U yh*@snddlZddlZddlmZmZddlmZmZmZm Z m Z m Z m Z ddl mmmmZddlmZmZddlmZddlmZddlmZmZmZddlmZdd l m!Z!dd l"m#Z#dd l$m%Z%d d dgZ&e%e'Z(eGdd d Z)Gdd d Z*e ee+dfe ee+dddZ,e!e e e+e e-fdddZ.e)e ee+dfe eee-efdddZ/dS)N) dataclassfield)AnyCallableDictListOptionalTupleUnion)eventsmetrics) WorkerSpec)LocalElasticAgent)DefaultLogsSpecs LogsSpecsSignalException)ChildFailedError)RendezvousParameters)parse_rendezvous_endpoint) get_logger LaunchConfigelastic_launch launch_agentc@seZdZUdZeed<eed<eed<dZeeed<dZ e ed<d Z e ed <dZ e ed <d Z e ed <eedZee efed<dZeed<dZeed<dZeed<dZe ed<dZee ed<eedZee e fed<dZee ed<ddZdS)raM 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. ..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_idZ default_rolerole rdzv_endpointZetcd rdzv_backend)default_factory rdzv_configs rdzv_timeout max_restartsg?monitor_intervalspawn start_methodlog_line_prefix_template metrics_cfg local_addrcCsFd}|jdkr|j|jd<nd|jkr0||jd<|jdkrBt|_dS)Nir$timeout)r%r#rr)selfZdefault_timeoutr0P/var/www/html/venv/lib/python3.8/site-packages/torch/distributed/launcher/api.py __post_init__Us    zLaunchConfig.__post_init__)__name__ __module__ __qualname____doc__int__annotations__rrrrstrrr r!rdictr#rrr%r'r(floatr*r+r,r-r2r0r0r0r1rs$ %        c@s2eZdZdZeeeedfdddZddZ dS)ra Launches an 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") N)config entrypointcCs||_||_dSN)_config _entrypoint)r/r<r=r0r0r1__init__|szelastic_launch.__init__cGst|j|jt|Sr>)rr?r@list)r/argsr0r0r1__call__szelastic_launch.__call__) r3r4r5r6rr rr9rArDr0r0r0r1ras  )r=rCreturncCsFt|tr|jSt|tr>|tjkr8tdd|DdS|SndSdS)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. css|]}|ddkr|VqdS)r-Nr0).0argr0r0r1 s z'_get_entrypoint_name..rN) isinstancerr3r9sys executablenext)r=rCr0r0r1_get_entrypoint_names   rN)rdzv_parametersrEcCsX|jdkrdS|j}|}|s(tdt|dd\}}|dkrPtd|d||fS)NZstatic)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)backendendpointstrip ValueErrorr)rOrR master_addr master_portr0r0r1_get_addr_and_ports  rW)r<r=rCrEc Cs|js&ttj}td|||_t||}td||j |j |j |j|j |j |j|j|j|jj|jd tf|j |j |j|j |j |jd|j}t|\}}t|j|j |t|t||j|j|||jd }t||j|j|jd} d} zzLt!"t!#|j| $} t%&| '| (r2t)|| j*d| j+WWjSt)k rTYnNt,k r|d } t%&| -Yn&t.k rt%&| -YnXW5| r|j XdS) Nz3config has no run_id, generated a random run_id: %saStarting elastic_operator with launch configs: entrypoint : %(entrypoint)s min_nodes : %(min_nodes)s max_nodes : %(max_nodes)s nproc_per_node : %(nproc_per_node)s run_id : %(run_id)s rdzv_backend : %(rdzv_backend)s rdzv_endpoint : %(rdzv_endpoint)s rdzv_configs : %(rdzv_configs)s max_restarts : %(max_restarts)s monitor_interval : %(monitor_interval)s log_dir : %(log_dir)s metrics_cfg : %(metrics_cfg)s ) r=rrrrr!r r#r'r(Zlog_dirr,)rQrRrrrr-) rZlocal_world_sizer=rC rdzv_handlerr'r(rUrVr-)specrr*r+T)namefailuresF)/rr9uuiduuid4r7loggerwarningrNinforrrr!r r#r'r(rZ root_log_dirr,rr-rWr rtuple rdzv_registryZget_rendezvous_handlerrr*r+rXshutdownr Zinitialize_metricsZ MetricsConfigrunr recordZget_event_succeededZ is_failedrr[Z return_valuesrZget_event_failed Exception) r<r=rCrZentrypoint_namerOrUrVrYZagentZ shutdown_rdzvresultr0r0r1rs       )0rKr\ dataclassesrrtypingrrrrrr r Z-torch.distributed.elastic.rendezvous.registry distributedZelasticZ rendezvousregistryrbZtorch.distributed.elasticr r Z*torch.distributed.elastic.agent.server.apir Z:torch.distributed.elastic.agent.server.local_elastic_agentrZ)torch.distributed.elastic.multiprocessingrrrZ0torch.distributed.elastic.multiprocessing.errorsrZ$torch.distributed.elastic.rendezvousrZ*torch.distributed.elastic.rendezvous.utilsrZ'torch.distributed.elastic.utils.loggingr__all__r3r^rrr9rNr7rWrr0r0r0r1 s:$       C(