import contextlib import io import logging import os import pickle import time import warnings from datetime import timedelta from typing import Callable, Dict, Optional, Tuple, Union import torch from torch._C._distributed_c10d import ( AllreduceCoalescedOptions, AllreduceOptions, AllToAllOptions, BarrierOptions, BroadcastOptions, GatherOptions, PrefixStore, ProcessGroup, ReduceOp, ReduceOptions, ReduceScatterOptions, ScatterOptions, Store, DebugLevel, get_debug_level, ) from torch._six import string_classes from .constants import default_pg_timeout from .rendezvous import register_rendezvous_handler, rendezvous # noqa: F401 # This module is wildcard imported from torch.distributed. # TODO: specify __all__ _MPI_AVAILABLE = True _NCCL_AVAILABLE = True _GLOO_AVAILABLE = True _pickler = pickle.Pickler _unpickler = pickle.Unpickler try: from torch._C._distributed_c10d import ProcessGroupMPI except ImportError: _MPI_AVAILABLE = False try: from torch._C._distributed_c10d import ProcessGroupNCCL except ImportError: _NCCL_AVAILABLE = False try: from torch._C._distributed_c10d import ProcessGroupGloo from torch._C._distributed_c10d import _ProcessGroupWrapper except ImportError: _GLOO_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(object): """ An enum-like class of available backends: GLOO, NCCL, 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" MPI = "mpi" TCP = "tcp" _plugins: Dict[str, Callable] = {} def __new__(cls, name: str): if not isinstance(name, string_classes): raise ValueError("Backend name must be a string, but got: {}".format(name)) value = getattr(Backend, name.upper(), Backend.UNDEFINED) if value == Backend.TCP: raise ValueError( "TCP backend has been deprecated. Please use " "Gloo or MPI backend for collective operations " "on CPU tensors." ) elif value == Backend.UNDEFINED: raise ValueError("Invalid backend: '{}'".format(name)) elif value != Backend.GLOO and value != Backend.NCCL and value != Backend.MPI: value = name.lower() return value @classmethod def register_backend(cls, name, func): """ 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``. .. note:: This support of 3rd party backend is experimental and subject to change. """ 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.upper()) Backend._plugins[name.upper()] = func # `_backend`, `dist_backend`, and `reduce_op` are here to maintain backward # compatibility with pre-c10d distributed package. # TODO: remove them when users are ready to take a hard dependency on PyTorch 1. _backend: str = Backend.UNDEFINED dist_backend = Backend class _reduce_op(object): 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.__members__.items(): setattr(self, k, v) self.__members__ = ReduceOp.__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 group(object): # Points to the default PG once initialized. WORLD: Optional[ProcessGroup] = None class GroupMember(object): # Alias to group.WORLD for backward compatibility WORLD = group.WORLD NON_GROUP_MEMBER = object() # 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) _pg_map: Dict[ProcessGroup, Tuple[str, Optional[Store]]] = {} # Process group's names, map from ProcessGroup to str _pg_names: Dict[ProcessGroup, str] = {} # Process group's global rank to local rank mapping _pg_group_ranks: Dict[ProcessGroup, Dict[int, int]] = {} # Default process group state _default_pg_init_method = None # Process group count for default naming _group_count = 0 STORE_BASED_BARRIER_PREFIX = "store_based_barrier_key" def _store_based_barrier(rank, store, timeout): """ 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 = "{}:{}".format(STORE_BASED_BARRIER_PREFIX, _group_count) store.add(store_key, 1) logger.info("Added key: {} to store for rank: {}".format(store_key, rank)) # Now wait for all workers to check in with the store. world_size = get_world_size() # Use 'add' instead of 'get' since for some store implementations 'add' # doesn't work well with 'get'. Ideally the store implementations should # be fixed, but for backward compatiblity reasons it is risky to change # the store implementations. Once, we completely migrate away from these # legacy stores, we can use 'get' here instead. worker_count = store.add(store_key, 0) start = time.time() log_time = time.time() while worker_count != world_size: time.sleep(0.01) worker_count = store.add(store_key, 0) # Print status periodically to keep track. if timedelta(seconds=(time.time() - log_time)) > timedelta(seconds=10): logger.info( "Waiting in store based barrier to initialize process group for " "rank: {}, key: {} (world_size={}, worker_count={}, timeout={})".format( rank, store_key, world_size, worker_count, timeout ) ) log_time = time.time() if timedelta(seconds=(time.time() - start)) > timeout: raise RuntimeError( "Timed out initializing process group in store based barrier on " "rank: {}, for key: {} (world_size={}, worker_count={}, timeout={})".format( rank, store_key, world_size, worker_count, timeout ) ) logger.info( f"Rank {rank}: Completed store-based barrier for key:{store_key} with {world_size} nodes." ) 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, rank): """ Helper that gets a given group's local rank in the group from a given global rank. """ if group is GroupMember.WORLD: raise RuntimeError( "group.WORLD does not have local rank to global " "rank mapping" ) if group not in _pg_group_ranks: raise RuntimeError("The given group does not exist") try: group_rank = _pg_group_ranks[group][rank] except KeyError: raise RuntimeError( f"The global rank {rank} is not part of the group {group}" ) from None return group_rank def _get_global_rank(group, group_rank): """ Helper that gets a given group's global rank from a given local rank in the group. """ if group is GroupMember.WORLD: raise RuntimeError( "group.WORLD does not have local rank to global " "rank mapping" ) group_rank_map = _pg_group_ranks[group] for rank, grp_rank in group_rank_map.items(): if grp_rank == group_rank: return rank raise RuntimeError("The group rank is not part of the group") 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( "Invalid function argument. Expected parameter `{}` " "to be of type torch.Tensor.".format(param_name) ) 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( "Invalid function argument. Expected parameter `{}` " "to be of type List[torch.Tensor].".format(param_name) ) 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 backend. """ 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``." ) backend = get_backend(p2p_op_list[0].group) if not all(backend == get_backend(p2p_op.group) for p2p_op in p2p_op_list): raise RuntimeError("All groups need to use the same backend.") def is_mpi_available(): """ Checks if the MPI backend is available. """ return _MPI_AVAILABLE def is_nccl_available(): """ Checks if the NCCL backend is available. """ return _NCCL_AVAILABLE def is_gloo_available(): """ Checks if the Gloo backend is available. """ return _GLOO_AVAILABLE def is_initialized(): """ Checking if the default process group has been initialized """ return GroupMember.WORLD is not None def is_torchelastic_launched(): """ 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 _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 = _pg_map[default_pg] return default_store def _update_default_pg(pg): GroupMember.WORLD = group.WORLD = pg def get_backend(group=None): """ 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 = _pg_map.get(pg, None) assert pg_store is not None return pg_store[0] def init_process_group( backend, init_method=None, timeout=default_pg_timeout, world_size=-1, rank=-1, store=None, group_name="", pg_options=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): The backend to use. Depending on build-time configurations, valid values include ``mpi``, ``gloo``, and ``nccl``. 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 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. 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. group_name (str, optional, deprecated): Group name. 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. """ global _pg_group_ranks 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://" backend = Backend(backend) if backend == Backend.MPI: if world_size != -1 or rank != -1: warnings.warn( "For MPI backend, world_size ({}) and rank ({}) " "are ignored since they are assigned by the " "MPI runtime.".format(world_size, rank) ) default_pg = _new_process_group_helper( -1, -1, [], Backend.MPI, 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) _pg_group_ranks[GroupMember.WORLD] = {i: i for i in range(GroupMember.WORLD.size())} # type: ignore[attr-defined, index] _backend = _pg_map[GroupMember.WORLD][0] # type: ignore[index] _default_pg_init_method = init_method # barrier at the end to ensure that once we return from this method, all # process groups including global variables are updated correctly on all # ranks. 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, timeout) # Set sequence numbers for gloo and nccl process groups. if get_backend(default_pg) in [Backend.GLOO, Backend.NCCL]: default_pg._set_sequence_number_for_group() def _new_process_group_helper( world_size, rank, group_ranks, backend, store, pg_options=None, group_name=None, timeout=default_pg_timeout, ): """ 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 ``group_ranks == []`` for the default group. """ global _pg_map global _group_count global _pg_names if not group_name: group_name = str(_group_count) _group_count += 1 if group_name in _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" ) # The list of group ranks is empty if we're creating the default group. is_default_group = len(group_ranks) == 0 backend = Backend(backend) pg: Union[ProcessGroupGloo, ProcessGroupMPI, ProcessGroupNCCL] if backend == 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." ) pg = ProcessGroupMPI.create(group_ranks) if not pg: return GroupMember.NON_GROUP_MEMBER _pg_map[pg] = (Backend.MPI, None) _pg_names[pg] = group_name else: # 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 group_ranks: return GroupMember.NON_GROUP_MEMBER # Use the group name as prefix in the default store, such that # a single store can be reused by multiple groups. prefix_store = PrefixStore(group_name, store) if backend == Backend.GLOO: if pg_options is not None: raise RuntimeError("GLOO options not supported") pg = ProcessGroupGloo(prefix_store, rank, world_size, timeout=timeout) # In debug mode and if GLOO is available, wrap in a wrapper PG that # enables enhanced collective checking for debugability. 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: pg = _create_process_group_wrapper( wrapped_pg=pg, store_prefix=group_name, store=store, rank=rank, world_size=world_size, timeout=timeout, ) _pg_map[pg] = (Backend.GLOO, store) _pg_names[pg] = group_name elif backend == 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 pg = ProcessGroupNCCL(prefix_store, rank, world_size, pg_options) # In debug mode and if GLOO is available, wrap in a wrapper PG that # enables enhanced collective checking for debugability. 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: pg = _create_process_group_wrapper( wrapped_pg=pg, store_prefix=group_name, store=store, rank=rank, world_size=world_size, timeout=timeout, ) _pg_map[pg] = (Backend.NCCL, store) _pg_names[pg] = group_name else: assert backend.upper() in Backend._plugins, ( f"unknown c10d backend type {backend.upper()}" ) pg = Backend._plugins[backend.upper()]( prefix_store, rank, world_size, timeout ) _pg_map[pg] = (backend, store) _pg_names[pg] = group_name return pg def destroy_process_group(group=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 _pg_map global _pg_names global _pg_group_ranks global _default_pg_init_method global _group_count if group == GroupMember.NON_GROUP_MEMBER: return if group is None: pg = GroupMember.WORLD else: pg = group assert pg is not None if _pg_map.get(pg, None) is None: raise RuntimeError("Invalid process group specified") if group is None or group == GroupMember.WORLD: _update_default_pg(None) _default_pg_init_method = None _pg_map.clear() _pg_names.clear() _pg_group_ranks.clear() # when process group doesn't have an explicit name (only WORLD (default) # process group can have an explicit name), we use global _group_counter # 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. _group_count = 0 else: del _pg_map[pg] del _pg_names[pg] del _pg_group_ranks[pg] def get_rank(group=None): """ 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=None): """ 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, dst, group=None, tag=0): """ Sends a tensor asynchronously. .. warning:: Modifying ``tensor`` before the request completes causes undefined behavior. 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, src=None, group=None, tag=0): """ Receives a tensor asynchronously. 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) def send(tensor, dst, group=None, tag=0): """ Sends a tensor synchronously. 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 """ _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() def recv(tensor, src=None, group=None, tag=0): """ 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 P2POp(object): """ 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 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, tensor, peer, group=None, tag=0): self.op = op self.tensor = tensor self.peer = peer self.group = group self.tag = tag def __new__(cls, op, tensor, peer, group=None, tag=0): _check_op(op) _check_single_tensor(tensor, "tensor") return object.__new__(cls) @contextlib.contextmanager def _batch_p2p_manager(backend): if backend == Backend.NCCL: ProcessGroupNCCL._group_start() try: yield finally: if backend == Backend.NCCL: ProcessGroupNCCL._group_end() 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 and Gloo 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: >>> 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) backend = get_backend(p2p_op_list[0].group) reqs = [] with _batch_p2p_manager(backend): for p2p_op in p2p_op_list: op = p2p_op.op tensor = p2p_op.tensor peer = p2p_op.peer curr_group = p2p_op.group tag = p2p_op.tag ret = op(tensor, peer, curr_group, tag) if ret is not None: reqs.append(ret) return reqs 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 """ 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() 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() 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 """ 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() 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: >>> # 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: default_pg = _get_default_group() work = default_pg.allreduce([tensor], opts) else: work = group.allreduce([tensor], opts) if async_op: return work else: work.wait() 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. """ _check_tensor_list(tensors, "tensor") 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() 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 """ 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() 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() 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 """ 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): 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) local_size = torch.LongTensor([byte_tensor.numel()]) return byte_tensor, local_size def _tensor_to_object(tensor, tensor_size): buf = tensor.numpy().tobytes()[:tensor_size] return _unpickler(io.BytesIO(buf)).load() def _check_for_nccl_backend(group): pg = group or _get_default_group() # Gate PG wrapper check on Gloo availability. if _GLOO_AVAILABLE: # It is not expected for PG to be wrapped many times, but support it just # in case while isinstance(pg, _ProcessGroupWrapper): pg = pg.wrapped_pg return ( is_nccl_available() and isinstance(pg, ProcessGroupNCCL) ) 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. object (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. Example:: >>> # 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 input_tensor, local_size = _object_to_tensor(obj) current_device = torch.device("cpu") is_nccl_backend = _check_for_nccl_backend(group) if is_nccl_backend: # See note about using torch.cuda.current_device() here in docstring. # We cannot simply use my_rank since rank == device is not necessarily # true. current_device = torch.device("cuda", torch.cuda.current_device()) input_tensor = input_tensor.to(current_device) local_size = local_size.to(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) 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. Example:: >>> # 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 appopriately. my_rank = get_rank() _validate_output_list_for_rank(my_rank, dst, object_gather_list) input_tensor, local_size = _object_to_tensor(obj) current_device = torch.device("cpu") is_nccl_backend = _check_for_nccl_backend(group) if is_nccl_backend: current_device = torch.device("cuda", torch.cuda.current_device()) input_tensor = input_tensor.to(current_device) local_size = local_size.to(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) if tensor.device != torch.device("cpu"): tensor = tensor.cpu() tensor_size = object_size_list[i] object_gather_list[i] = _tensor_to_object(tensor, tensor_size) 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 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()``. .. 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. Example:: >>> # 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 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) for obj in object_list]) object_sizes_tensor = torch.cat(size_list) else: object_sizes_tensor = torch.empty(len(object_list), dtype=torch.long) # 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. is_nccl_backend = _check_for_nccl_backend(group) current_device = None if device is not None: if is_nccl_backend and device.type != "cuda": raise ValueError("device type must be cuda for nccl backend") current_device = device else: current_device = torch.device("cpu") if is_nccl_backend: # See note about using torch.cuda.current_device() here in # docstring. We cannot simply use my_rank since rank == device is # not necessarily true. current_device = torch.device("cuda", torch.cuda.current_device()) if is_nccl_backend: object_sizes_tensor = object_sizes_tensor.to(current_device) # Broadcast object sizes broadcast(object_sizes_tensor, src=src, group=group) # Concatenate and broadcast serialized object tensors if my_rank == src: 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, ) if is_nccl_backend: object_tensor = object_tensor.to(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) 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. .. note:: Note that this API does not support the NCCL backend, as the tensor-based scatter collective is not supported by ProcessGroupNCCL. .. 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. Example:: >>> # 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(group) if my_rank == src: tensor_list, tensor_sizes = zip( *[_object_to_tensor(obj) 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) broadcast(max_tensor_size, src=src, group=group) # Scatter actual serialized objects output_tensor = torch.empty(max_tensor_size.item(), dtype=torch.uint8) 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) 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) 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: >>> # 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") 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() 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 Examples: >>> # All tensors below are of torch.int64 dtype. >>> # We have 2 process groups, 2 ranks. >>> output_tensor = torch.zeros(2, dtype=torch.int64) >>> output_tensor [tensor([0, 0])] # Rank 0 and 1 >>> tensor = torch.arange(1, dtype=torch.int64) + 1 + rank >>> tensor tensor([1]) # Rank 0 tensor([2]) # Rank 1 >>> dist.all_gather_base(output_tensor, tensor) >>> output_tensor tensor([1,2]) # Rank 0 tensor([1,2]) # Rank 1 .. warning:: `_all_gather_base` is experimental and subject to change. It is the caller's responsibility to ensure the output_tensor is correctly sized. """ _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_base") 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) ) if group is None: default_pg = _get_default_group() work = default_pg._allgather_base(output_tensor, input_tensor) else: work = group._allgather_base(output_tensor, input_tensor) if async_op: return work else: work.wait() 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. """ # 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, "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") 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." ) 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 = [] 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() 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 """ _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 = [] 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() 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. """ 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() 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. 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") 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() 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. """ _check_single_tensor(output, "output") _check_single_tensor(input, "input") if _rank_not_in_group(group): _warn_not_in_group("_reduce_scatter_base") return opts = ReduceScatterOptions() opts.reduceOp = op if group is None: default_pg = _get_default_group() work = default_pg._reduce_scatter_base(output, input, opts) else: work = group._reduce_scatter_base(output, input, opts) if async_op: return work else: work.wait() 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 cancatenated 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: >>> 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") 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() 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: >>> 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") 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() 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. Valid only for NCCL backend. 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() if device_ids is not None: if get_backend(group) != Backend.NCCL: raise RuntimeError( "Function argument device_ids not supported " "for the selected backend {}".format(get_backend(group)) ) 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 purposees, 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:: >>> # 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 new_group(ranks=None, timeout=default_pg_timeout, backend=None, pg_options=None): """ 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. Returns: A handle of distributed group that can be given to collective calls. """ global _pg_group_ranks default_pg = _get_default_group() default_backend, default_store = _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 # 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 backend = Backend(backend) pg = _new_process_group_helper( group_world_size, group_rank, ranks, backend, default_store, pg_options=pg_options, timeout=timeout, ) # Create the global rank to group rank mapping _pg_group_ranks[pg] = { global_rank: group_rank for group_rank, global_rank in enumerate(ranks) } # barrier at the end to ensure that once we return from this method, all # process groups including global variables are updated correctly on all # ranks. if backend == Backend.MPI: # MPI doesn't have 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(global_rank, default_store, timeout) # Set sequence numbers for gloo and nccl process groups. if pg != GroupMember.NON_GROUP_MEMBER and get_backend(pg) in [ Backend.GLOO, Backend.NCCL, ]: pg._set_sequence_number_for_group() return pg def new_subgroups( group_size=None, group=None, timeout=default_pg_timeout, backend=None, pg_options=None, ): """ Creates GPU 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 CUDA 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:: This API only works when CUDA is available. .. warning:: If ``group_size`` is passed in, the world size must be divisible by ``group_size``. If no ``group_size`` is passed in, and not all the machines have the same number of devices, the subgroup division will be different across nodes and can cause unexpected behaviors. .. 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. >>> 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 not torch.cuda.is_available(): raise ValueError("Subgroups can only be created when CUDA is available") if group_size is None: group_size = torch.cuda.device_count() world_size = get_world_size() if world_size < group_size: raise ValueError("The arg 'group_size' must not exceed the 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 {} is assigned to subgroup {}".format(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 GPU 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. >>> 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 not torch.cuda.is_available(): raise ValueError("Subgroups can only be created when CUDA is available") 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") world_size = get_world_size() 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( "Rank {} has appeared in both subgroup {} and {}".format( rank, rank_to_ranks_dict[rank], ranks ) ) rank_to_ranks_dict[rank] = ranks if my_rank == rank: cur_subgroup = subgroup logger.info("Rank {} is assigned to subgroup {}".format(rank, ranks)) return cur_subgroup, subgroups