sg^4ddlZddlmZmZddlmZddlmZmZm Z m Z m Z ddl Z ddl mZgdZGddZGdd eZGd d e ZGd d Zy)N)ABCabstractmethod) TracebackType)AnyList NamedTupleOptionalType)JoinHookJoinableJoinc(eZdZdZddZdeddfdZy)r a This defines a join hook, which provides two entry points in the join context manager. Entry points : a main hook, which is called repeatedly while there exists a non-joined process, and a post-hook, which is called once all processes have joined. To implement a join hook for the generic join context manager, define a class that inherits from :class:`JoinHook` and override ``main_hook()`` and ``post_hook()`` as appropriate. returnNcy)zCall this hook while there exists a non-joined process to shadow collective communications in a training iteration. Training iteration i.e., in one forward pass, backward pass, and optimizer step. Nselfs T/var/www/html/venv/lib/python3.12/site-packages/torch/distributed/algorithms/join.py main_hookzJoinHook.main_hookis_last_joinercy)aK Call hook after all processes have joined. It is passed an additional ``bool`` argument ``is_last_joiner``, which indicates if the rank is one of the last to join. Arguments: is_last_joiner (bool): ``True`` if the rank is one of the last to join; ``False`` otherwise. Nr)rrs r post_hookzJoinHook.post_hook rrrN)__name__ __module__ __qualname____doc__rboolrrrrr r s       rr ceZdZdZedfd ZedefdZeede jfdZ eede fdZ xZS)r a_ This defines an abstract base class for joinable classes. A joinable class (inheriting from :class:`Joinable`) should implement :meth:`join_hook`, which returns a :class:`JoinHook` instance, in addition to :meth:`join_device` and :meth:`join_process_group` that return device and process group information, respectively. rcTt|tj|_yN)super__init__ _JoinConfigconstruct_disabled_join_config _join_config)r __class__s rr%zJoinable.__init__7s 'FFHrc y)a Return a :class:`JoinHook` instance for the given :class:`Joinable`. Arguments: kwargs (dict): a :class:`dict` containing any keyword arguments to modify the behavior of the join hook at run time; all :class:`Joinable` instances sharing the same join context manager are forwarded the same value for ``kwargs``. Nr)rkwargss r join_hookzJoinable.join_hook<s rcy)zeReturn the device from which to perform collective communications needed by the join context manager.Nrrs r join_devicezJoinable.join_deviceI rcy)zfReturns the process group for the collective communications needed by the join context manager itself.Nrrs rjoin_process_groupzJoinable.join_process_groupOr/rr)rrrrrr%r r,propertytorchdevicer.rr1 __classcell__)r)s@rr r ,sII  X     U\\   C  rr c@eZdZUdZeed<eed<eed<edZy)r&zdThis includes all fields needed from a :class:`Joinable` instance for the join context manager side.enablethrow_on_early_terminationis_first_joinablectdddS)zReturn a :class:`_JoinConfig` instance indicating that join-related logic should be disabled. e.g. if the caller is not in a join context manager. Fr7r8r9)r&rrrr'z*_JoinConfig.construct_disabled_join_config]s Ue  rN)rrrrr __annotations__ staticmethodr'rrrr&r&Vs(o L $$  rr&ceZdZdZ ddeededefdZddZddZ d Z d e e e d e e d e efd ZdZdZedefdZy)r a This class defines the generic join context manager, which allows custom hooks to be called after a process joins. These hooks should shadow the collective communications of non-joined processes to prevent hanging and erroring and to ensure algorithmic correctness. Refer to :class:`JoinHook` for details about the hook definition. .. warning:: The context manager requires each participating :class:`Joinable` to call the method :meth:`notify_join_context()` before its own per- iteration collective communications to ensure correctness. .. warning:: The context manager requires that all ``process_group`` attributes in the :class:`JoinHook` objects are the same. If there are multiple :class:`JoinHook` objects, then the ``device`` of the first is used. The process group and device information is used for checking for non- joined processes and for notifying processes to throw an exception if ``throw_on_early_termination`` is enabled, both of which using an all- reduce. Arguments: joinables (List[Joinable]): a list of the participating :class:`Joinable` s; their hooks are iterated over in the given order. enable (bool): a flag enabling uneven input detection; setting to ``False`` disables the context manager's functionality and should only be set when the user knows the inputs will not be uneven (default: ``True``). throw_on_early_termination (bool): a flag controlling whether to throw an exception upon detecting uneven inputs (default: ``False``). Example:: >>> import os >>> import torch >>> import torch.distributed as dist >>> import torch.multiprocessing as mp >>> # xdoctest: +SKIP >>> import torch.nn.parallel.DistributedDataParallel as DDP >>> import torch.distributed.optim.ZeroRedundancyOptimizer as ZeRO >>> from torch.distributed.algorithms.join import Join >>> >>> # On each spawned worker >>> def worker(rank): >>> dist.init_process_group("nccl", rank=rank, world_size=2) >>> model = DDP(torch.nn.Linear(1, 1).to(rank), device_ids=[rank]) >>> optim = ZeRO(model.parameters(), torch.optim.Adam, lr=0.01) >>> # Rank 1 gets one more input than rank 0 >>> inputs = [torch.tensor([1.]).to(rank) for _ in range(10 + rank)] >>> with Join([model, optim]): >>> for input in inputs: >>> loss = model(input).sum() >>> loss.backward() >>> optim.step() >>> # All ranks reach here without hanging/erroring joinablesr7r8c t|dk(r td||_|jDcgc]}|jdi|c}|_||_||_|j|jycc}w)Nrz7The join context manager requires at least one joinabler) len ValueError _joinablesr, _join_hooks_enable_throw_on_early_termination_set_joinable_configs_extract_dist_info)rr?r7r8r+joinables rr%z Join.__init__s~ y>Q VW W#9= -5 H   ( (  +E( ""$ !  sA?Nct|jdkDsJd}|jD]+}t|j|j||_d}-y)zESet the :class:`_JoinConfig` of each participating :class:`Joinable`.rTr;FN)rArCr&rErFr()rr9rIs rrGzJoin._set_joinable_configssZ4??#a'''  &H$/||+/+K+K"3%H ! !&   &rcd}d}|jD]:}| |j}n||jk7r td|/|j}<||_t j |j|_||_y)a Extract the process group and device information from the joinables. If there are multiple joinables, then the context manager uses the first specified device. Preconditions: ``self._joinables`` is not ``None`` and is non-empty. Raises: ValueError If there are multiple conflicting ``process_group`` attributes among the ``Joinable`` objects. Nz7Using join context manager with multiple process groups) rCr1rBr._process_groupdistget_rank_rank_device)r process_groupr4rIs rrHzJoin._extract_dist_infos  .H$ ( ; ; ("="== M~!-- .,]]4#6#67  rcyr#rrs r __enter__zJoin.__enter__s rtypevalue tracebackc |jr|ryd}d}d}d}tjd|s||kDr)tjd|d|jd |d |j }|dk(rd}nD|j r|j|jD]} | jd}|d z }|s|jD]} | j|y) z Repeatedly runs the main hooks until all processes join; then, runs the post-hooks. Raises: RuntimeError If ``throw_on_early_termination=True``. NFTrioncez+Detected uneven input skew of greater than z. This means that rank z has at least zz fewer inputs than other currently-active ranks. This level of skew could lead to performance degradation during training.) rEwarnings simplefilterwarnrO_get_num_nonjoined_procsrF_notify_procs_to_terminaterDrr) rrTrUrVall_procs_joinedriWARN_THRESHOLDnum_nonjoined_procsr,s r__exit__z Join.__exit__s||t   f%">! A%&&=zzl.0@A33#'"?"?"A "a'#' 33335"&!1!1*I'')*"'Q/#4)) 0I    / 0rctjd|j}tj||j |j S)zaReturn the number of non-joined processes by shadowing an all-reduce in the non-joined processes.rYr4group)r3zerosrPrM all_reducerLitem)rrbs rr]zJoin._get_num_nonjoined_procss;#kk!DLLA +43F3FG"''))rctjd|j}tj||j t d|jd)zSchedule an all-reduce to notify non-joined processes to terminate. Also raise a ``RuntimeError`` indicating that the current process has exhausted its inputs. rYrerfzRank z exhausted all inputs.)r3onesrPrMrirL RuntimeErrorrO)rrls rr^zJoin._notify_procs_to_terminatesE zz!DLL1 D$7$78U4::,.DEFFrrIct|dsJdt|d|j}|jr |jsy|j }|j }tjd|}tj||d}|jrKtjd|}tj|| |j}|r td |S) aH Notifies the join context manager that the calling process has not yet joined. Then, if ``throw_on_early_termination=True``, checks if uneven inputs have been detected (i.e. if one process has already joined) and throws an exception if so. This method should be called from a :class:`Joinable` object before its per-iteration collective communications. For example, this should be called at the beginning of the forward pass in :class:`DistributedDataParallel`. Only the first :class:`Joinable` object passed into the context manager performs the collective communications in this method, and for the others, this method is vacuous. Arguments: joinable (Joinable): the :class:`Joinable` object calling this method. Returns: An async work handle for the all-reduce meant to notify the context manager that the process has not yet joined if ``joinable`` is the first one passed into the context manager; ``None`` otherwise. r(zCheck that the z/ constructor calls the ``Joinable`` constructorNrYreT)rgasync_oprfzLDetected at least one rank that exhausted inputs. Throwing across all ranks.)hasattrrTr(r9r7r.r1r3rlrMrir8rhrjrm)rI join_configr4rQrlworkrh should_throws rnotify_join_contextzJoin.notify_join_context(s4x0 d8n-.' ' 0 ++ ,,K4F4F%% 33 zz!F+t=4H  1 1KK&1E OOE 7 ::""%) "$ &< 20tM*+20 &20M* 20h* G4h44rr )rZabcrrtypesrtypingrrrr r r3torch.distributed distributedrM__all__r r r&r rrrr|sP#88  +  <' s' T * $uur