o H hV$@sddlZddlmZddlmZmZmZddlZddlm Z ddl m Z ddl m Z ddlmZmZdZGd d d e Zd ed ed efddZGdddZeed dde jdeeej jd e jfddZde jd efddZdS)N)Iterable)AnyNoReturnOptional)_State)DistributedDataParallel) _get_registrycontractc seZdZUejed<dfdd Zefdej de ej de ej d e ddf d d Z dd d Zdej de ej ddfddZdddZdddZdej deedfdee efdefddZdej deejdejdejfddZZS)_ReplicateState _ddp_weakrefreturnNcsNtt|_d|_t|_|j|_g|_d|_ d|_ i|_ g|_ dS)NF) super__init__nn ParameterListmodulehas_initialized _param_list _orig_module _param_names_no_sync _init_args _init_kwargs_comm_hook_argsself __class__[/var/www/vscode/kcb/lib/python3.10/site-packages/torch/distributed/_composable/replicate.pyrs    z_ReplicateState.__init__rignored_modulesignored_paramsprefixc Cst|rdS||vr dS|tkr|dnt}|jddD]\}}||vr5|j||j||q|D]\}} |j| ||||dq:dS)N.F)recurse)r$)_is_fully_sharded_ROOT_MODULE_PREFIXnamed_parametersrappendrnamed_children_collect_params) rrr"r#r$recurse_prefixnpname child_moduler r r!r,"s&  z_ReplicateState._collect_paramscs"tjddfdd}|dS)NT) recursivecs:jdusJjjijd_i_dS)Nr )rinitrregister_comm_hookr rr r! _lazy_initCs  z-_ReplicateState.lazy_init.._lazy_init)torch_disable_dynamo)rr5r rr! lazy_initBs  z_ReplicateState.lazy_initcKs|jrdSd|_||_dd|D}|D] }t|r#||qddlm}|||d||||d|vrd|ddur[|d}t |t j rU|j dkrUd|d <n |g|d <nd|d <| dt|jfi||_t|jt|j_dS) NTcSsh|] }|D]}|qqSr ) parameters).0mr/r r r! Xsz'_ReplicateState.init..r)_localize_dtensor)r# device_idcpu device_ids)rrmodulesr'updater9%torch.distributed.tensor.parallel.ddpr=r, isinstancer6devicetypepoprr_ddpweakrefref replicatestater )rrr"kwargsr# submoduler=r>r r r!r3Ms,       z_ReplicateState.initcCs0|jD] \}}|jj|i|q|jdSN)rrHr4clear)r comm_args comm_kwargsr r r!r4wsz"_ReplicateState.register_comm_hookcOs||_||_dSrO)rrrargsrMr r r!record_init_args|s z _ReplicateState.record_init_argsrT.rMcCs2|js|jr ||j |j_|jj|i|SrO)rrr8rrHrequire_backward_grad_sync _pre_forward)rrrTrMr r r!forward_pre_hooks  z _ReplicateState.forward_pre_hookinputoutputcCs |j|SrO)rH _post_forward)rrrYrZr r r!forward_post_hooks z!_ReplicateState.forward_post_hookrN)__name__ __module__ __qualname__rIrJ__annotations__rr(rModuleset Parameterstrr,r8r3r4rUtuplerdictrXr6Tensorr\ __classcell__r r rr!r sV    *    r rTrMrcOstd)NzGDDP does not support deepcopy. Please use state dict for serialization.)AssertionError)rTrMr r r!unimplemented_deepcopysrkc@s0eZdZddZdeddfddZd dd ZdS) DDPcOs"|jd}|j|g|Ri|S)z Override ``__new__`` to remove the DDP class and directly construct the original class for cases like indexing into a container module. )__mro____new__)clsrTrMorig_clsr r r!ros z DDP.__new__requires_gradient_syncrNcCs| t|_dS)a Sets if the module should sync gradients. This can be used to implement gradient accumulation without communication. Args: requires_gradient_sync (bool): Whether to reduce gradients for the module's parameters. N)rKrLr)rrrr r r!set_requires_gradient_syncs zDDP.set_requires_gradient_synccOst|j||fdSrO)rKrLrr*rSr r r!r4szDDP.register_comm_hookr])r^r_r`roboolrsr4r r r r!rls  rl) state_clsrr"c Ks$tjdd|vrt|dttjfstdt|dt|r'td|dur.i}nt |}t |}|j |j dd|dd}|durjd d lm}||}||krjd d lm}m}| |||||j|j||fi||j} d ti} td | jt| f| } | |_|S)zReplicates a module Args: module (torch.nn.Module): module to replicate Example:: >>> # xdoctest: +REQUIRES(module:torch._C._distributed_c10d) >>> module = nn.Linear(3, 3) >>> replicate(module) ztorch.distributed.replicater>z6Expected device_id to be int or torch.device, but got zGCannot apply `replicate()` on a Module already managed by `fully_shard`NT) with_kwargs device_meshr)_mesh_resources)r=_reconstruct_dtensor __deepcopy__rl)r6_C_log_api_usage_oncerDintrE RuntimeErrorrFr'rcrKrLregister_forward_pre_hookrXgettorch.distributed.device_meshrx get_root_meshrCr=ryregister_forward_hookr\rUrrkr^rl) rr"rMrLrwrx root_meshr=ryrpdctnew_clsr r r!rKs@         rKcCst|}|dur dSd|vS)z+Check if module is marked with fully_shard.NF fully_shard)r )rregistryr r r!r'sr'rO)rIcollections.abcrtypingrrrr6torch.nnr#torch.distributed._composable_statertorch.nn.parallelrr r r(r rkrlrbrKrtr'r r r r!s.     H