^jA ddlZddlZddlZddlmZddlmZmZddlm Z ddl Z ddl m cm cmcmZddlmZmZddlmZmZddlmZddlmZmZdd lmZmZmZdd l m!Z!m"Z"m#Z#dd l$m%Z%dd l&m'Z'dd l(m)Z)ddl*m+Z+ddl,m-Z-gdZ.e+e/Z0eGddZ1GddZ2dee3zdzde4e de3fdZ5de'de6e3dze7dzffdZ8 dde1dee3zdzde4e dedzde9e7e ff dZ:y)N)Callable) dataclassfield)Any)get_default_numa_optionsjustknobs_check)eventsmetrics) WorkerSpec)create_healthcheck_serverHealthCheckServer)_AliveCallbackProxyLocalElasticAgentTORCHELASTIC_HEALTH_CHECK_PORT)DefaultLogsSpecs LogsSpecsSignalException)ChildFailedError)RendezvousParameters)parse_rendezvous_endpoint) get_logger) NumaOptions) LaunchConfigelastic_launch launch_agentceZdZUdZeed<eed<eed<dZedzed<dZe ed<d Z e ed <dZ e ed <d Z e ed <e eZee efed<dZeed<dZeed<dZeed<dZe ed<dZe dzed<e eZee e fed<dZe dzed<dZe ed<dZedzed<dZe ed<dZee dzed <dZee dzed!<d"Z e!ed#<dZ"edzed$<d%Z#y)&ra Creates a rendezvous config. Args: min_nodes: Minimum amount of nodes that the user function will be launched on. Elastic agent ensures that the user function start only when the min_nodes amount enters the rendezvous. max_nodes: Maximum amount of nodes that the user function will be launched on. nproc_per_node: On each node the elastic agent will launch this amount of workers that will execute user defined function. rdzv_backend: rdzv_backend to use in the rendezvous (zeus-adapter, etcd). rdzv_endpoint: The endpoint of the rdzv sync. storage. rdzv_configs: Key, value pair that specifies rendezvous specific configuration. rdzv_timeout: Legacy argument that specifies timeout for the rendezvous. It is going to be removed in future versions, see the note below. The default timeout is 900 seconds. run_id: The unique run id of the job (if not passed a unique one will be deduced from run environment - flow workflow id in flow - or auto generated). role: User defined role of the worker (defaults to "trainer"). max_restarts: The maximum amount of restarts that elastic agent will conduct on workers before failure. monitor_interval: The interval in seconds that is used by the elastic_agent as a period of monitoring workers. start_method: The method is used by the elastic agent to start the workers (spawn, fork, forkserver). metrics_cfg: configuration to initialize metrics. local_addr: address of the local node if any. If not set, a lookup on the local machine's FQDN will be performed. local_ranks_filter: ranks for which to show logs in console. If not set, show from all. event_log_handler: name of the event logging handler as registered in `elastic/events/handlers.py `_. duplicate_stdout_filters: If non-empty, duplicates stdout to a file containing only lines that match _any_ of the filter strings. duplicate_stderr_filters: If non-empty, duplicates stderr to a file containing only lines that match _any_ of the filter strings. virtual_local_rank: Enable virtual local rank mode for workers (defaults to False). When enabled, LOCAL_RANK is set to 0 for all workers and CUDA_VISIBLE_DEVICES is adjusted so each worker accesses its assigned GPU at device index 0. shutdown_timeout: Time in seconds to wait for graceful shutdown of workers before sending SIGKILL. Can also be set via TORCH_ELASTIC_SHUTDOWN_TIMEOUT environment variable. Defaults to 30 seconds. .. note:: `rdzv_timeout` is a legacy argument that will be removed in future. Set the timeout via `rdzv_configs['timeout']` min_nodes max_nodesnproc_per_nodeN logs_specsrun_id default_rolerole rdzv_endpointetcd rdzv_backend)default_factory rdzv_configs rdzv_timeout max_restartsg?monitor_intervalspawn start_methodlog_line_prefix_template metrics_cfg local_addrnullevent_log_handler numa_optionszSIGTERM,SIGINT,SIGHUP,SIGQUITsignals_to_handleduplicate_stdout_filtersduplicate_stderr_filtersFvirtual_local_rankshutdown_timeoutcd}|jdk7r|j|jd<nd|jvr||jd<|jt|_|jxt j jrZt j j|jk(r/t|_tjd|j|j/ttj j#dd|_ y|jdkrt%d|jy) Nir*timeoutzUsing default numa options = %rTORCH_ELASTIC_SHUTDOWN_TIMEOUT30rz+shutdown_timeout must be non-negative, got )r+r)r rr6torchcuda is_available device_countrrloggerinfor;intosenvironget ValueError)selfdefault_timeouts i/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/torch/distributed/launcher/api.py __post_init__zLaunchConfig.__post_init__}s    "+/+<+S>S=TU ')$__name__ __module__ __qualname____doc__rF__annotations__r rr"strr$r%r'rdictr)rr+r-r.floatr0r1r2r3r5r6rr7r8listr9r:boolr;rNrOrMrr/s12hNN#'J D 'FCD#M3L##(#>L$sCx.>L#L#!e!L#+/cDj/"'"=Kc3h=!Jd !#s#'+L+$+<s<15d3i$.515d3i$.5$$#'cDj'rOrc@eZdZdZ ddedeezdzdedzfdZdZ y) ra Launches a torchelastic agent on the container that invoked the entrypoint. 1. Pass the ``entrypoint`` arguments as non ``kwargs`` (e.g. no named parameters)/ ``entrypoint`` can be a function or a command. 2. The return value is a map of each worker's output mapped by their respective global rank. Usage :: def worker_fn(foo): # ... def main(): # entrypoint is a function. outputs = elastic_launch(LaunchConfig, worker_fn)(foo) # return rank 0's output return outputs[0] # entrypoint is a command and ``script.py`` is the python module. outputs = elastic_launch(LaunchConfig, "script.py")(args) outputs = elastic_launch(LaunchConfig, "python")("script.py") Nconfig entrypointhealth_check_serverc.||_||_||_yN)_config _entrypoint_health_check_server)rKr\r]r^s rM__init__zelastic_launch.__init__s  %$7!rOcnt|j|jt||jS)N)r^)rrarbrXrc)rKargss rM__call__zelastic_launch.__call__s/ LL    J $ 9 9   rOr`) rPrQrRrSrrrUr rdrgrZrOrMrrs@<9= 88sNT)8/5 8 rOrr]rfreturnct|tr |jSt|tr(|tj k(rt d|DdS|Sy)aRetrieve entrypoint name with the rule: 1. If entrypoint is a function, use ``entrypoint.__qualname__``. 2. If entrypoint is a string, check its value: 2.1 if entrypoint equals to ``sys.executable`` (like "python"), use the first element from ``args`` which does not start with hifen letter (for example, "-u" will be skipped). 2.2 otherwise, use ``entrypoint`` value. 3. Otherwise, return empty string. c32K|]}|ddk7s |yw)r-NrZ).0args rM z'_get_entrypoint_name..s>A# >s r!) isinstancerrPrUsys executablenext)r]rfs rM_get_entrypoint_namerssL*h'""" J $  '>>C C rOrdzv_parametersc|jdk7ry|j}|j}|s tdt |d\}}|dk(rtd|d||fS)Nstatic)NNzKEndpoint is missing in endpoint. Try to add --master-addr and --master-portr*) default_portzport is missing in endpoint: z. Try to specify --master-port)backendendpointstriprJr)rtry master_addr master_ports rM_get_addr_and_portr}s(*''H~~H  Y   9PRSKb+H:5S T    %%rOr\r^c |jsDttjj}t j d|||_t||}t jdid|d|jd|jd|jd|jd|jd |jd |jd |jd |j d |j"j$d|j&d|j(d|j*d|j,d|j.d|j0t3d"|j|j|j|j|j|j4d|j}t7|\}}|j,t8j:d<|nt9j<t>} | StAddrF tCtEt | d}|jGt jd| tK|jL|j|tO|tQjR||j|j |||j4|j(|j*|j.|j0|jT} tW| |j"|jX|jZ|j\| } |7|j^} ta| tDr| jc| jdd} tgjhtgjj|j&| jm}tojp| js|j(|jurtw||jx!|jz| r| j|jSS#tH$rt j ddd}YwxYw#tv$rt$r2d} tojp| j|j(tH$r0tojp| j|j(wxYw#| r| j|jwwxYw)#Nz3config has no run_id, generated a random run_id: %saRStarting elastic_operator with launch configs: entrypoint : %(entrypoint)s min_nodes : %(min_nodes)s max_nodes : %(max_nodes)s nproc_per_node : %(nproc_per_node)s run_id : %(run_id)s rdzv_backend : %(rdzv_backend)s rdzv_endpoint : %(rdzv_endpoint)s rdzv_configs : %(rdzv_configs)s max_restarts : %(max_restarts)s monitor_interval : %(monitor_interval)s log_dir : %(log_dir)s metrics_cfg : %(metrics_cfg)s event_log_handler : %(event_log_handler)s numa_options : %(numa_options)s signals_to_handle : %(signals_to_handle)s duplicate_stdout_filters : %(duplicate_stdout_filters)s duplicate_stderr_filters : %(duplicate_stderr_filters)s r]rrrr"r'r%r)r-r.log_dirr2r5r6r7r8r9)rxryr"rrr3TORCHELASTIC_SIGNALS_TO_HANDLEzNai_infra/pytorch_distributed:torchelastic_enable_healthcheck_before_rendezvousF)default<)alive_callbackportr=z>Started early health check server on port %s before rendezvousz)Failed to start early health check serverT)exc_info)r$local_world_sizer]rf rdzv_handlerr-r.r{r|r3r5r6r8r9r:)specr r0r1r;r^)namefailuresrZ)Br"rUuuiduuid4rFrDwarningrsrErrrr'r%r)r-r.r root_log_dirr2r5r6r7r8r9rr3r}rGrHgetenvrrr rstart Exceptionr r$tuple rdzv_registryget_rendezvous_handlerr:rr0r1r;rro set_delegate_get_alive_timer initialize_metrics MetricsConfigrunr recordget_event_succeeded is_failedrr return_valuesrshutdownrget_event_failed)r\r]rfr^r"entrypoint_namertr{r|healthcheck_portragentcb shutdown_rdzvresults rMrrs ==TZZ\%%&LfU *:tAPB"P;"P87P8;A;R66R99Sr`);rGrprcollections.abcr dataclassesrrtypingrr@-torch.distributed.elastic.rendezvous.registry distributedelastic rendezvousregistryrtorch._utils_internalrrtorch.distributed.elasticr r *torch.distributed.elastic.agent.server.apir :torch.distributed.elastic.agent.server.health_check_serverr r :torch.distributed.elastic.agent.server.local_elastic_agentrrr)torch.distributed.elastic.multiprocessingrrr0torch.distributed.elastic.multiprocessing.errorsr$torch.distributed.elastic.rendezvousr*torch.distributed.elastic.rendezvous.utilsr'torch.distributed.elastic.utils.loggingrtorch.numa.bindingr__all__rPrDrrrUrXrsrrFr}rVrrZrOrMrs6 $( EEK5A  NEP>* = H  ii iX+ + \X^d%:$s)PS(&)& 3:sTz !"&.59 b) b)3%b) s)b)+T1 b)  #s(^ b)rO