_interval in which we print out # the status of the group or time out officially, throwing runtime error store.wait([last_worker_key], logging_interval) break except RuntimeError as e: worker_count = store.add(store_key, 0) # Print status periodically to keep track. logger.info( "Waiting in store based barrier to initialize process group for " "rank: %s, key: %s (world_size=%s, num_workers_joined=%s, timeout=%s)", rank, store_key, world_size, worker_count, timeout ) if timedelta(seconds=(time.time() - start)) > timeout: raise RuntimeError( "Timed out initializing process group in store based barrier on " "rank {}, for key: {} (world_size={}, num_workers_joined={}, timeout={})".format( rank, store_key, world_size, worker_count, timeout ) ) logger.info( "Rank %s: Completed store-based barrier for key:%s with %s nodes.", rank, store_key, world_size ) def _rank_not_in_group(group: ProcessGroup): """ Helper that checks if the current process's rank is not in a given group. """ if group is None: return False return group == GroupMember.NON_GROUP_MEMBER def _warn_not_in_group(op_name): global_rank = -1 if GroupMember.WORLD is None else GroupMember.WORLD.rank() warnings.warn( f"Running {op_name} on global rank {global_rank} which does not " "belong to the given group." ) def get_group_rank(group: ProcessGroup, global_rank: int) -> int: """ Translate a global rank into a group rank. ``global_rank`` must be part of ``group`` otherwise this raises RuntimeError. Args: group (ProcessGroup): ProcessGroup to find the relative rank. global_rank (int): Global rank to query. Returns: Group rank of ``global_rank`` relative to ``group`` N.B. calling this function on the default process group returns identity """ if group is GroupMember.WORLD: return global_rank if group not in _world.pg_group_ranks: raise RuntimeError(f"Group {group} is not registered, please create group with torch.distributed.new_group API") group_ranks = _world.pg_group_ranks[group] if global_rank not in group_ranks: raise RuntimeError(f"Global rank {global_rank} is not part of group {group}") return group_ranks[global_rank] def get_global_rank(group: ProcessGroup, group_rank: int) -> int: """ Translate a group rank into a global rank. ``group_rank`` must be part of `group` otherwise this raises RuntimeError. Args: group (ProcessGroup): ProcessGroup to find the global rank from. group_rank (int): Group rank to query. Returns: Global rank of ``group_rank`` relative to ``group`` N.B. calling this function on the default process group returns identity """ if group is GroupMember.WORLD: return group_rank if group not in _world.pg_group_ranks: raise RuntimeError(f"Group {group} is not registered, please create group with torch.distributed.new_group API") for rank, grp_rank in _world.pg_group_ranks[group].items(): if grp_rank == group_rank: return rank raise RuntimeError(f"Group rank {group_rank} is not part of group {group}") # TODO: remove this once the ecosystem moves away from it. def _get_global_rank(group, rank): """ This method is deprecated, please use get_global_rank. """ warnings.warn( "torch.distributed.distributed_c10d._get_global_rank is deprecated " "please use torch.distributed.distributed_c10d.get_global_rank instead" ) return get_global_rank(group, rank) def get_process_group_ranks(group: ProcessGroup): """ Get all ranks associated with ``group``. Args: group (ProcessGroup): ProcessGroup to get all ranks from. Returns: List of global ranks ordered by group rank. """ return list(_world.pg_group_ranks[group].keys()) def _get_group_size(group): """ Helper that gets a given group's world size. """ if group is GroupMember.WORLD or group is None: default_pg = _get_default_group() return default_pg.size() return group.size() def _check_single_tensor(param, param_name): """ Helper to check that the parameter ``param_name`` is a single tensor. """ if not isinstance(param, torch.Tensor): raise RuntimeError( f"Invalid function argument. Expected parameter `{param_name}` to be of type torch.Tensor." ) def _check_tensor_list(param, param_name): """ Helper to check that the parameter ``param_name`` is a list of tensors. """ if not isinstance(param, list) or not all( isinstance(p, torch.Tensor) for p in param ): raise RuntimeError( f"Invalid function argument. Expected parameter `{param_name}` to be of type List[torch.Tensor]." ) def _as_iterable(obj) -> collections.abc.Iterable: return obj if isinstance(obj, list) else (obj,) def _ensure_all_tensors_same_dtype(*tensors) -> None: last_dtype = None for tensor in itertools.chain(*map(_as_iterable, tensors)): tensor_dtype = tensor.dtype # Mixing complex and its element type is allowed if tensor_dtype.is_complex: tensor_dtype = torch.float32 if tensor_dtype == torch.complex64 else torch.complex128 if last_dtype is None: last_dtype = tensor_dtype else: if last_dtype != tensor_dtype: raise RuntimeError( "Invalid usage of tensors with different dtypes" f"Found {last_dtype} and {tensor.dtype}" ) def _check_op(op): """ Helper to check that the ``op`` is either isend or irecv. """ if op not in [isend, irecv]: raise RuntimeError( "Invalid ``op``. Expected ``op`` " "to be of type ``torch.distributed.isend`` or " "``torch.distributed.irecv``." ) def _check_p2p_op_list(p2p_op_list): """ Helper to check that the ``p2p_op_list`` is a list of P2POp instances and all ops use the same group. """ if not isinstance(p2p_op_list, list) or not all( isinstance(p2p_op, P2POp) for p2p_op in p2p_op_list ): raise RuntimeError( "Invalid ``p2p_op_list``. Each op is expected to " "to be of type ``torch.distributed.P2POp``." ) group = p2p_op_list[0].group if not all(group == p2p_op.group for p2p_op in p2p_op_list): raise RuntimeError("All ops need to use the same group.") def is_mpi_available() -> bool: """ Checks if the MPI backend is available. """ return _MPI_AVAILABLE def is_nccl_available() -> bool: """ Checks if the NCCL backend is available. """ return _NCCL_AVAILABLE def is_gloo_available() -> bool: """ Checks if the Gloo backend is available. """ return _GLOO_AVAILABLE def is_ucc_available() -> bool: """ Checks if the UCC backend is available. """ return _UCC_AVAILABLE def is_backend_available(backend: str) -> bool: """ Checks if the given backend is available and supports the built-in backends or third-party backends through function ``Backend.register_backend``. Args: backend (str): Backend name. Returns: bool: Returns true if the backend is available otherwise false. """ # If the backend has an ``is_backend_available`` function, return the result of that function directly available_func = getattr(torch.distributed, f"is_{backend.lower()}_available", None) if available_func: return available_func() return backend.lower() in Backend.backend_list def is_initialized() -> bool: """ Checking if the default process group has been initialized """ return GroupMember.WORLD is not None def is_torchelastic_launched() -> bool: """ Checks whether this process was launched with ``torch.distributed.elastic`` (aka torchelastic). The existence of ``TORCHELASTIC_RUN_ID`` environment variable is used as a proxy to determine whether the current process was launched with torchelastic. This is a reasonable proxy since ``TORCHELASTIC_RUN_ID`` maps to the rendezvous id which is always a non-null value indicating the job id for peer discovery purposes.. """ return os.getenv("TORCHELASTIC_RUN_ID") is not None def _is_barrier_after_init() -> int: # Environment variable to control whether process group should perform a # barrier after its init. Default value is 0, i.e. no barrier. If you # experience issue with this setting, you may set # `TORCH_DIST_INIT_BARRIER=1` to add the barrier. return int(os.getenv("TORCH_DIST_INIT_BARRIER", "0")) def _get_default_group(): """ Getting the default process group created by init_process_group """ if not is_initialized(): raise RuntimeError( "Default process group has not been initialized, " "please make sure to call init_process_group." ) return GroupMember.WORLD def _get_default_store(): """ Getting the default store created by init_process_group """ if not is_initialized(): raise RuntimeError( "Default process group has not been initialized, " "please make sure to call init_process_group." ) default_pg = _get_default_group() _, default_store = _world.pg_map[default_pg] return default_store def _update_default_pg(pg): _world.default_pg = pg def get_backend_config(group: Optional[ProcessGroup] = None) -> str: if group is None: pg = _get_default_group() else: pg = group if _rank_not_in_group(pg): raise RuntimeError("Invalid process group specified") backend_config = _world.pg_backend_config.get(pg) assert backend_config is not None return str(backend_config) def get_backend(group: Optional[ProcessGroup] = None) -> str: """ Returns the backend of the given process group. Args: group (ProcessGroup, optional): The process group to work on. The default is the general main process group. If another specific group is specified, the calling process must be part of :attr:`group`. Returns: The backend of the given process group as a lower case string. """ if group is None: pg = _get_default_group() else: pg = group if _rank_not_in_group(pg): raise RuntimeError("Invalid process group specified") pg_store = _world.pg_map[pg] if pg in _world.pg_map else None assert pg_store is not None return pg_store[0] _exception_logger @_time_logger def init_process_group( backend: Union[str, Backend] = None, init_method: Optional[str] = None, timeout: timedelta = default_pg_timeout, world_size: int = -1, rank: int = -1, store: Optional[Store] = None, group_name: str = "", pg_options: Optional[Any] = None, ): """ Initializes the default distributed process group, and this will also initialize the distributed package. There are 2 main ways to initialize a process group: 1. Specify ``store``, ``rank``, and ``world_size`` explicitly. 2. Specify ``init_method`` (a URL string) which indicates where/how to discover peers. Optionally specify ``rank`` and ``world_size``, or encode all required parameters in the URL and omit them. If neither is specified, ``init_method`` is assumed to be "env://". Args: backend (str or Backend, optional): The backend to use. Depending on build-time configurations, valid values include ``mpi``, ``gloo``, ``nccl``, and ``ucc``. If the backend is not provided, then both a ``gloo`` and ``nccl`` backend will be created, see notes below for how multiple backends are managed. This field can be given as a lowercase string (e.g., ``"gloo"``), which can also be accessed via :class:`Backend` attributes (e.g., ``Backend.GLOO``). If using multiple processes per machine with ``nccl`` backend, each process must have exclusive access to every GPU it uses, as sharing GPUs between processes can result in deadlocks. ``ucc`` backend is experimental. init_method (str, optional): URL specifying how to initialize the process group. Default is "env://" if no ``init_method`` or ``store`` is specified. Mutually exclusive with ``store``. world_size (int, optional): Number of processes participating in the job. Required if ``store`` is specified. rank (int, optional): Rank of the current process (it should be a number between 0 and ``world_size``-1). Required if ``store`` is specified. store(Store, optional): Key/value store accessible to all workers, used to exchange connection/address information. Mutually exclusive with ``init_method``. timeout (timedelta, optional): Timeout for operations executed against the process group. Default value equals 30 minutes. This is applicable for the ``gloo`` backend. For ``nccl``, this is applicable only if the environment variable ``NCCL_BLOCKING_WAIT`` or ``NCCL_ASYNC_ERROR_HANDLING`` is set to 1. When ``NCCL_BLOCKING_WAIT`` is set, this is the duration for which the process will block and wait for collectives to complete before throwing an exception. When ``NCCL_ASYNC_ERROR_HANDLING`` is set, this is the duration after which collectives will be aborted asynchronously and the process will crash. ``NCCL_BLOCKING_WAIT`` will provide errors to the user which can be caught and handled, but due to its blocking nature, it has a performance overhead. On the other hand, ``NCCL_ASYNC_ERROR_HANDLING`` has very little performance overhead, but crashes the process on errors. This is done since CUDA execution is async and it is no longer safe to continue executing user code since failed async NCCL operations might result in subsequent CUDA operations running on corrupted data. Only one of these two environment variables should be set. For ``ucc``, blocking wait is supported similar to NCCL. However, async error handling is done differently since with UCC we have progress thread and not watch-dog thread. group_name (str, optional, deprecated): Group name. This argument is ignored pg_options (ProcessGroupOptions, optional): process group options specifying what additional options need to be passed in during the construction of specific process groups. As of now, the only options we support is ``ProcessGroupNCCL.Options`` for the ``nccl`` backend, ``is_high_priority_stream`` can be specified so that the nccl backend can pick up high priority cuda streams when there're compute kernels waiting. .. note:: To enable ``backend == Backend.MPI``, PyTorch needs to be built from source on a system that supports MPI. .. note:: Support for multiple backends is experimental. Currently when no backend is specified, both ``gloo`` and ``nccl`` backends will be created. The ``gloo`` backend will be used for collectives with CPU tensors and the ``nccl`` backend will be used for collectives with CUDA tensors. A custom backend can be specified by passing in a string with format ":,:", e.g. "cpu:gloo,cuda:custom_backend". """ global _world global _backend global _default_pg_init_method if not isinstance(timeout, timedelta): raise RuntimeError( "Expected timeout argument to be of type datetime.timedelta" ) if GroupMember.WORLD is not None: raise RuntimeError("trying to initialize the default process group twice!") assert (store is None) or ( init_method is None ), "Cannot specify both init_method and store." if store is not None: assert world_size > 0, "world_size must be positive if using store" assert rank >= 0, "rank must be non-negative if using store" elif init_method is None: init_method = "env://" if backend: backend = Backend(backend) else: backend = Backend("undefined") """ Group name is not visible to users unless they access internals of c10d. This means we can ignore the value they provide as it not exposed in a public way. """ group_name = _process_group_name([], use_hashed_name=False) if backend == Backend.MPI: if world_size != -1 or rank != -1: warnings.warn( f"For MPI backend, world_size ({world_size}) and rank ({rank}) " "are ignored since they are assigned by the " "MPI runtime." ) default_pg, _ = _new_process_group_helper( -1, -1, [], backend, None, group_name=group_name, timeout=timeout ) _update_default_pg(default_pg) else: # backward compatible API if store is None: rendezvous_iterator = rendezvous( init_method, rank, world_size, timeout=timeout ) store, rank, world_size = next(rendezvous_iterator) store.set_timeout(timeout) # Use a PrefixStore to avoid accidental overrides of keys used by # different systems (e.g. RPC) in case the store is multi-tenant. store = PrefixStore("default_pg", store) default_pg, _ = _new_process_group_helper( world_size, rank, [], backend, store, pg_options=pg_options, group_name=group_name, timeout=timeout ) _update_default_pg(default_pg) _world.pg_group_ranks[GroupMember.WORLD] = {i: i for i in range(GroupMember.WORLD.size())} # type: ignore[attr-defined, index] _backend = _world.pg_map[GroupMember.WORLD][0] # type: ignore[index] _default_pg_init_method = init_method if _is_barrier_after_init() == 1: # barrier at the end to ensure that once we return from this method, all # process groups including global variables (if any) are updated # correctly on all ranks. # Update 04/2023: for large-scale runs, this barrier (esp. store-based # barrier) may be costly and/or unscalable. Also, in a lot of cases, # these barriers may be unnecessary, as proven by a green CI after # removal. An environment variable `TORCH_DIST_INIT_BARRIER` has been # added which enables this barrier only when set to 1. logger.info( "Performing barrier after ProcessGroup initialization since " "TORCH_DIST_INIT_BARRIER = 1" ) if backend == Backend.MPI: # MPI backend doesn't use store. barrier() else: # Use store based barrier here since barrier() used a bunch of # default devices and messes up NCCL internal state. _store_based_barrier(rank, store, group_name, world_size, timeout) def _new_process_group_helper( group_size, group_rank, global_ranks_in_group, backend, store, pg_options=None, group_name=None, timeout=default_pg_timeout, pg_tag=None ): """ Create a new distributed process group. This function must be called by ALL processes in the global group, even if the calling process is not part of the newly created group. In that case, this function returns GroupMember.NON_GROUP_MEMBER. This function is called with ``global_ranks_in_group == []`` for the default group. """ global _world if group_name in _world.pg_names.values(): raise RuntimeError( "The specified group name has already been " "created, please use a different group name" ) if not isinstance(timeout, timedelta): raise RuntimeError( "Expected timeout argument to be of type datetime.timedelta" ) if pg_tag not in [None, ""]: # creating with the same tag and rank set results in the same underlying PG existing_group = _find_pg_by_ranks_and_tag(pg_tag, global_ranks_in_group) if existing_group: _, prefix_store = _world.pg_map[existing_group] return existing_group, prefix_store # The list of group ranks is empty if we're creating the default group. is_default_group = len(global_ranks_in_group) == 0 # If this is a subgroup (which means group_ranks is specified), # we check if the current process is a member of the new group. if not is_default_group: global_rank = _get_default_group().rank() if global_rank not in global_ranks_in_group: return GroupMember.NON_GROUP_MEMBER, None prefix_store = PrefixStore(f"{group_name}/", store) base_pg_options = ProcessGroup.Options(backend=str(backend)) base_pg_options._timeout = timeout pg: ProcessGroup = ProcessGroup(prefix_store, group_rank, group_size, base_pg_options) backend_config = BackendConfig(backend) for device, backend_str in backend_config.get_device_backend_map().items(): # Use the group name as prefix in the default store, such that # a single store can be reused by multiple groups. backend_prefix_store = PrefixStore(f"{device}/", prefix_store) if backend_str == Backend.MPI: if not is_mpi_available(): raise RuntimeError( "Distributed package doesn't have MPI built in." " MPI is only included if you build PyTorch from" " source on a host that has MPI installed." ) backend_class = ProcessGroupMPI.create(global_ranks_in_group) backend_type = ProcessGroup.BackendType.MPI if not backend_class: return GroupMember.NON_GROUP_MEMBER # create new process group with accurate rank and size if pg.rank() == -1 and pg.size() == -1: pg = ProcessGroup(backend_prefix_store, backend_class.rank(), backend_class.size(), base_pg_options) elif backend_str == Backend.GLOO: # TODO: remove this check after lazy initialization is supported # if pg_options is not None: # raise RuntimeError("GLOO options not supported") backend_class = ProcessGroupGloo(backend_prefix_store, group_rank, group_size, timeout=timeout) backend_type = ProcessGroup.BackendType.GLOO elif backend_str == Backend.NCCL: if not is_nccl_available(): raise RuntimeError("Distributed package doesn't have NCCL built in") if pg_options is not None: assert isinstance( pg_options, ProcessGroupNCCL.Options ), "Expected pg_options argument to be of type ProcessGroupNCCL.Options" else: # default pg_options for NCCL pg_options = ProcessGroupNCCL.Options() pg_options.is_high_priority_stream = False pg_options._timeout = timeout backend_class = ProcessGroupNCCL(backend_prefix_store, group_rank, group_size, pg_options) backend_type = ProcessGroup.BackendType.NCCL elif backend_str == Backend.UCC and is_ucc_available(): # TODO: once UCC plugin is fully deprecated, remove # is_ucc_available() from above elif-condition and raise # RuntimeError if is_ucc_available() returns false. backend_class = ProcessGroupUCC(backend_prefix_store, group_rank, group_size, timeout=timeout) backend_type = ProcessGroup.BackendType.UCC else: assert backend_str.upper() in Backend._plugins, ( f"Unknown c10d backend type {backend_str.upper()}" ) backend_plugin = Backend._plugins[backend_str.upper()] creator_fn = backend_plugin.creator_fn extended_api = backend_plugin.extended_api backend_type = ProcessGroup.BackendType.CUSTOM if not extended_api: backend_class = creator_fn(backend_prefix_store, group_rank, group_size, timeout) else: dist_backend_opts = _DistributedBackendOptions() dist_backend_opts.store = backend_prefix_store dist_backend_opts.group_rank = group_rank dist_backend_opts.group_size = group_size dist_backend_opts.timeout = timeout dist_backend_opts.group_id = group_name dist_backend_opts.global_ranks_in_group = global_ranks_in_group backend_class = creator_fn(dist_backend_opts, pg_options) # Set sequence numbers for gloo and nccl backends. if backend_str in [Backend.GLOO, Backend.NCCL]: backend_class._set_sequence_number_for_group() # If the type is a subclass of ProcessGroup then return this process group immediately # TODO: This defaults to the old behavior for PythonProcessGroups which overwrites the # ProcessGroup instance if issubclass(type(backend_class), ProcessGroup): pg = backend_class break # Process group wrapper initialization for supported PGs when TORCH_DISTRIBUTED_DEBUG is set if backend_str in [Backend.GLOO, Backend.NCCL, Backend.UCC]: # In debug mode and if GLOO is available, wrap in a wrapper PG that # enables enhanced collective checking for debuggability. if get_debug_level() == DebugLevel.DETAIL: if not _GLOO_AVAILABLE: logger.info( """TORCH_DISTRIBUTED_DEBUG was set to DETAIL, but GLOO is not available. Build with Gloo to create a wrapper process group in debug mode to aid collective desynchronization debugging.""" ) else: backend_class = _create_process_group_wrapper( wrapped_pg=backend_class, store_prefix=group_name, store=backend_prefix_store, rank=group_rank, world_size=group_size, timeout=timeout, ) # register only a single backend when all get_device_backend_map values are the same if len(set(backend_config.get_device_backend_map().values())) == 1: for device in backend_config.get_device_backend_map().keys(): pg._register_backend(torch.device(device), backend_type, backend_class) # break out of outer loop to not create any more backends break pg._register_backend(torch.device(device), backend_type, backend_class) # update global state _world.pg_map[pg] = (backend, prefix_store) _world.pg_names[pg] = group_name _world.pg_backend_config[pg] = str(backend_config) # "" is the default tag for user PGs if pg_tag in [None, ""]: pg_tag = f"ptd:{group_name}" _world.tags_to_pg.setdefault("", []).append(pg) else: pg_tag = f"user:{pg_tag}" _world.tags_to_pg.setdefault(pg_tag, []).append(pg) _world.pg_to_tag[pg] = pg_tag return pg, prefix_store def destroy_process_group(group: Optional[ProcessGroup] = None): """ Destroy a given process group, and deinitialize the distributed package Args: group (ProcessGroup, optional): The process group to be destroyed, if group.WORLD is given, all process groups including the default one will be destroyed. """ global _world if group == GroupMember.NON_GROUP_MEMBER: return if group is None: pg = GroupMember.WORLD else: pg = group assert pg is not None if _world.pg_map.get(pg, None) is None: raise RuntimeError("Invalid process group specified") # When users register Python onCompletion hooks, those hooks will run on a # different thread than the main thread. Today, the ProcessGroup dtor does # wait for that thread. However, the dtor might finish after the Python # Interpreter exits. After that grabbing the GIL for the Python hook will crash. # We can either revive the interpreter when running hooks or keep the main one # alive until all works and hooks are done. The current implementation does the # latter. Therefore, we explicitly call _wait_for_pending_works() here to wait # for the pending hooks to finish. if pg.name().lower() == "nccl" and pg._has_hooks(): pg._wait_for_pending_works() if group is None or group == GroupMember.WORLD: _update_default_pg(None) _world.pg_map.clear() _world.pg_names.clear() _world.pg_group_ranks.clear() _world.pg_backend_config.clear() _world.pg_to_tag.clear() _world.tags_to_pg.clear() _world.pg_coalesce_state.clear() _world.pg_default_device.clear() # when process group doesn't have an explicit name (only WORLD (default) # process group can have an explicit name), we use global _world.group_count # to generate the name. We need to reset the counter on destruction to # allow consistent value to be generated when we re-create process # groups after some trainers recover from failure # # We only reset this when WORLD is being destroyed because if this # process group is in good state, we aren't dealing with failures. _world.group_count = 0 else: del _world.pg_map[pg] del _world.pg_names[pg] del _world.pg_group_ranks[pg] del _world.pg_backend_config[pg] if pg in _world.pg_default_device: del _world.pg_default_device[pg] if pg in _world.pg_coalesce_state.keys(): warnings.warn( "Some coalesced collectives haven't been launched when " "ProcessGroup is destroyed. They will be cleaned." ) del _world.pg_coalesce_state[pg] tag = _world.pg_to_tag.get(pg) del _world.pg_to_tag[pg] if tag is not None: try: _world.tags_to_pg[tag].remove(pg) if tag.startswith("ptd:"): _world.tags_to_pg[""].remove(pg) except Exception: pass def get_rank(group: Optional[ProcessGroup] = None) -> int: """ Returns the rank of the current process in the provided ``group`` or the default group if none was provided. Rank is a unique identifier assigned to each process within a distributed process group. They are always consecutive integers ranging from 0 to ``world_size``. Args: group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Returns: The rank of the process group -1, if not part of the group """ if _rank_not_in_group(group): return -1 default_pg = _get_default_group() if group is None or group is GroupMember.WORLD: return default_pg.rank() return get_group_rank(group, default_pg.rank()) def get_world_size(group: Optional[ProcessGroup] = None) -> int: """ Returns the number of processes in the current process group Args: group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Returns: The world size of the process group -1, if not part of the group """ if _rank_not_in_group(group): return -1 return _get_group_size(group) def isend(tensor: torch.Tensor, dst: int, group: Optional[ProcessGroup] = None, tag: int = 0) -> Work: """ Sends a tensor asynchronously. .. warning:: Modifying ``tensor`` before the request completes causes undefined behavior. .. warning:: ``tag`` is not supported with the NCCL backend. Args: tensor (Tensor): Tensor to send. dst (int): Destination rank. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. tag (int, optional): Tag to match send with remote recv Returns: A distributed request object. None, if not part of the group """ _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("isend") return if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() return default_pg.send([tensor], dst, tag) else: group_dst_rank = get_group_rank(group, dst) return group.send([tensor], group_dst_rank, tag) def irecv(tensor: torch.Tensor, src: Optional[int] = None, group: Optional[ProcessGroup] = None, tag: int = 0) -> Work: """ Receives a tensor asynchronously. .. warning:: ``tag`` is not supported with the NCCL backend. Args: tensor (Tensor): Tensor to fill with received data. src (int, optional): Source rank. Will receive from any process if unspecified. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. tag (int, optional): Tag to match recv with remote send Returns: A distributed request object. None, if not part of the group """ _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("irecv") return if group is None or group is GroupMember.WORLD: pg = _get_default_group() else: pg = group if src is None: return pg.recv_anysource([tensor], tag) else: if pg is GroupMember.WORLD: return pg.recv([tensor], src, tag) else: group_src_rank = get_group_rank(pg, src) return pg.recv([tensor], group_src_rank, tag) @_exception_logger def send(tensor: torch.Tensor, dst: int, group: Optional[ProcessGroup] = None, tag: int = 0) -> None: """ Sends a tensor synchronously. Args: tensor (Tensor): Tensor to send. dst (int): Destination rank. Destination rank should not be the same as the rank of the current process. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. tag (int, optional): Tag to match send with remote recv """ if get_rank() == dst: raise ValueError( "Invalid destination rank: destination rank should not be the same as " "the rank of the current process." ) _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("send") return if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() default_pg.send([tensor], dst, tag).wait() else: group_dst_rank = get_group_rank(group, dst) group.send([tensor], group_dst_rank, tag).wait() @_exception_logger def recv(tensor: torch.Tensor, src: Optional[int] = None, group: Optional[ProcessGroup] = None, tag: int = 0) -> int: """ Receives a tensor synchronously. Args: tensor (Tensor): Tensor to fill with received data. src (int, optional): Source rank. Will receive from any process if unspecified. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. tag (int, optional): Tag to match recv with remote send Returns: Sender rank -1, if not part of the group """ _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("recv") return -1 if group is None: pg = _get_default_group() else: pg = group if src is None: work = pg.recv_anysource([tensor], tag) work.wait() src_rank = work._source_rank() if group is None or group is GroupMember.WORLD: return src_rank else: return get_global_rank(pg, src_rank) else: if group is None or group is GroupMember.WORLD: pg.recv([tensor], src, tag).wait() else: group_src_rank = get_group_rank(pg, src) pg.recv([tensor], group_src_rank, tag).wait() return src class _IllegalWork(Work): def __getattribute__(self, name): if name in ["is_success", "exception", "wait", "source_rank", "_source_rank", "result", "synchronize"]: raise RuntimeError(f"Illegal to call {name} on IllegalWork object") class _CoalescingManager: def __init__(self): self.works: List[Work] = [] def append(self, work: Work): if work: self.works.append(work) def wait(self): for work in self.works: work.wait() @contextlib.contextmanager def _coalescing_manager( group: Optional[ProcessGroup] = None, device: Optional[torch.device] = None, async_ops: Optional[bool] = False, ): """ A context manager used to coalesce collectives or P2P operations when possible. Args: group (`ProcessGroup`, optional): The process group to work on. If None, the default process group will be used. device (`torch.device`, optional): Default is None, set to a device if there isn't a `**_coalesced` implementation by the backend. async_ops (`bool`, optional): whether the coalesced ops are async ops. Examples: >>> # xdoctest: +SKIP("no rank") >>> # Synchronous ops >>> with _coalescing_manager(): >>> for i in range(num_colls): >>> dist.all_reduce(tensors[i]) >>> # Asynchronous ops >>> with _coalescing_manager(async_ops=True) as cm: >>> for i in range(num_colls): >>> dist.all_reduce(tensors[i]) >>> cm.wait() .. warning:: :func:`_coalescing_manager` currently do not support coalescing all-reduces with different reduce operators, e.g. `ReduceOp.SUM` mixed with `ReduceOp.PRODUCT`. """ group = group or _get_default_group() op_list = _world.pg_coalesce_state.setdefault(group, []) if op_list: raise RuntimeError("ProcessGroup has non-empty op list at the start of coalescing") if device: group._start_coalescing(device) cm = _CoalescingManager() yield cm op_list = _world.pg_coalesce_state.pop(group) if op_list: # Collectives supporting "Fast Path" coalescing are captured. # See implementation in corresponding collective APIs. # Currently supported: # - coalesced `all_reduce` # - coalesced `all_gather_into_tensor` # - coalesced `reduce_scatter_tensor` op0 = op_list[0].op if op0 == all_reduce: tensors = [] for op in op_list: tensors.append(op.tensor) opts = AllreduceCoalescedOptions() opts.reduceOp = op_list[0].redop work = group.allreduce_coalesced(tensors, opts) elif op0 == all_gather_into_tensor: inputs = [] outputs = [] for op in op_list: inputs.append(op.tensor) outputs.append(op.dst_tensor) work = group.allgather_into_tensor_coalesced(outputs, inputs) elif op0 == reduce_scatter_tensor: inputs = [] outputs = [] for op in op_list: inputs.append(op.tensor) outputs.append(op.dst_tensor) opts = ReduceScatterOptions() opts.reduceOp = op_list[0].redop work = group.reduce_scatter_tensor_coalesced(outputs, inputs, opts) else: raise AssertionError( f"Coalescing manager does not support fast-path coalescing of {op0}, " f"yet {op0} is still recorded in op list. This is an internal error of c10d." ) if device: # Old style of letting each coll inside the context manager to call into C++ counterpart via python binding work = group._end_coalescing(device) if async_ops: cm.append(work) else: work.wait() def batch_isend_irecv(p2p_op_list): """ Send or Receive a batch of tensors asynchronously and return a list of requests. Process each of the operations in ``p2p_op_list`` and return the corresponding requests. NCCL, Gloo, and UCC backend are currently supported. Args: p2p_op_list: A list of point-to-point operations(type of each operator is ``torch.distributed.P2POp``). The order of the isend/irecv in the list matters and it needs to match with corresponding isend/irecv on the remote end. Returns: A list of distributed request objects returned by calling the corresponding op in the op_list. Examples: >>> # xdoctest: +SKIP("no rank") >>> send_tensor = torch.arange(2) + 2 * rank >>> recv_tensor = torch.randn(2) >>> send_op = dist.P2POp(dist.isend, send_tensor, (rank + 1)%world_size) >>> recv_op = dist.P2POp(dist.irecv, recv_tensor, (rank - 1 + world_size)%world_size) >>> reqs = batch_isend_irecv([send_op, recv_op]) >>> for req in reqs: >>> req.wait() >>> recv_tensor tensor([2, 3]) # Rank 0 tensor([0, 1]) # Rank 1 .. note:: Note that when this API is used with the NCCL PG backend, users must set the current GPU device with `torch.cuda.set_device`, otherwise it will lead to unexpected hang issues. In addition, if this API is the first collective call in the ``group`` passed to ``dist.P2POp``, all ranks of the ``group`` must participate in this API call; otherwise, the behavior is undefined. If this API call is not the first collective call in the ``group``, batched P2P operations involving only a subset of ranks of the ``group`` are allowed. """ _check_p2p_op_list(p2p_op_list) group = p2p_op_list[0].group device = p2p_op_list[0].tensor.device if device.type == "cuda": # NCCL style coalescing with _coalescing_manager(group, device, async_ops=True) as cm: for p2p_op in p2p_op_list: p2p_op.op(p2p_op.tensor, p2p_op.peer, p2p_op.group, p2p_op.tag) return cm.works else: # Backward support for Gloo reqs = [] for p2p_op in p2p_op_list: work = p2p_op.op(p2p_op.tensor, p2p_op.peer, p2p_op.group, p2p_op.tag) if work: reqs.append(work) return reqs @_exception_logger def broadcast_multigpu(tensor_list, src, group=None, async_op=False, src_tensor=0): """ Broadcasts the tensor to the whole group with multiple GPU tensors per node. ``tensor`` must have the same number of elements in all the GPUs from all processes participating in the collective. each tensor in the list must be on a different GPU Only nccl and gloo backend are currently supported tensors should only be GPU tensors Args: tensor_list (List[Tensor]): Tensors that participate in the collective operation. If ``src`` is the rank, then the specified ``src_tensor`` element of ``tensor_list`` (``tensor_list[src_tensor]``) will be broadcast to all other tensors (on different GPUs) in the src process and all tensors in ``tensor_list`` of other non-src processes. You also need to make sure that ``len(tensor_list)`` is the same for all the distributed processes calling this function. src (int): Source rank. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op src_tensor (int, optional): Source tensor rank within ``tensor_list`` Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ warnings.warn( "torch.distributed.broadcast_multigpu will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#multi-gpu-collective-functions" ) if _rank_not_in_group(group): _warn_not_in_group("broadcast_multigpu") return opts = BroadcastOptions() opts.rootRank = src opts.rootTensor = src_tensor if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() work = default_pg.broadcast(tensor_list, opts) else: group_src_rank = get_group_rank(group, src) opts.rootRank = group_src_rank work = group.broadcast(tensor_list, opts) if async_op: return work else: work.wait() @_exception_logger def broadcast(tensor, src, group=None, async_op=False): """ Broadcasts the tensor to the whole group. ``tensor`` must have the same number of elements in all processes participating in the collective. Args: tensor (Tensor): Data to be sent if ``src`` is the rank of current process, and tensor to be used to save received data otherwise. src (int): Source rank. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("broadcast") return opts = BroadcastOptions() opts.rootRank = src opts.rootTensor = 0 if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() work = default_pg.broadcast([tensor], opts) else: group_src_rank = get_group_rank(group, src) opts.rootRank = group_src_rank work = group.broadcast([tensor], opts) if async_op: return work else: work.wait() @_exception_logger def all_reduce_multigpu(tensor_list, op=ReduceOp.SUM, group=None, async_op=False): r""" Reduces the tensor data across all machines in such a way that all get the final result. This function reduces a number of tensors on every node, while each tensor resides on different GPUs. Therefore, the input tensor in the tensor list needs to be GPU tensors. Also, each tensor in the tensor list needs to reside on a different GPU. After the call, all ``tensor`` in ``tensor_list`` is going to be bitwise identical in all processes. Complex tensors are supported. Only nccl and gloo backend is currently supported tensors should only be GPU tensors Args: tensor_list (List[Tensor]): List of input and output tensors of the collective. The function operates in-place and requires that each tensor to be a GPU tensor on different GPUs. You also need to make sure that ``len(tensor_list)`` is the same for all the distributed processes calling this function. op (optional): One of the values from ``torch.distributed.ReduceOp`` enum. Specifies an operation used for element-wise reductions. group (ProcessGroup, optional): The process group to work on. If ``None``, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ warnings.warn( "torch.distributed.all_reduce_multigpu will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#multi-gpu-collective-functions" ) if _rank_not_in_group(group): return tensor_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in tensor_list ] opts = AllreduceOptions() opts.reduceOp = op if group is None: default_pg = _get_default_group() work = default_pg.allreduce(tensor_list, opts) else: work = group.allreduce(tensor_list, opts) if async_op: return work else: work.wait() @_exception_logger def all_reduce(tensor, op=ReduceOp.SUM, group=None, async_op=False): """ Reduces the tensor data across all machines in such a way that all get the final result. After the call ``tensor`` is going to be bitwise identical in all processes. Complex tensors are supported. Args: tensor (Tensor): Input and output of the collective. The function operates in-place. op (optional): One of the values from ``torch.distributed.ReduceOp`` enum. Specifies an operation used for element-wise reductions. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group Examples: >>> # xdoctest: +SKIP("no rank") >>> # All tensors below are of torch.int64 type. >>> # We have 2 process groups, 2 ranks. >>> tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank >>> tensor tensor([1, 2]) # Rank 0 tensor([3, 4]) # Rank 1 >>> dist.all_reduce(tensor, op=ReduceOp.SUM) >>> tensor tensor([4, 6]) # Rank 0 tensor([4, 6]) # Rank 1 >>> # All tensors below are of torch.cfloat type. >>> # We have 2 process groups, 2 ranks. >>> tensor = torch.tensor([1+1j, 2+2j], dtype=torch.cfloat) + 2 * rank * (1+1j) >>> tensor tensor([1.+1.j, 2.+2.j]) # Rank 0 tensor([3.+3.j, 4.+4.j]) # Rank 1 >>> dist.all_reduce(tensor, op=ReduceOp.SUM) >>> tensor tensor([4.+4.j, 6.+6.j]) # Rank 0 tensor([4.+4.j, 6.+6.j]) # Rank 1 """ _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("all_reduce") return if tensor.is_complex(): if not supports_complex(op): raise RuntimeError(f"all_reduce does not support {op} on complex tensors") tensor = torch.view_as_real(tensor) opts = AllreduceOptions() opts.reduceOp = op if group is None: group = _get_default_group() if group in _world.pg_coalesce_state.keys(): # We are in coalescing context, do not issue single operation, just append a collective representation coll = _CollOp(all_reduce, tensor, None, op, None) _world.pg_coalesce_state[group].append(coll) if async_op: return _IllegalWork() else: return None work = group.allreduce([tensor], opts) if async_op: return work else: work.wait() @_exception_logger def all_reduce_coalesced(tensors, op=ReduceOp.SUM, group=None, async_op=False): """ WARNING: at this time individual shape checking is not implemented across nodes. For example, if the rank 0 node passes [torch.rand(4), torch.rand(2)] and the rank 1 node passes [torch.rand(2), torch.rand(2), torch.rand(2)], the allreduce operation will proceed without complaint and return erroneous outputs. This lack of shape checking results in significant performance improvements but users of this function should take extra care to ensure that each node passes in tensors whose shapes match across nodes. Reduces each tensor in tensors (residing on the same device) across all machines in such a way that all get the final result. After the call each tensor in tensors is going to bitwise identical in all processes. Complex tensors are supported. Args: tensors (List[Tensor]): Input and output of the collective. The function operates in-place. op (Optional[ReduceOp]): One of the values from ``torch.distributed.ReduceOp`` enum. Specifies an operation used for element-wise reductions. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (Optional[bool]): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. """ warnings.warn( "torch.distributed.all_reduce_coalesced will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#collective-functions" ) _check_tensor_list(tensors, "tensor") _ensure_all_tensors_same_dtype(tensors) if _rank_not_in_group(group): _warn_not_in_group("all_reduce_coalesced") return if any(t.is_complex() for t in tensors) and not supports_complex(op): raise RuntimeError(f"all_reduce does not support {op} on complex tensors") tensors = [t if not t.is_complex() else torch.view_as_real(t) for t in tensors] opts = AllreduceCoalescedOptions() opts.reduceOp = op if group is None: default_pg = _get_default_group() work = default_pg.allreduce_coalesced(tensors, opts) else: work = group.allreduce_coalesced(tensors, opts) if async_op: return work.get_future() else: work.wait() @_exception_logger def reduce_multigpu( tensor_list, dst, op=ReduceOp.SUM, group=None, async_op=False, dst_tensor=0 ): """ Reduces the tensor data on multiple GPUs across all machines. Each tensor in ``tensor_list`` should reside on a separate GPU Only the GPU of ``tensor_list[dst_tensor]`` on the process with rank ``dst`` is going to receive the final result. Only nccl backend is currently supported tensors should only be GPU tensors Args: tensor_list (List[Tensor]): Input and output GPU tensors of the collective. The function operates in-place. You also need to make sure that ``len(tensor_list)`` is the same for all the distributed processes calling this function. dst (int): Destination rank op (optional): One of the values from ``torch.distributed.ReduceOp`` enum. Specifies an operation used for element-wise reductions. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op dst_tensor (int, optional): Destination tensor rank within ``tensor_list`` Returns: Async work handle, if async_op is set to True. None, otherwise """ warnings.warn( "torch.distributed.reduce_multigpu will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#multi-gpu-collective-functions" ) if _rank_not_in_group(group): _warn_not_in_group("reduce_multigpu") return opts = ReduceOptions() opts.reduceOp = op opts.rootRank = dst opts.rootTensor = dst_tensor if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() work = default_pg.reduce(tensor_list, opts) else: group_dst_rank = get_group_rank(group, dst) opts.rootRank = group_dst_rank work = group.reduce(tensor_list, opts) if async_op: return work else: work.wait() @_exception_logger def reduce(tensor, dst, op=ReduceOp.SUM, group=None, async_op=False): """ Reduces the tensor data across all machines. Only the process with rank ``dst`` is going to receive the final result. Args: tensor (Tensor): Input and output of the collective. The function operates in-place. dst (int): Destination rank op (optional): One of the values from ``torch.distributed.ReduceOp`` enum. Specifies an operation used for element-wise reductions. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ _check_single_tensor(tensor, "tensor") if _rank_not_in_group(group): _warn_not_in_group("reduce") return opts = ReduceOptions() opts.reduceOp = op opts.rootRank = dst if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() work = default_pg.reduce([tensor], opts) else: group_dst_rank = get_group_rank(group, dst) opts.rootRank = group_dst_rank work = group.reduce([tensor], opts) if async_op: return work else: work.wait() @_exception_logger def all_gather_multigpu( output_tensor_lists, input_tensor_list, group=None, async_op=False ): """ Gathers tensors from the whole group in a list. Each tensor in ``tensor_list`` should reside on a separate GPU Only nccl backend is currently supported tensors should only be GPU tensors Complex tensors are supported. Args: output_tensor_lists (List[List[Tensor]]): Output lists. It should contain correctly-sized tensors on each GPU to be used for output of the collective, e.g. ``output_tensor_lists[i]`` contains the all_gather result that resides on the GPU of ``input_tensor_list[i]``. Note that each element of ``output_tensor_lists`` has the size of ``world_size * len(input_tensor_list)``, since the function all gathers the result from every single GPU in the group. To interpret each element of ``output_tensor_lists[i]``, note that ``input_tensor_list[j]`` of rank k will be appear in ``output_tensor_lists[i][k * world_size + j]`` Also note that ``len(output_tensor_lists)``, and the size of each element in ``output_tensor_lists`` (each element is a list, therefore ``len(output_tensor_lists[i])``) need to be the same for all the distributed processes calling this function. input_tensor_list (List[Tensor]): List of tensors(on different GPUs) to be broadcast from current process. Note that ``len(input_tensor_list)`` needs to be the same for all the distributed processes calling this function. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ warnings.warn( "torch.distributed.all_gather_multigpu will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#multi-gpu-collective-functions" ) if _rank_not_in_group(group): _warn_not_in_group("all_gather_multigpu") return output_tensor_lists = [ [t if not t.is_complex() else torch.view_as_real(t) for t in l] for l in output_tensor_lists ] input_tensor_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in input_tensor_list ] if group is None: default_pg = _get_default_group() work = default_pg.allgather(output_tensor_lists, input_tensor_list) else: work = group.allgather(output_tensor_lists, input_tensor_list) if async_op: return work else: work.wait() def _object_to_tensor(obj, device): f = io.BytesIO() _pickler(f).dump(obj) byte_storage = torch.ByteStorage._from_buffer(f.getvalue()) # type: ignore[attr-defined] # Do not replace `torch.ByteTensor` or `torch.LongTensor` with torch.tensor and specifying dtype. # Otherwise, it will casue 100X slowdown. # See: https://github.com/pytorch/pytorch/issues/65696 byte_tensor = torch.ByteTensor(byte_storage).to(device) local_size = torch.LongTensor([byte_tensor.numel()]).to(device) return byte_tensor, local_size def _tensor_to_object(tensor, tensor_size): tensor = tensor.cpu() buf = tensor.numpy().tobytes()[:tensor_size] return _unpickler(io.BytesIO(buf)).load() @_exception_logger def all_gather_object(object_list, obj, group=None): """ Gathers picklable objects from the whole group into a list. Similar to :func:`all_gather`, but Python objects can be passed in. Note that the object must be picklable in order to be gathered. Args: object_list (list[Any]): Output list. It should be correctly sized as the size of the group for this collective and will contain the output. obj (Any): Pickable Python object to be broadcast from current process. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is ``None``. Returns: None. If the calling rank is part of this group, the output of the collective will be populated into the input ``object_list``. If the calling rank is not part of the group, the passed in ``object_list`` will be unmodified. .. note:: Note that this API differs slightly from the :func:`all_gather` collective since it does not provide an ``async_op`` handle and thus will be a blocking call. .. note:: For NCCL-based processed groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by ``torch.cuda.current_device()`` and it is the user's responsiblity to ensure that this is set so that each rank has an individual GPU, via ``torch.cuda.set_device()``. .. warning:: :func:`all_gather_object` uses ``pickle`` module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust. .. warning:: Calling :func:`all_gather_object` with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using :func:`all_gather` instead. Example:: >>> # xdoctest: +SKIP("need process group init") >>> # Note: Process group initialization omitted on each rank. >>> import torch.distributed as dist >>> # Assumes world_size of 3. >>> gather_objects = ["foo", 12, {1: 2}] # any picklable object >>> output = [None for _ in gather_objects] >>> dist.all_gather_object(output, gather_objects[dist.get_rank()]) >>> output ['foo', 12, {1: 2}] """ if _rank_not_in_group(group): _warn_not_in_group("all_gather_object") return current_device = _get_pg_default_device(group) input_tensor, local_size = _object_to_tensor(obj, current_device) # Gather all local sizes. This is so that we can find the max size, and index # until the correct size when deserializing the tensors. group_size = get_world_size(group=group) object_sizes_tensor = torch.zeros( group_size, dtype=torch.long, device=current_device ) object_size_list = [ object_sizes_tensor[i].unsqueeze(dim=0) for i in range(group_size) ] # Allgather tensor sizes all_gather(object_size_list, local_size, group=group) max_object_size = int(max(object_size_list).item()) # type: ignore[type-var] # Resize tensor to max size across all ranks. input_tensor.resize_(max_object_size) coalesced_output_tensor = torch.empty( max_object_size * group_size, dtype=torch.uint8, device=current_device ) # Output tensors are nonoverlapping views of coalesced_output_tensor output_tensors = [ coalesced_output_tensor[max_object_size * i : max_object_size * (i + 1)] for i in range(group_size) ] all_gather(output_tensors, input_tensor, group=group) # Deserialize outputs back to object. for i, tensor in enumerate(output_tensors): tensor = tensor.type(torch.uint8) if tensor.device != torch.device("cpu"): tensor = tensor.cpu() tensor_size = object_size_list[i] object_list[i] = _tensor_to_object(tensor, tensor_size) @_exception_logger def gather_object(obj, object_gather_list=None, dst=0, group=None): """ Gathers picklable objects from the whole group in a single process. Similar to :func:`gather`, but Python objects can be passed in. Note that the object must be picklable in order to be gathered. Args: obj (Any): Input object. Must be picklable. object_gather_list (list[Any]): Output list. On the ``dst`` rank, it should be correctly sized as the size of the group for this collective and will contain the output. Must be ``None`` on non-dst ranks. (default is ``None``) dst (int, optional): Destination rank. (default is 0) group: (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is ``None``. Returns: None. On the ``dst`` rank, ``object_gather_list`` will contain the output of the collective. .. note:: Note that this API differs slightly from the gather collective since it does not provide an async_op handle and thus will be a blocking call. .. note:: For NCCL-based processed groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by ``torch.cuda.current_device()`` and it is the user's responsiblity to ensure that this is set so that each rank has an individual GPU, via ``torch.cuda.set_device()``. .. warning:: :func:`gather_object` uses ``pickle`` module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust. .. warning:: Calling :func:`gather_object` with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using :func:`gather` instead. Example:: >>> # xdoctest: +SKIP("need process group init") >>> # Note: Process group initialization omitted on each rank. >>> import torch.distributed as dist >>> # Assumes world_size of 3. >>> gather_objects = ["foo", 12, {1: 2}] # any picklable object >>> output = [None for _ in gather_objects] >>> dist.gather_object( ... gather_objects[dist.get_rank()], ... output if dist.get_rank() == 0 else None, ... dst=0 ... ) >>> # On rank 0 >>> output ['foo', 12, {1: 2}] """ if _rank_not_in_group(group): _warn_not_in_group("gather_object") return # Ensure object_gather_list is specified appropriately. my_rank = get_rank() _validate_output_list_for_rank(my_rank, dst, object_gather_list) current_device = _get_pg_default_device(group) input_tensor, local_size = _object_to_tensor(obj, current_device) # Gather all local sizes. This is so that we can find the max size, and index # until the correct size when deserializing the tensors. group_size = get_world_size(group=group) object_sizes_tensor = torch.zeros( group_size, dtype=torch.long, device=current_device ) object_size_list = [ object_sizes_tensor[i].unsqueeze(dim=0) for i in range(group_size) ] # Allgather tensor sizes. An all-gather is needed here despite this being a # gather, since each rank needs to broadcast a tensor of the same (maximal) # size. all_gather(object_size_list, local_size, group=group) max_object_size = int(max(object_size_list).item()) # type: ignore[type-var] # Resize tensor to max size across all ranks. input_tensor.resize_(max_object_size) # Avoid populating output tensors if the result won't be gathered on this rank. if my_rank == dst: coalesced_output_tensor = torch.empty( max_object_size * group_size, dtype=torch.uint8, device=current_device ) # Output tensors are nonoverlapping views of coalesced_output_tensor output_tensors = [ coalesced_output_tensor[max_object_size * i : max_object_size * (i + 1)] for i in range(group_size) ] # All ranks call gather with equal-sized tensors. gather( input_tensor, gather_list=output_tensors if my_rank == dst else None, dst=dst, group=group, ) if my_rank != dst: return for i, tensor in enumerate(output_tensors): tensor = tensor.type(torch.uint8) tensor_size = object_size_list[i] object_gather_list[i] = _tensor_to_object(tensor, tensor_size) @_exception_logger def broadcast_object_list(object_list, src=0, group=None, device=None): """ Broadcasts picklable objects in ``object_list`` to the whole group. Similar to :func:`broadcast`, but Python objects can be passed in. Note that all objects in ``object_list`` must be picklable in order to be broadcasted. Args: object_list (List[Any]): List of input objects to broadcast. Each object must be picklable. Only objects on the ``src`` rank will be broadcast, but each rank must provide lists of equal sizes. src (int): Source rank from which to broadcast ``object_list``. group: (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is ``None``. device (``torch.device``, optional): If not None, the objects are serialized and converted to tensors which are moved to the ``device`` before broadcasting. Default is ``None``. Returns: ``None``. If rank is part of the group, ``object_list`` will contain the broadcasted objects from ``src`` rank. .. note:: For NCCL-based process groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by ``torch.cuda.current_device()`` and it is the user's responsibility to ensure that this is set so that each rank has an individual GPU, via ``torch.cuda.set_device()``. .. note:: Note that this API differs slightly from the :func:`all_gather` collective since it does not provide an ``async_op`` handle and thus will be a blocking call. .. warning:: :func:`broadcast_object_list` uses ``pickle`` module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust. .. warning:: Calling :func:`broadcast_object_list` with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using :func:`broadcast` instead. Example:: >>> # xdoctest: +SKIP("need process group init") >>> # Note: Process group initialization omitted on each rank. >>> import torch.distributed as dist >>> if dist.get_rank() == 0: >>> # Assumes world_size of 3. >>> objects = ["foo", 12, {1: 2}] # any picklable object >>> else: >>> objects = [None, None, None] >>> # Assumes backend is not NCCL >>> device = torch.device("cpu") >>> dist.broadcast_object_list(objects, src=0, device=device) >>> objects ['foo', 12, {1: 2}] """ if _rank_not_in_group(group): _warn_not_in_group("broadcast_object_list") return # Current device selection. # To preserve backwards compatibility, ``device`` is default to ``None`` # in which case we run current logic of device selection, i.e. # ``current_device`` is CUDA if backend is NCCL otherwise CPU device. In the # case it is not ``None`` we move the size and object tensors to be # broadcasted to this device. current_device = device or _get_pg_default_device(group) my_rank = get_rank() # Serialize object_list elements to tensors on src rank. if my_rank == src: tensor_list, size_list = zip(*[_object_to_tensor(obj, current_device) for obj in object_list]) object_sizes_tensor = torch.cat(size_list) else: object_sizes_tensor = torch.empty(len(object_list), dtype=torch.long, device=current_device) # Broadcast object sizes broadcast(object_sizes_tensor, src=src, group=group) # Concatenate and broadcast serialized object tensors # Note: torch.cat will do an extra memory copy to the current device, if the tensor_list # has only one element, we can skip the copy. if my_rank == src: if len(tensor_list) == 1: object_tensor = tensor_list[0] else: object_tensor = torch.cat(tensor_list) else: object_tensor = torch.empty( # type: ignore[call-overload] torch.sum(object_sizes_tensor).item(), # type: ignore[arg-type] dtype=torch.uint8, device=current_device ) broadcast(object_tensor, src=src, group=group) # Deserialize objects using their stored sizes. offset = 0 if my_rank != src: for i, obj_size in enumerate(object_sizes_tensor): obj_view = object_tensor[offset : offset + obj_size] obj_view = obj_view.type(torch.uint8) if obj_view.device != torch.device("cpu"): obj_view = obj_view.cpu() offset += obj_size object_list[i] = _tensor_to_object(obj_view, obj_size) @_exception_logger def scatter_object_list( scatter_object_output_list, scatter_object_input_list, src=0, group=None ): """ Scatters picklable objects in ``scatter_object_input_list`` to the whole group. Similar to :func:`scatter`, but Python objects can be passed in. On each rank, the scattered object will be stored as the first element of ``scatter_object_output_list``. Note that all objects in ``scatter_object_input_list`` must be picklable in order to be scattered. Args: scatter_object_output_list (List[Any]): Non-empty list whose first element will store the object scattered to this rank. scatter_object_input_list (List[Any]): List of input objects to scatter. Each object must be picklable. Only objects on the ``src`` rank will be scattered, and the argument can be ``None`` for non-src ranks. src (int): Source rank from which to scatter ``scatter_object_input_list``. group: (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. Default is ``None``. Returns: ``None``. If rank is part of the group, ``scatter_object_output_list`` will have its first element set to the scattered object for this rank. .. note:: Note that this API differs slightly from the scatter collective since it does not provide an ``async_op`` handle and thus will be a blocking call. .. warning:: :func:`scatter_object_list` uses ``pickle`` module implicitly, which is known to be insecure. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling. Only call this function with data you trust. .. warning:: Calling :func:`scatter_object_list` with GPU tensors is not well supported and inefficient as it incurs GPU -> CPU transfer since tensors would be pickled. Please consider using :func:`scatter` instead. Example:: >>> # xdoctest: +SKIP("need process group init") >>> # Note: Process group initialization omitted on each rank. >>> import torch.distributed as dist >>> if dist.get_rank() == 0: >>> # Assumes world_size of 3. >>> objects = ["foo", 12, {1: 2}] # any picklable object >>> else: >>> # Can be any list on non-src ranks, elements are not used. >>> objects = [None, None, None] >>> output_list = [None] >>> dist.scatter_object_list(output_list, objects, src=0) >>> # Rank i gets objects[i]. For example, on rank 2: >>> output_list [{1: 2}] """ if _rank_not_in_group(group): _warn_not_in_group("scatter_object_list") return if ( not isinstance(scatter_object_output_list, list) or len(scatter_object_output_list) < 1 ): raise RuntimeError( "Expected argument scatter_object_output_list to be a list of size at least 1." ) my_rank = get_rank() pg_device = _get_pg_default_device(group) if my_rank == src: tensor_list, tensor_sizes = zip( *[_object_to_tensor(obj, pg_device) for obj in scatter_object_input_list] ) tensor_list, tensor_sizes = list(tensor_list), list(tensor_sizes) # Src rank broadcasts the maximum tensor size. This is because all ranks are # expected to call into scatter() with equal-sized tensors. if my_rank == src: max_tensor_size = max(tensor_sizes) for tensor in tensor_list: tensor.resize_(max_tensor_size) else: max_tensor_size = torch.tensor([0], dtype=torch.long, device=pg_device) broadcast(max_tensor_size, src=src, group=group) # Scatter actual serialized objects output_tensor = torch.empty(max_tensor_size.item(), dtype=torch.uint8, device=pg_device) scatter( output_tensor, scatter_list=None if my_rank != src else tensor_list, src=src, group=group, ) # Scatter per-object sizes to trim tensors when deserializing back to object obj_tensor_size = torch.tensor([0], dtype=torch.long, device=pg_device) scatter( obj_tensor_size, scatter_list=None if my_rank != src else tensor_sizes, src=src, group=group, ) # Deserialize back to object scatter_object_output_list[0] = _tensor_to_object(output_tensor, obj_tensor_size) @_exception_logger def all_gather(tensor_list, tensor, group=None, async_op=False): """ Gathers tensors from the whole group in a list. Complex tensors are supported. Args: tensor_list (list[Tensor]): Output list. It should contain correctly-sized tensors to be used for output of the collective. tensor (Tensor): Tensor to be broadcast from current process. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group Examples: >>> # xdoctest: +SKIP("need process group init") >>> # All tensors below are of torch.int64 dtype. >>> # We have 2 process groups, 2 ranks. >>> tensor_list = [torch.zeros(2, dtype=torch.int64) for _ in range(2)] >>> tensor_list [tensor([0, 0]), tensor([0, 0])] # Rank 0 and 1 >>> tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank >>> tensor tensor([1, 2]) # Rank 0 tensor([3, 4]) # Rank 1 >>> dist.all_gather(tensor_list, tensor) >>> tensor_list [tensor([1, 2]), tensor([3, 4])] # Rank 0 [tensor([1, 2]), tensor([3, 4])] # Rank 1 >>> # All tensors below are of torch.cfloat dtype. >>> # We have 2 process groups, 2 ranks. >>> tensor_list = [torch.zeros(2, dtype=torch.cfloat) for _ in range(2)] >>> tensor_list [tensor([0.+0.j, 0.+0.j]), tensor([0.+0.j, 0.+0.j])] # Rank 0 and 1 >>> tensor = torch.tensor([1+1j, 2+2j], dtype=torch.cfloat) + 2 * rank * (1+1j) >>> tensor tensor([1.+1.j, 2.+2.j]) # Rank 0 tensor([3.+3.j, 4.+4.j]) # Rank 1 >>> dist.all_gather(tensor_list, tensor) >>> tensor_list [tensor([1.+1.j, 2.+2.j]), tensor([3.+3.j, 4.+4.j])] # Rank 0 [tensor([1.+1.j, 2.+2.j]), tensor([3.+3.j, 4.+4.j])] # Rank 1 """ _check_tensor_list(tensor_list, "tensor_list") _check_single_tensor(tensor, "tensor") _ensure_all_tensors_same_dtype(tensor_list, tensor) if _rank_not_in_group(group): _warn_not_in_group("all_gather") return tensor_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in tensor_list ] tensor = tensor if not tensor.is_complex() else torch.view_as_real(tensor) if group is None: default_pg = _get_default_group() work = default_pg.allgather([tensor_list], [tensor]) else: work = group.allgather([tensor_list], [tensor]) if async_op: return work else: work.wait() @_exception_logger def all_gather_into_tensor(output_tensor, input_tensor, group=None, async_op=False): """ Gather tensors from all ranks and put them in a single output tensor. Args: output_tensor (Tensor): Output tensor to accommodate tensor elements from all ranks. It must be correctly sized to have one of the following forms: (i) a concatenation of all the input tensors along the primary dimension; for definition of "concatenation", see ``torch.cat()``; (ii) a stack of all the input tensors along the primary dimension; for definition of "stack", see ``torch.stack()``. Examples below may better explain the supported output forms. input_tensor (Tensor): Tensor to be gathered from current rank. Different from the ``all_gather`` API, the input tensors in this API must have the same size across all ranks. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group Examples: >>> # xdoctest: +SKIP("need process group init") >>> # All tensors below are of torch.int64 dtype and on CUDA devices. >>> # We have two ranks. >>> device = torch.device(f'cuda:{rank}') >>> tensor_in = torch.arange(2, dtype=torch.int64, device=device) + 1 + 2 * rank >>> tensor_in tensor([1, 2], device='cuda:0') # Rank 0 tensor([3, 4], device='cuda:1') # Rank 1 >>> # Output in concatenation form >>> tensor_out = torch.zeros(world_size * 2, dtype=torch.int64, device=device) >>> dist.all_gather_into_tensor(tensor_out, tensor_in) >>> tensor_out tensor([1, 2, 3, 4], device='cuda:0') # Rank 0 tensor([1, 2, 3, 4], device='cuda:1') # Rank 1 >>> # Output in stack form >>> tensor_out2 = torch.zeros(world_size, 2, dtype=torch.int64, device=device) >>> dist.all_gather_into_tensor(tensor_out2, tensor_in) >>> tensor_out2 tensor([[1, 2], [3, 4]], device='cuda:0') # Rank 0 tensor([[1, 2], [3, 4]], device='cuda:1') # Rank 1 .. warning:: The Gloo backend does not support this API. """ _check_single_tensor(input_tensor, "input_tensor") _check_single_tensor(output_tensor, "output_tensor") if _rank_not_in_group(group): _warn_not_in_group("all_gather_into_tensor") return output_tensor = ( output_tensor if not output_tensor.is_complex() else torch.view_as_real(output_tensor) ) input_tensor = ( input_tensor if not input_tensor.is_complex() else torch.view_as_real(input_tensor) ) group = group or _get_default_group() if group in _world.pg_coalesce_state.keys(): # We are in coalescing context, do not issue single operation, just append a collective representation coll = _CollOp(all_gather_into_tensor, input_tensor, output_tensor) _world.pg_coalesce_state[group].append(coll) if async_op: return _IllegalWork() else: return None work = group._allgather_base(output_tensor, input_tensor) if async_op: return work else: work.wait() @_exception_logger def _all_gather_base(output_tensor, input_tensor, group=None, async_op=False): """ Single tensor all gather. Gathers a single tensor from all ranks, and puts them in a single output tensor. Args: output_tensor (Tensor): Output tensor. It should contain correctly-sized tensors to be used for output of the collective. input_tensor (Tensor): Tensor to be broadcast from current process. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group .. warning:: `_all_gather_base` is a private function. Users should use `all_gather_into_tensor` instead. """ warnings.warn( "torch.distributed._all_gather_base is a private function and will be " "deprecated. Please use torch.distributed.all_gather_into_tensor " "instead." ) return all_gather_into_tensor(output_tensor, input_tensor, group, async_op) @_exception_logger def all_gather_coalesced( output_tensor_lists, input_tensor_list, group=None, async_op=False ): """ Gathers input tensors from the whole group in a list in a coalesced manner. Complex tensors are supported. Args: output_tensor_lists (list[list[Tensor]]): Output list. It should contain correctly-sized tensors to be used for output of the collective. input_tensor_list (list[Tensor]): Tensors to be broadcast from current process. At least one tensor has to be non empty. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group Example: we have 2 process groups, 2 ranks. rank 0 passes: input_tensor_list = [[[1, 1], [1, 1]], [2], [3, 3]] output_tensor_lists = [[[[-1, -1], [-1, -1]], [-1], [-1, -1]], [[[-1, -1], [-1, -1]], [-1], [-1, -1]]] rank 1 passes: input_tensor_list = [[[3, 3], [3, 3]], [5], [1, 1]] output_tensor_lists = [[[[-1, -1], [-1, -1]], [-1], [-1, -1]], [[[-1, -1], [-1, -1]], [-1], [-1, -1]]] both rank 0 and 1 get: output_tensor_lists = [[[1, 1], [1, 1]], [2], [3, 3]], [[3, 3], [3, 3]], [5], [1, 1]]]. WARNING: at this time individual shape checking is not implemented across nodes. For example, if the rank 0 node passes [torch.rand(4), torch.rand(2)] and the rank 1 node passes [torch.rand(2), torch.rand(2), torch.rand(2)], the all_gather_coalesced operation will proceed without complaint and return erroneous outputs. This lack of shape checking results in significant performance improvements but users of this function should take extra care to ensure that each node passes in tensors whose shapes match across nodes. """ warnings.warn( "torch.distributed.all_gather_coalesced will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#collective-functions" ) # We only check basic compatibility with C++ params here, C++ code will # do shape and type checking. if _rank_not_in_group(group): _warn_not_in_group("all_gather_coalesced") return _check_tensor_list(input_tensor_list, "input_tensor_list") _ensure_all_tensors_same_dtype(input_tensor_list) if not isinstance(output_tensor_lists, list): raise RuntimeError( "Invalid function argument: output_tensor_lists should be a list" ) for output_tensor_list in output_tensor_lists: _check_tensor_list(output_tensor_list, "output_tensor_lists") _ensure_all_tensors_same_dtype(output_tensor_list) output_tensor_lists = [ [t if not t.is_complex() else torch.view_as_real(t) for t in l] for l in output_tensor_lists ] input_tensor_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in input_tensor_list ] if group is None: default_pg = _get_default_group() work = default_pg.allgather_coalesced(output_tensor_lists, input_tensor_list) else: work = group.allgather_coalesced(output_tensor_lists, input_tensor_list) if async_op: return work.get_future() else: work.wait() def _validate_output_list_for_rank(my_rank, dst, gather_list): if dst == my_rank: if not gather_list: raise ValueError( "Argument ``gather_list`` must be specified on destination rank." ) elif gather_list: raise ValueError( "Argument ``gather_list`` must NOT be specified " "on non-destination ranks." ) @_exception_logger def gather(tensor, gather_list=None, dst=0, group=None, async_op=False): """ Gathers a list of tensors in a single process. Args: tensor (Tensor): Input tensor. gather_list (list[Tensor], optional): List of appropriately-sized tensors to use for gathered data (default is None, must be specified on the destination rank) dst (int, optional): Destination rank (default is 0) group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ _check_single_tensor(tensor, "tensor") # Parameter ``gather_list`` may be left unspecified on non-dst ranks. if gather_list: _check_tensor_list(gather_list, "gather_list") else: gather_list = [] _ensure_all_tensors_same_dtype(tensor, gather_list) if _rank_not_in_group(group): _warn_not_in_group("gather") return my_rank = get_rank() _validate_output_list_for_rank(my_rank, dst, gather_list) output_tensors = [gather_list] if dst == my_rank else [] input_tensors = [tensor] opts = GatherOptions() opts.rootRank = dst if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() work = default_pg.gather(output_tensors, input_tensors, opts) else: group_dst_rank = get_group_rank(group, dst) opts.rootRank = group_dst_rank work = group.gather(output_tensors, input_tensors, opts) if async_op: return work else: work.wait() @_exception_logger def scatter(tensor, scatter_list=None, src=0, group=None, async_op=False): """ Scatters a list of tensors to all processes in a group. Each process will receive exactly one tensor and store its data in the ``tensor`` argument. Complex tensors are supported. Args: tensor (Tensor): Output tensor. scatter_list (list[Tensor]): List of tensors to scatter (default is None, must be specified on the source rank) src (int): Source rank (default is 0) group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group .. note:: Note that all Tensors in scatter_list must have the same size. Example:: >>> # xdoctest: +SKIP("need process group init") >>> # Note: Process group initialization omitted on each rank. >>> import torch.distributed as dist >>> tensor_size = 2 >>> t_ones = torch.ones(tensor_size) >>> t_fives = torch.ones(tensor_size) * 5 >>> output_tensor = torch.zeros(tensor_size) >>> if dist.get_rank() == 0: >>> # Assumes world_size of 2. >>> # Only tensors, all of which must be the same size. >>> scatter_list = [t_ones, t_fives] >>> else: >>> scatter_list = None >>> dist.scatter(output_tensor, scatter_list, src=0) >>> # Rank i gets scatter_list[i]. For example, on rank 1: >>> output_tensor tensor([5., 5.]) """ _check_single_tensor(tensor, "tensor") # Parameter ``scatter_list`` may be left unspecified on non-src ranks. if scatter_list: _check_tensor_list(scatter_list, "scatter_list") else: scatter_list = [] _ensure_all_tensors_same_dtype(tensor, scatter_list) if _rank_not_in_group(group): _warn_not_in_group("scatter") return scatter_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in scatter_list ] tensor = tensor if not tensor.is_complex() else torch.view_as_real(tensor) my_rank = get_rank() if src == my_rank: if not scatter_list: raise ValueError( "Argument ``scatter_list`` must be specified on source rank." ) input_tensors = [scatter_list] output_tensors = [tensor] else: if scatter_list: raise ValueError( "Argument ``scatter_list`` must NOT be specified " "on non-source ranks." ) input_tensors = [] output_tensors = [tensor] opts = ScatterOptions() opts.rootRank = src if group is None or group is GroupMember.WORLD: default_pg = _get_default_group() work = default_pg.scatter(output_tensors, input_tensors, opts) else: group_src_rank = get_group_rank(group, src) opts.rootRank = group_src_rank work = group.scatter(output_tensors, input_tensors, opts) if async_op: return work else: work.wait() @_exception_logger def reduce_scatter_multigpu( output_tensor_list, input_tensor_lists, op=ReduceOp.SUM, group=None, async_op=False ): """ Reduce and scatter a list of tensors to the whole group. Only nccl backend is currently supported. Each tensor in ``output_tensor_list`` should reside on a separate GPU, as should each list of tensors in ``input_tensor_lists``. Args: output_tensor_list (List[Tensor]): Output tensors (on different GPUs) to receive the result of the operation. Note that ``len(output_tensor_list)`` needs to be the same for all the distributed processes calling this function. input_tensor_lists (List[List[Tensor]]): Input lists. It should contain correctly-sized tensors on each GPU to be used for input of the collective, e.g. ``input_tensor_lists[i]`` contains the reduce_scatter input that resides on the GPU of ``output_tensor_list[i]``. Note that each element of ``input_tensor_lists`` has the size of ``world_size * len(output_tensor_list)``, since the function scatters the result from every single GPU in the group. To interpret each element of ``input_tensor_lists[i]``, note that ``output_tensor_list[j]`` of rank k receives the reduce-scattered result from ``input_tensor_lists[i][k * world_size + j]`` Also note that ``len(input_tensor_lists)``, and the size of each element in ``input_tensor_lists`` (each element is a list, therefore ``len(input_tensor_lists[i])``) need to be the same for all the distributed processes calling this function. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. """ warnings.warn( "torch.distributed.reduce_scatter_multigpu will be deprecated. If you must " "use it, please revisit our documentation later at " "https://pytorch.org/docs/master/distributed.html#multi-gpu-collective-functions" ) if _rank_not_in_group(group): _warn_not_in_group("reduce_scatter_multigpu") return opts = ReduceScatterOptions() opts.reduceOp = op if group is None: default_pg = _get_default_group() work = default_pg.reduce_scatter(output_tensor_list, input_tensor_lists, opts) else: work = group.reduce_scatter(output_tensor_list, input_tensor_lists, opts) if async_op: return work else: work.wait() @_exception_logger def reduce_scatter(output, input_list, op=ReduceOp.SUM, group=None, async_op=False): """ Reduces, then scatters a list of tensors to all processes in a group. Args: output (Tensor): Output tensor. input_list (list[Tensor]): List of tensors to reduce and scatter. op (optional): One of the values from ``torch.distributed.ReduceOp`` enum. Specifies an operation used for element-wise reductions. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. """ _check_single_tensor(output, "output") _check_tensor_list(input_list, "input_list") _ensure_all_tensors_same_dtype(output, input_list) if _rank_not_in_group(group): _warn_not_in_group("reduce_scatter") return opts = ReduceScatterOptions() opts.reduceOp = op if group is None: default_pg = _get_default_group() work = default_pg.reduce_scatter([output], [input_list], opts) else: work = group.reduce_scatter([output], [input_list], opts) if async_op: return work else: work.wait() @_exception_logger def reduce_scatter_tensor(output, input, op=ReduceOp.SUM, group=None, async_op=False): """ Reduces, then scatters a tensor to all ranks in a group. Args: output (Tensor): Output tensor. It should have the same size across all ranks. input (Tensor): Input tensor to be reduced and scattered. Its size should be output tensor size times the world size. The input tensor can have one of the following shapes: (i) a concatenation of the output tensors along the primary dimension, or (ii) a stack of the output tensors along the primary dimension. For definition of "concatenation", see ``torch.cat()``. For definition of "stack", see ``torch.stack()``. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. Examples: >>> # xdoctest: +SKIP("need process group init") >>> # All tensors below are of torch.int64 dtype and on CUDA devices. >>> # We have two ranks. >>> device = torch.device(f'cuda:{rank}') >>> tensor_out = torch.zeros(2, dtype=torch.int64, device=device) >>> # Input in concatenation form >>> tensor_in = torch.arange(world_size * 2, dtype=torch.int64, device=device) >>> tensor_in tensor([0, 1, 2, 3], device='cuda:0') # Rank 0 tensor([0, 1, 2, 3], device='cuda:1') # Rank 1 >>> dist.reduce_scatter_tensor(tensor_out, tensor_in) >>> tensor_out tensor([0, 2], device='cuda:0') # Rank 0 tensor([4, 6], device='cuda:1') # Rank 1 >>> # Input in stack form >>> tensor_in = torch.reshape(tensor_in, (world_size, 2)) >>> tensor_in tensor([[0, 1], [2, 3]], device='cuda:0') # Rank 0 tensor([[0, 1], [2, 3]], device='cuda:1') # Rank 1 >>> dist.reduce_scatter_tensor(tensor_out, tensor_in) >>> tensor_out tensor([0, 2], device='cuda:0') # Rank 0 tensor([4, 6], device='cuda:1') # Rank 1 .. warning:: The Gloo backend does not support this API. """ _check_single_tensor(output, "output") _check_single_tensor(input, "input") if _rank_not_in_group(group): _warn_not_in_group("reduce_scatter_tensor") return opts = ReduceScatterOptions() opts.reduceOp = op group = group or _get_default_group() # Check if we are in coalescing context # If we are, do not issue single operation, just append a collective representation if group in _world.pg_coalesce_state.keys(): coll = _CollOp(reduce_scatter_tensor, input, output, op, None) _world.pg_coalesce_state[group].append(coll) if async_op: return _IllegalWork() else: return None work = group._reduce_scatter_base(output, input, opts) if async_op: return work else: work.wait() def _reduce_scatter_base(output, input, op=ReduceOp.SUM, group=None, async_op=False): """ Reduces, then scatters a flattened tensor to all processes in a group. Args: output (Tensor): Output tensor. input (Tensor): Input tensor that is of size output tensor size times world size group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. .. warning:: `_reduce_scatter_base` is a private function. Users should use `reduce_scatter_tensor` instead. """ warnings.warn( "torch.distributed._reduce_scatter_base is a private function and will " "be deprecated. Please use torch.distributed.reduce_scatter_tensor " "instead." ) return reduce_scatter_tensor(output, input, op, group, async_op) @_exception_logger def all_to_all_single( output, input, output_split_sizes=None, input_split_sizes=None, group=None, async_op=False, ): """ Each process splits input tensor and then scatters the split list to all processes in a group. Then concatenate the received tensors from all the processes in the group and return single output tensor. Complex tensors are supported. Args: output (Tensor): Gathered concatenated output tensor. input (Tensor): Input tensor to scatter. output_split_sizes: (list[Int], optional): Output split sizes for dim 0 if specified None or empty, dim 0 of ``output`` tensor must divide equally by ``world_size``. input_split_sizes: (list[Int], optional): Input split sizes for dim 0 if specified None or empty, dim 0 of ``input`` tensor must divide equally by ``world_size``. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. .. warning:: `all_to_all_single` is experimental and subject to change. Examples: >>> # xdoctest: +SKIP("Undefined rank") >>> input = torch.arange(4) + rank * 4 >>> input tensor([0, 1, 2, 3]) # Rank 0 tensor([4, 5, 6, 7]) # Rank 1 tensor([8, 9, 10, 11]) # Rank 2 tensor([12, 13, 14, 15]) # Rank 3 >>> output = torch.empty([4], dtype=torch.int64) >>> dist.all_to_all_single(output, input) >>> output tensor([0, 4, 8, 12]) # Rank 0 tensor([1, 5, 9, 13]) # Rank 1 tensor([2, 6, 10, 14]) # Rank 2 tensor([3, 7, 11, 15]) # Rank 3 >>> # Essentially, it is similar to following operation: >>> scatter_list = list(input.chunk(world_size)) >>> gather_list = list(output.chunk(world_size)) >>> for i in range(world_size): >>> dist.scatter(gather_list[i], scatter_list if i == rank else [], src = i) >>> # Another example with uneven split >>> input tensor([0, 1, 2, 3, 4, 5]) # Rank 0 tensor([10, 11, 12, 13, 14, 15, 16, 17, 18]) # Rank 1 tensor([20, 21, 22, 23, 24]) # Rank 2 tensor([30, 31, 32, 33, 34, 35, 36]) # Rank 3 >>> input_splits [2, 2, 1, 1] # Rank 0 [3, 2, 2, 2] # Rank 1 [2, 1, 1, 1] # Rank 2 [2, 2, 2, 1] # Rank 3 >>> output_splits [2, 3, 2, 2] # Rank 0 [2, 2, 1, 2] # Rank 1 [1, 2, 1, 2] # Rank 2 [1, 2, 1, 1] # Rank 3 >>> output = ... >>> dist.all_to_all_single(output, input, output_splits, input_splits) >>> output tensor([ 0, 1, 10, 11, 12, 20, 21, 30, 31]) # Rank 0 tensor([ 2, 3, 13, 14, 22, 32, 33]) # Rank 1 tensor([ 4, 15, 16, 23, 34, 35]) # Rank 2 tensor([ 5, 17, 18, 24, 36]) # Rank 3 >>> # Another example with tensors of torch.cfloat type. >>> input = torch.tensor([1+1j, 2+2j, 3+3j, 4+4j], dtype=torch.cfloat) + 4 * rank * (1+1j) >>> input tensor([1+1j, 2+2j, 3+3j, 4+4j]) # Rank 0 tensor([5+5j, 6+6j, 7+7j, 8+8j]) # Rank 1 tensor([9+9j, 10+10j, 11+11j, 12+12j]) # Rank 2 tensor([13+13j, 14+14j, 15+15j, 16+16j]) # Rank 3 >>> output = torch.empty([4], dtype=torch.int64) >>> dist.all_to_all_single(output, input) >>> output tensor([1+1j, 5+5j, 9+9j, 13+13j]) # Rank 0 tensor([2+2j, 6+6j, 10+10j, 14+14j]) # Rank 1 tensor([3+3j, 7+7j, 11+11j, 15+15j]) # Rank 2 tensor([4+4j, 8+8j, 12+12j, 16+16j]) # Rank 3 """ if _rank_not_in_group(group): _warn_not_in_group("all_to_all_single") return opts = AllToAllOptions() _check_single_tensor(output, "output") _check_single_tensor(input, "input") _ensure_all_tensors_same_dtype(output, input) if input.is_complex(): input = torch.view_as_real(input) if output.is_complex(): output = torch.view_as_real(output) output_split_sizes = [] if output_split_sizes is None else output_split_sizes input_split_sizes = [] if input_split_sizes is None else input_split_sizes if group is None: default_pg = _get_default_group() work = default_pg.alltoall_base( output, input, output_split_sizes, input_split_sizes, opts ) else: work = group.alltoall_base( output, input, output_split_sizes, input_split_sizes, opts ) if async_op: return work else: work.wait() @_exception_logger def all_to_all(output_tensor_list, input_tensor_list, group=None, async_op=False): """ Each process scatters list of input tensors to all processes in a group and return gathered list of tensors in output list. Complex tensors are supported. Args: output_tensor_list (list[Tensor]): List of tensors to be gathered one per rank. input_tensor_list (list[Tensor]): List of tensors to scatter one per rank. group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group. .. warning:: `all_to_all` is experimental and subject to change. Examples: >>> # xdoctest: +SKIP("Undefined rank") >>> input = torch.arange(4) + rank * 4 >>> input = list(input.chunk(4)) >>> input [tensor([0]), tensor([1]), tensor([2]), tensor([3])] # Rank 0 [tensor([4]), tensor([5]), tensor([6]), tensor([7])] # Rank 1 [tensor([8]), tensor([9]), tensor([10]), tensor([11])] # Rank 2 [tensor([12]), tensor([13]), tensor([14]), tensor([15])] # Rank 3 >>> output = list(torch.empty([4], dtype=torch.int64).chunk(4)) >>> dist.all_to_all(output, input) >>> output [tensor([0]), tensor([4]), tensor([8]), tensor([12])] # Rank 0 [tensor([1]), tensor([5]), tensor([9]), tensor([13])] # Rank 1 [tensor([2]), tensor([6]), tensor([10]), tensor([14])] # Rank 2 [tensor([3]), tensor([7]), tensor([11]), tensor([15])] # Rank 3 >>> # Essentially, it is similar to following operation: >>> scatter_list = input >>> gather_list = output >>> for i in range(world_size): >>> dist.scatter(gather_list[i], scatter_list if i == rank else [], src=i) >>> input tensor([0, 1, 2, 3, 4, 5]) # Rank 0 tensor([10, 11, 12, 13, 14, 15, 16, 17, 18]) # Rank 1 tensor([20, 21, 22, 23, 24]) # Rank 2 tensor([30, 31, 32, 33, 34, 35, 36]) # Rank 3 >>> input_splits [2, 2, 1, 1] # Rank 0 [3, 2, 2, 2] # Rank 1 [2, 1, 1, 1] # Rank 2 [2, 2, 2, 1] # Rank 3 >>> output_splits [2, 3, 2, 2] # Rank 0 [2, 2, 1, 2] # Rank 1 [1, 2, 1, 2] # Rank 2 [1, 2, 1, 1] # Rank 3 >>> input = list(input.split(input_splits)) >>> input [tensor([0, 1]), tensor([2, 3]), tensor([4]), tensor([5])] # Rank 0 [tensor([10, 11, 12]), tensor([13, 14]), tensor([15, 16]), tensor([17, 18])] # Rank 1 [tensor([20, 21]), tensor([22]), tensor([23]), tensor([24])] # Rank 2 [tensor([30, 31]), tensor([32, 33]), tensor([34, 35]), tensor([36])] # Rank 3 >>> output = ... >>> dist.all_to_all(output, input) >>> output [tensor([0, 1]), tensor([10, 11, 12]), tensor([20, 21]), tensor([30, 31])] # Rank 0 [tensor([2, 3]), tensor([13, 14]), tensor([22]), tensor([32, 33])] # Rank 1 [tensor([4]), tensor([15, 16]), tensor([23]), tensor([34, 35])] # Rank 2 [tensor([5]), tensor([17, 18]), tensor([24]), tensor([36])] # Rank 3 >>> # Another example with tensors of torch.cfloat type. >>> input = torch.tensor([1+1j, 2+2j, 3+3j, 4+4j], dtype=torch.cfloat) + 4 * rank * (1+1j) >>> input = list(input.chunk(4)) >>> input [tensor([1+1j]), tensor([2+2j]), tensor([3+3j]), tensor([4+4j])] # Rank 0 [tensor([5+5j]), tensor([6+6j]), tensor([7+7j]), tensor([8+8j])] # Rank 1 [tensor([9+9j]), tensor([10+10j]), tensor([11+11j]), tensor([12+12j])] # Rank 2 [tensor([13+13j]), tensor([14+14j]), tensor([15+15j]), tensor([16+16j])] # Rank 3 >>> output = list(torch.empty([4], dtype=torch.int64).chunk(4)) >>> dist.all_to_all(output, input) >>> output [tensor([1+1j]), tensor([5+5j]), tensor([9+9j]), tensor([13+13j])] # Rank 0 [tensor([2+2j]), tensor([6+6j]), tensor([10+10j]), tensor([14+14j])] # Rank 1 [tensor([3+3j]), tensor([7+7j]), tensor([11+11j]), tensor([15+15j])] # Rank 2 [tensor([4+4j]), tensor([8+8j]), tensor([12+12j]), tensor([16+16j])] # Rank 3 """ if _rank_not_in_group(group): _warn_not_in_group("all_to_all") return opts = AllToAllOptions() _check_tensor_list(output_tensor_list, "output_tensor_list") _check_tensor_list(input_tensor_list, "input_tensor_list") _ensure_all_tensors_same_dtype(output_tensor_list, input_tensor_list) input_tensor_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in input_tensor_list ] output_tensor_list = [ t if not t.is_complex() else torch.view_as_real(t) for t in output_tensor_list ] if group is None: default_pg = _get_default_group() work = default_pg.alltoall(output_tensor_list, input_tensor_list, opts) else: work = group.alltoall(output_tensor_list, input_tensor_list, opts) if async_op: return work else: work.wait() @_exception_logger def barrier(group=GroupMember.WORLD, async_op=False, device_ids=None): """ Synchronizes all processes. This collective blocks processes until the whole group enters this function, if async_op is False, or if async work handle is called on wait(). Args: group (ProcessGroup, optional): The process group to work on. If None, the default process group will be used. async_op (bool, optional): Whether this op should be an async op device_ids ([int], optional): List of device/GPU ids. Returns: Async work handle, if async_op is set to True. None, if not async_op or if not part of the group """ if _rank_not_in_group(group): _warn_not_in_group("barrier") return opts = BarrierOptions() opts.device = _get_pg_default_device(group) if device_ids is not None: if isinstance(device_ids, list): opts.device_ids = device_ids else: raise RuntimeError( "Invalid function argument: device_ids type should be List[int]" ) if group is None: default_pg = _get_default_group() work = default_pg.barrier(opts=opts) else: work = group.barrier(opts=opts) if async_op: return work else: work.wait() def monitored_barrier(group=GroupMember.WORLD, timeout=None, wait_all_ranks=False): """ Synchronizes all processes similar to ``torch.distributed.barrier``, but takes a configurable timeout and is able to report ranks that did not pass this barrier within that timeout. Specifically, for non-zero ranks, will block until a send/recv is processed from rank 0. Rank 0 will block until all send /recv from other ranks are processed, and will report failures for ranks that failed to respond in time. Note that if one rank does not reach the monitored_barrier (for example due to a hang), all other ranks would fail in monitored_barrier. This collective will block all processes/ranks in the group, until the whole group exits the function successfully, making it useful for debugging and synchronizing. However, it can have a performance impact and should only be used for debugging or scenarios that require full synchronization points on the host-side. For debugging purposes, this barrier can be inserted before the application's collective calls to check if any ranks are desynchronized. .. note:: Note that this collective is only supported with the GLOO backend. Args: group (ProcessGroup, optional): The process group to work on. If ``None``, the default process group will be used. timeout (datetime.timedelta, optional): Timeout for monitored_barrier. If ``None``, the default process group timeout will be used. wait_all_ranks (bool, optional): Whether to collect all failed ranks or not. By default, this is ``False`` and ``monitored_barrier`` on rank 0 will throw on the first failed rank it encounters in order to fail fast. By setting ``wait_all_ranks=True`` ``monitored_barrier`` will collect all failed ranks and throw an error containing information about all failed ranks. Returns: ``None``. Example:: >>> # xdoctest: +SKIP("need process group init") >>> # Note: Process group initialization omitted on each rank. >>> import torch.distributed as dist >>> if dist.get_rank() != 1: >>> dist.monitored_barrier() # Raises exception indicating that >>> # rank 1 did not call into monitored_barrier. >>> # Example with wait_all_ranks=True >>> if dist.get_rank() == 0: >>> dist.monitored_barrier(wait_all_ranks=True) # Raises exception >>> # indicating that ranks 1, 2, ... world_size - 1 did not call into >>> # monitored_barrier. """ # Need to call rank not in group before using the group, otherwise # "Invalid process group" error is raised. if _rank_not_in_group(group): _warn_not_in_group("monitored_barrier") return if get_backend(group) != Backend.GLOO: raise RuntimeError("monitored_barrier is only implemented for GLOO backend.") if timeout is None: timeout = default_pg_timeout group_to_use = _get_default_group() if group is None else group return group_to_use.monitored_barrier(timeout, wait_all_ranks=wait_all_ranks) def _create_process_group_wrapper( wrapped_pg: ProcessGroup, store_prefix: str, store: Store, rank: int, world_size: int, timeout: timedelta = default_pg_timeout, ): # Create a separate prefix store for the helper process group. prefix = f"{PG_WRAPPER_STORE_PREFIX}:{store_prefix}" store = PrefixStore(prefix, store) helper_pg = ProcessGroupGloo(store, rank, world_size, timeout=timeout) # Wrap the underlying pg with ProcessGroupWrapper. wrapped_pg = _ProcessGroupWrapper(wrapped_pg, helper_pg) return wrapped_pg def _process_group_name(ranks, use_hashed_name): global _world if use_hashed_name: pg_name = hashlib.sha1(bytes("_".join(map(str, ranks)), "utf-8")).hexdigest() while pg_name in _world.pg_names.values(): pg_name = hashlib.sha1(bytes(pg_name + "_", "utf-8")).hexdigest() else: pg_name = str(_world.group_count) _world.group_count += 1 return pg_name def _get_backend_from_str(backend: Optional[str] = None) -> Backend: # Default to the same backend as the global process group # if backend is not specified. if not backend: backend = get_backend(_get_default_group()) return Backend(backend) @_time_logger def new_group(ranks=None, timeout=default_pg_timeout, backend=None, pg_options=None, use_local_synchronization=False): """ Creates a new distributed group. This function requires that all processes in the main group (i.e. all processes that are part of the distributed job) enter this function, even if they are not going to be members of the group. Additionally, groups should be created in the same order in all processes. .. warning:: Using multiple process groups with the ``NCCL`` backend concurrently is not safe and the user should perform explicit synchronization in their application to ensure only one process group is used at a time. This means collectives from one process group should have completed execution on the device (not just enqueued since CUDA execution is async) before collectives from another process group are enqueued. See `Using multiple NCCL communicators concurrently `_ for more details. Args: ranks (list[int]): List of ranks of group members. If ``None``, will be set to all ranks. Default is ``None``. timeout (timedelta, optional): Timeout for operations executed against the process group. Default value equals 30 minutes. This is applicable for the ``gloo`` backend. For ``nccl``, this is applicable only if the environment variable ``NCCL_BLOCKING_WAIT`` or ``NCCL_ASYNC_ERROR_HANDLING`` is set to 1. When ``NCCL_BLOCKING_WAIT`` is set, this is the duration for which the process will block and wait for collectives to complete before throwing an exception. When ``NCCL_ASYNC_ERROR_HANDLING`` is set, this is the duration after which collectives will be aborted asynchronously and the process will crash. ``NCCL_BLOCKING_WAIT`` will provide errors to the user which can be caught and handled, but due to its blocking nature, it has a performance overhead. On the other hand, ``NCCL_ASYNC_ERROR_HANDLING`` has very little performance overhead, but crashes the process on errors. This is done since CUDA execution is async and it is no longer safe to continue executing user code since failed async NCCL operations might result in subsequent CUDA operations running on corrupted data. Only one of these two environment variables should be set. backend (str or Backend, optional): The backend to use. Depending on build-time configurations, valid values are ``gloo`` and ``nccl``. By default uses the same backend as the global group. This field should be given as a lowercase string (e.g., ``"gloo"``), which can also be accessed via :class:`Backend` attributes (e.g., ``Backend.GLOO``). If ``None`` is passed in, the backend corresponding to the default process group will be used. Default is ``None``. pg_options (ProcessGroupOptions, optional): process group options specifying what additional options need to be passed in during the construction of specific process groups. i.e. for the ``nccl`` backend, ``is_high_priority_stream`` can be specified so that process group can pick up high priority cuda streams. use_local_synchronization (bool, optional): perform a group-local barrier at the end of the process group creation. This is different in that non-member ranks don't need to call into API and don't join the barrier. Returns: A handle of distributed group that can be given to collective calls or None if the rank is not part of ``ranks``. N.B. use_local_synchronization doesn't work with MPI. N.B. While use_local_synchronization=True can be significantly faster with larger clusters and small process groups, care must be taken since it changes cluster behavior as non-member ranks don't join the group barrier(). N.B. use_local_synchronization=True can lead to deadlocks when each rank creates multiple overlaping process groups. To avoid that, make sure all ranks follow the same global creation order. """ return _new_group_with_tag(ranks, timeout, backend, pg_options, None, use_local_synchronization=use_local_synchronization) def _new_group_with_tag( ranks=None, timeout=default_pg_timeout, backend=None, pg_options=None, pg_tag=None, use_local_synchronization=False ): """ This is a variant of ``new_group`` that exposes tag creation. :: N.B. The mechanism is experimental and tied to the functional collectives effort, see ``torch.distributed._functional_collectives`` for reference on how to use it. """ global _world default_pg = _get_default_group() default_backend, default_store = _world.pg_map[default_pg] global_rank = default_pg.rank() global_world_size = default_pg.size() # Default to the same backend as the global process group # if the backend is not specified. if not backend: backend = default_backend backend = Backend(backend) if use_local_synchronization: # MPI backend doesn't have have a way for us to perform a partial sync if backend == Backend.MPI: raise RuntimeError("MPI backend doesn't support use_local_synchronization=True") if ranks is not None and get_rank() not in ranks: return None # checks the input ranks if ranks is not None: ranks = sorted(ranks) group_world_size = len(ranks) if group_world_size > global_world_size: raise RuntimeError( "the new group's world size should be less or " "equal to the world size set by " "init_process_group" ) # check ranks' sanity for rank in ranks: if rank < 0 or rank >= global_world_size: raise RuntimeError( "The new group's rank should be within the " "the world_size set by init_process_group" ) if global_rank in ranks: group_rank = ranks.index(global_rank) else: group_rank = None else: ranks = list(range(global_world_size)) group_world_size = global_world_size group_rank = global_rank group_name = _process_group_name(ranks, use_hashed_name=use_local_synchronization) pg, pg_store = _new_process_group_helper( group_world_size, group_rank, ranks, backend, default_store, group_name=group_name, pg_options=pg_options, timeout=timeout, pg_tag=pg_tag ) # Create the global rank to group rank mapping _world.pg_group_ranks[pg] = { global_rank: group_rank for group_rank, global_rank in enumerate(ranks) } if _is_barrier_after_init() == 1: # barrier at the end to ensure that once we return from this method, all # process groups including global variables (if any) are updated # correctly on all ranks. # Update 04/2023: for large-scale runs, this barrier (esp. store-based # barrier) may be costly and/or unscalable. Also, in a lot of cases, # these barriers may be unnecessary, as proven by a green CI after # removal. An environment variable `TORCH_DIST_INIT_BARRIER` has been # added which enables this barrier only when set to 1. logger.info( "Performing barrier after ProcessGroup initialization since " "TORCH_DIST_INIT_BARRIER = 1" ) if backend == Backend.MPI: # MPI doesn't have store. barrier() else: barrier_store = pg_store if use_local_synchronization else default_store world_size = len(ranks) if use_local_synchronization else get_world_size() # Use store based barrier here since barrier() used a bunch of # default devices and messes up NCCL internal state. _store_based_barrier(global_rank, barrier_store, group_name, world_size, timeout) return pg def new_subgroups( group_size=None, group=None, timeout=default_pg_timeout, backend=None, pg_options=None, ): """ Creates subgroups of equal size. By default, it creates intra-machine subgroups, where each of which contains all the ranks of a machine, based on the assumption that each machine has the same number of devices. This is a convenience API that calls ``new_group`` to generate multiple subgroups. It requires that all processes in the main group (i.e. all processes that are part of the distributed job) enter this function, even if they are not going to be members of the group. .. warning:: If ``group_size`` is passed in, the world size must be divisible by ``group_size``. If no ``group_size`` is passed in, it believe that you are creating a group based on CUDA and determining the group size by number of CUDA devices, and if not all the machines have the same number of devices, the subgroup division will be different across nodes and can cause unexpected behaviors. Therefore, if you are creating a subgroup that does not depend on CUDA (such as Gloo on CPU), please pass in ``group_size`` correctly. .. warning:: Using multiple process groups with the ``NCCL`` backend concurrently is not safe and the user should perform explicit synchronization in their application to ensure only one process group is used at a time. This means collectives from one process group should have completed execution on the device (not just enqueued since CUDA execution is async) before collectives from another process group are enqueued. See `Using multiple NCCL communicators concurrently `_ for more details. Args: group_size (int, optional): The size of each subgroup. If ``None``, the default subgroup size is equal to the number of devices on each machine, based on the assumption that each machine has exactly the same number of devices. Default is ``None``. timeout (timedelta, optional): Timeout for operations executed against the process group. Default value equals 30 minutes. This is applicable for the ``gloo`` backend. For ``nccl``, this is applicable only if the environment variable ``NCCL_BLOCKING_WAIT`` or ``NCCL_ASYNC_ERROR_HANDLING`` is set to 1. When ``NCCL_BLOCKING_WAIT`` is set, this is the duration for which the process will block and wait for collectives to complete before throwing an exception. When ``NCCL_ASYNC_ERROR_HANDLING`` is set, this is the duration after which collectives will be aborted asynchronously and the process will crash. ``NCCL_BLOCKING_WAIT`` will provide errors to the user which can be caught and handled, but due to its blocking nature, it has a performance overhead. On the other hand, ``NCCL_ASYNC_ERROR_HANDLING`` has very little performance overhead, but crashes the process on errors. This is done since CUDA execution is async and it is no longer safe to continue executing user code since failed async NCCL operations might result in subsequent CUDA operations running on corrupted data. Only one of these two environment variables should be set. backend (str or Backend, optional): The backend to use. Depending on build-time configurations, valid values are ``gloo`` and ``nccl``. By default uses the same backend as the global group. This field should be given as a lowercase string (e.g., ``"gloo"``), which can also be accessed via :class:`Backend` attributes (e.g., ``Backend.GLOO``). If ``None`` is passed in, the backend corresponding to the default process group will be used. Default is ``None``. pg_options (ProcessGroupOptions, optional): process group options specifying what additional options need to be passed in during the construction of specific process groups. i.e. for the ``nccl`` backend, ``is_high_priority_stream`` can be specified so that process group can pick up high priority cuda streams. Returns: The subgroup containing the current rank, and all the subgroups used for cleanup. Examples: >>> # Create intra-machine subgroups. >>> # xdoctest: +SKIP("need process group init") >>> cur_subgroup, subgroups = dist.new_subgroups() >>> # Allreduce within the machine. >>> rank = dist.get_rank() >>> tensor = torch.ones(1, device=rank) * rank >>> dist.all_reduce(tensor, group=cur_subgroup) >>> tensor tensor([8]) # Assume 8 is the number of CUDA devices per machine. >>> # Cleanup. >>> for subgroup in subgroups: >>> dist.destroy_process_group(subgroup) """ if group_size is None: if not torch.cuda.is_available(): raise ValueError("Default group size only takes effect when CUDA is available." "If your subgroup using a backend that does not depend on CUDA," "please pass in 'group_size' correctly.") group_size = torch.cuda.device_count() if group_size <= 0: raise ValueError(f"The arg 'group_size' ({group_size}) must be positive") world_size = get_world_size() if world_size < group_size: raise ValueError(f"The arg 'group_size' ({group_size}) must not exceed the world size ({world_size})") if world_size % group_size != 0: raise ValueError("The world size must be divisible by 'group_size'") subgroups = [] cur_subgroup = None for subgroup_id in range(world_size // group_size): start_rank = subgroup_id * group_size end_rank = start_rank + group_size ranks_in_subgroup = list(range(start_rank, end_rank)) subgroup = new_group( ranks=ranks_in_subgroup, timeout=timeout, backend=backend, pg_options=pg_options, ) subgroups.append(subgroup) rank = get_rank() if rank in ranks_in_subgroup: cur_subgroup = subgroup logger.info( "Rank %s is assigned to subgroup %s", rank, ranks_in_subgroup ) return cur_subgroup, subgroups def new_subgroups_by_enumeration( ranks_per_subgroup_list, timeout=default_pg_timeout, backend=None, pg_options=None, ): """ Creates subgroups by dividing the global world, where the division is specified by a nested list of ranks. The subgroups cannot have overlap, and some ranks may not have to be in any subgroup. This is a convenience API that calls ``new_group`` to generate multiple subgroups. It requires that all processes in the main group (i.e. all processes that are part of the distributed job) enter this function, even if they are not going to be members of the group. .. warning:: Using multiple process groups with the ``NCCL`` backend concurrently is not safe and the user should perform explicit synchronization in their application to ensure only one process group is used at a time. This means collectives from one process group should have completed execution on the device (not just enqueued since CUDA execution is async) before collectives from another process group are enqueued. See `Using multiple NCCL communicators concurrently `_ for more details. Args: ranks_per_subgroup_list (list[list[int]]): A nested list of ranks of group members. timeout (timedelta, optional): Timeout for operations executed against the process group. Default value equals 30 minutes. This is applicable for the ``gloo`` backend. For ``nccl``, this is applicable only if the environment variable ``NCCL_BLOCKING_WAIT`` or ``NCCL_ASYNC_ERROR_HANDLING`` is set to 1. When ``NCCL_BLOCKING_WAIT`` is set, this is the duration for which the process will block and wait for collectives to complete before throwing an exception. When ``NCCL_ASYNC_ERROR_HANDLING`` is set, this is the duration after which collectives will be aborted asynchronously and the process will crash. ``NCCL_BLOCKING_WAIT`` will provide errors to the user which can be caught and handled, but due to its blocking nature, it has a performance overhead. On the other hand, ``NCCL_ASYNC_ERROR_HANDLING`` has very little performance overhead, but crashes the process on errors. This is done since CUDA execution is async and it is no longer safe to continue executing user code since failed async NCCL operations might result in subsequent CUDA operations running on corrupted data. Only one of these two environment variables should be set. backend (str or Backend, optional): The backend to use. Depending on build-time configurations, valid values are ``gloo`` and ``nccl``. By default uses the same backend as the global group. This field should be given as a lowercase string (e.g., ``"gloo"``), which can also be accessed via :class:`Backend` attributes (e.g., ``Backend.GLOO``). If ``None`` is passed in, the backend corresponding to the default process group will be used. Default is ``None``. pg_options (ProcessGroupOptions, optional): process group options specifying what additional options need to be passed in during the construction of specific process groups. i.e. for the ``nccl`` backend, ``is_high_priority_stream`` can be specified so that process group can pick up high priority cuda streams. Returns: The subgroup containing the current rank, and all the subgroups used for cleanup. Examples: >>> # Create two subgroups, where each has 2 processes. >>> # xdoctest: +SKIP("need process group init") >>> cur_subgroup, subgroups = dist.new_subgroups(ranks=[[0, 2], [1, 3]]) >>> rank = dist.get_rank() >>> tensor = torch.ones(1, device=rank) * rank >>> dist.all_reduce(tensor, group=cur_subgroup) >>> tensor tensor([2]) # Subgroup 0: ranks 0 and 2 tensor([4]) # Subgroup 1: ranks 1 and 3 """ if ranks_per_subgroup_list is None or len(ranks_per_subgroup_list) == 0: raise ValueError("The arg 'ranks_per_subgroup_list' cannot be empty") subgroups = [] cur_subgroup = None # Create a mapping from rank to subgroup to check if there is any subgroup overlap. rank_to_ranks_dict = {} # type: ignore[var-annotated] for ranks in ranks_per_subgroup_list: subgroup = new_group( ranks=ranks, timeout=timeout, backend=backend, pg_options=pg_options, ) subgroups.append(subgroup) my_rank = get_rank() for rank in ranks: if rank in rank_to_ranks_dict: raise ValueError( f"Rank {rank} has appeared in both subgroup {rank_to_ranks_dict[rank]} and {ranks}" ) rank_to_ranks_dict[rank] = ranks if my_rank == rank: cur_subgroup = subgroup logger.info("Rank %s is assigned to subgroup %s", rank, ranks) return cur_subgroup, subgroups def _find_pg_by_ranks_and_tag(tag: str, ranks: List[int]) -> ProcessGroup: if len(tag) > 0 and not tag.startswith("ptd:") and not tag.startswith("user:"): tag = f"user:{tag}" for group in _world.tags_to_pg.get(tag, []): if group.size() != len(ranks): continue group_ranks = get_process_group_ranks(group) good = all(r in group_ranks for r in ranks) if good: return group return None def _find_or_create_pg_by_ranks_and_tag(tag: str, ranks: List[int], stride: int) -> ProcessGroup: assert len(ranks) % stride == 0, f"Ranks length ({len(ranks)}) must be divisible by stride ({stride})" my_rank = get_rank() my_ranks = None if stride == len(ranks): my_ranks = ranks.copy() assert my_rank in my_ranks, "rankset doesn't include the current node" else: for i in range(0, len(ranks), stride): rank_set = ranks[i : i + stride] if my_rank in rank_set: my_ranks = rank_set assert my_ranks is not None, "rankset doesn't include the current node" my_ranks.sort() pg = _find_pg_by_ranks_and_tag(tag, my_ranks) if pg is not None: return pg if tag == "": raise ValueError("Cannot automatically create PG with empty tag") # TODO copy settings and timeout from default PG return _new_group_with_tag(my_ranks, pg_tag=tag) def _get_group_tag(pg: ProcessGroup) -> str: """ Returns the tag associated with ``pg``. """ tag = _world.pg_to_tag[pg] if tag.startswith("user:"): tag = tag[5:] return tag def _get_process_group_name(pg: ProcessGroup) -> str: return _world.pg_names[pg] # This ops are not friently to TorchDynamo. So, we decide to disallow these ops # in FX graph, allowing them to run them on eager, with torch.compile. dynamo_unsupported_distributed_c10d_ops = [ all_reduce_multigpu, recv, all_gather_object, all_gather_coalesced, all_to_all_single, all_reduce, gather_object, all_to_all, all_reduce_coalesced, gather, broadcast_object_list, barrier, reduce_multigpu, scatter, scatter_object_list, reduce, reduce_scatter_multigpu, all_gather, broadcast_multigpu, all_gather_multigpu, reduce_scatter, all_gather_into_tensor, broadcast, reduce_scatter_tensor, send, ]