def main(args=None): # If ``args`` not passed, defaults to ``sys.argv[:1]`` args = parse_args(args) min_nodes, max_nodes = parse_min_max_nnodes(args.nnodes) assert 0 < min_nodes <= max_nodes assert args.max_restarts >= 0 elastic_agent = None if args.standalone: etcd_server = EtcdServer() etcd_server.start() args.rdzv_backend = "etcd" args.rdzv_endpoint = etcd_server.get_endpoint() args.rdzv_id = str(uuid.uuid4()) log.info(f"\n**************************************\n" f"Rendezvous info:\n" f"--rdzv_backend={args.rdzv_backend} " f"--rdzv_endpoint={args.rdzv_endpoint} " f"--rdzv_id={args.rdzv_id}\n" f"**************************************\n") nproc_per_node = determine_local_world_size(args.nproc_per_node) if "OMP_NUM_THREADS" not in os.environ and nproc_per_node > 1: omp_num_threads = 1 print( f"*****************************************\n" f"Setting OMP_NUM_THREADS environment variable for each process to be " f"{omp_num_threads} in default, to avoid your system being overloaded, " f"please further tune the variable for optimal performance in " f"your application as needed. \n" f"*****************************************") # This env variable will be passed down to the subprocesses os.environ["OMP_NUM_THREADS"] = str(omp_num_threads) with_python = not args.no_python cmd = [] if with_python: cmd = [sys.executable, "-u"] if args.module: cmd.append("-m") else: if args.module: raise ValueError("Don't use both the '--no_python' flag" " and the '--module' flag at the same time.") cmd.append(args.training_script) cmd.extend(args.training_script_args) rdzv_parameters = RendezvousParameters( backend=args.rdzv_backend, endpoint=args.rdzv_endpoint, run_id=args.rdzv_id, min_nodes=min_nodes, max_nodes=max_nodes, **_parse_rendezvous_config(args.rdzv_conf), ) rdzv_handler = rdzv_registry.get_rendezvous_handler(rdzv_parameters) try: spec = WorkerSpec( role=args.role, local_world_size=nproc_per_node, entrypoint=cmd[0], args=(*cmd[1:], ), rdzv_handler=rdzv_handler, max_restarts=args.max_restarts, monitor_interval=args.monitor_interval, redirects=Std.from_str(args.redirects), tee=Std.from_str(args.tee), ) metrics.initialize_metrics() elastic_agent = LocalElasticAgent(spec=spec, start_method=args.start_method, log_dir=args.log_dir) run_result = elastic_agent.run(spec.role) events.record( elastic_agent.get_agent_status_event(WorkerState.SUCCEEDED)) if run_result.is_failed(): # ChildFailedError is treated specially by @record # if the error files for the failed children exist # @record will copy the first error (root cause) # to the error file of the launcher process raise ChildFailedError( name=args.training_script, failures=run_result.failures, ) except ChildFailedError: raise except Exception: if elastic_agent: events.record( elastic_agent.get_agent_status_event(WorkerState.FAILED)) else: events.record(_construct_event(args)) raise finally: rdzv_handler.shutdown() if args.standalone: etcd_server.stop()
def launch_agent( config: LaunchConfig, entrypoint: Union[Callable, str, None], args: List[Any], ) -> Dict[int, Any]: if not config.run_id: run_id = str(uuid.uuid4().int) logger.warning(f"config has no run_id, generate a new one: {run_id}") config.run_id = run_id entrypoint_name = _get_entrypoint_name(entrypoint, args) logger.info( f"Starting elastic_operator with launch configs:\n" f" entrypoint : {entrypoint_name}\n" f" min_nodes : {config.min_nodes}\n" f" max_nodes : {config.max_nodes}\n" f" nproc_per_node : {config.nproc_per_node}\n" f" run_id : {config.run_id}\n" f" rdzv_backend : {config.rdzv_backend}\n" f" rdzv_endpoint : {config.rdzv_endpoint}\n" f" rdzv_configs : {config.rdzv_configs}\n" f" max_restarts : {config.max_restarts}\n" f" monitor_interval : {config.monitor_interval}\n" f" log_dir : {config.log_dir}\n" f" metrics_cfg : {config.metrics_cfg}\n" ) rdzv_parameters = RendezvousParameters( backend=config.rdzv_backend, endpoint=config.rdzv_endpoint, run_id=config.run_id, min_nodes=config.min_nodes, max_nodes=config.max_nodes, **config.rdzv_configs, ) agent = None rdzv_handler = rdzv_registry.get_rendezvous_handler(rdzv_parameters) try: spec = WorkerSpec( role=config.role, local_world_size=config.nproc_per_node, entrypoint=entrypoint, args=tuple(args), rdzv_handler=rdzv_handler, max_restarts=config.max_restarts, monitor_interval=config.monitor_interval, redirects=config.redirects, tee=config.tee, ) cfg = metrics.MetricsConfig(config.metrics_cfg) if config.metrics_cfg else None metrics.initialize_metrics(cfg) agent = LocalElasticAgent( spec=spec, start_method=config.start_method, log_dir=config.log_dir ) result = agent.run() events.record(agent.get_agent_status_event(WorkerState.SUCCEEDED)) if result.is_failed(): # ChildFailedError is treated specially by @record # if the error files for the failed children exist # @record will copy the first error (root cause) # to the error file of the launcher process. raise ChildFailedError( name=entrypoint_name, failures=result.failures, ) else: return result.return_values except ChildFailedError: raise except Exception: if agent: events.record(agent.get_agent_status_event(WorkerState.FAILED)) else: events.record(_construct_event(config)) raise finally: rdzv_handler.shutdown()