sgddlZddlZddlZddlZddlZddlZddlZddlZddlm Z ddl m Z ddl m Z mZddlmZddlmZmZmZmZmZmZmZddlmcmcmZddlmcmcmcm Z!ddl"m#Z#m$Z$m%Z%ddl&m'Z'm(Z(dd l)m*Z*m+Z+dd lm,Z,dd l-m.Z.gd Z/d Z0dZ1e.e2Z3e GddZ4GddZ5Gdde6eZ7GddZ8GddZ9e GddZ:de6fdZ;GddejxZ=Gdd e=Z>y)!N) defaultdict)contextmanager) dataclassfield)Enum)AnyCallableDictListOptionalTupleUnion)Event EventSourcerecord)prof put_metric)ProcessFailureSignalException)RendezvousGracefulExitError) get_logger) WorkerSpecWorker WorkerState WorkerGroup RunResult ElasticAgentSimpleElasticAgentz!torchelastic/agent/terminal_statedefaultceZdZUdZeed<eed<ejed<dZ e e ed<dZ e e edfed<dZeed <d Zeed <d Zeed <dZe eed<dZe eed<dZe eed<dZdZy)raiBlueprint information about a particular type of worker. For a given role, there must only exist a single worker spec. Worker spec is expected to be homogeneous across all nodes (machine), that is each node runs the same number of workers for a particular spec. Args: role: user-defined role for the workers with this spec local_world_size: number local workers to run fn: (deprecated use entrypoint instead) entrypoint: worker function or command args: arguments to pass to ``entrypoint`` rdzv_handler: handles rdzv for this set of workers max_restarts: number of max retries for the workers monitor_interval: monitor status of workers every ``n`` seconds master_port: fixed port to run the c10d store on rank 0 if not specified then will chose a random free port master_addr: fixed master_addr to run the c10d store on rank 0 if not specified then will chose hostname on agent rank 0 redirects: redirect std streams to a file, selectively redirect for a particular local rank by passing a map tee: tees the specified std stream(s) to console + file, selectively tee for a particular local rank by passing a map, takes precedence over ``redirects`` settings. rolelocal_world_size rdzv_handlerNfn entrypointargs max_restartsg?monitor_interval master_port master_addr local_addrc|jdkDsJ|jdkDsJ|jr,tjdt |j|_|j sJy)NrzJWorkerSpec.fn will be deprecated, please use WorkerSpec.entrypoint instead)category)r"r*r$warningswarnDeprecationWarningr%selfs ]/var/www/html/venv/lib/python3.12/site-packages/torch/distributed/elastic/agent/server/api.py __post_init__zWorkerSpec.__post_init__Zs^$$q((($$q((( 77 MM<+  #ggDOct|jtr)tjj |jS|jJ|jj S)zGet the entry point name. If the entrypoint is a function (e.g. ``Callable``) returns its ``__qualname__`` else if the entrypoint is a binary (e.g. ``str``), returns the binary name. ) isinstancer%strospathbasename __qualname__r3s r5get_entrypoint_namezWorkerSpec.get_entrypoint_namegsK doos +77##DOO4 4??. ..??// /r7)__name__ __module__r>__doc__r:__annotations__intrdzvRendezvousHandlerr$r r r%rr'r r)r*floatr+r,r-r6r?r&r7r5rr/s8 I(((!B!-1JhT)*1D%L#!e!!%K#%!%K#% $J $  0r7rc JeZdZdZgdZ d dededededef dZd Zd Zy ) rahA worker instance. Contrast this with ``WorkerSpec`` that represents the specifications of a worker. A ``Worker`` is created from a ``WorkerSpec``. A ``Worker`` is to a ``WorkerSpec`` as an object is to a class. The ``id`` of the worker is interpreted by the specific implementation of ``ElasticAgent``. For a local agent, it could be the ``pid (int)`` of the worker, for a remote agent it could be encoded as ``host:port (string)``. Args: id (Any): uniquely identifies a worker (interpreted by the agent) local_rank (int): local rank of the worker global_rank (int): global rank of the worker role_rank (int): rank of the worker across all workers that have the same role world_size (int): number of workers (globally) role_world_size (int): number of workers that have the same role id local_rank global_rank role_rank world_sizerole_world_sizerKrLrMrNrOcXd|_||_||_||_||_||_yNrI)r4rKrLrMrNrOs r5__init__zWorker.__init__s: * !, ( *%4r7c d|jd|jd|jd|jd|j S)Nz local_rank=z ,global_rank=z ,role_rank=z ,world_size=z,role_world_size=rKrLrMrNrOr3s r5__str__zWorker.__str__sJ$//*-8H8H7I$..)doo5F 4 45 7 r7ct|SrQ)r:r3s r5__repr__zWorker.__repr__s 4yr7N)rXrXrX) r@rAr>rB __slots__rDrRrUrWr&r7r5rrts\(I! 444 4  4  4B r7rcFeZdZdZdZdZdZdZdZdZ dZ e d dd e fd Z y ) ra A state of the ``WorkerGroup``. Workers in a worker group change state as a unit. If a single worker in a worker group fails the entire set is considered failed:: UNKNOWN - agent lost track of worker group state, unrecoverable INIT - worker group object created not yet started HEALTHY - workers running and healthy UNHEALTHY - workers running and unhealthy STOPPED - workers stopped (interrupted) by the agent SUCCEEDED - workers finished running (exit 0) FAILED - workers failed to successfully finish (exit !0) A worker group starts from an initial ``INIT`` state, then progresses to ``HEALTHY`` or ``UNHEALTHY`` states, and finally reaches a terminal ``SUCCEEDED`` or ``FAILED`` state. Worker groups can be interrupted and temporarily put into ``STOPPED`` state by the agent. Workers in ``STOPPED`` state are scheduled to be restarted in the near future by the agent. Some examples of workers being put into ``STOPPED`` state are: 1. Worker group failure|unhealthy observed 2. Membership change detected When actions (start, stop, rdzv, retry, etc) on worker group fails and results in the action being partially applied to the worker group the state will be ``UNKNOWN``. Typically this happens on uncaught/unhandled exceptions during state change events on the agent. The agent is not expected to recover worker groups in ``UNKNOWN`` state and is better off self terminating and allowing the job manager to retry the node. UNKNOWNINITHEALTHY UNHEALTHYSTOPPED SUCCEEDEDFAILEDstatereturncF|tjtjhvS)zReturn the state of the Worker. Returns: True if the worker state represents workers still running (e.g. that the process exists but not necessarily healthy). )rr]r^)rbs r5 is_runningzWorkerState.is_runnings ,,k.C.CDDDr7N)r@rAr>rBr[r\r]r^r_r`ra staticmethodboolrer&r7r5rrsQ DG DGIGI FE-EDEEr7rc$eZdZdZgdZdefdZy)raA set of ``Worker`` instances. The class defines a set of ``Worker`` instances for the given ``WorkerSpec`` managed by ``ElasticAgent``. Whether the worker group contains cross instance workers or not depends on the implementation of the agent. )specworkersstore group_rankgroup_world_sizerbr,r+ric||_t|jjDcgc]}t|c}|_d|_d|_d|_d|_d|_ tj|_ ycc}w)N)rK) riranger"rrjrkrlrmr,r+rr\rb)r4riis r5rRzWorkerGroup.__init__sj 6;DIIrBrYrrRr&r7r5rrs I &Z &r7rc eZdZdZgdZdededefdZdefdZ e d efd Z e defd Z e d e dedeeeffd Zy)_RoleInstanceInfoa The class is used by the agent to exchange the information with other agents. The information is used to determine the rank of the workers that agent manages in heterogeneous environments, where different agents can have different number of workers. r!rankr"r!rtr"c.||_||_||_y)zInitialize the agent class instance. Args: role (str): user-defined role for the workers with this spec rank (int): the rank of the agent local_world_size (int): number of local workers to run Nrs)r4r!rtr"s r5rRz_RoleInstanceInfo.__init__s  0r7rcc|j|j|jd}tj|j dS)NrsUTF-8encoding)r!rtr"jsondumpsencode)r4 dict_datas r5 serializez_RoleInstanceInfo.serialize*s?IIII $ 5 5 zz)$++W+==r7datacxtj|jd}t|d|d|dS)Nrwrxr!rtr")rzloadsdecoderr)rr}s r5 deserializez_RoleInstanceInfo.deserialize2s?JJt{{G{<=  f y0)rBrYr:rDrRbytesr~rfrrr r rr&r7r5rrrrs5I 1S 1 1s 1>5> %  s$$$c$eCHo$$r7rrcteZdZUdZeed<eeZe e e fed<eeZ e e e fed<defdZy) raReturn results of the worker executions. Run results follow an "all-or-nothing" policy where the run is successful if and only if ALL local workers managed by this agent complete successfully. If the result is successful (e.g. ``is_failed() = False``) then the ``return_values`` field contains the outputs (return values) of the workers managed by THIS agent mapped by their GLOBAL ranks. That is ``result.return_values[0]`` is the return value of global rank 0. .. note:: ``return_values`` are only meaningful for when the worker entrypoint is a function. Workers specified as a binary entrypoint do not canonically have a return value and the ``return_values`` field is meaningless and may be empty. If ``is_failed()`` returns ``True`` then the ``failures`` field contains the failure information, again, mapped by the GLOBAL rank of the worker that failed. The keys in ``return_values`` and ``failures`` are mutually exclusive, that is, a worker's final state can only be one of: succeeded, failed. Workers intentionally terminated by the agent according to the agent's restart policy, are not represented in either ``return_values`` nor ``failures``. rb)default_factory return_valuesfailuresrcc<|jtjk(SrQ)rbrrar3s r5 is_failedzRunResult.is_failedkszz[////r7N)r@rAr>rBrrCrdictrr rDrrrrgrr&r7r5rrMsK0 $)$$?M4S>?*/*EHd3&'E040r7rrccPtjtjSrQ)socketgetfqdn gethostnamer&r7r5_get_fq_hostnameros >>&,,. //r7cteZdZdZej efdedefdZ ej efdede fdZ y)ra0An agent process responsible for managing one or more worker processes. The worker processes are assumed to be regular distributed PyTorch scripts. When the worker process is created by the agent, the agent provides the necessary information for the worker processes to properly initialize a torch process group. The exact deployment topology and ratio of agent-to-worker is dependent on the specific implementation of the agent and the user's job placement preferences. For instance, to run a distributed training job on GPU with 8 trainers (one per GPU) one can: 1. Use 8 x single GPU instances, place an agent per instance, managing 1 worker per agent. 2. Use 4 x double GPU instances, place an agent per instance, managing 2 workers per agent. 3. Use 2 x quad GPU instances, place an agent per instance, managing 4 workers per agent. 4. Use 1 x 8 GPU instance, place an agent per instance, managing 8 workers per agent. Usage :: group_result = agent.run() if group_result.is_failed(): # workers failed failure = group_result.failures[0] logger.exception("worker 0 failed with exit code : %s", failure.exit_code) else: return group_result.return_values[0] # return rank 0's results r!rcct)alRun the agent. Supports retrying the worker group on failures up to ``max_restarts``. Returns: The result of the execution, containing the return values or failure details for each worker mapped by the worker's global rank. Raises: Exception - any other failures NOT related to worker process NotImplementedErrorr4r!s r5runzElasticAgent.runs "!r7ct)a'Return the ``WorkerGroup`` for the given ``role``. Note that the worker group is a mutable object and hence in a multi-threaded/process environment it may change state. Implementors are encouraged (but not required) to return a defensive read-only copy. rrs r5get_worker_groupzElasticAgent.get_worker_group "!r7N) r@rAr>rBabcabstractmethod DEFAULT_ROLEr:rrrrr&r7r5rrss^ D * " "y " " +7"S"K""r7rc eZdZdZd+dedefdZefdede fdZ e jde de eeffd Ze j d,de d edd fd Ze jde defdZe jej,d fdej.d edd fdZede dd fdZededededeefdZede dd fdZede dd fdZeefdedefdZ de!fdZ"de!fdZ#dedd fdZ$dededefdZ%e&defdZ' d-ded e(de)ed!e)ed"e)ede!f d#Z*d$efd%Z+d&Z,d,d'efd(Z-efdedefd)Z.d*Z/y ).rzAn ``ElasticAgent`` that manages one particular type of worker role. An ``ElasticAgent`` that manages workers (``WorkerGroup``) for a single ``WorkerSpec`` such as one particular type of worker role. riexit_barrier_timeoutct||_|jjj|_d|_||_d|_yNr)r _worker_grouprir)_remaining_restarts_store_exit_barrier_timeout_total_execution_time)r4rirs r5rRzSimpleElasticAgent.__init__s@(.#'#5#5#:#:#G#G  %9"%&"r7r!rcc|jSrQ)rrs r5rz#SimpleElasticAgent.get_worker_groups!!!r7 worker_groupct)zStart ``worker_group.spec.local_world_size`` number of workers. This is according to worker spec for the worker group . Returns a map of ``local_rank`` to worker ``id``. rr4rs r5_start_workersz!SimpleElasticAgent._start_workerss "!r7F is_restartNct)zStop all workers in the given worker group. Implementors must deal with workers in all states defined by ``WorkerState``. That is, it must gracefully handle stopping non-existent workers, unhealthy (stuck) workers, etc. r)r4rrs r5 _stop_workersz SimpleElasticAgent._stop_workerss "!r7ct)z~Check on the workers for the ``worker_group``. This function also returns the new state of the worker group. rrs r5_monitor_workersz#SimpleElasticAgent._monitor_workerss "!r7 death_sigct)zClean up any resources that were allocated during the agent's work. Args: death_sig: Signal to send to the child process, SIGTERM is default r)r4rrs r5 _shutdownzSimpleElasticAgent._shutdownrr7c|j}|jd5|jj}dddj}|j }|j }|jxs|jj}|jxs|jj}||_ |jd5|j||||} ddd |_ ||_||_ ||_||_||_ |j|j z } t"j%d|j&| ||||| D cgc]} | j(c} | D cgc]} | j*c} | D cgc]} | j,c} | D cgc]} | j.c} | D cgc]} | j c} d y#1swYxYw#1swYxYwcc} wcc} wcc} wcc} wcc} w)zRun rendezvous for the workers specified by the worker spec. Assigns workers a new global rank and world size. Updates the rendezvous store for the worker group. RENDEZVOUSNASSIGN_WORKER_RANKSa[%(role)s] Rendezvous complete for workers. Result: restart_count=%(restart_count)s master_addr=%(master_addr)s master_port=%(master_port)s group_rank=%(group_rank)s group_world_size=%(group_world_size)s local_ranks=%(local_ranks)s role_ranks=%(role_ranks)s global_ranks=%(global_ranks)s role_world_sizes=%(role_world_sizes)s global_world_sizes=%(global_world_sizes)s ) r! restart_countr,r+rlrm local_ranks role_ranks global_ranksrole_world_sizesglobal_world_sizes)rirecord_durationr#next_rendezvousrkrtrNr,bootstrap_store_infor+r_assign_worker_ranksrjrlrmr)rloggerinfor!rKrMrLrO) r4rri rdzv_inforkrlrmr,r+rjrworkers r5 _rendezvouszSimpleElasticAgent._rendezvouss    ! !, / <))99;I <^^ $//&&T)*H*H*T*T &&T)*H*H*T*T   ! !"7 8 //z#3TG  ' " ", (8 %#.  #.  ))D,D,DD   < !.**($4@GHf 1 1H>EFFv//FBI J!3!3 JJQ$RV%;%;$RGN&OVv'8'8&O  5 < <  @ IF J$R&Os5G G G#"G(;G-G2-G7 GG rlrmc 8d}d}t|j||j}|j|||j |dk(rP|j t |Dcgc]}d| c}} | D cgc]} tj| } } td} d} | D]2}| |jxx|jz cc<| |jz } 4d}td}g}g}t| D]\}}|j|||jtj|| ||j| |jg||jz }||jxx|jz cc<|j||tj|j||\}}}}g}t |jD])}t!|||z||z||}|j|+|Scc}wcc} w)aoDetermine proper ranks for worker processes. The rank assignment is done according to the following algorithm: 1. Each agent writes its configuration(group_rank, group_world_size , num_workers) to the common store. 2. The rank 0 agent reads all the role_info from the store and determines each agents worker ranks. 3. Determine the global rank: the global rank of the workers is computed by cumulative sum of the local_world_size for all workers in front of it. For efficiency reasons each worker is assigned a base global rank such that it's workers are in the range [base_global_rank, base_global_rank + local_world_size). 4. Determine the role rank: The role rank is determined using the algorithms in the point 3 with the exception that the ranks are calculated with respect to the role name. 5. The rank 0 agent writes the assigned ranks to the store. 6. Each agent reads the assigned ranks from the store. Time complexity: each worker O(1), rank0 O(n), overall O(n) ztorchelastic/role_info/ztorchelastic/assigned_ranks/rcyrr&r&r7r5z9SimpleElasticAgent._assign_worker_ranks..Vr7cyrr&r&r7r5rz9SimpleElasticAgent._assign_worker_ranks..]rr7rT)rrr!r"setr~ multi_getrorrrappendrzr{ multi_setrgetr)r4rkrlrmriROLE_INFO_PREFIXASSIGNED_RANKS_PREFIXagent_role_inforprole_infos_bytes info_bytes role_infos role_sizes global_sizerbase_global_rankrkeysvaluesglobal_world_sizebase_role_rankrOrjrKrs r5rz'SimpleElasticAgent._assign_worker_ranks*sT45 >+ IIz4#8#8   %&zl3_5N5N5PQ ?$8=>N8OP1*1#.P   #3"--j9J %Y/JK' : 9>>*i.H.HH*y999  : ! $Y/JDF )* 5 I 9 45aS9: JJ,'&y~~6&y~~6  !I$>$>> 9>>*i.H.HH* I OOD& ) JJuyy$9#::,!GH I      5 56 #J%,z9(:5, / F NN6 " #mQs ( H?HcR|jj}tjd||j |tjd||j |}|j D]\}}|j|}||_tj|_ y)aStart a fresh set of workers for the worker_group. Essentially, a rendezvous followed by a ``start_workers``. The caller should first call ``_stop_workers()`` to stop running workers prior to calling this method. Optimistically sets the state of the worker group that just started as ``HEALTHY`` and delegates the actual monitoring of state to ``_monitor_workers()`` method z [%s] Rendezvous'ing worker groupz[%s] Starting worker groupN) rir!rrrritemsrjrJrr]rb)r4rr! worker_idsrKw_idrs r5_initialize_workersz&SimpleElasticAgent._initialize_workerss  %% 6= & 0$7((6 * 0 0 2  J!))*5FFI )00 r7c|jj}tjd||j |dt j |_|j|y)zCRestart (stops, rendezvous, starts) all local workers in the group.z[%s] Stopping worker groupT)rN) rir!rrrrr_rbr)r4rr!s r5_restart_workersz#SimpleElasticAgent._restart_workerssS  %% 0$7  >  NNLahh W NN188 $"O    ?# ),T^^-= -J)KD &# ),T^^-= -J)KD &s6AB)) D2C E D>DDE9Fcj|jdtjtjS)Nra)rbsource raw_error)_construct_eventrAGENT traceback format_excr3s r5get_event_failedz#SimpleElasticAgent.get_event_faileds2$$$$**,%  r7cD|jdtjS)Nr`)rbr)rrrr3s r5get_event_succeededz&SimpleElasticAgent.get_event_succeededs&$$$$%  r7rc D|jjD]}|jj|j}|j ||}|rt j|jnd}t|j|tj||yrQ) rrjrrrL_get_worker_staterzr{error_file_datarrrWORKER)r4rrfailurerbrs r5rz(SimpleElasticAgent._record_worker_eventss((00 XFoo))&*<*<=G//?E?F 7#:#:;DI 4(( 0B0BFIV W  Xr7rc:|jj|j}|jtj tj hvr|sy|s|j|jvr|jjStd|j)N TERMINATEDzUnknown worker: ) rrrLrbrr^rarvalue ValueError)r4rrr s r5rz$SimpleElasticAgent._get_worker_states~//%%f&8&89 <V>V>X    !3!3400(( %%++{/?/?? ?? JJ' (..t/A/ABJ$$E',D   $ $':;T=U=U V $q)9)9);(<=q A --- F..  ""$!!;00+2D2DEE++a/KK5 00)),,1,))$*<*<=&&t'9'9:/:/A/AD&&,%%+---$0$B$B$D!!//:: $q(KK4)" ))$*<*<=v/ |6Bgr7cFtjd|jj|jt j } t j|j|jjt|jtjdt j |z y#t$r&}tjd|jd}~wt$r.tjdt j |z YywxYw)aS Define a barrier that keeps the agent process alive until all workers finish. Wait for ``exit_barrier_timeout`` seconds for all agents to finish executing their local workers (either successfully or not). This acts as a safety guard against user scripts that terminate at different times. zOLocal worker group finished (%s). Waiting %s seconds for other agents to finish)rkrN key_prefixbarrier_timeoutz2Done waiting for other agents. Elapsed: %s secondszGot termination signal: %sNz2Error waiting on exit barrier. Elapsed: %s seconds)rrrrbrr store_utilbarrierrrm_TERMINAL_STATE_SYNC_IDrrrr8 exception)r4startrs r5r6z SimpleElasticAgent._exit_barriers   <    $ $  & &       kk-->>2 $ : :   KKD e#   NN7 B     D e#  s A0B<< D !C&&7D D )i,)F)NNN)0r@rAr>rBrrGrRrr:rrrrr rDrrrgrrrsignalSIGTERMSignalsrrrr rrrrrrrrrrrrrr rrr)r(rr6r&r7r5rrs 'Z'u',8"S"K" ";"4S>"" rVs #%(DDD33::GG>UL> > H  A0A0 A0HGGT3E#t3El&&@6$6$r 00 0B0#0;"377;"|}}r7