^jhddlZddlZddlZddlZddlmZmZmZmZddl m Z m Z m Z ddl Z ddlmZddlmcmcmZddlmcmcmZddlmcmcmZddlmZddlmZddlm Z ddl!m"Z"ddl#m$Z$ddl%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+m,Z,dd l-m.Z.m/Z/m0Z0m1Z1dd l2m3Z3dd l4m5Z5m6Z6m7Z7m8Z8m9Z9m:Z:m;Z;mZ>dd l?m@Z@ddlAmBZBddlCmDZDe rddlEmFZFdZG ddlHmIZImJZJdZLdZMeNejejfZPejePzdzZQe:je1je:je1je:je1je:je1je:je1jiZWe:je:jgZXe:je:jfZYe d`de'deQde:de>dzde"dzde'f dZZe de'deQde"de'fdZ[e de de\fdZ]e de"de\fdZ^e d e_dejfd!Z`e d"ejd e_dejfd#Zad"ejd e_deNejejffd$Zbe d`de'd%ejd&ee j8jdzd'ee j8jee j8jzdzde'f d(Zed'efe d)e\ddfd*Zge de'd%ejd+ehejd,e_e jzdzde'f d-Zje de'd%ejde'fd.Zke de'de:dzd/e9dzd0e6dzd1e\d2e\d3e_d4e_de'fd5Zle de'de'fd6Zme de'd7e5d8e\de'fd9Zne d`de'de"de'fd:Zoe de'de'fd;Zpd%ejdejd,e_e jzdzd?eejgdfdzd@e\de'f dAZre de'dejfdBZsdCejdDee j8jdzdehejfdEZt d`dCe j8jd&ehe j8jdFee j8jdzdehe j8jfdGZudCe j8jd&ehe j8jdehevfdHZwdCejdehevfdIZxd%ejd+ehejd,e_e jzdzddfdJZyd,e_e jzdzdKe_dLe&de jdzfdMZzd%ejd+ehejd&ehejdeNe\e\ffdNZ{dCejd?eejgdfd&ehejddfdOZ|dCejdPe jdzd&ehejdLe&fdQZ}dCejd&ehejdefejfdRZ~d%ejd+ehejdSehe jdPe jdzddf dTZd{?O?O>PQ   *, -N 5//<<>. **// *- 89F 5E !5GGKMGZF[]  ;**E LrDcjt|txr"t|dk(xrtd|DS)Nc3PK|]}t|tj ywN) isinstancedist ProcessGroup).0pgs rB z1_is_valid_hybrid_shard_pg_type..sJb 2t001J$&)rXtuplelenallrGs rBrMrMs: =%( K  ! # K JMJ JrDcDt|txr|jdk(S)NrU)rXr rI)r.s rBrHrHs k: . H;3C3Cq3HHrDnum_devices_per_nodec6tj|\}}|S)aU Return a process group across the current node. For example, given each row is a distinct node: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 This API would return an intra-node subgroup across [0, 1, ..., 7] or [8, 9, ..., 15] depending on the process's rank. For example, rank 3 would get [0, 1, ..., 7]. )rY new_subgroups)rcintra_node_subgroup_s rB_init_intra_node_process_grouprhs!"//0DE rDglobal_process_groupc\d}tj|}tj|}||z}tj||z}t |D]?}t |Dcgc] }|||zz } }tj | |} ||k(s>| }A|t |d|Scc}w)a Return an inter-node process group where each contained rank has the same local rank. For example, given each row is a distinct node: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 This API would return inter-node process group [0, 8], [1, 9], [2, 10], and so forth depending on the process's rank. For example, rank 1 would get [1, 9], rank 5 would get [5, 13]. N)ranksbackendz. expected to assign inter-node pg, but did not)rY get_backendget_world_sizeget_rankrange new_groupAssertionError) rirc inter_node_pgsharding_backendr: num_nodes my_local_rank local_rankiranks_for_inter_groupgrps rB_init_inter_node_process_groupr{s M''(<=$$%9:J22IMM"67:NNM01 =B9=M! 89J!22 3! ! nn#8BRS  &M oK L   ! s%B)c0t|t||fS)a Initialize intra and inter-node process groups and return the ones corresponding to this process's rank. This function can be used to initialize process groups for ``HYBRID_SHARD`` or ``_HYBRID_SHARD_ZERO2`` in FSDP. This function assumes each node has an equal number of CUDA-enabled devices. Returns: Tuple[dist.ProcessGroup, dist.ProcessGroup]: Intra and inter-node process group. )rhr{)rircs rBrJrJs# '';<&';=QR rDmoduleignored_modulesignored_statesc| | tdd}|du}|rt|}t|dng}t| t|ngdt|dkDr"t |dt j r|}n|}t|||_t||j||_ t||j|_ |S)NzfCannot pass both ignored_modules and ignored_states at the same time. Please just pass ignored_states.TFr) r3list_check_ignored_statesr`rXnn Parameter_get_ignored_modules_ignored_modules_get_ignored_params_ignored_params_get_ignored_buffer_names_ignored_buffer_names)r*r}r~rignored_parameterspassed_as_ignored_statesignored_states_lists rB_init_ignored_module_statesrs"~'A :  -T9">2148 %4%@D !b%  !# )!,bll ;!4 1O1&/JE/ E #< #E LrDrct|dk(ry|r`td|D}td|D}|s9|s6t|Dchc] }t|c}t}t d|yytd|Ds6t|Dchc] }t|c}t}t d|ycc}wcc}w) z Check that the ignored states are uniformly parameters or uniformly modules. We may remove this check in the future if we permit mixing. rNc3PK|]}t|tj ywrW)rXrrr[r*s rBr]z(_check_ignored_states..KsUUE2<<8Ur^c3PK|]}t|tj ywrWrXrModulers rBr]z(_check_ignored_states..LsS5*UBII6Sr^)keyzUignored_states expects all nn.Parameter or all nn.Module list elements but got types c3PK|]}t|tj ywrWrrs rBr]z(_check_ignored_states..UsLE:eRYY/Lr^z>ignored_modules expects nn.Module list elements but got types )r`rasortedrNreprr3)rr all_params all_modulesr* sorted_typess rBrr@s >aUnUU SNSS +!N"K54;"KQUVL**69 #.zL^LL!N"K54;"KQUVL%( M #L#Ls B; Cignored_params device_idc6d}|1t|tjr|ntj|}|t||D]|}|jjdvr| |j}+|jj|jk7sOt d|jd|jj|xstj j}|jdk(r t dtj||_ |S)a= Determine device handle used for initializing FSDP. If a device is specified by ``device_id``, then returns device handle corresponds to that device type. Otherwise, If the module is already on a non-CPU device, then the device type is that non-CPU device type. If the module is on CPU or meta, then the device type is the current accelerator device. See the :ref:`Accelerators` for details. This method will be called once ignored parameters was determined, as the device handle maybe needed for other initialization. N>cpumetazLFSDP does not support modules with different device types but got params on z and rzOFSDP needs a non-CPU accelerator device, but no accelerator device is detected.) rXtorchdevice_get_orig_paramsrN RuntimeError_C_get_acceleratorr from_devicerK)r*r}rrdetermined_deviceparams rB_init_device_handler]s()U\\2 i(   %fn= E||  O3 ($)LL!<<$$(9(>(>>&-->-C-C,DE%,,J[J[I\^ .L1J1J1L  ! !U *a -889JKE LrDct||_i}|jD]\}}t|}|j||<!||_|SrW)_get_buffer_names _buffer_names named_buffersrdtype_buffer_name_to_orig_dtype)r*r}r buffer_namebuffers rB_init_buffer_staters_ ,F3E :<%335? V' 4 28,,";/?(BE$ LrDmixed_precision cpu_offloadlimit_all_gathersuse_orig_paramsbackward_prefetch_limitforward_prefetch_limitc|jdk(rQ|tjk7r-tjd|xstj ddtj}n/|tjk(rtjdt d|xstj |_|xs t|_ |5tjjdt|jtjj!t"d d k(|_|xs t'|_||_||_t.j0|_d|_t7|_t;j<|_tAjB|j>|||_"d|_#i}||_$d} | |_%g} | |_&|S) NrFz/FSDP is switching to use `NO_SHARD` instead of z since the world size is 1.rU stacklevelzoThe `NO_SHARD` sharding strategy is deprecated. If having issues, please use `DistributedDataParallel` instead.z'torch.distributed.fsdp.mixed_precision.1)'r:rNO_SHARDwarningswarn FULL_SHARD FutureWarningr,rrrr_log_api_usage_oncestrosenvirongetr_use_full_prec_in_evalrrr_use_orig_paramsrIDLEtraining_state_is_rootr_free_event_queuerYget_debug_level _debug_levelexec_order_utils_ExecOrderData_exec_order_data_unshard_event_fully_sharded_module_to_handle_handleparams) r*r,rrrrrrrrrs rB_init_core_staters 1  0 9 9 9 MMA$C(8(C(CDE''   -55 .77 7  <   0N3C3N3NE+?~/?E" $$5c%:O:O6P5Q R  2B73>  $3z|E/E,E(--EEN-/E--/E-<< E  EIK#,KE)'+GEM"$FEL LrDcfg}||_g}||_g}||_d|_d|_d|_|S)NT)_root_pre_forward_handles_pre_forward_handles_post_forward_handles_sync_gradients _comm_hook_comm_hook_state)r*rrrs rB_init_runtime_statersK8:&?E#24!5E35"7E EE!E LrDbackward_prefetchforward_prefetchc"||_||_|SrW)rr)r*rrs rB_init_prefetching_staters 0E-E LrDc||jnd}|r+||jk7rt|j|_|Sd|_|SrW)_get_root_meshr6r#rK_fsdp_extension)r*r. root_meshs rB_init_extensionrsT 1<0G **,TIyE$6$66 1%2F2F G L!% LrDctj|_t}t |_||_i}||_|SrW)r!FULL_STATE_DICT_state_dict_typerr_optim_state_dict_config_state_dict_config_unshard_params_ctx)r*state_dict_configunshard_params_ctxs rB_init_state_dict_staters?*::E)<)>%=%?E"0E57 2E LrDrc|D]X}t|jdk(sd}|jD] \}}||us |}n|s tdt d|dy)z Verify if the parameters are accepted by FSDP. The only restriction now is that the parameter cannot be a scalar tensor (param.shape == []). rrzExpected param_name to be setz/FSDP doesn't support scalar parameters. Change z& to a 1D tensor with numel equal to 1.N)r`shapenamed_parametersrrr3)r}rr param_namenameparam_s rB_verify_managed_paramsrs   u{{ q J & 7 7 9  fF?!%J $%DEE$%KM  rDfully_sharded_module param_init_fnsync_module_statesct|j|t|jj}t |jj \}}|s|r|t||j nA|r#t||j jn|rtj|fdj D chc]}|jD]} | } }} t|j| |t|j|jj_tt!|j} t#|| |rwj$t&vrNt)|| j*|jD]$} t-| t.st1| t.d&t)|| j2t5| |Scc} }w)zHInitialize a ``FlatParamHandle`` from a module ``fully_sharded_module``.c>t|duxr|jvSrW)rr) submoduler*s rBz0_init_param_handle_from_module..Qs('=i'HD'P(8!7!77rD)check_fnF)_check_single_device_moduler_get_device_from_device_idr8rK_need_to_materialize_moduler_materialize_with_param_init_fn_materialize_meta_moduler'materialize_modulebuffers_move_module_to_device_get_compute_devicecompute_devicerrrr,r4_sync_module_params_and_buffersr;hasattr FSDP_SYNCEDsetattrr+_init_param_handle_from_params) r*rrrrdevice_from_device_idis_meta_moduleis_torchdistX_deferred_initignored_modulerignored_buffersmanaged_paramss ` rB_init_param_handle_from_moduler2s 4e6K6KYW65::u333Ne33U5K5K3/N/ 5=;T' -1G1G     " "   %(( 8 $44 $,,.  O    /    E*+?AVAVWXN/@  " "&@ @ ,$ne6J6J /668 86;/FK7 8 ( .%2E2E #5.:NO LYs G*c t|dk(ryt|||jt|j|j j |jj|jj|jj|j|j|j }|j|jr t!d|j"j%|j&||_||j(|j*<t-j.d}|j j r,|j&j.|k7r|j1|yyy)Nr)fsdp_extensionz!Expected state._handle to be Noner)r`rr SHARDING_STRATEGY_MAPr,roffload_paramsr param_dtype reduce_dtypekeep_low_precision_gradsr+rrshardrrrrappend flat_paramr_fully_sharded_modulerr flat_param_to)r*rrhandle cpu_devices rBrrs+  6{a  e556 (( )) ** 66  ,, F LLN }}@AA LL))*EMJPE))&*F*FGe$J ''F,=,=,D,D ,RZ(-S'rD root_modulerc 2d} | t|n t}|D]V}t|tj j st|dt|zt|sMtd|jD])}tj|r|j|+|Dchc]3}|jD]}t|tjs| 5}}}||vrtj dd|jD]B}t|} | t#| d s t%d |j'| j(D|S#t$r }t|dt|z|d}~wwxYwcc}}w) ah Check that ``_ignored_modules`` is an iterable of ``nn.Module`` s without any FSDP instances. Return the modules contained in their module subtrees as a :class:`set`. Nested FSDP instances are excluded, but their already-computed ignored modules are included. ``_ignored_modules`` represents the argument passed by the user to FSDP. z>`ignored_modules` should be an iterable of `torch.nn.Module`s Nzbut got zbut got an iterable with z1`ignored_modules` should not include FSDP moduleszTrying to ignore the top-level module passed into the FSDP constructor itself will result in all parameters being ignored and is not well-supported: rUrrz?Expected optional_fsdp_state to have _ignored_modules attribute)set TypeErrorrNrXrrrrr3modulestraversal_utils _composableadd fsdp_fileFullyShardedDataParallelrrrrrupdater) r'r msg_prefixignored_root_moduleser}childr~roptional_fsdp_states rBrrsRJQ%5%AC !su  'R&%((//2J+DT&\N)SST T !& )PQ Q R%%'-**62 $ $V ,-+ ^^% %!C!CD  O o%  228 ;  !((*I 4Y?  *.0BC$U  " "#6#G#G HI Q Q x5E0F/G%HHIqPQ$sE'78F' F0F  Frct}|Dchc]%}|jD]}t|r|'}}}|j||,|Dchc]}t|r|}}|j||j D]B}t |} | t | ds td|j| jD|Scc}}wcc}w)z Return the parameters of the modules in ``ignored_modules`` and the parameters in ``ignored_parameters``. :class:`FlatParameter` s are excluded from the result. rz>Expected optional_fsdp_state to have _ignored_params attribute) r) parametersrr1r+rrrrr) r'r~rall_ignored_paramsmpparams_in_ignored_modulesparams_in_ignored_parametersrr6s rBrrs36%#!ALLN!'(BTUVBW! !!78%)( 1CA1FA( $(  !!">?!((*K 4Y?  *.0AB$T  % %&9&I&I JK -!( s#CCC%Cc t}|Dchc]}|jD]}|}}}|j|jDchc]\}}||vr t |c}}|j D]B}t |}|t|ds td|j|jD|Scc}}wcc}}w)z6Return the cleaned buffer FQNs in ``ignored_modules``.rzDExpected optional_fsdp_state to have _ignored_buffer_names attribute) r)r r1rrr+rrrrr) r'r~all_ignored_buffer_namesr:rbuffers_in_ignored_modulesrrr6s rBrrs *-("aiik",2"""##(3'@'@'B # V33 k * !((*W 4Y?  *.0GH$Z % + +,?,U,U VW $#-"  s CC cf|jDchc]\}}t|c}}Scc}}w)zrReturn the fully prefixed names of all buffers in the module hierarchy rooted at ``root_module`` as a class:`set`.)rr)r'rrgs rBrr&s5>I=V=V=X +9;+&  s-ct||Dchc]}|j}}t|dk(r%tjd|vr| t dyt|dkDrt d|ycc}w)z Raise an error if ``module`` has original parameters on multiple devices, ignoring the parameters in ``ignored_params``. Thus, after this method, the module must be either fully on the CPU or fully on a non-CPU device. rUrNzTTo support a module with both CPU and GPU params, please pass in device_id argument.rFz;FSDP only supports single device modules but got params on )rrr`rr)r}rrrdevicess rBrr-s*:&.)QRu||RGR 7|qU\\%0G;  5   W I' S   SsA4r8 device_handlec b|yt|tjr|ntj|}|jdk7rk|j_t j d|d|d|jd|jd d tj|j}|S) z Return a ``torch.device`` for the specified ``device_id``. Processes ``device_id`` and returns either the corresponding device or ``None`` if ``device_id`` is ``None``. Nrz"FSDP got the argument `device_id` z on rank zJ, which does not have an explicit index. FSDP will use the current device z6. If this is incorrect, please explicitly call `torch.zk.set_device()` before FSDP initialization or pass in the explicit device index as the `device_id` argument.rUr)rXrrrNindexrrcurrent_device)rr8rDrs rBrrKs 5<<8 ell9>U {{e 4 0 9f00=0L0L0N/OPCCI;;-P1 1  m::<= MrDctt||}td|D}|jD]-}||vr|j dD]}||j z}/| xrt xrtd|D}||fS)z Return if ``module`` has parameters on meta device and if ``module`` is using torchdistX deferred initialization. At most of the returned bools can be ``True``. If either is ``True``, then ``module`` needs to be materialized. c34K|]}|jywrW)is_metar[rs rBr]z._need_to_materialize_module..vsC5CsFrecursec3FK|]}tj|ywrW)r(is_fakerKs rBr]z._need_to_materialize_module..s@ U#@s!)rranyr+r rJ_TORCHDISTX_AVAIL)r}rr~rrrbufrs rBrris*6>BCNCNCCN^^%*  ' $$U$3 *C ckk )N **  A  A @@ @ 6 66rDct|std|dt|t||}|D] }|| y)Nz Expected z to be callable but got )callabler3rN_get_modules_to_materialize)r'rr~modules_to_materializer}s rBrrsV M " &>tM?R>S T  9oV(frDrc >|xs#tj|j}t||}d} tj5|D]u}t j |jd|jd}tt|dkD}|sS|j|d|jw dddy#1swYyxYw#t$r7} tjdt!| dt#|dd | d} ~ wwxYw) NFrLr)rrMzIUnable to call `reset_parameters()` for module on meta device with error z(. Please ensure that your module oftype z* implements a `reset_parameters()` method.rUr)rrrGrUno_grad itertoolschainr8r r`rto_emptyreset_parameters BaseExceptionrrrrN) r'rr~rDmaterialization_devicerVr}module_state_iterhas_module_statesr4s rBrrs$3ell$$&79oV F]]_ .0 .%.OO%%e%4NN5N1%! %(->(?$@1$D!$OO+A5OQ++- . . . .   !!$Q)L>!K M  s<C AC!%CCCCC D%2DDc"g}tj|g}|h}|rq|j}|j||j D]:}||vst |||vs|j ||j|<|rq|SrW) collectionsdequepopleftr!childrenrr.)r'r~rVqueuevisited_modulesr} child_modules rBrUrUs /1   {m ,E'2mO %%f-"OO- +LO3*<8@ 7##L1 \* +  "!rDrc tjd | tj}|j |g}g}|r|j }|j fd|jdD|j fd|jdD|jD].}t|tjr|j |0|r|D cgc] } | |vs|  } } |D cgc] } | |vs|  } } t| | |ytt||d} | | j k(r t!yyycc} wcc} w)a Move ``module`` depending on ``device_from_device_id`` and its current device. This includes moving ignored modules' parameters. - If ``device_from_device_id`` is not ``None``, then this moves ``module`` to the device. - If ``device_from_device_id`` is ``None``, then this does not move ``module`` but warns the user if it is on CPU. Precondition: ``_check_single_device_module()``. rNc3@K|]}|jk(r|ywrWr)r[rr&s rBr]z)_move_module_to_device..s%<<:-FrLc3@K|]}|jk(r|ywrWrk)r[rr&s rBr]z)_move_module_to_device..s%==J.rl)rrrbrcr!rdextendr8r rerXr/r0_move_states_to_devicenextr_warn_cpu_init)r}rrrrfrr  curr_modulerr;params_to_move bufs_to_moverr&s @rBr r s\$e$J(/:.?.?.A V%'&(--/K MM(33E3B  NN)11%1@  )113 , !)Y-O-OPLL+ ,!&&,Gq/F!GG#*Gaa.FG G~|=RS !&.94 @E U\\Z78 HGs1 E;E EEr c6t|dk(rt|dk(ryt|dkDr|dj}nt|dkDr|dj}tjd}||D]k}tj5|j ||_|j *|j j ||j _dddm|D]}|j ||_y|k(r tyy#1swYxYw)z Move states to the specified device. Precondition: ``_check_single_device_module()`` and module's parameters and buffers have been materialized if needed. rNr)r`rrrXtodatagradrq)rr rrGr&rrs rBroros  6{aCLA- 6{Q)) W  **e$J( KE K"XX&;< ::)&+jjmm4I&JEJJO K K K  ;F ))$9:FK ; : % &  K Ks A DD c2tjddy)NamThe passed-in `module` is on CPU and will thus have FSDP's sharding initialization run on CPU, which may be slower than on GPU. We recommend passing in the `device_id` argument for FSDP to move `module` to GPU for the sharding initialization. `module` must also be on GPU device to work with the `sync_module_states=True` flag since that requires GPU communication.rUr)rrrDrBrqrq's MM 1 rDctt||d}|&|jjdk7r |j}n#t j|j }|||k7rt d|d|d||S)a) Determine and return this FSDP instance's compute device. If the module is already on a non-CPU device, then the compute device is that non-CPU device. If the module is on CPU, then the compute device is the current device. Since this method should be called after materializing the module, any non-CPU device should not be meta device. For now, the compute device is always a CUDA or CUDA-like device with its explicit index. Precondition: ``_check_single_device_module()`` and ``_move_module_to_device()``. Nrz4Inconsistent compute device and `device_id` on rank z: z vs )rprrrNrrGr3)r}rrr8rDrr s rBr r 3s* !&.94 @E U\\..%7m&B&B&DE(^?T-TB4&d#8"9 ;   rDcVg}|jD]}t|tdrt|td|j }t |ry|j \}}|D]`}t||xxtjdx\} |j| 9xtdx\I} tdt| |j||D]} | j } t | ry| j \}}|D]`}t| |xxtjdx\} |j| 9xtdx\I} tdt| |j| t|t||tdy)z Synchronize module states (i.e. parameters ``params`` and all not-yet-synced buffers) by broadcasting from rank 0 to all ranks. Precondition: ``sync_module_states == True`` and ``self.process_group`` has been set. FTrzNz#expected Tensor or OpaqueBase, got r)src)r getattrrrdetachr%__tensor_flatten__rTensorr!r rrrN+_check_module_states_for_sync_module_statesr$PARAM_BROADCAST_BUCKET_SIZE) r}rr+ module_statesrdetached_bufferattrsrgattrv unexpectedrdetached_params rBr r Vs)+M.."6v{E2 FK .$mmoO,_=+==?q! D!/480U\\^0)0031)Z\ *'"0"Ed:FVEW X# $$_5)6,1 ( 8%88:HE1 nd3,,%,,Q/-%&#,A$zBRAST    01"0 ># rDrcD|rtd|Dr tdyy)Nc3`K|]&}|jtjdk((yw)rN)rr)r[tensors rBr]z>_check_module_states_for_sync_module_states..s'17 e,,s,.zThe module has CPU parameters or buffers when `sync_module_states=True`, which requires them to be on GPU. Please specify the `device_id` argument or move the module to GPU before passing it to FSDP.)rPr3)rs rBrrs7;H C  }rDc#K|j} t|}||vrt|s|#t$rYywxYww)aD Return an iterator over the original parameters in ``module``. The iterator does not return the parameters in ``ignored_params``, any ``FlatParameter`` s (which may be present due to nested FSDP wrapping), or any original parameters already flattened (only relevant when ``use_orig_params=True``). N)r8rpr StopIteration)r}r param_genrs rBrrsT!!#I OEN*3Ee3L     sA 4 AAAAc t|D]A\}}||vs t|rtd|d|jd|jy)a5 Check that original parameters in ``fsdp_module`` have been flattened. The flattened parameters are made invisible to ``named_parameters()`` for the module hierarchy rooted at ``fsdp_module``. This should be called as a sanity check after flattening the wrapped module's parameters. z Found an unflattened parameter: z;  N)rrrr9 __class__) fsdp_modulerrrs rB_check_orig_params_flattenedrs^?{K E  &/A%/H2:,b::<.%//!24 rDch|tjk(rtjStjSrW)rrr allreduce_hookreduce_scatter_hook)r,s rB_get_default_comm_hookrs3  0 9 9 9 $$ . .rDc.tj|S)NrG)r r<rGs rBrOrOs  % %M BBrDrW)rbrYrrcollections.abcrrrrtypingrrr rtorch.distributed distributedrY(torch.distributed.fsdp._exec_order_utilsfsdp_exec_order_utilsr'torch.distributed.fsdp._traversal_utils_traversal_utilsr,2torch.distributed.fsdp.fully_sharded_data_parallelfully_sharded_data_parallelr/torch.nnrtorch._opaque_baser (torch.distributed.algorithms._comm_hooksr torch.distributed.device_meshr "torch.distributed.distributed_c10dr $torch.distributed.fsdp._common_utilsrrrrrrr"torch.distributed.fsdp._flat_paramrrrr%torch.distributed.fsdp._limiter_utilsrtorch.distributed.fsdp.apirrrrrrr r!torch.distributed.fsdp.wrapr"&torch.distributed.tensor.parallel.fsdpr#torch.distributed.utilsr$torch.utils._python_dispatchr%torch.utils.hooksr&rQ torchdistxr'r( ImportErrorrrr_rZHybridShardProcessGroupTypeProcessGroupTyperr SHARD_GRAD_OP HYBRID_SHARD_HYBRID_SHARD_ZERO2rr4#NO_RESHARD_AFTER_FORWARD_STRATEGIESrCr5boolrMrHintrhr{rJrrrrrr)rrrrrrrrrrrrrrrrrrrrrrUrr rorqr r rrrrr<rOrzrDrBrs  CC44 CCAAFF)B4A B   0D<F1.0 #D$5$5t7H7H$HI$$'BBTI 5>>!7!B!B""$:$H$H!!#9#F#F((*@*T*T !!(( ""(('# &* 0 0#0(0 dN 0 d" 0  00f( (#(( ((V#$IJI4II  ARAR  "++"" ""J++ 4  d// /0& + + II+ehhoo.5+UXX//0uxx   +++\I9= :- - II- %-U\\!D( -  --`  II"? ?'$.?$d*?d" ?  ?  ?!? ???D         '     : J *  *299d2<<6HT(N N))NU\\!D(NRYYK-.5 N  N  NNb) )  )))))>::uxx/$6: ^:@?C"")"!!3!34t;"    "J$$)$ X$@299S II  % U\\!D(   <U\\!D( % \\D <7 II7 %7^7 4: 7<  RYYK-. ^  ## <<$.#^#% #L""-0^" "))_",3 II3 %3&3!<<$. 3  3l   %,, !<<$. @  II %!<<$.   %   \\ F9 II9  9$$9 9x   %      II  % bll , % (.>C$$CCq$s];;^^