^jLddlZddlZddlZddlZddlmZddlmZddlm Z ddl m Z m Z ddl mZddlZddlmZddlmZddlmZdd lmZdd lmZdd lmZdd lmZdd lm Z ddl!m"Z"m#Z#ddl$m%Z%m&Z&m'Z'ddl(m)Z)ddl*m+Z+m,Z,ddl-m.Z.ddl/m0Z0m1Z1m2Z2e rddl3m4Z4gdZ5Gdde Z6ede7 d2dede+dejpdzde9d e:d!e#dzd"e fd#Z;ed$%e0dddddd$d&ded'eeGd*d+Z?ed$%dddde6jddd$d,ded'eTlog_exceptions checkpoint_idr1r5r2r4use_collectivesr=r>c ~tjjd|xs,tj xstj }|rt jddt5ttt||d}tt||||||cdddS#1swYyxYw) a< Save a distributed model in SPMD style. This function is different from ``torch.save()`` as it handles ``ShardedTensor`` , and ``DTensor`` by having each rank only save their local shards. For each ``Stateful`` object (having both a ``state_dict`` and a ``load_state_dict``), save will call ``state_dict`` before serialization. .. warning:: There are no guarantees of Backwards Compatibility across PyTorch versions for saved state_dicts. .. warning:: If using the `process_group` argument, make sure that only its ranks call `save_state_dict` and that all data in state_dict belong to it. .. note:: When saving checkpoint for FSDP's `ShardingStrategy.HYBRID_SHARD`, only one of the shard_group should be calling `save_state_dict` and the corresponding process group needs to be passed in. .. note:: If no process group is available, this function assumes the intention is to save the state_dict in the local process. .. note: Rank 0 is assumed to be the coordinator rank. Args: state_dict (Dict[str, Any]): The state_dict to save. checkpoint_id (Union[str, os.PathLike, None]): The ID of this checkpoint instance. The meaning of the checkpoint_id depends on the storage. It can be a path to a folder or to a file. It can also be a key if the storage is a key-value store. (Default: ``None``) storage_writer (Optional[StorageWriter]): Instance of StorageWriter used to perform writes. If this is not specified, DCP will automatically infer the writer based on the checkpoint_id. If checkpoint_id is also None, an exception will be raised. (Default: ``None``) planner (Optional[SavePlanner]): Instance of SavePlanner. If this is not specified, the default planner will be used. (Default: ``None``) process_group (Optional[ProcessGroup]): ProcessGroup to be used for cross-rank synchronization. (Default: ``None``) no_dist (bool): If ``True``, this function will assume the intent is to load a checkpoint on a single rank/process. (Default: ``False``) use_collectives (bool): If ``False``, this function will assume the intent is to save a checkpoint without using cross-rank synchronization. (Default: ``True``) This configuration is experimental and should be used with caution. It will change the format of the saved checkpoint and may not be backward compatible. Returns: Metadata: Metadata object for the saved checkpoint. Example: >>> # xdoctest: +SKIP >>> my_model = MyModule() >>> state_dict = {"model": my_model} >>> fs_storage_writer = torch.distributed.checkpoint.FileSystemWriter( ... "/checkpoint/1" ... ) >>> torch.distributed.checkpoint.save( >>> state_dict=state_dict, >>> storage_writer=fs_storage_writer, >>> ) .. note:: save_state_dict uses collectives to coordinate writes across ranks. For NCCL-based process groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by ``torch.cuda.current_device()`` and it is the user's responsibility to ensure that this is set so that each rank has an individual GPU, via ``torch.cuda.set_device()``. z!torch.distributed.checkpoint.savezptorch.distributed is disabled, unavailable or uninitialized, assuming the intent is to save in a single process. stacklevelF)reader)r0r1r2r4r5r>N)torch_C_log_api_usage_oncedist is_availableis_initializedwarningswarnrrrr r9_stateful_to_state_dict)r0r=r1r5r2r4r>s r.rrWs~ HH  !DEQd//11Q4;N;N;P7PG ~    >.-PUV  .z:)'+     s 36B33B<c2eZdZUdZeded<eded<y)r"a!This class contains futures for staging and upload completion. It is returned by async_save(). staging_completion is a future that indicates when local copy of state_dict is complete. upload_completion is a future that indicates when a checkpoint completed saving. Nstaging_completionupload_completion)r&r'r(r)r__annotations__r,r-r.r"r"st $d|#r-r")r=r1r5r2async_checkpointer_type async_stagerr4r>rQrRc htjjdtjrNtj r:|xs t } tjd| jvr tdd.|t|tr|nttddddtddfd td dt t"t"zffd } t%j&5} | j)| } |t*j,k(r t/n t1} | j3| |||||| }$|j5fd | j7dddt t rh| }t!}|fd t t"dt dfd}|j9s|j5|n|j;dt=|Std fd}|S#1swYxYw)a Asynchronous version of ``save``. This code first de-stages the state_dict on to the staging storage (defaults to CPU memory), and then calls the `save` in a separate thread. .. warning:: This feature is experimental and subject to change. If you provide an ``async_stager``, call ``close()`` after the last checkpoint is saved. Internally created default stagers are closed automatically. Args: state_dict (Dict[str, Any]): The state_dict to save. checkpoint_id (Union[str, os.PathLike, None]): The ID of this checkpoint instance. The meaning of the checkpoint_id depends on the storage. It can be a path to a folder or to a file. It can also be a key if the storage is a key-value store. (Default: ``None``) storage_writer (Optional[StorageWriter]): Instance of StorageWriter used to perform 'stage' and 'save'. If this is not specified, DCP will automatically infer the writer based on the checkpoint_id. If checkpoint_id is also None, an exception will be raised. (Default: ``None``) planner (Optional[SavePlanner]): Instance of SavePlanner. If this is not specified, the default planner will be used. (Default: ``None``) process_group (Optional[ProcessGroup]): ProcessGroup to be used for cross-rank synchronization. (Default: ``None``) async_checkpointer_type (AsyncCheckpointerType): whether to do checkpoint in separate thread or process (Default: ``AsyncCheckpointerType.THREAD``) async_stager (AsyncStager): provides staging implementation. If ``storage_writer`` implements ``AsyncStager`` and ``async_stager`` is not provided, ``storage_writer`` will be used for staging. User-provided stagers remain owned by the caller and should be closed after the last checkpoint is saved. no_dist (bool): If ``True``, this function will assume the intent is to save a checkpoint on a single rank/process. (Default: ``False``) use_collectives: If False, Save the checkpoint without rank coordination. (Default: ``True``) This configuration is experimental and should be used with caution. It will change the format of the saved checkpoint and may not be backward compatible. Returns: Future: A future holding the resultant Metadata object from `save`. Example: >>> # xdoctest: +SKIP >>> my_model = MyModule() >>> state_dict = {"model": my_model} >>> fs_storage_writer = torch.distributed.checkpoint.FileSystemWriter( ... "/checkpoint/1" ... ) >>> checkpoint_future = torch.distributed.checkpoint.async_save( >>> state_dict=state_dict, >>> storage_writer=fs_storage_writer, >>> ) >>> >>> # ... do some work ... >>> >>> checkpoint_future.result() z'torch.distributed.checkpoint.async_savecpuzfA CPU backend must be enabled for async save; try initializing process group with 'cpu:gloo,cuda:nccl'NFr6c6sdjyyy)NT)close)owned_async_stagerowned_async_stager_closedsr.maybe_close_owned_async_stagerz2async_save..maybe_close_owned_async_stagerGs'  )2K(, %  $ $ &3L )r-Tr:c&jSN)stage)rRr0sr.stage_state_dictz$async_save..stage_state_dictMs!!*--r-r<cSr[r,)_rYs r.zasync_save..hs 6T6Vr-original_staging_futurereturn_staging_futurec |j|jdy#t$r}|j|Yd}~yd}~wwxYwr[)result set_result Exception set_exception)rarbes r.callbackzasync_save..callbackpsB 7'..0%006 7%33A66 7s!$ AAA)rNrOc@jrjyyr[) should_synchronize_after_executesynchronize_staging)rRsr.maybe_synchronize_stagingz-async_save..maybe_synchronize_stagings<<002=r-)r6N)rDrErFrGrHrIrdevice _device_typesAssertionError isinstancerrrrLrrr contextlib ExitStackrir!r+r r execute_saveadd_done_callbackpop_alldonerer")r0r=r1r5r2rQrRr4r>pgr]stackstaging_future_or_state_dictupload_executor upload_futurestaging_futurerbrirmrYrWrXs` ` @@@r.r r sA\ HH  !JK t224  202 << b&6&6 6 x .2  %*^[*Q)L( L". (4J %' t,.f_5G.-.    5  ) NN9 :'7'9$'*?*G*GG 1 246  !0 < < (')'+!=!   )  + +,V W MMO36.75.4h3H 7%+O%< 7#)$< 7""$  , ,X 6 ! , ,T 2!4   4 0 3 1 3 "#us BH((H1ci}|jD]-\}}d}d||_td||||</|S)z]Creates a shallow copy of `state_dict` where `state_dict` is called for each Stateful object.cFt|tr|jS|Sr[)rqrr0)elems r._elem_to_state_dictz4_stateful_to_state_dict.._elem_to_state_dicts(24(B4??$ L Lr-z_stateful_to_state_dict.Tr:)itemsr&r)r0stateful_state_dictkeyrrs r.rLrLso%%'  T M*B#'G$$ #J#5T#J $ $C   r-c`tjjdt|| | t  t ddi}t ddx}||d<j|d<td i|fd} td i|fd} drjd| | n| } | | g} | dtd i|fd } td i|fd }rjd | |}|S| }||g}j|S) Nz,torch.distributed.checkpoint.save_state_dictplanner is Noner=r2c\ tdj}dtjjj vr4t jddj jnj |jdtjjj vr)jjjnjjj}j|}|S) Nr storage_metazThe function definition for SavePlanner.set_up_planner has been updated to include the storage_meta argument. Please update your implementation to include this parameter.r@rA)r0ris_coordinatorkwargs)rankr>) rprinspect signatureset_up_planner parametersrJrKrset_up_storage_writerrcreate_local_planprepare_local_plan)r local_plandistWr5r0r1r>s r. local_stepz$_save_state_dict..local_steps ? !23 3%224 !2!273I3I!J!U!U U MM.    " ":u/C/C D  " "%)$33 #    !E!EFQQ R  0 0$$ZZ / 1   0 01E1E F..0 #66zB r-cl tdj|\}j|}|S)Nr)rpcreate_global_planprepare_global_plan)all_local_plansglobal_metadatar5r1s r. global_stepz%_save_state_dict..global_stepsA ? !23 3+2+E+Eo+V((<<_Mr-planrc td tdj}j|}|j|j S)Nrzcentral_plan is None)rp finish_plan write_datawaitvalue)final_local_plan all_writes central_planr5r1s r.rz$_save_state_dict..write_datasc ? !23 3   !78 8"..|<#../?I !!r-cH tdj|S)Nzglobal_metadata is None)metadataresults)rpfinish) all_resultsrr1s r.finish_checkpointz+_save_state_dict..finish_checkpoints.  " !:; ; Lr-writer,) rDrErFrr rpgetattrgrouprreduce_scatter all_reducebarrier)r0r1r2r3r4r5r> ckpt_kwargsckpt_idrrr global_planrrr write_resultsrrrs`` `` @@@r.r9r9sh HH  !OP G 5E FE$&.//OK>?DAAN'. O$',{{ O$&+&!'!F&+&'%)L++FJ L )| &1:,&? "1~ &+& "' "&+&' ##GZ9JK O ,6< $m_5  Or-)NrFN)NrFNT)DrrrosrJconcurrent.futuresr dataclassesrenumrtypingrrtyping_extensionsrrDtorch.distributed distributedrG#torch.distributed._state_dict_utilsr 4torch.distributed.checkpoint._async_process_executorr 3torch.distributed.checkpoint._async_thread_executorr +torch.distributed.checkpoint._storage_utilsr ,torch.distributed.checkpoint.default_plannerr #torch.distributed.checkpoint.loggerr%torch.distributed.checkpoint.metadatar$torch.distributed.checkpoint.plannerrr$torch.distributed.checkpoint.stagingrrr%torch.distributed.checkpoint.statefulr$torch.distributed.checkpoint.storagerr"torch.distributed.distributed_c10drutilsrrr,torch.distributed.checkpoint._async_executorr__all__r! FutureWarning ProcessGroupintboolrstrPathLikerr"r*r rLr9r,r-r.rs@ %!&( ?GKB:F ;KA88U D ! /3"&   ! $$t+     4      .4(/3+/"&.2 r r $t+r "D( r 4  r $$t+ r r r r )r j  $ $  $4(/3+/"&.25J5Q5Q'+ nn$t+n"D( n 4  n $$t+ n3n$nnn n)nb4(  O ) &/3"& hh!h$$t+h h  h 4  hhhr-