o hs @sddlZddlZddlZddlZddlZddlmZmZddlm Z m Z ddl m Z ddl mZmZmZmZmZmZddlZddlmZddlmZddlmZmZddlmZdd lm Z d d l!m"Z"d d l#m$Z$m%Z%m&Z&d d l'm(Z(er|ddlm)Z)gdZ*e+e,Z-Gddde Z.e.j/Z/e.j0Z0e.j1Z1e.j2Z2e.j3Z3e.j4Z4e.j5Z5e.j6Z6e.j7Z7e.j8Z8e/Z9e0Z:e1Z;e8ZGdddeZ? dPde@eAeBee?fdeeAdeCfddZDGdddeZEdPdeBejFdeeCfddZG dPdeBejFdeeCde@eAej)ffd d!ZHGd"d#d#eEZIGd$d%d%eIZJGd&d'd'eIZKGd(d)d)eIZL *dQd+eBee?d,eAdeBe?fd-d.ZMd+eBee?deBe?fd/d0ZNd+e@eAeBe?fd1eeAgeAfd2eAde@eAeBe?ffd3d4ZOd5e@eAeBee?fd6eAd2eAd7eAde@eAeAff d8d9ZPGd:d;d;eEZQGdd?d?eQZS @dRdAdBZTGdCdDdDeQZUGdEdFdFeQZVGdGdHdHeQZWdIeCfdJdKZXd1eeAgeAfd2eAfdLdMZYdNdOZZdS)SN)ABCabstractmethod)Counter defaultdict)Enum)AnyCallable NamedTupleOptional TYPE_CHECKINGUnion)OptimizedModule) FSDPModule UnshardHandle)_Loss)record_function)generate_stage_to_rank_mapping) merge_chunkssplit_args_kwargs_into_chunksTensorChunkSpec)_PipelineStageBase)Work) get_schedule_classPipelineScheduleSinglePipelineScheduleMulti Schedule1F1B ScheduleGPipeScheduleInterleaved1F1BScheduleLoopedBFSScheduleInterleavedZeroBubbleScheduleZBVZeroBubblec@sHeZdZdZdZdZdZdZdZdZ dZ d Z d Z d d Z ed dZdS)_ComputationTyper cCsHtjdtjdtjdtjdtjdtjdtjdtjdtj d tj d i }||S) NFIWUNSHARDRESHARDSEND_FRECV_FSEND_BRECV_BB) r"FORWARDBACKWARD_INPUTBACKWARD_WEIGHTr/r0r1r2r3r4 FULL_BACKWARD)selfstr_mapr<z/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/torch/distributed/pipelining/schedules.py__str__9s z_ComputationType.__str__cCs|dkrtjS|dkrtjS|dkrtjS|dkrtjS|dkr#tjS|dkr*tjS|dkr1tjS|dkr8tjS|d kr?tj S|d krFtj St d |) Nr,r-r.r/r0r1r2r3r4r5zInvalid computation type ) r"r6r7r8r/r0r1r2r3r4r9 RuntimeErroractionr<r<r=from_strHs*z_ComputationType.from_strN)__name__ __module__ __qualname__r6r7r8r/r0r1r2r3r4r9r> staticmethodrBr<r<r<r=r",sr"z?(\d+)(F|I|B|W|UNSHARD|RESHARD|SEND_F|RECV_F|SEND_B|RECV_B)(\d*)c@sHeZdZUeed<eed<dZeeed<ddZe de fdd Z dS) _Action stage_indexcomputation_typeNmicrobatch_indexcCs4t|j}|t|j7}|jdur|t|j7}|SN)strrHrIrJ)r:reprr<r<r=__repr__~s  z_Action.__repr__ action_stringcCsh|}t|}r&|\}}}tt|t|t|r#t|SdS|dkr,dSt d|d)z Reverse of __repr__ String should be formatted as [stage][action type][(microbatch)] e.g. `2F0`, `1UNSHARD`, `3SEND_F1` NzInvalid action string: zD, should be formatted as [stage][action type][(microbatch)] e.g. 2F0) strip _action_regexmatchgroupsrGintr"rBlenr?)rOrSrHrIrJr<r<r=rBs z_Action.from_str) rCrDrErU__annotations__r"rJr rNrFrLrBr<r<r<r=rGys rGpipeline_ordererror_step_numberreturnc s6tD]}tt|D]}||dur!d||<qqtddDfddtD}fddtD}ttj |ddi}t}d dt|D}d dt |g|RDd t|d d d fddt |D} fddt ||D} | dd | d} | S)z Formats the pipeline order in a timestep (row) x rank (column) grid of actions and returns the formatted string. If `error_step_number` is passed in, an additional label will be added to signify which step that it is erroring on. NrPcss|]}t|VqdSrK)rV).0actionsr<r<r= sz)_format_pipeline_order..c s*g|]}dt|ttdqS)zStep r)rLzfillrVr[i) num_stepsr<r= sz*_format_pipeline_order..csg|] }|dgqS)rP)get)r[key)rarXr<r=rbs fillvaluecSsg|]}dt|qS)zRank rLr_r<r<r=rbcSsg|] }tdd|DqS)css(|]}|durtt|ndVqdS)Nr)rVrL)r[itemr<r<r=r]s&4_format_pipeline_order...)max)r[colr<r<r=rbs rr#c3s&|]\}}|d|VqdS.N Expecting z losses but got ) isinstancer|anyrVrrr?clearextend)r:stageslossescontains_last_stager<r<r=_update_lossess   z _PipelineSchedule._update_lossesarg_mbs kwarg_mbsrrcCt)z Run one iteration of the pipeline schedule with list of microbatches. Will go through all the microbatches according to the schedule implementation. Args: microbatches: list of microbatch args. NotImplementedError)r:rrrrr<r<r=_step_microbatches%sz$_PipelineSchedule._step_microbatchestargetrcOr) Run one iteration of the pipeline schedule with *whole-batch* input. Will chunk the input into microbatches automatically, and go through the microbatches according to the schedule implementation. args: positional arguments to the model (as in non-pipeline case). kwargs: keyword arguments to the model (as in non-pipeline case). target: target for the loss function. losses: a list to store the losses for each microbatch. r)r:rrargskwargsr<r<r=step7s z_PipelineSchedule.stepcsdtffdd }|dur||dndgj}|dur#||dnigj}|dur2||d|durDt|tsDtd t|||fS) z* Pre-process/check inputs namecsRt|tst|dt|t|jkr'tdjd|dt|dS)Nz must be a list but got a rrlz but got )rr| TypeErrortyperVr ValueError)mbsrr:r<r=check_type_and_lenPs z;_PipelineSchedule._check_inputs..check_type_and_lenNrr<rrz losses must be a list but got a )rLrrr|rr)r:rrrrrr<rr= _check_inputsEs       z_PipelineSchedule._check_inputscCs |||SrK)r)r:rrr<r<r=rks z_PipelineSchedule._compute_lossrrcCsB|s|rt|||j|j|j\}}||fSdg|jig|jfS)zj Splits a full-batch input into chunks (i.e. microbatches) and returns the chunks r<)rrrr)r:rr args_split kwargs_splitr<r<r= _split_inputsns z_PipelineSchedule._split_inputs output_chunksrZcCs t||jS)z Merge output chunks back to a batch state. If output_merge_spec is None, the utility will merge output chunks by dimension 0 (batch dim). )rr)r:rr<r<r=_merge_outputssz _PipelineSchedule._merge_outputsNNNNTNNNNrK)rCrDrErUr rtorchTensortuplerdictrLr rboolrrrrrr|rrrrrrr<r<r<r=rsv "   &  rp2p_opsdesccCs>t|dkrdS|r|dnd}td||t|S)zt Simple wrapper over batch_isend_irecv from torch.distributed, which just adds a descriptive logger on top. rNz, rPzbatch_p2p %s%s)rVrdebugdistbatch_isend_irecvpop)rrdesc_strr<r<r= _batch_p2ps rcCs`tt}i}t|dkr|S|D] }||j|qt|D] \}}t||d||<q!|S)z Sorts the list of P2P ops by the peer rank, and then calls batch_isend_irecv. Return a dictionary of works by peer rank. This function helps us avoid hangs in case of skip connections. rr)rr|rVpeerrr{itemsr)rr ops_by_peer work_by_peeropropsr<r<r=_sorted_batch_p2ps  rcseZdZdZ     ddededeedeee dfd ee e e fd ee e e e fee fd effd d ZddZddddeefddZZS)ra Base class for single-stage schedules. Implements the `step` method. Derived classes should implement `_step_microbatches`. Gradients are scaled by num_microbatches depending on the `scale_grads` argument, defaulting to True. This setting should match the configuration of your loss_fn, which may either average losses (scale_grads=True) or sum losses (scale_grads=False). NTrrrr.rrrcs\tj||||||d||_|j|_|j|j_d|_||jkr,td|d|jddS)NrrrrrrFzNumber of microbatches (z9) must be greater than or equal to the number of stages (z).) superr_stage num_stages _num_stagesr has_backward_stage_initializedr)r:rrrrrrrrr<r=rs&    zPipelineScheduleSingle.__init__cCs0|j|j|||jr|j|jd|_dSNT)r_prepare_forward_infrarr_prepare_backward_infrar)r:rrr<r<r=_initialize_stages z(PipelineScheduleSingle._initialize_stagerrcOsd|j|||\}}|durtt||j}nd}||||||jjr0| |jj SdSrN) rclear_runtime_statesrr|r tensor_splitrrrrr)r:rrrrrr targets_splitr<r<r=rs zPipelineScheduleSingle.stepr)rCrDrE__doc__rrUr rrrrrLr rrrrr|r __classcell__r<r<rr=rs0 "rc @DeZdZdZ    d deedeedeedeefddZdS) _ScheduleForwardOnlyzo The forward-only schedule. Will go through all the microbatches and perform only the forward pass Nrrrrc Cs|dus|dur td|||||\}}|js#||d|dg}t|jD]X}td|>|j|}t |dd}| D]} | qD|j ||||||j |}t |dd}|| Wdn1stwYtd|jj|q*|D]} | qdS) z< Run one iteration of the pipeline schedule Nz7Forward-only schedule does not support loss computationrForward fwd_recvrfwd_send[%s] Forwarded microbatch %s)r?rrrryrrrget_fwd_recv_opsrrzwaitforward_one_chunkget_fwd_send_opsrrrrH) r:rrrrfwd_sends_to_waitr`rworksworkr<r<r=rs.         z'_ScheduleForwardOnly._step_microbatchesrrCrDrErr r|rr<r<r<r=r rc @r) rz^ The GPipe schedule. Will go through all the microbatches in a fill-drain manner. Nrrrrc Cs0|||||\}}|js||d|dg}t|jD]a}td|>|j|}t|dd}| D]} | q8|j |||||} |j |}t|dd}| | Wdn1shwYtd|jj|||j| ||q|D]} | q|jsdSg} t|jD]a}td|G|j|}t|d d}| D]} | q||j|} |jj|| ||jd kd |j|}t|d d}| | Wdn1swYtd |jj|q|jj|jr|jnd d||j|| D]} | qdS)z Run one iteration of the pipeline schedule with list of microbatches. Will go through all the microbatches according to the GPipe schedule. Args: microbatches: list of microbatch args. rrrrrNrz Backward bwd_recvrr last_backwardbwd_sendz[%s] Backwarded microbatch %sgrad_scale_factor)rrrryrrrrrrzrrrrrrrHrrget_bwd_recv_opsrbackward_one_chunkget_bwd_send_opsrr) r:rrrrrr`rrrrbwd_sends_to_waitrr<r<r=rBs\                z ScheduleGPipe._step_microbatchesrrr<r<r<r=r<rrc @r) rzo The 1F1B schedule. Will perform one forward and one backward on the microbatches in steady state. NrrrrcCs|||||\}}|js||d|dt|j|j|jj}d}d}d}g} t|D]E} |j |} t | dd} rB| |j |||||} |rT| |j |} ||dkrft | dd}||j| |||d7}q. |j|}t | |dd}r| ||j|}|jj||||jdkd |j|}|d7}||jkrn4|j |} t || d d}r| |j |||||} ||j| |||j |} |d7}qut |d d}||jkr2|j|}t |d d} r| ||j|}|jj||||jdkd |r| |j|}t |d d}|d7}||jks|jj|jr<|jndd |rG| ||j|dS)z Run one iteration of the pipeline schedule with list of microbatches. Will go through all the microbatches according to the 1F1B schedule. Args: microbatches: list of microbatch args. rNrrrrTfwd_send_bwd_recvrbwd_send_fwd_recvrrr)rrrminrrrrHryrrrrrrr rr r rr)r:rrrr warmup_chunks fwd_mb_index bwd_mb_index send_work fwd_sends_ fwd_recvs recv_workr bwd_recvs fuse_workr bwd_sendsr<r<r=rs              ,      zSchedule1F1B._step_microbatchesrrr<r<r<r=rrrr$compute_actionsmax_active_stagesc sdtdtttdttfdd}tgdtffdd }dtffd d }t|D]?\}}|d ur7q.||||d ttfd d }ttfdd }|D]} || qX|D]} || qa|q.S)aRGiven a basic schedule involving only compute actions (F,B,W), add UNSHARD/RESHARD actions for FSDP. UNSHARD refers to fetching the full contents of an FSDP-sharded layer, requiring an all-gather operation. RESHARD does the opposite, releasing memory (but doing no commmunication) We abandon the "timestep lock" during lowering max_active_stages controls how many prefetches we allow. It should be measured in mb and tuneable but in practice 3 stages is probably the thing we want? (to account for having one f and one b active, and something else prefetching?) count next_actionsrZcSsTt}g}|D] }|dur'|j|vr'||j||jt||kr'|Sq|S)zdRemove duplicates (same stage, different microbatch), find next 'count' stages that will do compute.N)setrHaddrrV)rrseenretar<r<r=next_stage_indicesAs   z0_add_unshard_reshard..next_stage_indicesrHc |t|tddSrK)r rrGr/rH active_stagesfsdp_aware_actionsr<r=_unshardS z&_add_unshard_reshard.._unshardcr%rK)removerrGr0r&r'r<r=_reshardWr+z&_add_unshard_reshard.._reshardNc|vSrKr<s)r(r<r=bz&_add_unshard_reshard..cr.rKr<r/)next_nr<r=r1dr2)rUr|r rGrrsfilterr) rrr$r*r-r`rAfetchevictrr<)r(r)r3r=_add_unshard_reshard1s.     r7cCsg}|r]|d}|durqt|r+|d}dur+|dt|r+|d}dus|jtkrV|durV|jtkrV|j|jkrV|j|jkrV|t|jt |j|dn|||s|S)a9Given a basic schedule involving only compute actions (F,I,W), merge adjacent I and W ops into B ops. (note: I = BACKWARD_INPUT, W = BACKWARD_WEIGHT, B = FULL_BACKWARD) B refers to running the whole backward (not separating grad_input and grad_weight), which can be more efficient in some cases. rN) rrVrIr7r8rHrJrrGr9)rmerged_actionsrA next_actionr<r<r= _merge_bwws(         r: stage_to_rankrc sdd|D}dd|D}dtdtffdd dtdtttfffdd }dttd ttdtffd d }|rd }t|D]x}t||dksZJd|dt||||d} || ||shqB| dur||| ||| | r|| \} } ||| ||| || j | || j | || dt||dkr||=d}qB|sJd|s<|S)NcSi|]}|gqSr<r<r[rr<r<r= sz"_add_send_recv..cSi|]}|tqSr<rr=r<r<r=r>rArZcsd|jtkr|jdko|jd|jkS|jttfvr0|jdko/|jd|jkSdS)NrrF)rIr,rHr7r9r@)rr;r<r= _has_commss  z"_add_send_recv.._has_commscsx|s J|d|j}|j}|j}t||tkrtnt|}|tkr(|dn|d}t||tkr4tnt|}||fS)Nz is not a valid comm actionr) rHrIrJrGr,r1r3r2r4)rA stage_idxctypemb_idxsendrecv_stage_idxrecv)rBr<r= _get_commssz"_add_send_recv.._get_comms prev_actionscs|durdS|jtkr,|jdks,t|jt|j|vrdSt|jdt|j|vr*dSdS|jttfvrd|jdksdt|jt|j|vrFdSt|jdt|j|vrTdSt|jdt|j|vrbdSdSdS)aWe don't put our own recv ops in the schedule, we let a sender on another rank put our recv ops in place. This helps ensure a sane (non-hanging) ordering of sends and recvs. But it also means we might not be able to schedule our next compute action yet. NTrrF) rIr,rHrGr2rJr7r9r4)rArJ)rr<r=_ready_to_schedules8z*_add_send_recv.._ready_to_scheduleFrrank=z, len(compute_actions[rank])=Tz6Malformed compute schedule, can't schedule sends/recvs) rGrrr rr{rVrr rHr) rr;r comm_actionsrJrIrKprogressrrArFrHr<)rBrr;r=_add_send_recvsH  ,    rOr\ pp_group_sizenum_microbatchesc Cs|t||ksJd|dt|t|D] }||vs#Jd|qddt|D}i}|D]}||D]}|dur>q7t|tsKJd|d|j}|j} |j} | tkrb||t| nb| t kr| ||tvsyJd|d | d ||t | nA| t kr| ||tvsJd |d | d ||t | n | t kr| ||t vsJd |d | d ||t | ||vr|||<q7||} || ksJd|d|d| q7q1|D]U}t||t} t||t } t||t }t||t }| |ksJd| dtd|d|| ||d|ks;Jd|d|d| d|d| q|S)Nz2Schedule has incorrect number of ranks - expected z , actual z%Schedule is missing actions for rank c Ss*i|]}|ttttttttiqSr<)r,rr5r-r.)r[stage_idr<r<r=r>sz&_validate_schedule..zGot an invalid action: z, expected instance of _Actionz Running Full Backward for stage z , microbatch z without first running Forwardz!Running Backward Input for stage z"Running Backward Weight for stage z% without first running Backward InputzStage z is assigned to both rank z and rank zGot rlz microbatches for stage z , expected r#z(Invalid backward microbatches for stage z : expected z( total backwards, but got B=z, I=z, W=) rVryrrGrHrIrJr,r r5r-r.)r\rPrrQr stage_actionsstage_index_to_rank_mappingrAs_idrDmb_id existing_rankf_mbb_mbi_mbw_mbr<r<r=_validate_schedulesz       " r\cs$eZdZdZ      d$deededeedee e dfd ee e e fd ee e e efe efd eed effd d Zde edffddZde eeeefddfddZddZd%ddZddddeefddZ    d&deed eed!eedeefd"d#ZZS)'raX Base class for multi-stage schedules. Implements the `step` method. Gradients are scaled by num_microbatches depending on the `scale_grads` argument, defaulting to True. This setting should match the configuration of your loss_fn, which may either average losses (scale_grads=True) or sum losses (scale_grads=False). NTrrrr.rruse_full_backwardrc stj||||||d||_|dj|_|dj|_|dj|_t |j|j|_ |jD]} |j | _ q,|jD]} |j | _ q6d|_ |jdufdd|_i|_|durZtddSdS)NrrFcs |joSrKr)rhas_lossr<r=r1 z0PipelineScheduleMulti.__init__..zDeprecation warning: 'use_full_backward' is no longer supported. Simply stop passing it, and everything should still work fine.)rr_stagesrr group_sizerP group_rankrrstage_index_to_group_rankrr_stages_initializedr_should_compute_lossrXrwarning) r:rrrrrrr]rrrr^r=rYs8         zPipelineScheduleMulti.__init__rcCsVt}|jD]}|jr||j||}n||j||}|jr%||jqd|_dSr)rrais_firstrrrrre)r:rrnext_stage_argsrr<r<r=_initialize_stagess   z(PipelineScheduleMulti._initialize_stagesr\rZcCs.t||j|j|j|_|jD]}|j|_qdS)z] Allocates the stage index to rank mapping which is needed for communication N)r\rPrrrdra)r:r\rr<r<r=_validate_and_set_stage_mappings  z5PipelineScheduleMulti._validate_and_set_stage_mappingcCsZt|ddd}t|}|jD] }||j|qWddS1s&wYdS)zQDump a CSV representation of the schedule into a file with the provided filename.wrPnewlineN)opencsvwriterrXwriterowr:filenamecsvfilerqrr<r<r= _dump_csvs   "zPipelineScheduleMulti._dump_csv compute_onlycCsx|dksJt|dd }t|}t|D]\}}dd|D|j|<qWdn1s/wY||jdS)zLoad a CSV representation of the schedule from a file with the provided filename. This API will most likely get renamed/refactored so is marked as internal for now. format must be "compute_only" for PipelineScheduleMulti. rwrPrmcSg|]}t|qSr<rGrBr[r0r<r<r=rbz3PipelineScheduleMulti._load_csv..N)rorpreaderrsrXrk)r:rtformatrur|rrur<r<r= _load_csvs  zPipelineScheduleMulti._load_csvrrc Osz|jD]}|q|||\}}|dur tt||j}nd}||||||jD] }|jr:| |j Sq-dSr) rarrr|rrrrrrr) r:rrrrrrrrr<r<r=rs  zPipelineScheduleMulti.steprrrc CsN|||||\}}|js||d|ddd|jD}t}t}|D]!}|dkr9||j|d||jdkrJ||j|dq)t } t |j |j D]\} } zg} | dur3| j } | j}| j}|dusuJd| tjkr||}||||||}|||||| ||n| tjkr||}|||}| |d7<| ||jk}|jr|jnd}|j||d|d|r||| ||nZ| tjkr||}|||}|j||d d d| ||n9| tjkr,||}| |d7<| ||jk}|jr|jnd}|j||d |r+||ntd | |D]Z}|j |}d}| t |krI|| }|dur|j } |j}|j}|dus`Jd| tjkr}|d|vr{||d}| |!|q5| tttfvrq5td | q5|D]Z}|j |}d}| t |kr|| }|dur|j } |j}|j}|dusJd| ttfvrƐq| ttfvr|d|vr||d}| |"|qtd | q| rt#| $WqVt%y}zt&'d |j |j(j)| | t&'d t*|j | d|d}~ww|+|j|dS) Operate on the microbatches for looped schedules (multiple stages on each rank). TODO: Does not use sorted_batch_isend_irecv(). As a result, this schedule does not support models with skip connections. rcSi|]}|j|qSr<r&rr<r<r=r>z.rNzCAll currently supported action types require valid microbatch_indexTr full_backwardrFrzUnknown computation type zy[Rank %s] pipeline schedule %s caught the following exception at time_step %s when running action %sz%srY),rrerjrarkeysr rdrrrsrXrrIrJrHr"r6rrrrr9rrrr r r7r8backward_weight_one_chunkrrVrr rr ExceptionrerrorrrCrr)r:rrrrstage_index_to_stageall_prev_ranksall_next_ranksrHbackward_counter time_steprArrIrrrrrr prev_rank prev_rank_opsprev_rank_action next_rank next_rank_opsnext_rank_actioner<r<r=rs"                        z(PipelineScheduleMulti._step_microbatches)NNNNNTrwr)rCrDrErr|rrUr rrrrrLr rrrrjrGrkrvr~rrrr<r<rr=rOs^  2  $rc seZdZdZ ddeeeeefde ffdd Z dde de ffdd Z de fd d Z d d Z    ddeedeedeedeefddZZS)_PipelineScheduleRuntimea% Provides a simple runtime that requires a 'schedule IR' including specified communication operations. Can be instantiated directly by creating _PipelineScheduleRuntime and calling load_csv, or can be subclassed and the subclass can be responsible for creating a schedule IR. rwr\r}cst|i_|dkr.|D]}gj|<||D]}|dus"Jj||qqdS|dkrP|D] }t||j|<q4tjfddjd_dStd|d) z Given an in-memory representation for a simple compute-only schedule, lower it to a complex schedule including communication actions. Stores the schedule in self, and must be called before running step_mo() compute_commsNrwc j|SrKrdr/rr<r=r1r`z8_PipelineScheduleRuntime._load_actions..)r;rformat= is not implemented)rrkpipeline_order_with_commsrr7rOrr)r:r\r}rrArrr= _load_actionss*       z&_PipelineScheduleRuntime._load_actionsrtcs|dkrt|||jdS|dkrOi}t|dd'}t|}t|D] \}}dd|D||<q(|j||dWddS1sHwYdStd |d ) aLoads a csv in simple format and then lowers it to include comunication actions format must be either "compute_only" or "compute_comms". If compute_only, the lowering passes will automatically be run to generate a compute_comms schedule. rwrrPrmcSrxr<ryrzr<r<r=rbr{z6_PipelineScheduleRuntime._load_csv..)r}Nrr) rr~rrXrorpr|rsr)r:rtr}r\rur|rrurr<r=r~s  "z"_PipelineScheduleRuntime._load_csvcCsl|jdus Jdt|ddd}t|}|jD] }||j|qWddS1s/wYdS)zaDump a CSV representation of the compute + comms schedule into a file with the provided filename.Nz6Must initialize compute_comms schedule before dump_csvrlrPrm)rrorprqrrrsr<r<r=rvs   "z"_PipelineScheduleRuntime._dump_csvcstjfddjS)NcrrKrr/rr<r=r1r`z4_PipelineScheduleRuntime._simulate..)_simulate_comms_computerrrr<rr= _simulates  z"_PipelineScheduleRuntime._simulateNrrrrc s@|||||\}}|js||d|ddd|jD}|jdus(Jdi}i}g}itdtffdd } t} t|j|j D]2\} } z| j } | j durZ| j nd }|dksn| t t fvsnJd | d | j}||}t|jt}|d |v}|d |v}td | | | tkr|t||n| tkr|t||n| tkr||f|vsJdt|||||f<n| tkr||f|vsJdt|||||f<nw| t kr|r|vr|vsJd|d|jjdd|<nT| t kr0|r.|vsJd|d|vs)Jd|d|jn+| t kr||r<| ||j!s[|s[||f|vsRJd| d|"||f#|$|||||}|%|||||r{||d &||n| t'kr|r| ||j(s|s||f|vsJd| d|"||f#|)||}| |d 7<| ||j*k}|j+r|j*nd }|j,||d|d|r|+||r||d -|.||nt| t/kr2|r| ||j(s|s||f|vs Jd| d|"||f#|)||}|j,||ddd|r1||d -|.||n)| t0krS|r>| || |d 7<|j1|| ||j*kdnt2d | dWqHt3y{}zt4d| | t5t6|j| d|d}~wwt7|r|"#t7|st7dksJd |8|j|dS)!rrcSrr<r&rr<r<r=r>rz?_PipelineScheduleRuntime._step_microbatches..Nz=Must call _load_actions() before calling _step_microbatches()rCcs>|vr||=||vsJd|dS)zQIf an unshard is active for `stage_idx`, wait() it and mark `stage_idx` unshared.z*Attempted to compute on sharded stage_idx=N)rr )rC unshard_opsunsharded_stagesr<r=_assert_unshardeds   zF_PipelineScheduleRuntime._step_microbatches.._assert_unshardedzaction=z missing mb_indexrz8_PipelineScheduleRuntime running time_step %d, action %szARecv twice for {stage_idx=} {mb_index=} without executing forwardzBRecv twice for {stage_idx=} {mb_index=} without executing backwardzUnsharding the same stage_idx=z twiceT)async_opzResharding stage_idx=z without unshardingz before finishing unshardzComputing action=z before receiving inputz Attempted to run compute action=rFrz is unknown or unsupportedz\_PipelineScheduleRuntime caught exception at step %s when running action %s. Full Schedule:rzUnused unshard operations)9rrerjrarrrUrrsrrIrJr/r0rHrsubmodrrrr1rrrr3r r2rr4r unshardreshardr6rhrrrrset_local_fwd_inputr9rrrrr set_local_bwd_inputget_local_bwd_outputr7r8rrrrprintrrVr)r:rrrrr bwd_recv_ops fwd_recv_opssend_opsrrrrA comp_typerrCrstage_uses_fsdpis_next_stage_on_this_rankis_prev_stage_on_this_rankrrrrrr<rr=rs                                     z+_PipelineScheduleRuntime._step_microbatchesrr)rCrDrErrrUr|r rGrLrr~rvrrrr<r<rr=rs0 #  rc sneZdZdZ   d deededeee e fdeee e e fee fdef fd d Zd d ZZS)rai Breadth-First Pipeline Parallelism. See https://arxiv.org/abs/2211.05953 for details. Simliar to Interleaved 1F1B, Looped BFS supports multiple stages per rank. What is different is that when microbatches are ready for multiple local stages, Loops BFS will prioritizes the earlier stage, running all available microbatches at once. NTrrrrrcsDtj|||||di|_t|jD] }||}||j|<qdS)N)rrrrr)rrrXryrP!_calculate_single_rank_operations)r:rrrrrrrank_opsrr<r=rs   zScheduleLoopedBFS.__init__cst|j}t||j||j}ddt|D}|D]|fddt|jDqd|jd|}|dg|t|D]|fddtt|jDqA|S)NcSg|]}dqSrKr<r[rr<r<r=rbzGScheduleLoopedBFS._calculate_single_rank_operations..c3|] }ttj|VqdSrK)rGr"r6r[rr&r<r=r]  zFScheduleLoopedBFS._calculate_single_rank_operations..r#rc3rrK)rGr"r9rr&r<r=r]!r)rVraryrPrrreversed)r:rn_local_stages stage_indicesrpost_warmup_opsr<r&r=r s     z3ScheduleLoopedBFS._calculate_single_rank_operations)NNT)rCrDrErr|rrUr r rrrrLrrrrrrr<r<rr=rs" rFc Cstt} tt} tt} ddt|D} ||d|d|||}| r-||d}|||}g}d}| r;tnt}t|D]}||krn||}| |}d| |<| t|tj|||dkrm| dg|qA||krz||krnn[||}| |}d| |<| t|tj|||}| |}d| |<| t|||||| r|||kr|||}| |}d| |<| t|tj ||d7}qA| s| d||}| |}d| |<| t|||||| r"|||kr"|||}| |}d| |<| t|tj ||d7}qA| rU|t |krU|||}| |}d| |<| t|tj ||d7}| rU|t |ks-| S)NcSrrKr<rr<r<r=rb;rz&_get_1f1b_rank_ops..r#rr) rrUryr7r9rrGr"r6rr8rV)rrP warmup_ops fwd_bwd_ops cooldown_opsrforward_stage_indexbackward_stage_indexnum_1f1b_microbatchesenable_zero_bubblefwd_stage_mb_indexbwd_stage_mb_indexweight_stage_mb_indexrr total_opsbackward_op_idsweight_op_countFULL_BACKWARD_OR_BACKWARD_INPUTrfwd_stage_indexrrbwd_stage_indexrweight_stage_indexweight_mb_indexr<r<r=_get_1f1b_rank_ops(s                       rceZdZdZ     ddeededeedee e dfd ee e e fd ee e e efe efd effd d ZdeeefddZZS)ra The Interleaved 1F1B schedule. See https://arxiv.org/pdf/2104.04473 for details. Will perform one forward and one backward on the microbatches in steady state and supports multiple stages per rank. When microbatches are ready for multiple local stages, Interleaved 1F1B prioritizes the earlier microbatch (also called "depth first"). This schedule is mostly similar to the original paper. It differs by being relaxing the requirement of num_microbatch % pp_size == 0. Using the flex_pp schedule, we will have num_rounds = max(1, n_microbatches // pp_group_size) and it works as long as n_microbatches % num_rounds is 0. As a few examples, support 1. pp_group_size = 4, n_microbatches = 10. We will have num_rounds = 2 and n_microbatches % 2 is 0. 2. pp_group_size = 4, n_microbatches = 3. We will have num_rounds = 1 and n_microbatches % 1 is 0. NTrrrr.rrrc s|dj|_tj|||||||dt||_|dj|_td||j|_ ||j |_ ||j dkr@t d|j d|di|_ t |jD] }||} | |j |<qHdS)Nrrrrrrrrrz_Interleaved 1F1B requires the number of microbatches to be a multiple of the number of rounds ( ), but got .)rbrPrrrVrrcrrjnumber_of_roundsmicrobatches_per_roundrrXryr) r:rrrrrrrrrrr<r=rs6     z ScheduleInterleaved1F1B.__init__rZc sfdd}|jj}|}||}||}td|||fdd}fdd}tjj||||S)Ncs<jdj}d}||jd|}t|jjS)Nrr#rrrPrrrwarmups_ops_last_stagemultiply_factorrrr<r=get_rank_warmup_ops zVScheduleInterleaved1F1B._calculate_single_rank_operations..get_rank_warmup_ops=rank %s, warmup_ops %s, 1f1b %s, cooldown_ops %s total_ops %sc|jj}|jSrKrrrPr local_indexrr:r<r=r zVScheduleInterleaved1F1B._calculate_single_rank_operations..forward_stage_indexc,jd|jj}|jSNrrrrPrrr:rr<r=r zWScheduleInterleaved1F1B._calculate_single_rank_operations..backward_stage_indexrrrrrrP) r:rrmicrobatch_opsrrrrrr<rr=rs4     z9ScheduleInterleaved1F1B._calculate_single_rank_operationsrrCrDrErr|rrUr rrrrrLr rrrrGrrr<r<rr=rs.&rcseZdZdZ     ddeededeedee e dfd ee e e fd ee e e efe efd effd d ZdeeefddZddZZS)r aw The Interleaved Zero Bubble schedule. See https://arxiv.org/pdf/2401.10241 for details. Will perform one forward and one backward on inputs for the microbatches in steady state and supports multiple stages per rank. Uses the backward for weights to fill in the pipeline bubble. In particular this is implementing the ZB1P schedule in the paper. NTrrrr.rrrc s|D] }t|jtrtdq|dj|_tj|||||||dt||_ |dj |_ t d||j|_ ||j |_||j dkrOtd|j d|di|_t|jD] } || } | |j| <qW||j |j|_dS)NzYThe Zero Bubble schedule is not supported with stage modules that have used torch.compilerrrzZZero bubble requires the number of microbatches to be a multiple of the number of rounds (rr)rrr r?rbrPrrrVrrcrrjrrrrXryr_add_bubbles_to_actions r:rrrrrrrrrrrr<r=r0sF        z&ScheduleInterleavedZeroBubble.__init__rZc sfdd}|jj}|}||}||}td|||fdd}fdd}} tjj||||| dd S) Ncs<jdj}d}||jd|}t|jjSrrrrr<r=rgrz\ScheduleInterleavedZeroBubble._calculate_single_rank_operations..get_rank_warmup_opsrcrrKrrrr<r=rrz\ScheduleInterleavedZeroBubble._calculate_single_rank_operations..forward_stage_indexcrrrrrr<r=rrz]ScheduleInterleavedZeroBubble._calculate_single_rank_operations..backward_stage_indexT)rr) r:rrrrrrrrrr<rr=rfs:     z?ScheduleInterleavedZeroBubble._calculate_single_rank_operationscCsr|j}dd}t}i}i}i}d}t|jD]} g|| <d|| <d|| <q d} t} t|jD]q} || } | t|| kr@q1d} || | dur|| | } | dusVJ| \}}}||||||s|| || | |durz| |||f|| d7<q1|| d|| d7<q1|| d7<|| dq1|| | rnq'|dkrt d|||S)NcSsh|tjkr|dkr|d||f|vrdSdS|tjkr2||dkr)|tj|f|vS|d||f|vSdS)NrrTF)r"r6r9)rr microbatchnum_stages_globalseen_opsr<r<r= need_bubbles  zJScheduleInterleavedZeroBubble._add_bubbles_to_actions..need_bubblerTFrz?Non zero bubbles added: total_bubbles_added=%s bubbles_added=%s) rXrryrPrVrr updaterrg)r:rr\rrresult next_pointer bubbles_addedtotal_bubbles_addedr should_stop temp_seen_ops timestamp temp_actionrHrrr<r<r=rsZ       "z5ScheduleInterleavedZeroBubble._add_bubbles_to_actionsr)rCrDrErr|rrUr rrrrrLr rrrrGrrrr<r<rr=r %s06@r cr)r!a The Zero Bubble schedule (ZBV variant). See https://arxiv.org/pdf/2401.10241 Section 6 for details. This schedules requires exactly two stages per rank. This schedule will perform one forward and one backward on inputs for the microbatches in steady state and supports multiple stages per rank. Uses backward with respect to weights to fill in the pipeline bubble. This ZB-V schedule would have the "zero bubble" property only if time forward == time backward input == time backward weights. In practice, this is not likely true for real models so alternatively a greedy scheduler could be implemented for unequal/unbalanced time. NTrrrr.rrrc s|dj|_tj|||||||dt|j|jdd|_|jD]}|j|_q t||_ |j dkr:t d|j d|dj |_ |dj |_ i|_t|jD] } || } | |j| <qNdS)Nrrv)styler#z0ZBV requires exactly 2 stages per rank, but got r)rbrPrrrrrdrarVrrrcrrrXryrrrr<r=rs:          zScheduleZBVZeroBubble.__init__rZcs>tdjdj}ddt|D}d\}}}}dj|d}|} jd|} t|D]} |t| t|d|d7}q0|} t| D]} |t| t|d|d7}|t| t|d|d7}qGj|} t| D](} |t| t|d|d7}|t| t|d|t| t |d|d7}qo||ks||kr||kr|t| t|d|d7}|t| t|d|t| t |d|d7}|t| t|d|d7}|t| t|d|t| t |d|d7}||ks||ks||}}|}t|D]} |t| t|d|d7}|t| t|d|d7}qj|}t|D]} |t| t|d|d7}|t| t |d|d7}q,||krd|t| t |d|d7}||ksQ||kr||t| t |d|d7}||ksi||kr||ksJ||kr||ksJfdd|D}|S)Nr#rcSrrKr<rr<r<r=rb) rzKScheduleZBVZeroBubble._calculate_single_rank_operations..)rrrr)rIrJcs2g|]}|dur|jdur|jjkr|ndqSrK)rJr)r[rArr<r=rb s  ) rjrPrryrrrGr,r-r.)r:rn_microrf0_cntf1_cntb0_cntb1_cnt warmup_n1stage_id_chunk0stage_id_chunk1r warmup_n2 warmup_n3w0_cntw1_cnt cooldown_n1 cooldown_n2r<rr=r% s                                    z7ScheduleZBVZeroBubble._calculate_single_rank_operationsrrr<r<rr=r!s.,r! schedule_namec Cs`ttttttttd}dd|D}| }||vr*t d|dt ||||S)z Maps a schedule name (case insensitive) to its corresponding class object. Args: schedule_name (str): The name of the schedule. )1F1BInterleaved1F1BGPipe LoopedBFSInterleavedZeroBubblerr ZBVZeroBubblecSsi|]}||qSr<)lower)r[kr<r<r=r> r{z&get_schedule_class..zUnknown schedule name 'z'. The valid options are ) rrrrr rrr!rrrr|)r schedule_maplowercase_keyslowercase_schedule_namer<r<r=r s   rc sfddtDddtDddDdtdttffdd }dttd tffd d }rd }tD]-}t|d krKq@|d }||rh|dur^||||d d}q@||dq@tddD] }t|d kr|=qttD]6}t|d krq|ddurq|d }||r|dur||d<|||d qtddD] }t|d krЈ|=q|stdt D]}td|d|d qt ds:S)aThis function dry-run simulates the actions in the schedule from the perspective of all ranks, and flags any deadlocks caused by missing or misordered communications. It also simulates any bubbles in time where a rank can not execute any action due to waiting for unmet dependencies. The total number of simulator steps can be used as a metric for unit tests involving IR optimization passes as reordering and merging of IR can reduce the number of simulated steps. The simulation is not high-fidelity and does not model overlapping of compute and communication, or cuda streams. Future work may be to enhance this and model the compute time, comms overlap, and even memory. cs i|] }|dd|DqS)cSsg|]}|dur|qSrKr<)r[r#r<r<r=rb rgz6_simulate_comms_compute...r<r=)rXr<r=r> sz+_simulate_comms_compute..cSr<r<r<r=r<r<r=r> scSr?r<r@r=r<r<r=r> rArrAcs,|||dur||dSdSrK)rr )rrA)_prev_ops_rank _scheduler<r=add_to_schedule sz0_simulate_comms_compute..add_to_schedulerZcs|durdS|j}|}|jtkr7|jdkrdSt|jt|j|vr'dSt|jdt|j|vr5dSdS|jttfvrq|jdkrGdSt|jt|j|vrSdSt|jdt|j|vradSt|jdt|j|vrodSdS|jt krxdS|jt krt|jt|j}||vS|jtkr|d}t|t |j}||vS|jt krt|jt|j}t|jt|j}||vp||vS|jtkr|d}t|t |j}||vSt d|)NTrrFzUnsupported action type ) rHrIr,rGr2rJr7r9r4r8r1r3r)rArCprev_ops expected_fpeer_stage_idx expected_send expected_b expected_bw)rrr;r<r=rK s`          z3_simulate_comms_compute.._ready_to_scheduleFrNT)reverserzWIP comms schedule: rLz next action= zSchedule is not progressing) r{rUr rGrrVrr rrr) rXr;rrrKrNrrAr`r<)rrrrXr;r=r s` :       -rc Csg}t|D]+}t||D]"\}}|durq|t||jtttfvr&dndd|||ddqqddl}t |d}| d |i|WddS1sOwYdS) a This function dumps a schedule IR into a chrometrace format so it can be visualized. It is currently very basic and only serves as a graphical alternative to dumping the schedule IR as text. As future work we may extend this to include more accurate heuristics for durations, or let users input durations, add 'flow events' to let the UI show the connection between sends and recvs, and model cuda streams for comm/compute as separate streams on the chrometrace view. N computation communicationXr)rcatphpidtidtsdurrrl traceEvents) r{rsrrLrIr,r5r.jsonrodump)schedulerteventsrtimesteprAr.fr<r<r=_dump_chrometrace; s,  "r4rK)r$)rF)[rwrpr}loggingreabcrr collectionsrrenumrtypingrrr r r r rtorch.distributed distributedr torch._dynamor torch.distributed.fsdprrtorch.nn.modules.lossrtorch.profilerr_utilsrrrrrrrr__all__ getLoggerrCrr"r6r7r8r/r0r1r2r3r4r9r,r-r.r5compilerRrGrrUr|rLrrP2POprrrrrrr7r:rOr\rrrrrr r!rrr4r<r<r<r=s          6% >7   S3[  F  #  i  LUHI  tE4