o j9:j"? @sddlZddlZddlZddlmZddlmZmZddlm Z ddl Z ddl m m mmZddlmZmZddlmZmZddlmZddlmZdd lmZmZmZdd lm 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.Z/eGdddZ0GdddZ1dee2BdBde3e de2fddZ4de&de5e2dBe6dBffddZ7de0dee2BdBde3e de8e6e ffddZ9dS) N)Callable) dataclassfield)Any)get_default_numa_optionsjustknobs_check)eventsmetrics) WorkerSpec)create_healthcheck_server)_AliveCallbackProxyLocalElasticAgentTORCHELASTIC_HEALTH_CHECK_PORT)DefaultLogsSpecs LogsSpecsSignalException)ChildFailedError)RendezvousParameters)parse_rendezvous_endpoint) get_logger) NumaOptions) LaunchConfigelastic_launch launch_agentc@sbeZdZUdZeed<eed<eed<dZedBed<dZe ed<d Z e ed <dZ e ed <d Z e ed <e edZee efed<dZeed<dZeed<dZeed<dZe ed<dZe dBed<e edZee e fed<dZe dBed<dZe ed<dZedBed<dZe ed<dZee dBed <dZee dBed!<d"Z e!ed#<dZ"edBed$<d%d&Z#dS)'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_timeoutcCsd}|jdkr|j|jd<n d|jvr||jd<|jdur!t|_|jdur>tjr>tj|j kr>t |_t d|j|j durOttjdd|_ dS|j dkr\td|j dS) Nir'timeoutzUsing default numa options = %rTORCH_ELASTIC_SHUTDOWN_TIMEOUT30rz+shutdown_timeout must be non-negative, got )r(r&rrr3torchcuda is_available device_countrrloggerinfor8intosenvironget ValueError)selfdefault_timeoutrIe/home/nk/hobo-godmode/plappi-mvp/.venv/lib/python3.10/site-packages/torch/distributed/launcher/api.py __post_init__|s,         zLaunchConfig.__post_init__)$__name__ __module__ __qualname____doc__rB__annotations__rrrstrr!r"r$rdictr&rr(r*r+floatr-r.r/r0r2r3rr4r5listr6r7boolr8rKrIrIrIrJr.s4 4            rc@s2eZdZdZdedeeBdBfddZddZdS) 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") config entrypointNcCs||_||_dSN)_config _entrypoint)rGrVrWrIrIrJ__init__s zelastic_launch.__init__cGst|j|jt|SrX)rrYrZrT)rGargsrIrIrJ__call__szelastic_launch.__call__) rLrMrNrOrrrQr[r]rIrIrIrJrs  rrWr\returncCs@t|tr|jSt|tr|tjkrtdd|DdS|SdS)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-NrI).0argrIrIrJ sz'_get_entrypoint_name..r) isinstancerrLrQsys executablenext)rWr\rIrIrJ_get_entrypoint_names   rgrdzv_parameterscCsX|jdkrdS|j}|}|stdt|dd\}}|dkr(td|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)backendendpointstriprFr)rhrl master_addr master_portrIrIrJ_get_addr_and_ports  rprVcCs|jsttj}td|||_t||}tdid|d|j d|j d|j d|jd|j d |j d |jd |jd |jd |jjd|jd|jd|jd|jd|jd|jtd"|j |j |j|j |j |jd|j}t|\}}|jtjd<d}d} tt} | durt dddrzt!} t"| t| dd}|#td| Wnt$ytjdddd}d} Ynwt%|j&|j |t'|t()||j|j|||j|j|j|j|j|j*d} t+| |j|j,|j-|j.|d } | dur| /| j0d} z^z0t12t13|j| 4}t56| 7|j|8r#t9||j:d!|j;WW| r0| j<=SSt9y9t>yLd} t56| ?|jt$y]t56| ?|jw| rg| j<=ww)#Nz3config has no run_id, generated a random run_id: %saRStarting elastic_operator with launch configs: entrypoint : %(entrypoint)s min_nodes : %(min_nodes)s max_nodes : %(max_nodes)s nproc_per_node : %(nproc_per_node)s run_id : %(run_id)s rdzv_backend : %(rdzv_backend)s rdzv_endpoint : %(rdzv_endpoint)s rdzv_configs : %(rdzv_configs)s max_restarts : %(max_restarts)s monitor_interval : %(monitor_interval)s log_dir : %(log_dir)s metrics_cfg : %(metrics_cfg)s event_log_handler : %(event_log_handler)s numa_options : %(numa_options)s signals_to_handle : %(signals_to_handle)s duplicate_stdout_filters : %(duplicate_stdout_filters)s duplicate_stderr_filters : %(duplicate_stderr_filters)s rWrrrrr$r"r&r*r+log_dirr/r2r3r4r5r6)rkrlrrrr0TORCHELASTIC_SIGNALS_TO_HANDLEzNai_infra/pytorch_distributed:torchelastic_enable_healthcheck_before_rendezvousF)default<)alive_callbackportr9z>Started early health check server on port %s before rendezvousz)Failed to start early health check serverT)exc_info)r!local_world_sizerWr\ rdzv_handlerr*r+rnror0r2r3r5r6r7)specrr-r.r8health_check_server)namefailuresrI)@rrQuuiduuid4rBr@warningrgrArrrr$r"r&r*r+r root_log_dirr/r2r3r4r5r6rr0rprCrDgetenvrrr r start Exceptionr r!tuple rdzv_registryget_rendezvous_handlerr7r r-r.r8 set_delegate_get_alive_timer initialize_metrics MetricsConfigrunrrecordget_event_succeeded is_failedrr} return_valuesryshutdownrget_event_failed)rVrWr\rentrypoint_namerhrnror{alive_callback_proxyhealthcheck_portrzagent shutdown_rdzvresultrIrIrJrs        (        r):rCrdr~collections.abcr dataclassesrrtypingrr<-torch.distributed.elastic.rendezvous.registry distributedelastic rendezvousregistryrtorch._utils_internalrrtorch.distributed.elasticrr *torch.distributed.elastic.agent.server.apir :torch.distributed.elastic.agent.server.health_check_serverr :torch.distributed.elastic.agent.server.local_elastic_agentr r r)torch.distributed.elastic.multiprocessingrrr0torch.distributed.elastic.multiprocessing.errorsr$torch.distributed.elastic.rendezvousr*torch.distributed.elastic.rendezvous.utilsr'torch.distributed.elastic.utils.loggingrtorch.numa.bindingr__all__rLr@rrrQrTrgrrBrprRrrIrIrIrJsN         l"'