Nj< ddlZddlZddlZddlZddlmZmZmZmZddl 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 ddlHmIZImJZJn #eK$rdZGYnwxYwdZLdZMeNejOejOfZPejOePzdzZQe:jRe1jRe:jSe1jSe:jTe1jTe:jUe1jUe:jVe1jViZWe:jUe:jVgZXe:jTe:jVfZYe d`de'deQde:de>dzde"dzde'f dZZe de'deQde"de'fdZ[e de de\fdZ]e de"de\fdZ^e d e_dejOfd!Z`e d"ejOd e_dejOfd#Zad"ejOd e_deNejOejOffd$Zbe d`de'd%ejcd&ee jjcdzd'ee jjdee jjczdzde'f d(Zed'efe d)e\ddfd*Zge de'd%ejcd+ehejdd,e_e jizdzde'f d-Zje de'd%ejcde'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%ejcdejcd,e_e jizdzd?eejcgdfdzd@e\de'f dAZre de'dejcfdBZsdCejcdDee jjcdzdehejcfdEZt d`dCe jjcd&ehe jjcdFee jjddzdehe jjdfdGZudCe jjcd&ehe jjcdehevfdHZwdCejcdehevfdIZxd%ejcd+ehejdd,e_e jizdzddfdJZyd,e_e jizdzdKe_dLe&de jidzfdMZzd%ejcd+ehejdd&ehejcdeNe\e\ffdNZ{dCejcd?eejcgdfd&ehejcddfdOZ|dCejcdPe jidzd&ehejcdLe&fdQZ}dCejcd&ehejcdefejcfdRZ~d%ejcd+ehejddSehe jdPe jidzddf dTZd||_|d|_|d|_nt d|j|Gt}t||j \}}||_||_n>t|r|\|_|_nt dt|t|j|_|S)Nrr1z,Expected device_mesh to have ndim=2 but got zmExpected process_group to be passed in as either None or Tuple[dist.ProcessGroup, dist.ProcessGroup] but got r+)"_is_valid_hybrid_shard_device_meshr6r7r;r+r3ndimr !_init_intra_and_inter_node_groups_device_handle device_count_is_valid_hybrid_shard_pg_typetype_get_default_comm_hook_state_inter_node_state)r*r+r. default_groupintra_node_groupinter_node_groups rBr5r5sJ  -k : : !,E $/#8#8!#8#D#DE "-"7"7"7"C"CE  Q{?OQQ   *,, -N 5/<<>>. . **// *- 8 8 9F 5E !5!5]GKMGZGZ]]  ;*E LrDct|to+t|dkotd|DS)Nc3JK|]}t|tjVdSN) isinstancedist ProcessGroup).0pgs rB z1_is_valid_hybrid_shard_pg_type..s/JJb 2t011JJJJJJrD)rXtuplelenallrGs rBrMrMsL =%(( K   ! # K JJMJJJ J JrDcBt|to |jdkS)NrU)rXr rI)r.s rBrHrHs k: . . H;3Cq3HHrDnum_devices_per_nodec4tj|\}}|S)aU Return a process group across the current node. For example, given each row is a distinct node: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 This API would return an intra-node subgroup across [0, 1, ..., 7] or [8, 9, ..., 15] depending on the process's rank. For example, rank 3 would get [0, 1, ..., 7]. )rY new_subgroups)rbintra_node_subgroup_s rB_init_intra_node_process_grouprgs!"/0DEE rDglobal_process_groupcV d}tj|}tj|}|z}tj|z}t D]< fdt |D}tj||} |kr|}=|t |d|S)a Return an inter-node process group where each contained rank has the same local rank. For example, given each row is a distinct node: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 This API would return inter-node process group [0, 8], [1, 9], [2, 10], and so forth depending on the process's rank. For example, rank 1 would get [1, 9], rank 5 would get [5, 13]. Nc g|] }|zz Srk)r[i local_rankrbs rB z2_init_inter_node_process_group..s1! ! ! 89J!22 3! ! ! rD)ranksbackendz. expected to assign inter-node pg, but did not)rY get_backendget_world_sizeget_rankrange new_groupAssertionError) rhrb inter_node_pgsharding_backendr: num_nodes my_local_rankranks_for_inter_groupgrprms ` @rB_init_inter_node_process_groupr}s M'(<==$%9::J22IM"677:NNM011  ! ! ! ! ! =B9=M=M! ! ! n#8BRSSS  & &M L L L    rDc@t|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. )rgr})rhrbs rBrJrJs) '';<<&';=QRR rDmoduleignored_modulesignored_statesc||tdd}|du}|r t|}t|dn#g}t|t|ngdt|dkr%t |dt jr|}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*rrrignored_parameterspassed_as_ignored_statesignored_states_lists rB_init_ignored_module_statesrs$"~'A :   -T9 ">22148888 %4%@D ! ! !b%    !## )!,bl ; ; 2!4  1O1&/JJE/ E #< ##E LrDrct|dkrdS|rltd|D}td|D}|s4|s4td|Dt}t d|dSdStd|Ds2td |Dt}t d |dS) z Check that the ignored states are uniformly parameters or uniformly modules. We may remove this check in the future if we permit mixing. rNc3JK|]}t|tjVdSrW)rXrrr[r*s rBr]z(_check_ignored_states..Ks.UUUE2<88UUUUUUrDc3JK|]}t|tjVdSrWrXrModulers rBr]z(_check_ignored_states..Ls.SS5*UBI66SSSSSSrDc,h|]}t|SrkrNrs rB z(_check_ignored_states..O"K"K"K54;;"K"K"KrD)keyzUignored_states expects all nn.Parameter or all nn.Module list elements but got types c3JK|]}t|tjVdSrWrrs rBr]z(_check_ignored_states..Us.LLE:eRY//LLLLLLrDc,h|]}t|Srkrrs rBrz(_check_ignored_states..VrrDz>ignored_modules expects nn.Module list elements but got types )r_r`sortedreprr3)rr all_params all_modules sorted_typess rBrr@s6 >aUUnUUUUU SSNSSSSS  + !"K"KN"K"K"KQUVVVL9*699     LL^LLLLL !"K"KN"K"K"KQUVVVL(%((   rDignored_params device_idcd}|0t|tjr|ntj|}|t||D]T}|jjdvr||j}|jj|jkr$t d|jd|jjU|ptj}|jdkrt 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*rrrdetermined_deviceparams rB_init_device_handler]s1()U\22 )IIi((   %fn==  E| O33 ($)L!!<$(9(>>>&^->-C^^JO,J[^^? .L1J1J1L1L  !U * *a -89JKKE LrDct||_i}|D]\}}t|}|j||<||_|SrW)_get_buffer_names _buffer_names named_buffersrdtype_buffer_name_to_orig_dtype)r*rr buffer_namebuffers rB_init_buffer_statersf ,F33E :<%3355?? V' 44 28,";//'AE$ LrDmixed_precision cpu_offloadlimit_all_gathersuse_orig_paramsbackward_prefetch_limitforward_prefetch_limitcb|jdkrC|tjkr&tjd|p tjddtj}n,|tjkrtjdt d|p tj|_|p t|_ |4tj dt|j tjt"d d k|_|p t'|_||_||_t.j|_d|_t7|_t;j|_tAj!|j|||_"d|_#i}||_$d} | |_%g} | |_&|S) NrFz/FSDP is switching to use `NO_SHARD` instead of z since the world size is 1.rU stacklevelzoThe `NO_SHARD` sharding strategy is deprecated. If having issues, please use `DistributedDataParallel` instead.z'torch.distributed.fsdp.mixed_precision.1)'r:rNO_SHARDwarningswarn FULL_SHARD FutureWarningr,rrrr_log_api_usage_oncestrosenvirongetr_use_full_prec_in_evalrrr_use_orig_paramsrIDLEtraining_state_is_rootr_free_event_queuerYget_debug_level _debug_levelexec_order_utils_ExecOrderData_exec_order_data_unshard_event_fully_sharded_module_to_handle_handleparams) r*r,rrrrrrrrrs rB_init_core_staters 1  0 9 9 9 M'$C(8(C'''      -5 .7 7 7  <      0N3C3NE+?~/?/?E" $$ Rc%:O6P6P R R    2B773>  $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||nd}|r%||jkrt|j|_nd|_|SrW)_get_root_meshr6r#rK_fsdp_extension)r*r. root_meshs rB_init_extensionrs[ 1<0G **,,,TI%yE$666 1%2F G G!% 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_statersB*:E)<)>)>%=%?%?E"0E57 2E LrDrc|D]b}t|jdkrHd}|D] \}}||ur|}n|stdt d|dcdS)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_parametersrvr3)rrr param_namenameparam_s rB_verify_managed_paramsrs    u{  q J & 7 7 9 9   fF??!%JE# F$%DEEEM$MMM  !  rDfully_sharded_module param_init_fnsync_module_statesctt|j|t|jj}t |jj\}}|s|r|t||jn:|rt||jjn|rtj |fddjD}t|j||t|j|jj_ tt|j} t!|| |r|jt$vrXt'|| j|D]-} t-| t.rt1| t.d.t'|| jt5| |S)zHInitialize a ``FlatParamHandle`` from a module ``fully_sharded_module``.Nc8t|duo|jvSrW)rr) submoduler*s rBz0_init_param_handle_from_module..Qs('=i'H'HD'P(8!77rD)check_fnc@h|]}|D]}|Srkbuffers)r[ignored_modulers rBrz1_init_param_handle_from_module..UsK $,,..  rDF)_check_single_device_moduler_get_device_from_device_idr8rK_need_to_materialize_moduler_materialize_with_param_init_fn_materialize_meta_moduler'materialize_module_move_module_to_device_get_compute_devicecompute_devicerrrr,r4_sync_module_params_and_buffersr;r hasattr FSDP_SYNCEDsetattrr+_init_param_handle_from_params) r*rrrrdevice_from_device_idis_meta_moduleis_torchdistX_deferred_initignored_buffersmanaged_paramsrs ` rB_init_param_handle_from_moduler!2sM 4e6KYWWW65:u33Ne3U5K33/N/  5 =;T' -1G          "      % ( 8888    #4O    /    E*+?AVWWXXN/@@@  "&@ @ @ ,$ne6J   /6688 8 86;//8FK777' .%2E   #5.:NOOO LrDc Lt|dkrdSt|||jt|j|jj|jj|jj |jj |j |j |j  }||jrt!d|j|j||_||j|j<t-jd}|jjr'|jj|kr||dSdSdS)Nr)fsdp_extensionz!Expected state._handle to be Noner)r_rrSHARDING_STRATEGY_MAPr,roffload_paramsr param_dtype reduce_dtypekeep_low_precision_gradsr+rrshardrrvrappend flat_paramr_fully_sharded_modulerr flat_param_to)r*rrhandle cpu_devices rBrrs+  6{{a  e56 ( ) * 6  ,   F LLNNN }B@AAA L)***EMJPE)&*FGe$$J ')F,=,D ,R,RZ((((()),R,RrD root_modulercd} |t|n t}n5#t$r(}t|dt|z|d}~wwxYw|D]a}t|tjjs"t|dt|zt|rtdb| D]+}tj |s| |,d|D}||vrtjd|d | D]L}t|}|9t|d st!d ||jM|S) 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 modulescth|]5}|D]}t|tj|6Srk)modulesrX fsdp_fileFullyShardedDataParallel)r[rchilds rBrz'_get_ignored_modules..s\ ^^%% %!CDD  rDzTrying 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 TypeErrorrNrXrrrrr3r3traversal_utils _composableaddrrrrvupdater) r0r msg_prefixignored_root_moduleserrroptional_fsdp_states rBrrs RJQ%5%AC ! ! !suu  QQQ %H5E0F0F%H%HHIIqPQ&RR&%(/22 UJ)ST&\\)S)SSTT T !& ) ) RPQQ Q R %%''--*622 - $ $V , , , *O o%%  ;28 ; ;     !((**II 4Y??  *.0BCC $U  " "#6#G H H H s$ A#AArclt}d|D}|||!d|D}|||D]L}t|}|9t |dst d||jM|S)z Return the parameters of the modules in ``ignored_modules`` and the parameters in ``ignored_parameters``. :class:`FlatParameter` s are excluded from the result. c^h|]*}|D]}t||+Srk) parametersr)r[mps rBrz&_get_ignored_params..sS!!!ALLNN!!'(BTUVBWBW! !!!!rDNc0h|]}t||Srk)r)r[rEs rBrz&_get_ignored_params..s6( ( ( 1CA1F1F( ( ( ( rDrz>Expected optional_fsdp_state to have _ignored_params attribute)r7r<r3rrrvr)r0rrall_ignored_paramsparams_in_ignored_modulesparams_in_ignored_parametersrr@s rBrrs36%%!!"!!!7888%( ( )( ( ( $ !!">???!((**KK 4Y??  *.0ABB $T  % %&9&I J J J rDcdt}d|D|fd|D|D]L}t |}|9t |dst d||jM|S)z6Return the cleaned buffer FQNs in ``ignored_modules``.c@h|]}|D]}|Srkr )r[rDrs rBrz,_get_ignored_buffer_names.. sA"""aiikk"",2""""rDc<h|]\}}|v t|Srkr)r[rrbuffers_in_ignored_moduless rBrz,_get_ignored_buffer_names..s;   # V333 k * *333rDNrzDExpected optional_fsdp_state to have _ignored_buffer_names attribute)r7r<rr3rrrvr)r0rall_ignored_buffer_namesrr@rNs @rBrrs *-""'"""##    '2'@'@'B'B   !((**WW 4Y??  *.0GHH $Z % + +,?,U V V V ##rDc>d|DS)zrReturn the fully prefixed names of all buffers in the module hierarchy rooted at ``root_module`` as a class:`set`.c2h|]\}}t|SrkrM)r[rrfs rBrz$_get_buffer_names..(s2   +9;+&&   rD)r)r0s rBrr&s/  =H=V=V=X=X   rDcdt||D}t|dkr)tjd|vr|t ddSt|dkrt d|dS)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. ch|] }|j Srkrr[rs rBrz._check_single_device_module..8sRRRu|RRRrDrUrNzTTo 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 )rr_rrr)rrrdevicess rBrr-sSR)9&.)Q)QRRRG 7||qU\%00G;;  5    W   S' S S     rDr8 device_handlec R|dSt|tjr|ntj|}|jdkrg|j`t jd|d|d|d|jd d tj|}|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)rr8rWrs rBrrKst 5<88U el9>U>U {e 4  1 1 1 1 10=0L0L0N0N 1 1DJ; 1 1 1     m::<<== MrDc:tt||}td|D}|D]*}||vr|dD] }||jz} +| ot otd|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. c3$K|] }|jV dSrW)is_metarUs rBr]z._need_to_materialize_module..vs$CC5CCCCCCrDFrecursec3>K|]}tj|VdSrW)r(is_fakerUs rBr]z._need_to_materialize_module..s,@@ U##@@@@@@rD)rranyr3r r]_TORCHDISTX_AVAIL)rrrr rrbufrs rBrris*6>BBCCNCCNCCCCCN^^%%**  ' ' $$U$33 * *C ck )NN *  A  A @@@@@ @ @ 6 66rDct|s"td|dt|t||}|D] }||dS)Nz Expected z to be callable but got )callabler3rN_get_modules_to_materialize)r0rrmodules_to_materializers rBrrs M " "  T T TtM?R?R T T   9oVV( frDrc |p%tj|}t||}d} tj5|D]}t j|d|d}tt|dk}|r+| |d|  ddddS#1swxYwYdS#t$r>} tjdt!| dt#|dd | d} ~ wwxYw) NFr^r)rr_zIUnable to call `reset_parameters()` for module on meta device with error z(. Please ensure that your module oftype z* implements a `reset_parameters()` method.rUr)rrrZrgno_grad itertoolschainrCr r_rto_emptyreset_parameters BaseExceptionrrrrN) r0rrrWmaterialization_devicerhrmodule_state_iterhas_module_statesr?s rBrrs3el$$&&779oVV F]__ . .0 . .%.O%%e%44NN5N11%%! %(->(?(?$@$@1$D!$.OO+A5OQQQ++--- . . . . . . . . . . . . . . . . . . .   M!$Q M MLL M M M     s<C:BC- C:-C11C:4C15C:: E9D==EcDg}tj|g}|h}|r|}|||D]C}||vr=t |.||vr*||||D||SrW) collectionsdequepopleftr*childrenrr;)r0rrhqueuevisited_modulesr child_modules rBrgrgs /1  {m , ,E'2mO  +%%f---"OO-- + +LO33*<88@ 77##L111 \***  + "!rDrc tjd |#tj}||g}g}|r|}| fd|dD| fd|dD| D]1}t|tj s||2|ƈfd|D} fd|D} t| | |dStt|d} | | j krt!dSdSdS) a Move ``module`` depending on ``device_from_device_id`` and its current device. This includes moving ignored modules' parameters. - If ``device_from_device_id`` is not ``None``, then this moves ``module`` to the device. - If ``device_from_device_id`` is ``None``, then this does not move ``module`` but warns the user if it is on CPU. Precondition: ``_check_single_device_module()``. rNc32K|]}|jk |VdSrWrT)r[rr/s rBr]z)_move_module_to_device..s<<:------rDFr^c32K|]}|jk |VdSrWrT)r[rr/s rBr]z)_move_module_to_device..s<=J......rDcg|]}|v| Srkrk)r[rErs rBrnz*_move_module_to_device..s#GGGq/F/F!/F/F/FrDcg|]}|v| Srkrk)r[rErs rBrnz*_move_module_to_device..s#GGGaa.F.F.F.F.FrD)rrrtrur*rvextendrCr rwrXr4r5_move_states_to_devicenextr_warn_cpu_init) rrrrrxrr  curr_modulerparams_to_move bufs_to_moverr/s `` @rBrrs$e$$J(/:.?.A.A V%'&( ,--//K MM(33E3BB    NN)11%1@@    )1133 , , !)Y-OPP,LL+++% ,&HGGGVGGGGGGG7GGG ~|=RSSS !&.994 @ @E U\Z7777rDr c\t|dkrt|dkrdSt|dkr|dj}n t|dkr |dj}tjd}||D]r}tj5|||_|j$|j||j_dddn #1swxYwYs|D]}|||_dS||krtdSdS)z Move states to the specified device. Precondition: ``_check_single_device_module()`` and module's parameters and buffers have been materialized if needed. rNr)r_rrrjtodatagradr)rr rrZr/rrs rBrrs 6{{aCLLA-- 6{{Q) W   *e$$J( K KE K K"XX&;<< :)&+jmm4I&J&JEJO K K K K K K K K K K K K K K K ; ;F ))$9::FKK ; ; : % % & %sAC))C- 0C- c2tjdddS)NamThe passed-in `module` is on CPU and will thus have FSDP's sharding initialization run on CPU, which may be slower than on GPU. We recommend passing in the `device_id` argument for FSDP to move `module` to GPU for the sharding initialization. `module` must also be on GPU device to work with the `sync_module_states=True` flag since that requires GPU communication.rUr)rrrkrDrBrr's/ M 1 rDctt||d}||jjdkr|j}n&t j|}|||krt d|d|d||S)a) Determine and return this FSDP instance's compute device. If the module is already on a non-CPU device, then the compute device is that non-CPU device. If the module is on CPU, then the compute device is the current device. Since this method should be called after materializing the module, any non-CPU device should not be meta device. For now, the compute device is always a CUDA or CUDA-like device with its explicit index. Precondition: ``_check_single_device_module()`` and ``_move_module_to_device()``. Nrz4Inconsistent compute device and `device_id` on rank z: z vs )rrrrNrrZr3)rrrr8rWrrs rBrr3s* !&.994 @ @E U\.%77m&B&B&D&DEE(^?T-T-T ;4 ; ; ; ;#8 ; ;    rDcg}|D]}t|tdst|td|}t |r|\}}|D]m}t||xxtjdx\} | | <xtdx\M} tdt| | ||D]} | } t | r| \}}|D]m}t| |xxtjdx\} | | <xtdx\M} tdt| | | t|t||tddS)z Synchronize module states (i.e. parameters ``params`` and all not-yet-synced buffers) by broadcasting from rank 0 to all ranks. Precondition: ``sync_module_states == True`` and ``self.process_group`` has been set. FTrkNz#expected Tensor or OpaqueBase, got r)src)r getattrrrdetachr%__tensor_flatten__rTensorr*r rvrN+_check_module_states_for_sync_module_statesr$PARAM_BROADCAST_BUCKET_SIZE) rrr+ module_statesrdetached_bufferattrsrfattrv unexpectedrdetached_params rBrrVsZ)+M..""66v{E22 6 FK . . .$mmooO,_== 6+==??q!  D!/48800U\^^^^^^^)00333310)Z\\\\\\ *'"0 Xd:FVFV X X## $$_55511 ( 8 8 1%88::HE1  nd33,,%,,Q////-,%&#,T$zBRBRTT    0 0 0 0/ >>># rDrc^|r(td|DrtddSdS)Nc3PK|]!}|jtjdkV"dS)rN)rr)r[tensors rBr]z>_check_module_states_for_sync_module_states..sD17 e,,,rDzThe module has CPU parameters or buffers when `sync_module_states=True`, which requires them to be on GPU. Please specify the `device_id` argument or move the module to GPU before passing it to FSDP.)rbr3)rs rBrrsg ;H  C       rDc#K|} t|}||vrt|s|V'#t$rYdSwxYw)aD Return an iterator over the original parameters in ``module``. The iterator does not return the parameters in ``ignored_params``, any ``FlatParameter`` s (which may be present due to nested FSDP wrapping), or any original parameters already flattened (only relevant when ``use_orig_params=True``). TN)rCrr StopIteration)rr param_genrs rBrrsy!!##I  OOEN**3Ee3L3L*        s(A A Ac t|D]G\}}||vr>t|s/td|d|d|jHdS)a5 Check that original parameters in ``fsdp_module`` have been flattened. The flattened parameters are made invisible to ``named_parameters()`` for the module hierarchy rooted at ``fsdp_module``. This should be called as a sanity check after flattening the wrapped module's parameters. z Found an unflattened parameter: z;  N)rrrr9 __class__) fsdp_modulerrrs rB_check_orig_params_flattenedrs?{KK E  & &/A%/H/H &4:44::<<44"'/44 rDcR|tjkr tjn tjSrW)rrr allreduce_hookreduce_scatter_hook)r,s rB_get_default_comm_hookrs)  0 9 9 9 $$  .rDc,tj|S)NrG)r r<rGs rBrOrOs  %M B B BBrDrW)rtrkrrcollections.abcrrrrtypingrrr rtorch.distributed distributedrY(torch.distributed.fsdp._exec_order_utilsfsdp_exec_order_utilsr'torch.distributed.fsdp._traversal_utils_traversal_utilsr92torch.distributed.fsdp.fully_sharded_data_parallelfully_sharded_data_parallelr4torch.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&rc torchdistxr'r( ImportErrorrrr^rZHybridShardProcessGroupTypeProcessGroupTyperr SHARD_GRAD_OP HYBRID_SHARD_HYBRID_SHARD_ZERO2r$r4#NO_RESHARD_AFTER_FORWARD_STRATEGIESrCr5boolrMrHintrgr}rJrrrrrr7rrrrrrrrrr!rrrrrrrrrrrrgrrrrrrrrrrr<rOrkrDrBrsR CCCCCCCCCCCC4444444444 CCCCCCCCCCCCAAAAAAAAAAAAFFFFFFFFFFFF))))))BBBBBB444444AAAAAA BAAAAA                    0/////DDDDDD<<<<<<FFFFFF2111111.........0 #D$5t7H$HI$'BBTI 5>!7!B"$:$H!#9#F(*@*T !( "('# &* 00 0#0(0 dN 0 d" 0  0000f( (#(( ((((V#$IJI4IIII  AR    "+"" """"J+ 4 d/ /0& ++ + I+eho.5+UX/0ux   +++++\I9= :- - I- %-U\!D( -  ----`  I"? ?'$.?$d*?d" ?  ?  ?!? ?????D           '        : J *    *29d2<6HT(N N)NU\!D(NRYK-.5 N  N  NNNNb) )  ))))))>::ux/$6: ^::::@?C""")"!!34t;"   """"J$$)$ X$$$$@29S I  % U\!D(      <U\!D( % \D <7 I7 %7^7 4: 7777<  RYK-. ^     ## <$.#^#% ####L""-0^" ")_"""",3 I3 %3&3!<$. 3  3333l   %, !<$. @    I %!<$.   %   \ F9 I9  9$9 9999x   %         I  % bl    , % (.>C$CCCCCCCsCC! C!