from __future__ import print_function, division, unicode_literals import errno import json import logging import os import os.path import re import signal import subprocess import sys import shutil import traceback from collections import defaultdict from copy import deepcopy, copy from datetime import datetime from distutils.spawn import find_executable from functools import partial, cmp_to_key from itertools import chain from tempfile import mkdtemp, NamedTemporaryFile from time import sleep, time from typing import Text, Optional, Any, Tuple import attr import psutil import six from clearml_agent.backend_api.services import queues as queues_api from clearml_agent.backend_api.services import tasks as tasks_api from clearml_agent.backend_api.services import workers as workers_api from pathlib2 import Path from pyhocon import ConfigTree, ConfigFactory from six.moves.urllib.parse import quote from clearml_agent.backend_config.defs import UptimeConf from clearml_agent.helper.check_update import start_check_update_daemon from clearml_agent.commands.base import resolve_names, ServiceCommandSection from clearml_agent.definitions import ( ENVIRONMENT_SDK_PARAMS, PROGRAM_NAME, DEFAULT_VENV_UPDATE_URL, ENV_DOCKER_IMAGE, ENV_TASK_EXECUTE_AS_USER, ENV_DOCKER_HOST_MOUNT, ENV_TASK_EXTRA_PYTHON_PATH, ENV_AGENT_GIT_USER, ENV_AGENT_GIT_PASS, ENV_WORKER_ID, ENV_DOCKER_SKIP_GPUS_FLAG, ) from clearml_agent.definitions import WORKING_REPOSITORY_DIR, PIP_EXTRA_INDICES from clearml_agent.errors import APIError, CommandFailedError, Sigterm from clearml_agent.helper.base import ( return_list, print_parameters, dump_yaml, warning, normalize_path, check_directory_path, select_for_platform, mkstemp as safe_mkstemp, print_table, safe_remove_file, is_windows_platform, rm_tree, is_conda, named_temporary_file, ExecutionInfo, HOCONEncoder, error, get_python_path, is_linux_platform, rm_file, add_python_path, safe_remove_tree, ) from clearml_agent.helper.console import ensure_text, print_text, decode_binary_lines from clearml_agent.helper.os.daemonize import daemonize_process from clearml_agent.helper.package.base import PackageManager from clearml_agent.helper.package.conda_api import CondaAPI from clearml_agent.helper.package.post_req import PostRequirement from clearml_agent.helper.package.external_req import ExternalRequirements, OnlyExternalRequirements from clearml_agent.helper.package.pip_api.system import SystemPip from clearml_agent.helper.package.pip_api.venv import VirtualenvPip from clearml_agent.helper.package.poetry_api import PoetryConfig, PoetryAPI from clearml_agent.helper.package.pytorch import PytorchRequirement from clearml_agent.helper.package.requirements import RequirementsManager from clearml_agent.helper.package.venv_update_api import VenvUpdateAPI from clearml_agent.helper.process import ( kill_all_child_processes, WorkerParams, ExitStatus, Argv, COMMAND_SUCCESS, Executable, get_bash_output, shutdown_docker_process, get_docker_id, commit_docker, terminate_process, check_if_command_exists, ) from clearml_agent.helper.package.priority_req import PriorityPackageRequirement, PackageCollectorRequirement from clearml_agent.helper.repo import clone_repository_cached, RepoInfo, VCS, fix_package_import_diff_patch from clearml_agent.helper.resource_monitor import ResourceMonitor from clearml_agent.helper.runtime_verification import check_runtime, print_uptime_properties from clearml_agent.session import Session from clearml_agent.helper.singleton import Singleton from .events import Events DOCKER_ROOT_CONF_FILE = "/root/clearml.conf" DOCKER_DEFAULT_CONF_FILE = "/root/default_clearml.conf" @attr.s class LiteralScriptManager(object): """ Manage notebook tasks """ venv_folder = attr.ib(type=str) @staticmethod def is_literal_script(task): # type: (tasks_api.Task) -> bool """ Returns whether a task object represents a notebook task """ script = task.script if not script: return False diff = script.diff if not diff: return False # test git diff prefix if diff.lstrip().lower().startswith("diff "): return False # test git submodule prefix # noinspection PyBroadException try: if diff.lstrip().lower().startswith("submodule ") and \ diff.splitlines()[1].lstrip().lower().startswith("diff "): return False except Exception: pass # none of the above return True @staticmethod def write(task, directory, entry_point=None): # type: (tasks_api.Task, str, Optional[str]) -> str """ Create notebook file for ``task`` in ``directory`` """ if entry_point: full_path = normalize_path(Text(directory), entry_point) if os.path.exists(full_path): return entry_point with open(full_path, "wt") as f: f.write(task.script.diff) return full_path with named_temporary_file( delete=False, prefix="script_", suffix=".py", dir=Text(directory), mode="wt" ) as f: f.write(task.script.diff) return f.name def create_notebook_file(self, task, execution, repo_info): # type: (tasks_api.Task, ExecutionInfo, Optional[RepoInfo]) -> Tuple[str, str] """ Create notebook file in appropriate location :return: directory and script path """ log = logging.getLogger(__name__) if repo_info and repo_info.root: location = Path(repo_info.root, execution.working_dir) else: if execution.working_dir and execution.working_dir.strip() != '.': log.warning( "found task with `script.working_dir` (`%s`) but without `script.repository`, ignoring", execution.working_dir, ) if not execution.entry_point: execution.entry_point = 'untitled.py' else: # ignore any folders in the entry point we only need the file name execution.entry_point = execution.entry_point.split(os.path.sep)[-1] location = None location = location or (repo_info and repo_info.root) if not location: location = Path(self.venv_folder, "code") location.mkdir(exist_ok=True) log.debug("selected execution directory: %s", location) return Text(location), self.write(task, location, execution.entry_point) def get_repo_auth_string(user, password): # type: (Text, Text) -> Text """ Return user:password only if user and password are valid :param user: username :param password: :return: URL authentication string """ if not (user and password): return "" return ":".join(map(quote, (user, password))) CONCAT_CMD = select_for_platform(linux=" && ", windows=" & ") class TaskStopReason(object): no_stop = 0 # type: TaskStopReason stopped = 1 # type: TaskStopReason reset = 2 # type: TaskStopReason status_changed = 3 # type: TaskStopReason exception = 4 # type: TaskStopReason def get_task(session, task_id, *args, **kwargs): return session.api_client.tasks.get_all(id=[task_id], *args, **kwargs)[0] def get_task_container(session, task_id): if session.check_min_api_version("2.13"): result = session.send_request( service='tasks', action='get_all', version='2.13', json={'id': [task_id], 'only_fields': ['container']}, method='get', async_enable=False, ) try: container = result.json()['data']['tasks'][0]['container'] if result.ok else {} container['arguments'] = str(container.get('arguments') or '').split(' ') except (ValueError, TypeError): container = {} else: response = get_task(session, task_id, only_fields=["execution.docker_cmd"]) task_docker_cmd_parts = str(response.execution.docker_cmd or '').strip().split(' ') try: container = dict( container=task_docker_cmd_parts[0], arguments=task_docker_cmd_parts[1:] if len (task_docker_cmd_parts[0])>1 else '' ) except (ValueError, TypeError): container = {} return container class TaskStopSignal(object): """ Follow task status and signals when it should be stopped """ _number_of_consecutive_reset_tests = 4 statuses = tasks_api.TaskStatusEnum unexpected_statuses = [ statuses.closed, statuses.stopped, statuses.failed, statuses.published, statuses.queued, ] default = TaskStopReason.no_stop stopping_message = "stopping" def __init__(self, command, session, events_service, task_id): # type: (Worker, Session, Events, Text) -> () """ :param command: workers command :param session: command session :param events_service: events service object :param task_id: followed task ID """ self.command = command self.session = session self.events_service = events_service self.worker_id = command.worker_id self._task_reset_state_counter = 0 self.task_id = task_id def test(self): # type: () -> TaskStopReason """ Returns whether task should stop and for what reason, returns TaskStopReason.no_stop if task shouldn't stop. Catches and logs exceptions. """ try: return self._test() except Exception as ex: self.command.log_traceback(ex) # make sure we break nothing return TaskStopSignal.default def _test(self): # type: () -> TaskStopReason """ "Unsafe" version of test() """ task_info = get_task( self.session, self.task_id, only_fields=["status", "status_message"] ) status = task_info.status message = task_info.status_message if status == self.statuses.in_progress and self.stopping_message in message: self.command.log( "task status_message has '%s', task will terminate", self.stopping_message, ) return TaskStopReason.stopped if status in self.unexpected_statuses: # ## and "worker" not in message: self.command.log("unexpected status change, task will terminate") return TaskStopReason.status_changed if status == self.statuses.created: if ( self._task_reset_state_counter >= self._number_of_consecutive_reset_tests ): self.command.log("task was reset, task will terminate") return TaskStopReason.reset self._task_reset_state_counter += 1 warning_msg = "Warning: Task {} was reset! if state is consistent we shall terminate ({}/{}).".format( self.task_id, self._task_reset_state_counter, self._number_of_consecutive_reset_tests, ) if self.events_service: self.events_service.send_log_events( self.worker_id, task_id=self.task_id, lines=[warning_msg], level="WARNING", ) print(warning_msg) else: self._task_reset_state_counter = 0 return TaskStopReason.no_stop class Worker(ServiceCommandSection): _pip_extra_index_url = PIP_EXTRA_INDICES _default_pip = tuple() _requirement_substitutions = ( PytorchRequirement, PriorityPackageRequirement, PostRequirement, ExternalRequirements, partial(PackageCollectorRequirement, collect_package=['trains']), partial(PackageCollectorRequirement, collect_package=['clearml']), ) # poll queues every _polling_interval seconds _polling_interval = 5.0 # machine status update intervals, seconds _machine_update_interval = 30.0 # message printed before starting task logging, # it will be parsed by services_mode, to identify internal docker logging start _task_logging_start_message = "Running task '{}'" # last message before passing control to the actual task _task_logging_pass_control_message = "Running task id [{}]:" _run_as_user_home = '/clearml_agent_home' _docker_fixed_user_cache = '/clearml_agent_cache' _temp_cleanup_list = [] @property def service(self): """ Worker command service endpoint """ return "workers" @property def _task_status_change_message(self): return "Changed by {} {}".format(PROGRAM_NAME, self.worker_id) @staticmethod def register_signal_handler(): def handler(*_): for f in Worker._temp_cleanup_list + [Singleton.get_pid_file()]: safe_remove_tree(f) raise Sigterm() signal.signal(signal.SIGTERM, handler) def __init__(self, *args, **kwargs): super(Worker, self).__init__(*args, **kwargs) self.monitor = None self.log = self._session.get_logger(__name__) self.register_signal_handler() self._worker_registered = False self.is_conda = is_conda(self._session.config) # type: bool # Add extra index url - system wide extra_url = None try: if self._session.config.get("agent.package_manager.extra_index_url", None): extra_url = self._session.config.get("agent.package_manager.extra_index_url", []) if not isinstance(extra_url, (tuple, list)): extra_url = [extra_url] # put external pip url before default ones, so we first look for packages there for e in reversed(extra_url): self._pip_extra_index_url.insert(0, e) except Exception: self.log.warning('Failed adding extra-index-url to pip environment: {}'.format(extra_url)) # update pip install command pip_install_cmd = ["pip", "install"] if self._pip_extra_index_url: pip_install_cmd.extend( chain.from_iterable( ("--extra-index-url", x) for x in self._pip_extra_index_url ) ) self.pip_install_cmd = tuple(pip_install_cmd) self.worker_id = self._session.config["agent.worker_id"] or "{}:{}".format( self._session.config["agent.worker_name"], os.getpid() ) self._last_stats = defaultdict(lambda: 0) self._last_report_timestamp = psutil.time.time() self.temp_config_path = None self.queues = () self.venv_folder = None # type: Optional[Text] self.package_api = None # type: Optional[PackageManager] self.global_package_api = None self.is_venv_update = self._session.config.agent.venv_update.enabled self.poetry = PoetryConfig(self._session) self.docker_image_func = None self._docker_image = None self._docker_arguments = None PackageManager.set_pip_version(self._session.config.get("agent.package_manager.pip_version", None)) self._extra_docker_arguments = self._session.config.get("agent.extra_docker_arguments", None) self._extra_shell_script = self._session.config.get("agent.extra_docker_shell_script", None) self._docker_force_pull = self._session.config.get("agent.docker_force_pull", False) self._daemon_foreground = None self._standalone_mode = None self._services_mode = None self._dynamic_gpus = None self._force_current_version = None self._redirected_stdout_file_no = None self._uptime_config = self._session.config.get("agent.uptime", None) self._downtime_config = self._session.config.get("agent.downtime", None) self._suppress_cr = self._session.config.get("agent.suppress_carriage_return", True) self._host_ssh_cache = None # True - supported # None - not initialized # str - not supported, version string indicates last server version self._runtime_props_support = None @classmethod def _verify_command_states(cls, kwargs): """ Conform and enforce command argument This is where you can automatically turn on/off switches based on different states. :param kwargs: :return: kwargs """ if kwargs.get('services_mode'): kwargs['cpu_only'] = True # kwargs['docker'] = kwargs.get('docker') or [] kwargs['gpus'] = None return kwargs def _get_requirements_manager(self, os_override=None, base_interpreter=None, requirement_substitutions=None): requirements_manager = RequirementsManager( self._session, base_interpreter=base_interpreter ) requirement_substitutions = requirement_substitutions or self._requirement_substitutions for requirement_cls in requirement_substitutions: if os_override and issubclass(requirement_cls, PytorchRequirement): requirement_cls = partial(requirement_cls, os_override=os_override) requirements_manager.register(requirement_cls) if not requirements_manager.found_cuda: self.warning( "could not find installed CUDA/CuDNN version using original requirements file" ) return requirements_manager def handle_user_abort(self, task_id): """ Set task status to appropriate value on user abort. """ try: task_status = self._session.send_api( tasks_api.GetByIdRequest(task_id) ).task.status if task_status == tasks_api.TaskStatusEnum.in_progress: print("\nUser abort - stopping task {}".format(task_id)) self._session.send_api(tasks_api.StoppedRequest(task_id)) except Exception: pass def run_one_task(self, queue, task_id, worker_args, docker=None): # type: (Text, Text, WorkerParams, Optional[Text]) -> () """ Run one task pulled from queue. :param queue: ID of queue that task was pulled from :param task_id: ID of task to run :param worker_args: Worker command line arguments :param docker: Docker image in which the execution task will run """ # start new process and execute task id # "Running task '{}'".format(task_id) print(self._task_logging_start_message.format(task_id)) # set task status to in_progress so we know it was popped from the queue try: self._session.send_api(tasks_api.StartedRequest(task=task_id, force=True)) except Exception: print("Warning: Could not start task id '{}', skipping".format(task_id)) return # setup console log temp_stdout_name = safe_mkstemp( suffix=".txt", prefix=".clearml_agent_out.", name_only=True ) # temp_stderr_name = safe_mkstemp(suffix=".txt", prefix=".clearml_agent_err.", name_only=True) temp_stderr_name = None print( "Storing stdout and stderr log to '{}', '{}'".format( temp_stdout_name, temp_stderr_name or temp_stdout_name ) ) docker_image = None worker_id = '{}:service:{}'.format(self.worker_id, task_id) \ if self._services_mode and not self._dynamic_gpus else self.worker_id if self.docker_image_func: # noinspection PyBroadException try: task_container = get_task_container(self._session, task_id) except Exception: task_container = {} if task_container: self.send_logs(task_id=task_id, lines=['Running Task {} inside docker: {}\n'.format(task_id, str(task_container))], level="INFO") docker_image = task_container.get('image') docker_arguments = task_container.get('arguments') docker_setup_script = task_container.get('setup_shell_script') else: self.send_logs(task_id=task_id, lines=['running Task {} inside default docker image: {} {}\n'.format( task_id, self._docker_image, self._docker_arguments or '')], level="INFO") docker_image = self._docker_image docker_arguments = self._docker_arguments docker_setup_script = None # Update docker command if self._services_mode: # if this is services mode, give the docker a unique worker id, as it will register itself. full_docker_cmd = self.docker_image_func( worker_id=worker_id, docker_image=docker_image, docker_arguments=docker_arguments, docker_bash_setup_script=docker_setup_script) else: full_docker_cmd = self.docker_image_func( docker_image=docker_image, docker_arguments=docker_arguments, docker_bash_setup_script=docker_setup_script) # if we are using the default docker, update back the Task: if not task_container: # noinspection PyBroadException try: self._session.send_api( tasks_api.EditRequest(task_id, force=True, execution=dict( # noqa docker_cmd=' '.join([docker_image] + docker_arguments) if docker_arguments else str(docker_image)))) except Exception: pass # if this is services_mode, change the worker_id to a unique name # abd use full-monitoring, ot it registers itself as a worker for this specific service. # notice, the internal agent will monitor itself once the docker is up and running full_docker_cmd[-1] = full_docker_cmd[-1] + 'execute {} {} --id {}'.format( '--full-monitoring' if self._services_mode else '--disable-monitoring', '--standalone-mode' if self._standalone_mode else '', task_id) # send the actual used command line to the backend self.send_logs(task_id=task_id, lines=['Executing: {}\n'.format(full_docker_cmd)], level="INFO") cmd = Argv(*full_docker_cmd) print('Running Docker:\n{}\n'.format(str(cmd))) else: cmd = worker_args.get_argv_for_command("execute") + ( '--full-monitoring' if self._services_mode else '--disable-monitoring', "--id", task_id, ) events_service = self.get_service(Events) stop_signal = TaskStopSignal( command=self, session=self._session, events_service=events_service, task_id=task_id, ) stop_signal_status = TaskStopSignal.default status = ExitStatus.interrupted try: # set WORKER ID ENV_WORKER_ID.set(worker_id) if self._docker_force_pull and docker_image: full_pull_cmd = ['docker', 'pull', docker_image] pull_cmd = Argv(*full_pull_cmd) status, stop_signal_status = self._log_command_output( task_id=task_id, cmd=pull_cmd, stdout_path=temp_stdout_name, stderr_path=temp_stderr_name, daemon=True, stop_signal=stop_signal, ) status, stop_signal_status = self._log_command_output( task_id=task_id, cmd=cmd, stdout_path=temp_stdout_name, stderr_path=temp_stderr_name, daemon=True, stop_signal=stop_signal, ) errors = temp_stderr_name and Path(temp_stderr_name).read_text() if errors: print("\nEncountered errors:\n\n{}\n".format(errors)) if status is None and self._services_mode: print( "Service bootstrap completed: task '{}'".format(task_id) ) elif status is None: print( "DONE: Running task '{}' (user aborted)".format(task_id) ) else: print("DONE: Running task '{}', exit status {}".format(task_id, status)) except KeyboardInterrupt: self.handle_user_abort(task_id) status = ExitStatus.interrupted finally: if self._services_mode and stop_signal_status is None: print('Service started, docker running in the background') else: self.handle_task_termination(task_id, status, stop_signal_status) # remove temp files after we sent everything to the backend if self.docker_image_func: shutdown_docker_process(docker_cmd_contains='--id {}\'\"'.format(task_id)) safe_remove_file(temp_stdout_name) safe_remove_file(temp_stderr_name) def run_tasks_loop(self, queues, worker_params, priority_order=True, gpu_indexes=None, gpu_queues=None): """ :summary: Pull and run tasks from queues. :description: 1. Go through ``queues`` by order. 2. Try getting the next task for each and run the first one that returns. 3. Go to step 1 :param list(str) queues: IDs of queues to pull tasks from :param worker_params worker_params: Worker command line arguments :param bool priority_order: If True pull order in priority manner. always from the first If False, pull from each queue once in a round robin manner :param list gpu_indexes: list of gpu_indexes. Needs special backend support :param list gpu_queues: list of pairs (queue_id, num_gpus). Needs special backend support """ if not self._daemon_foreground: print('Starting infinite task polling loop...') _last_machine_update_ts = 0 if self._services_mode: max_num_instances = int(self._services_mode) else: max_num_instances = None # store in runtime configuration, if max_num_instances and not self.set_runtime_properties(key='max_num_instances', value=max_num_instances): warning('Maximum number of service instance not supported, removing limit.') max_num_instances = -1 # get current running instances available_gpus = None if gpu_indexes and gpu_queues: available_gpus, gpu_queues = self._setup_dynamic_gpus(gpu_queues) # multi instance support self._services_mode = True # last 64 tasks list_task_gpus_ids = {} try: while True: queue_tags = None runtime_props = None if max_num_instances and max_num_instances > 0: # make sure we do not have too many instances to run if len(Singleton.get_running_pids()) >= max_num_instances: if self._daemon_foreground or worker_params.debug: print("Reached max number of services, sleeping for {:.1f} seconds".format( self._polling_interval)) sleep(self._polling_interval) continue # update available gpus if gpu_queues: available_gpus = self._dynamic_gpu_get_available(gpu_indexes) # if something went wrong or we have no free gpus # start over from the highest priority queue if not available_gpus: if self._daemon_foreground or worker_params.debug: print("All GPUs allocated, sleeping for {:.1f} seconds".format(self._polling_interval)) sleep(self._polling_interval) continue # iterate over queues (priority style, queues[0] is highest) for queue in queues: if queue_tags is None or runtime_props is None: queue_tags, runtime_props = self.get_worker_properties(queues) if not self.should_be_currently_active(queue_tags[queue], runtime_props): continue if gpu_queues: # peek into queue # get next task in queue try: response = self._session.send_api(queues_api.GetByIdRequest(queue=queue)) except Exception: # if something went wrong start over from the highest priority queue break if not len(response.queue.entries): continue # check if we have enough available gpus if gpu_queues[queue] > len(available_gpus): # not enough available_gpus, we should sleep and start over if self._daemon_foreground or worker_params.debug: print("Not enough free GPUs {}/{}, sleeping for {:.1f} seconds".format( len(available_gpus), gpu_queues[queue], self._polling_interval)) sleep(self._polling_interval) break # get next task in queue try: response = self._session.send_api( queues_api.GetNextTaskRequest(queue=queue) ) except Exception as e: print( "Warning: Could not access task queue [{}], error: {}".format( queue, e ) ) continue else: try: task_id = response.entry.task except AttributeError: if self._daemon_foreground or worker_params.debug: print("No tasks in queue {}".format(queue)) continue # clear output log if we start a new Task if not worker_params.debug and self._redirected_stdout_file_no is not None and \ self._redirected_stdout_file_no > 2: # noinspection PyBroadException try: os.lseek(self._redirected_stdout_file_no, 0, 0) os.ftruncate(self._redirected_stdout_file_no, 0) except: pass self.report_monitor(ResourceMonitor.StatusReport(queues=queues, queue=queue, task=task_id)) org_gpus = os.environ.get('NVIDIA_VISIBLE_DEVICES') worker_id = self.worker_id # the following is only executed in dynamic gpus mode if gpu_queues and gpu_queues.get(queue): # pick the first available GPUs gpus = available_gpus[:gpu_queues.get(queue)] available_gpus = available_gpus[gpu_queues.get(queue):] self.set_runtime_properties( key='available_gpus', value=','.join(str(g) for g in available_gpus)) os.environ['CUDA_VISIBLE_DEVICES'] = \ os.environ['NVIDIA_VISIBLE_DEVICES'] = ','.join(str(g) for g in gpus) list_task_gpus_ids.update({str(g): task_id for g in gpus}) self.worker_id = ':'.join(self.worker_id.split(':')[:-1] + ['gpu'+','.join(str(g) for g in gpus)]) self.send_logs( task_id=task_id, lines=["task {} pulled from {} by worker {}\n".format(task_id, queue, self.worker_id)], level="INFO", ) self.run_one_task(queue, task_id, worker_params) if gpu_queues: self.worker_id = worker_id os.environ['CUDA_VISIBLE_DEVICES'] = \ os.environ['NVIDIA_VISIBLE_DEVICES'] = org_gpus self.report_monitor(ResourceMonitor.StatusReport(queues=self.queues)) queue_tags = None runtime_props = None # if we are using priority start pulling from the first always, # if we are doing round robin, pull from the next one if priority_order: break else: # sleep and retry polling if self._daemon_foreground or worker_params.debug: print("No tasks in Queues, sleeping for {:.1f} seconds".format(self._polling_interval)) sleep(self._polling_interval) if self._session.config["agent.reload_config"]: self.reload_config() finally: # if we are in dynamic gpus mode, shutdown all active runs if self.docker_image_func: for t_id in set(list_task_gpus_ids.values()): if shutdown_docker_process(docker_cmd_contains='--id {}\'\"'.format(t_id)): self.handle_task_termination(task_id=t_id, exit_code=0, stop_reason=TaskStopReason.stopped) else: # if we are here, just kill all sub processes kill_all_child_processes() for t_id in set(list_task_gpus_ids.values()): # check if Task is running, task_info = get_task( self._session, t_id, only_fields=["status"] ) # this is a bit risky we might have rerun it again after it already completed # basically we are not removing completed tasks from the list, hence the issue if str(task_info.status) == "in_progress": self.handle_task_termination( task_id=t_id, exit_code=0, stop_reason=TaskStopReason.stopped) def _dynamic_gpu_get_available(self, gpu_indexes): # noinspection PyBroadException try: response = self._session.send_api(workers_api.GetAllRequest(last_seen=60)) except Exception: return None worker_name = self._session.config["agent.worker_name"] + ':gpu' our_workers = [ w.id for w in response.workers if w.id.startswith(worker_name) and w.id != self.worker_id] gpus = [] for w in our_workers: gpus += [int(g) for g in w.split(':')[-1].lower().replace('gpu', '').split(',')] available_gpus = list(set(gpu_indexes) - set(gpus)) return available_gpus def _setup_dynamic_gpus(self, gpu_queues): available_gpus = self.get_runtime_properties() if available_gpus is None: raise ValueError("Dynamic GPU allocation is not supported by the ClearML-server") available_gpus = [prop["value"] for prop in available_gpus if prop["key"] == 'available_gpus'] if available_gpus: available_gpus = [int(g) for g in available_gpus[-1].split(',')] if not isinstance(gpu_queues, dict): gpu_queues = dict(gpu_queues) if not self.set_runtime_properties( key='available_gpus', value=','.join(str(g) for g in available_gpus)): raise ValueError("Dynamic GPU allocation is not supported by the ClearML-server") return available_gpus, gpu_queues def get_worker_properties(self, queue_ids): queue_tags = { q.id: {'name': q.name, 'tags': q.tags} for q in self._session.send_api( queues_api.GetAllRequest(id=queue_ids, only_fields=["id", "tags"]) ).queues } runtime_props = self.get_runtime_properties() return queue_tags, runtime_props def get_runtime_properties(self): if self._runtime_props_support is not True: # either not supported or never tested if self._runtime_props_support == self._session.api_version: # tested against latest api_version, not supported return None if not self._session.check_min_api_version(UptimeConf.min_api_version): # not supported due to insufficient api_version self._runtime_props_support = self._session.api_version return None try: res = self.get("get_runtime_properties", worker=self.worker_id)["runtime_properties"] # definitely supported self._runtime_props_support = True return res except APIError: self._runtime_props_support = self._session.api_version return None def set_runtime_properties(self, key, value): if self._runtime_props_support is not True: # either not supported or never tested if self._runtime_props_support == self._session.api_version: # tested against latest api_version, not supported return False if not self._session.check_min_api_version(UptimeConf.min_api_version): # not supported due to insufficient api_version self._runtime_props_support = self._session.api_version return False # noinspection PyBroadException try: self.post("set_runtime_properties", json={ 'runtime_properties': [{'key': key, 'value': value}], 'worker': self.worker_id}) # definitely supported self._runtime_props_support = True return True except APIError: self._runtime_props_support = self._session.api_version except Exception as ex: # not sure what happened pass return False def should_be_currently_active(self, current_queue, runtime_properties): """ Checks if a worker is active according to queue tags, worker's runtime properties and uptime schedule. """ runtime_properties = runtime_properties or [] if UptimeConf.queue_tag_off in current_queue['tags']: self.log.debug("Queue {} is tagged '{}', worker will not pull tasks".format( current_queue['name'], UptimeConf.queue_tag_off) ) return False if UptimeConf.queue_tag_on in current_queue['tags']: self.log.debug("Queue {} is tagged '{}', worker will pull tasks".format( current_queue['name'], UptimeConf.queue_tag_on) ) return True force_flag = next( (prop for prop in runtime_properties if prop["key"] == UptimeConf.worker_key), None ) if force_flag: if force_flag["value"].lower() in UptimeConf.worker_value_off: self.log.debug("worker has the following runtime property: '{}'. worker will not pull tasks".format( force_flag) ) return False elif force_flag["value"].lower() in UptimeConf.worker_value_on: self.log.debug("worker has the following runtime property: '{}'. worker will pull tasks".format( force_flag) ) return True else: print( "Warning: invalid runtime_property '{}: {}' supported values are: '{}/{}', ignoring".format( force_flag["key"], force_flag["value"], UptimeConf.worker_value_on, UptimeConf.worker_value_off ) ) if self._uptime_config: self.log.debug("following uptime configurations") return check_runtime(self._uptime_config) if self._downtime_config: self.log.debug("following downtime configurations") return check_runtime(self._downtime_config, is_uptime=False) return True def reload_config(self): try: reloaded = self._session.reload() except Exception as ex: self.log("Failed reloading config file") self.log_traceback(ex) else: if reloaded: self.log( 'Config file change detected, reloading and updating "{.temp_config_path}"'.format( self ) ) self.dump_config(self.temp_config_path) def check(self, **_): try: check_directory_path(str(Path(".").resolve()), check_whitespace_in_path=False) except OSError as e: if e.errno == errno.ENOENT: raise CommandFailedError("current working directory does not exist") raise for key in "agent.venvs_dir", "sdk.storage.cache.default_base_dir": try: value = self._session.config.get(key, None) if value: check_directory_path(value) except CommandFailedError as e: raise CommandFailedError( 'Invalid config key "{}": {.message}'.format(key, e) ) if is_windows_platform(): # if not self.is_conda: # self.warning("Worker on Windows without Conda are not supported") if self._session.config.agent.venv_update: # self.warning("venv-update is not supported on Windows") self.is_venv_update = False self._session.print_configuration() def daemon(self, queues, log_level, foreground=False, docker=False, detached=False, order_fairness=False, **kwargs): # check that we have docker command if we need it if docker not in (False, None) and not check_if_command_exists("docker"): raise ValueError("Running in Docker mode, 'docker' command was not found") self._standalone_mode = kwargs.get('standalone_mode', False) self._services_mode = kwargs.get('services_mode', False) # must have docker in services_mode if self._services_mode: kwargs = self._verify_command_states(kwargs) self._uptime_config = kwargs.get('uptime', None) or self._uptime_config self._downtime_config = kwargs.get('downtime', None) or self._downtime_config if self._uptime_config and self._downtime_config: self.log.error( "Both uptime and downtime were specified when only one of them could be used. Both will be ignored." ) self._uptime_config = None self._downtime_config = None # support --dynamic-gpus dynamic_gpus, gpu_indexes, queues = self._parse_dynamic_gpus(kwargs, queues) if self._services_mode and dynamic_gpus: raise ValueError("Combining --dynamic-gpus and --services-mode is not supported") # if we do not need to create queues, make sure they are valid # match previous behaviour when we validated queue names before everything else queues = self._resolve_queue_names(queues, create_if_missing=kwargs.get('create_queue', False)) # We are not running a daemon we are killing one. # find the pid send termination signal and leave if kwargs.get('stop', False): return 1 if not self._kill_daemon(dynamic_gpus=dynamic_gpus) else 0 queues_info = [ q.to_dict() for q in self._session.send_api( queues_api.GetAllRequest(id=queues) ).queues ] if kwargs.get('status', False): runtime_properties = self.get_runtime_properties() if self._downtime_config: print_uptime_properties(self._downtime_config, queues_info, runtime_properties, is_uptime=False) else: print_uptime_properties(self._uptime_config, queues_info, runtime_properties) return 1 # make sure we only have a single instance, # also make sure we set worker_id properly and cache folders self._singleton(dynamic_gpus=bool(dynamic_gpus)) if dynamic_gpus: self._register_dynamic_gpus(gpu_indexes) # check if we have the latest version start_check_update_daemon() self.check(**kwargs) self.log.debug("starting resource monitor thread") print("Worker \"{}\" - ".format(self.worker_id), end='') columns = ("id", "name", "tags") print("Listening to queues:") print_table(queues_info, columns=columns, titles=columns) # register worker self._register(queues) # create temp config file with current configuration self.temp_config_path = NamedTemporaryFile( suffix=".cfg", prefix=".clearml_agent.", mode='w+t').name # print docker image if docker is not False and docker is not None: self._force_current_version = kwargs.get('force_current_version', False) self.set_docker_variables(docker) else: self.dump_config(self.temp_config_path) # only in none docker we have to make sure we have CUDA setup # make sure we have CUDA set if we have --gpus if kwargs.get('gpus') and self._session.config.get('agent.cuda_version', None) in (None, 0, '0'): message = 'Running with GPUs but no CUDA version was detected!\n' \ '\tSet OS environemnt CUDA_VERSION & CUDNN_VERSION to the correct version\n' \ '\tExample: export CUDA_VERSION=10.1 or (Windows: set CUDA_VERSION=10.1)' if is_conda(self._session.config): self._unregister(queues) safe_remove_file(self.temp_config_path) raise ValueError(message) else: warning(message+'\n') if self._services_mode: print('ClearML-Agent running in services mode') self._daemon_foreground = foreground if not foreground: out_file, name = safe_mkstemp( prefix=".clearml_agent_daemon_out", suffix=".txt", open_kwargs={ "buffering": self._session.config.get("agent.log_files_buffering", 1) }, ) print( "Running CLEARML-AGENT daemon in background mode, writing stdout/stderr to {}".format( name ) ) if not self._session.debug_mode: self._temp_cleanup_list.append(name) # on widows we do nothing if detached and is_windows_platform(): print('Detached not supported on Windows, ignoring --detached') detached = False if not detached: # redirect std out/err to new file sys.stdout = sys.stderr = out_file else: # in detached mode # fully detach stdin.stdout/stderr and leave main process, running in the background daemonize_process(out_file.fileno()) self._redirected_stdout_file_no = out_file.fileno() # make sure we update the singleton lock file to the new pid Singleton.update_pid_file() # reprint headers to std file (we are now inside the daemon process) print("Worker \"{}\" :".format(self.worker_id)) self._session.print_configuration() print_table(queues_info, columns=columns, titles=columns) try: while True: try: self.new_monitor(ResourceMonitor.StatusReport(queues=queues)) self.run_tasks_loop( queues, worker_params=WorkerParams( log_level=log_level, config_file=self.temp_config_path, debug=self._session.debug_mode, trace=self._session.trace, ), priority_order=not order_fairness, gpu_indexes=gpu_indexes, gpu_queues=dynamic_gpus, ) except Exception: tb = six.text_type(traceback.format_exc()) print("FATAL ERROR:") print(tb) crash_file, name = safe_mkstemp(prefix=".clearml_agent-crash", suffix=".log") try: with crash_file: crash_file.write(tb) except Exception: print( "Could not write crash log to {}\nException:\n{}".format( name, tb ) ) sleep(1) finally: self._unregister(queues) safe_remove_file(self.temp_config_path) def _parse_dynamic_gpus(self, kwargs, queues): dynamic_gpus = kwargs.get('dynamic_gpus', None) if not dynamic_gpus: return None, None, queues queue_names = [q.name for q in queues] if not all('=' in q for q in queue_names): raise ValueError("using --dynamic-gpus, --queues [{}], " "queue must be in format =".format(queue_names)) gpu_indexes = kwargs.get('gpus') # test gpus were passed correctly if not gpu_indexes or len(gpu_indexes.split('-')) > 2 or (',' in gpu_indexes and '-' in gpu_indexes): raise ValueError('--gpus must be provided, in one of two ways: ' 'comma separated \'0,1,2,3\' or range \'0-3\'') try: if '-' in gpu_indexes: gpu_indexes = list(range(int(gpu_indexes.split('-')[0]), 1 + int(gpu_indexes.split('-')[1]))) else: gpu_indexes = [int(g) for g in gpu_indexes.split(',')] except Exception: raise ValueError( 'Failed parsing --gpus "{}". ' '--dynamic_gpus must be use with ' 'specific gpus for example "0-7" or "0,1,2,3"'.format(kwargs.get('gpus'))) dynamic_gpus = [(s[:-1 - len(s.split('=')[-1])], int(s.split('=')[-1])) for s in queue_names] queue_names = [q for q, _ in dynamic_gpus] # resolve queue ids dynamic_gpus_q = self._resolve_queue_names( queue_names, create_if_missing=kwargs.get('create_queue', False)) dynamic_gpus = list(zip(dynamic_gpus_q, [i for _, i in dynamic_gpus])) # maintain original priority order queues = [q for q, _ in dynamic_gpus] self._dynamic_gpus = True # order queue priority based on the combination we have dynamic_gpus = sorted( dynamic_gpus, reverse=True, key=cmp_to_key( lambda x, y: -1 if x[1] < y[1] or x[1] == y[1] and dynamic_gpus_q.index(x[0]) > dynamic_gpus_q.index(y[0]) else +1)) return dynamic_gpus, gpu_indexes, queues def _register_dynamic_gpus(self, gpu_indexes): # test server support available_gpus = self._dynamic_gpu_get_available(gpu_indexes) if not self.set_runtime_properties( key='available_gpus', value=','.join(str(g) for g in available_gpus)): raise ValueError("Dynamic GPU allocation is not supported by the ClearML-server") def report_monitor(self, report): if not self.monitor: self.new_monitor(report=report) else: self.monitor.set_report(report) self.monitor.send_report() def stop_monitor(self): if self.monitor: self.monitor.stop() self.monitor = None def new_monitor(self, report=None): self.stop_monitor() self.monitor = ResourceMonitor( session=self._session, worker_id=self.worker_id, first_report_sec=3.0, report_frequency_sec=self._machine_update_interval) self.monitor.set_report(report) self.monitor.start() return self.monitor def dump_config(self, filename, config=None): def to_json(config): return json.dumps(config.as_plain_ordered_dict(), cls=HOCONEncoder, indent=4) # noinspection PyBroadException try: current_content = Path(filename).read_text() except Exception: current_content = None # noinspection PyBroadException try: new_content = six.text_type(self._session.to_json() if config is None else to_json(config)) # Overwrite file only if the content is different, because we are mounting the same file into # multiple containers in services mode, and we don't want to change it if we do not have to. if new_content != current_content: Path(filename).write_text(new_content) except Exception: return False return True def _log_command_output( self, task_id, # type: Text cmd, # type: Executable stdout_path=None, # type: Text stderr_path=None, # type: Optional[Text] daemon=False, # type: bool cwd=None, # type: Text stop_signal=None, # type: Optional[TaskStopSignal] **kwargs # type: Any ): # type: (...) -> Tuple[Optional[int], Optional[TaskStopReason]] def _print_file(file_path, prev_pos=0): with open(file_path, "rb") as f: f.seek(prev_pos) binary_text = f.read() pos = f.tell() # skip the previously printed lines, blines = binary_text.split(b'\n') if binary_text else [] if not blines: return blines, pos return ( decode_binary_lines(blines if blines[-1] else blines[:-1], replace_cr=not self._suppress_cr, overwrite_cr=self._suppress_cr), pos ) stdout = open(stdout_path, "wt") stderr = open(stderr_path, "wt") if stderr_path else stdout stdout_line_count, stdout_pos_count, stdout_last_lines = 0, 0, [] stderr_line_count, stderr_pos_count, stderr_last_lines = 0, 0, [] service_mode_internal_agent_started = None stopping = False status = None try: _last_machine_update_ts = time() stop_reason = None process = cmd.call_subprocess( subprocess.Popen, stdout=stdout, stderr=stderr, cwd=cwd and str(cwd), **kwargs ) while status is None and not stopping: stop_reason = stop_signal.test() if stop_signal else TaskStopSignal.default if stop_reason != TaskStopSignal.default: # mark quit loop stopping = True if daemon: self.send_logs( task_id=task_id, lines=["User aborted: stopping task ({})\n".format(str(stop_reason))], level="ERROR", ) kill_all_child_processes(process.pid) else: sleep(self._polling_interval) status = process.poll() # flush stdout and stderr buffers if stdout: stdout.flush() if stderr: stderr.flush() # get diff from previous poll printed_lines, stdout_pos_count = _print_file(stdout_path, stdout_pos_count) if self._services_mode and not stopping and not status: # if the internal agent started, we stop logging, it will take over logging. # if the internal agent started running the task itself, it will return status==0, # then we can quit the monitoring loop of this process printed_lines, service_mode_internal_agent_started, status = self._check_if_internal_agent_started( printed_lines, service_mode_internal_agent_started, task_id) if status is not None: stop_reason = 'Service started' stdout_line_count += self.send_logs(task_id, printed_lines) if stderr_path: printed_lines, stderr_pos_count = _print_file(stderr_path, stderr_pos_count) stderr_line_count += self.send_logs(task_id, printed_lines) except subprocess.CalledProcessError as ex: # non zero return code stop_reason = TaskStopReason.exception status = ex.returncode except KeyboardInterrupt: # so someone else will catch us raise except Exception: # we should not get here, but better safe than sorry printed_lines, stdout_pos_count = _print_file(stdout_path, stdout_pos_count) stdout_line_count += self.send_logs(task_id, printed_lines) if stderr_path: printed_lines, stderr_pos_count = _print_file(stderr_path, stderr_pos_count) stderr_line_count += self.send_logs(task_id, printed_lines) stop_reason = TaskStopReason.exception status = -1 # if running in services mode, keep the file open # in case the docker was so quick it started and finished, check the stop reason if self._services_mode and service_mode_internal_agent_started and stop_reason == 'Service started': return None, None stdout.close() if stderr_path: stderr.close() # Send last lines printed_lines, stdout_pos_count = _print_file(stdout_path, stdout_pos_count) stdout_line_count += self.send_logs(task_id, printed_lines) if stderr_path: printed_lines, stderr_pos_count = _print_file(stderr_path, stderr_pos_count) stderr_line_count += self.send_logs(task_id, printed_lines) return status, stop_reason def _check_if_internal_agent_started(self, printed_lines, service_mode_internal_agent_started, task_id): log_start_msg = self._task_logging_start_message.format(task_id) log_control_end_msg = self._task_logging_pass_control_message.format(task_id) filter_lines = printed_lines if not service_mode_internal_agent_started else [] for i, line in enumerate(printed_lines): if not service_mode_internal_agent_started and line.startswith(log_start_msg): service_mode_internal_agent_started = True filter_lines = printed_lines[:i+1] elif line.startswith(log_control_end_msg): return filter_lines, service_mode_internal_agent_started, 0 return filter_lines, service_mode_internal_agent_started, None def send_logs(self, task_id, lines, level="DEBUG"): """ Send output lines as log events to backend :param task_id: ID of task to send logs for :type task_id: Text :param lines: lines to send :type lines: [Text] :param str level: log level, default DEBUG :return: number of lines sent :rtype: int """ if not lines: return 0 print_text("".join(lines), newline=False) # remove backspaces from the text log, they look bad. for i, l in enumerate(lines): lines[i] = l.replace('\x08', '') events_service = self.get_service(Events) try: events_service.send_log_events( self.worker_id, task_id=task_id, lines=lines, level=level ) return len(lines) except Exception as e: print("\n### Error sending log: %s ###\n" % e) # revert number of sent lines (we will try next time) return 0 def _update_commit_id(self, task_id, execution, repo_info): """ If commit ID is not set, set it to the currently running version of the repository """ if not repo_info.commit or execution.version_num: return self.log("Updating task commit ID: %s", repo_info.commit) try: self._session.send_api( tasks_api.EditRequest( task_id, script=dict(version_num=repo_info.commit), force=True ) ) except Exception: pass def apply_diff(self, task, vcs, execution_info, repo_info): # type: (Any, VCS, ExecutionInfo, RepoInfo) -> None """ Apply task diff if present """ diff = task.script and task.script.diff if not diff: return print("Applying uncommitted changes") try: success = vcs.patch(normalize_path(repo_info.root), diff) except Exception as ex: self.log.warning("could not apply diff: %s", ex) success = False if not success: raise ValueError("Failed applying git diff:\n{}\n\n" "ERROR! Failed applying git diff, see diff above.".format(diff)) @resolve_names def build( self, task_id, target=None, python_version=None, docker=None, entry_point=None, install_globally=False, **_ ): if not task_id: raise CommandFailedError("Worker build must have valid task id") self._session.print_configuration() if docker is not False and docker is not None: return self._build_docker(docker, target, task_id, entry_point) current_task = self._session.api_client.tasks.get_by_id(task_id) execution = self.get_execution_info(current_task) if self._session.config.get("agent.package_manager.force_repo_requirements_txt", False): requirements = None print("[package_manager.force_repo_requirements_txt=true] " "Skipping requirements, using repository \"requirements.txt\" ") else: try: requirements = current_task.script.requirements except AttributeError: requirements = None if not python_version: try: python_version = current_task.script.binary python_version = python_version.split('/')[-1].replace('python', '') # if we can cast it, we are good python_version = '{:.1f}'.format(float(python_version)) except: python_version = None venv_folder, requirements_manager, is_cached = self.install_virtualenv( venv_dir=target, requested_python_version=python_version, execution_info=execution, cached_requirements=requirements) if not is_cached and self._default_pip: if install_globally and self.global_package_api: self.global_package_api.install_packages(*self._default_pip) else: self.package_api.install_packages(*self._default_pip) directory, vcs, repo_info = self.get_repo_info(execution, current_task, venv_folder.as_posix()) cwd = vcs.location if vcs and vcs.location else directory if is_cached: # reinstalling git / local packages package_api = copy(self.package_api) OnlyExternalRequirements.cwd = package_api.cwd = cwd package_api.requirements_manager = self._get_requirements_manager( base_interpreter=package_api.requirements_manager.get_interpreter(), requirement_substitutions=[OnlyExternalRequirements], ) # make sure we run the handlers cached_requirements = \ {k: package_api.requirements_manager.replace(requirements[k] or '') for k in requirements} package_api.load_requirements(cached_requirements) # make sure we call the correct freeze requirements_manager = package_api.requirements_manager else: self.install_requirements( execution, repo_info, requirements_manager=requirements_manager, cached_requirements=requirements, cwd=cwd, package_api=self.global_package_api if install_globally else None, ) freeze = self.freeze_task_environment( task_id=task_id, requirements_manager=requirements_manager, update_requirements=False) script_dir = directory # Summary print("Restoring running environment of task id [%s]:" % task_id) if freeze: print("Summary - installed python packages:") print(dump_yaml(freeze)) else: print("No freeze information available") print("Virtual environment: {}".format(venv_folder / 'bin')) print("Source code: {}".format(repo_info.root if repo_info else execution.entry_point)) print("Entry point: {}".format(Path(script_dir) / execution.entry_point)) return 0 def _build_docker(self, docker, target, task_id, entry_point=None, standalone_mode=True): self.temp_config_path = safe_mkstemp( suffix=".cfg", prefix=".clearml_agent.", text=True, name_only=True ) if not target: target = "task_id_{}".format(task_id) temp_config, docker_image_func = self.get_docker_config_cmd(docker) self.dump_config(self.temp_config_path, config=temp_config) self.docker_image_func = docker_image_func try: task_container = get_task_container(self._session, task_id) except Exception: task_container = {} if task_container: docker_image = task_container.get('image') docker_arguments = task_container.get('arguments') docker_setup_script = task_container.get('setup_shell_script') else: docker_image = self._docker_image docker_arguments = self._docker_arguments docker_setup_script = None print('Building Task {} inside docker image: {} {} setup_script={}\n'.format( task_id, docker_image, docker_arguments or '', docker_setup_script or '')) full_docker_cmd = self.docker_image_func( docker_image=docker_image, docker_arguments=docker_arguments, docker_bash_setup_script=docker_setup_script) end_of_build_marker = "build.done=true" docker_cmd_suffix = ' build --id {task_id} --install-globally; ' \ 'echo "" >> {conf_file} ; ' \ 'echo {end_of_build_marker} >> {conf_file} ; ' \ 'bash'.format( task_id=task_id, end_of_build_marker=end_of_build_marker, conf_file=DOCKER_ROOT_CONF_FILE ) full_docker_cmd[-1] = full_docker_cmd[-1] + docker_cmd_suffix cmd = Argv(*full_docker_cmd) # we will be checking the configuration file for changes temp_config = Path(self.temp_config_path) base_time_stamp = temp_config.stat().st_mtime # start the docker print('Starting docker build') cmd.call_subprocess(subprocess.Popen) # now we need to wait until the line shows on our configuration file. while True: while temp_config.stat().st_mtime == base_time_stamp: sleep(5.0) with open(temp_config.as_posix()) as f: lines = [l.strip() for l in f.readlines()] if 'build.done=true' in lines: break base_time_stamp = temp_config.stat().st_mtime print('\nDocker build done') # get the docker id. docker_id = get_docker_id(docker_cmd_contains='--id {} '.format(task_id)) if not docker_id: print("Error: cannot locate docker for storage") return if entry_point == "clone_task" or entry_point == "reuse_task": change = 'ENTRYPOINT if [ ! -s "{trains_conf}" ] ; then ' \ 'cp {default_trains_conf} {trains_conf} ; ' \ ' fi ; clearml-agent execute --id {task_id} --standalone-mode {clone}'.format( default_trains_conf=DOCKER_DEFAULT_CONF_FILE, trains_conf=DOCKER_ROOT_CONF_FILE, task_id=task_id, clone=("--clone" if entry_point == "clone_task" else ""), ) else: change = 'ENTRYPOINT []' print('Committing docker container to: {}'.format(target)) print(commit_docker(container_name=target, docker_id=docker_id, apply_change=change)) shutdown_docker_process(docker_id=docker_id) return @resolve_names def execute( self, task_id, log_level, optimization=0, disable_monitoring=False, full_monitoring=False, require_queue=False, log_file=None, standalone_mode=None, docker=False, clone=False, **_ ): self._standalone_mode = standalone_mode if not task_id: raise CommandFailedError("Worker execute must have valid task id") try: current_task = self._session.api_client.tasks.get_by_id(task_id) if not current_task.id: pass except AttributeError: raise ValueError( "Could not find task id={} (for host: {})".format( task_id, self._session.config.get("api.host", "") ) ) except Exception as ex: raise ValueError( "Could not find task id={} (for host: {})\nException: {}".format( task_id, self._session.config.get("api.host", ""), ex ) ) if clone: try: print("Cloning task id={}".format(task_id)) current_task = self._session.api_client.tasks.get_by_id( self._session.send_api( tasks_api.CloneRequest(task=current_task.id, new_task_name='Clone of {}'.format(current_task.name)) ).id ) print("Task cloned, new task id={}".format(current_task.id)) except Exception: raise CommandFailedError("Cloning failed") else: # make sure this task is not stuck in an execution queue, it shouldn't have been, but just in case. try: res = self._session.api_client.tasks.dequeue(task=current_task.id) if require_queue and res.meta.result_code != 200: raise ValueError("Execution required enqueued task, " "but task id={} is not queued.".format(current_task.id)) except Exception: if require_queue: raise if docker is not False and docker is not None: self.set_docker_variables(docker) # We expect the same behaviour in case full_monitoring was set, and in case docker mode is used if full_monitoring or docker is not False: if full_monitoring: if not (ENV_WORKER_ID.get() or '').strip(): self._session.config["agent"]["worker_id"] = '' # make sure we support multiple instances if we need to self._singleton() self.temp_config_path = self.temp_config_path or safe_mkstemp( suffix=".cfg", prefix=".clearml_agent.", text=True, name_only=True ) self.dump_config(self.temp_config_path) self._session._config_file = self.temp_config_path worker_params = WorkerParams( log_level=log_level, config_file=self._session.config_file, debug=self._session.debug_mode, trace=self._session.trace, ) self.report_monitor(ResourceMonitor.StatusReport(task=current_task.id)) self.run_one_task(queue='', task_id=current_task.id, worker_args=worker_params, docker=docker) self.stop_monitor() self._unregister() if full_monitoring and self.temp_config_path: safe_remove_file(self._session.config_file) Singleton.close_pid_file() return self._session.print_configuration() # now mark the task as started self._session.api_client.tasks.started( task=current_task.id, status_reason="worker started execution", status_message=self._task_status_change_message, ) if not disable_monitoring: self.log.debug("starting resource monitor") self.report_monitor(ResourceMonitor.StatusReport(task=current_task.id)) execution = self.get_execution_info(current_task) if self._session.config.get("agent.package_manager.force_repo_requirements_txt", False): requirements = None print("[package_manager.force_repo_requirements_txt=true] " "Skipping requirements, using repository \"requirements.txt\" ") else: try: requirements = current_task.script.requirements except AttributeError: requirements = None try: python_ver = current_task.script.binary python_ver = python_ver.split('/')[-1].replace('python', '') # if we can cast it, we are good python_ver = '{:.1f}'.format(float(python_ver)) except: python_ver = None venv_folder, requirements_manager, is_cached = self.install_virtualenv( standalone_mode=standalone_mode, requested_python_version=python_ver, execution_info=execution, cached_requirements=requirements) if not is_cached and not standalone_mode: if self._default_pip: self.package_api.install_packages(*self._default_pip) print("\n") directory, vcs, repo_info = self.get_repo_info( execution, current_task, venv_folder ) print("\n") cwd = vcs.location if vcs and vcs.location else directory if is_cached and not standalone_mode: # reinstalling git / local packages package_api = copy(self.package_api) OnlyExternalRequirements.cwd = package_api.cwd = cwd package_api.requirements_manager = self._get_requirements_manager( base_interpreter=package_api.requirements_manager.get_interpreter(), requirement_substitutions=[OnlyExternalRequirements] ) # make sure we run the handlers cached_requirements = \ {k: package_api.requirements_manager.replace(requirements[k] or '') for k in requirements} if str(cached_requirements.get('pip', '')).strip() \ or str(cached_requirements.get('conda', '')).strip(): package_api.load_requirements(cached_requirements) # make sure we call the correct freeze requirements_manager = package_api.requirements_manager elif not is_cached and not standalone_mode: self.install_requirements( execution, repo_info, requirements_manager=requirements_manager, cached_requirements=requirements, cwd=cwd, ) # do not update the task packages if we are using conda, # it will most likely make the task environment unreproducible skip_freeze_update = self.is_conda and not self._session.config.get( "agent.package_manager.conda_full_env_update", False) freeze = self.freeze_task_environment( task_id=current_task.id, requirements_manager=requirements_manager, add_venv_folder_cache=venv_folder, execution_info=execution, update_requirements=not skip_freeze_update, ) script_dir = (directory if isinstance(directory, Path) else Path(directory)).absolute().as_posix() # run code # print("Running task id [%s]:" % current_task.id) print(self._task_logging_pass_control_message.format(current_task.id)) extra = ['-u', ] if optimization: extra.append( WorkerParams(optimization=optimization).get_optimization_flag() ) # check if this is a module load, then load it. try: if current_task.script.binary and current_task.script.binary.startswith('python') and \ execution.entry_point and execution.entry_point.split()[0].strip() == '-m': # we need to split it import shlex extra.extend(shlex.split(execution.entry_point)) else: extra.append(execution.entry_point) except: extra.append(execution.entry_point) command = self.package_api.get_python_command(extra) print("[{}]$ {}".format(execution.working_dir, command.pretty())) if freeze: print("Summary - installed python packages:") print(dump_yaml(freeze)) else: print("No freeze information available") print("Environment setup completed successfully\n") sdk_env = { # config_file updated in session.py "task_id": current_task.id, "log_level": log_level, "log_to_backend": "0", "config_file": self._session.config_file, # The config file is the tmp file that clearml_agent created } os.environ.update( { sdk_key: str(value) for key, value in sdk_env.items() for sdk_key in ENVIRONMENT_SDK_PARAMS[key] } ) if repo_info: self._update_commit_id(current_task.id, execution, repo_info) # get Task Environments and update the process if self._session.config.get('agent.enable_task_env', None): hyper_params = self._get_task_os_env(current_task) if hyper_params: os.environ.update(hyper_params) # Add the script CWD to the python path python_path = get_python_path(script_dir, execution.entry_point, self.package_api, is_conda_env=self.is_conda) if ENV_TASK_EXTRA_PYTHON_PATH.get(): python_path = add_python_path(python_path, ENV_TASK_EXTRA_PYTHON_PATH.get()) if python_path: os.environ['PYTHONPATH'] = python_path # check if we want to run as another user, only supported on linux if ENV_TASK_EXECUTE_AS_USER.get() and is_linux_platform(): command, script_dir = self._run_as_user_patch( command, self._session.config_file, script_dir, venv_folder, self._session.config.get('sdk.storage.cache.default_base_dir'), ENV_TASK_EXECUTE_AS_USER.get()) use_execv = False else: use_execv = is_linux_platform() and not isinstance(self.package_api, (PoetryAPI, CondaAPI)) print("Starting Task Execution:\n".format(current_task.id)) exit_code = -1 try: if disable_monitoring: try: sys.stdout.flush() sys.stderr.flush() os.chdir(script_dir) if use_execv: os.execv(command.argv[0].as_posix(), tuple([command.argv[0].as_posix()])+command.argv[1:]) else: exit_code = command.check_call(cwd=script_dir) exit(exit_code) except subprocess.CalledProcessError as ex: # non zero return code exit_code = ex.returncode if not use_execv: exit(exit_code) except Exception as ex: if not use_execv: exit(-1) raise ex else: # store stdout/stderr into file, and send to backend temp_stdout_fname = log_file or safe_mkstemp( suffix=".txt", prefix=".clearml_agent_out.", name_only=True ) print("Storing stdout and stderr log into [%s]" % temp_stdout_fname) exit_code, _ = self._log_command_output( task_id=current_task.id, cmd=command, stdout_path=temp_stdout_fname, cwd=script_dir, ) except KeyboardInterrupt: self.handle_user_abort(current_task.id) raise except Exception as e: self.log.warning(str(e)) self.log_traceback(e) exit_code = -1 # kill leftover processes kill_all_child_processes() # if we return ExitStatus.interrupted==2, # it means user aborted, KeyboardInterrupt should have caught it, # that cannot happen when running with disable monitoring exit_code = exit_code if exit_code != ExitStatus.interrupted else -1 if not disable_monitoring: # we need to change task status according to exit code self.handle_task_termination(current_task.id, exit_code, TaskStopReason.no_stop) self.stop_monitor() # unregister the worker self._unregister() return 1 if exit_code is None else exit_code def _get_task_os_env(self, current_task): if not self._session.check_min_api_version('2.9'): return None # noinspection PyBroadException try: hyper_params = self._session.get( service="tasks", action="get_hyper_params", tasks=[current_task.id]) hyper_params = { str(p['name']): str(p['value']) for p in hyper_params['params'][0]['hyperparams'] if p['section'] == 'Environment'} return hyper_params except Exception: return None def set_docker_variables(self, docker): temp_config, docker_image_func = self.get_docker_config_cmd(docker) self.dump_config(self.temp_config_path, config=temp_config) self.docker_image_func = docker_image_func def get_execution_info(self, current_task): # type: (...) -> ExecutionInfo try: execution = ExecutionInfo.from_task(current_task) except Exception as e: self.error("Could not parse task execution info: {}".format(e.args[0])) current_task.failed( status_reason=e.args[0], status_message=self._task_status_change_message ) self.exit(e.args[0]) if "\\" in execution.working_dir: warning( 'Working dir "{}" contains backslashes. ' "All path separators must be forward slashes.".format( execution.working_dir ) ) print("Executing task id [%s]:" % current_task.id) for pair in attr.asdict(execution).items(): print("{} = {}".format(*pair)) print() return execution def get_repo_info(self, execution, task, venv_folder): # type: (ExecutionInfo, tasks_api.Task, str) -> Tuple[str, Optional[VCS], Optional[RepoInfo]] literal_script = LiteralScriptManager(venv_folder) has_repository = bool(execution.repository) is_literal_script = literal_script.is_literal_script(task) if not has_repository and not is_literal_script: raise CommandFailedError( "Can not run task without repository or literal script in `script.diff`" ) repo_info = None directory = None vcs = None script_file = None if has_repository: vcs, repo_info = self._get_repo_info(execution, task, venv_folder) directory = Path(repo_info.root, execution.working_dir or ".") for pair in attr.asdict(repo_info).items(): print("{}: {}".format(*pair)) if not is_literal_script: self.apply_diff( task=task, vcs=vcs, execution_info=execution, repo_info=repo_info ) script_file = Path(directory) / execution.entry_point if is_literal_script: self.log.info("found literal script in `script.diff`") directory, script = literal_script.create_notebook_file( task, execution, repo_info ) execution.entry_point = script script_file = Path(execution.entry_point) else: # in case of no literal script, there is not difference between empty working dir and `.` execution.working_dir = execution.working_dir or "." # fix our import patch (in case we have __future__) if script_file and script_file.is_file(): fix_package_import_diff_patch(script_file.as_posix()) if is_literal_script and not has_repository: return directory, None, None if not directory: assert False, "unreachable code" return directory, vcs, repo_info def _get_repo_info(self, execution, task, venv_folder): try: self._session.config.put("agent.standalone_mode", self._standalone_mode) vcs, repo_info = clone_repository_cached( session=self._session, execution=execution, destination=Path(venv_folder) / WORKING_REPOSITORY_DIR, ) except CommandFailedError: raise except Exception as ex: print('Repository cloning failed: {}'.format(ex)) task.failed( status_reason="failed cloning repository", status_message=self._task_status_change_message, ) if self._session.debug_mode: raise raise CommandFailedError( "Failed cloning repository. \n" "1) Make sure you pushed the requested commit:\n{}\n" "2) Check if remote-worker has valid credentials [see worker configuration file]".format( str(execution).replace('ExecutionInfo', '', 1).replace('version_num', 'commit_id', 1)) ) return vcs, repo_info def handle_task_termination(self, task_id, exit_code, stop_reason): # type: (Text, int, TaskStopReason) -> None try: if stop_reason == TaskStopReason.stopped: self.log("Stopping - tasks.stop was called for task") self.send_logs(task_id, ["Process aborted by user"]) self._session.api_client.tasks.stopped( task=task_id, status_reason="task was stopped by tasks.stop", status_message=self._task_status_change_message, ) elif stop_reason == TaskStopReason.status_changed: try: task_status = get_task( self._session, task_id, only_fields=["status"] ).status self.log( "Task status changed unexpectedly (status: {}), " "terminating task process.".format(task_status) ) except Exception as ex: self.log( "Task status changed unexpectedly. Was not able to obtain the current status: " "{}: {}".format(type(ex), ex) ) self.log_traceback(ex) elif stop_reason == TaskStopReason.reset: self.log("Task was reset unexpectedly") elif stop_reason == TaskStopReason.no_stop: self.handle_task_process_termination(task_id, exit_code) else: self.log( "INTERNAL ERROR: unidentified task stop reason: {}".format( stop_reason ) ) except Exception as e: # task probably set its own status self.log( "Warning: could not update task id '{}' status. Task exit code {}".format( task_id, exit_code ) ) self.log_traceback(e) def handle_task_process_termination(self, task_id, exit_code): # type: (Text, int) -> None self.log("Task process terminated") if exit_code == COMMAND_SUCCESS: self.log("Task success: completing") self.send_logs(task_id, ["Process completed successfully"]) self._session.api_client.tasks.completed( task=task_id, status_reason="worker execution done", status_message=self._task_status_change_message, ) elif exit_code in (ExitStatus.interrupted, 256+ExitStatus.interrupted): self.log("Task interrupted: stopping") self.send_logs(task_id, ["Process terminated by user"]) self._session.api_client.tasks.stopped( task=task_id, status_reason="user abort", status_message=self._task_status_change_message, ) else: self.log("Task failure: setting status to 'failed'") self.send_logs(task_id, ["Process failed, exit code {}".format(exit_code)]) self._session.api_client.tasks.failed( task=task_id, status_reason="worker execution exit code {}".format(exit_code), status_message=self._task_status_change_message, ) def freeze_task_environment(self, task_id=None, requirements_manager=None, add_venv_folder_cache=None, execution_info=None, update_requirements=False): try: freeze = self.package_api.freeze() except Exception as e: print("Could not freeze installed packages") self.log_traceback(e) return None if requirements_manager: freeze = requirements_manager.replace_back(freeze) if not task_id: return freeze # get original requirements and update with the new frozen requirements previous_reqs = {} # noinspection PyBroadException try: current_task = get_task(self._session, task_id, only_fields=["script.requirements"]) requirements = current_task.script.requirements previous_reqs = dict(**requirements) # replace only once. if requirements.get('pip') and not requirements.get('org_pip'): requirements['org_pip'] = requirements.pop('pip') if requirements.get('conda') and not requirements.get('org_conda'): requirements['org_conda'] = requirements.pop('conda') requirements.update(freeze) except Exception: requirements = freeze # add to cache print('Adding venv into cache: {}'.format(add_venv_folder_cache)) if add_venv_folder_cache: self.package_api.add_cached_venv( requirements=[freeze, previous_reqs], docker_cmd=execution_info.docker_cmd if execution_info else None, python_version=getattr(self.package_api, 'python', ''), cuda_version=self._session.config.get("agent.cuda_version"), source_folder=add_venv_folder_cache, exclude_sub_folders=['task_repository', 'code']) # If do not update back requirements if not update_requirements: return freeze request = tasks_api.SetRequirementsRequest(task=task_id, requirements=requirements) try: self._session.send_api(request) except APIError as e: print("Could not set task requirements") self.log_traceback(e) return freeze def _install_poetry_requirements(self, repo_info): # type: (Optional[RepoInfo]) -> Optional[PoetryAPI] if not repo_info: return None try: if not self.poetry.enabled: return None self.poetry.initialize(cwd=repo_info.root) api = self.poetry.get_api(repo_info.root) if api.enabled: print('Poetry Enabled: Ignoring requested python packages, using repository poetry lock file!') api.install() return api except Exception: self.log.error("failed installing poetry requirements:") return None def install_requirements( self, execution, repo_info, requirements_manager, cached_requirements=None, cwd=None, package_api=None ): ExternalRequirements.cwd = cwd return self.install_requirements_for_package_api(execution, repo_info, requirements_manager, cached_requirements=cached_requirements, cwd=cwd, package_api=package_api if package_api else self.package_api) def install_requirements_for_package_api( self, execution, repo_info, requirements_manager, cached_requirements=None, cwd=None, package_api=None, ): # type: (ExecutionInfo, RepoInfo, RequirementsManager, Optional[dict], Optional[str], Optional[Any]) -> None """ :summary: Install requirements for task script using pip. :description: A file named "requirements.txt" is looked for in each containing folder between the repository root and the directory containing the script. For each such file, CUDA based packages are looked for and replaced if necessary. :param execution: task execution information :param repo_info: repository information :param requirements_manager: requirements manager for task :param cached_requirements: cached requirements from previous run :param cwd: current folder :param package_api: package_api to be used when installing requirements """ if package_api: package_api.cwd = cwd api = self._install_poetry_requirements(repo_info) if api: package_api = api return package_api.upgrade_pip() package_api.set_selected_package_manager() # always install cython, # if we have a specific version in the requirements, # the PriorityPackageRequirement(SimpleSubstitution) will reinstall cython with the specific version if not self.is_conda: package_api.out_of_scope_install_package('Cython') cached_requirements_failed = False if cached_requirements and (cached_requirements.get('pip') is not None or cached_requirements.get('conda') is not None): self.log("Found task requirements section, trying to install") try: package_api.load_requirements(cached_requirements) except Exception as e: self.log_traceback(e) cached_requirements_failed = True raise ValueError("Could not install task requirements!\n{}".format(e)) else: self.log("task requirements installation passed") return if not repo_info: if cached_requirements_failed: raise ValueError("Could not install task requirements!") self.log("no repository to install requirements from") return repo_requirements_installed = False for parent in reversed( (Path(execution.working_dir) / execution.entry_point).parents ): requirements_file = ( repo_info.root / parent / "requirements.txt" ) # type: Path if not requirements_file.is_file(): continue print("Trying pip install: {}".format(requirements_file)) requirements_text = requirements_file.read_text() new_requirements = requirements_manager.replace(requirements_text) temp_file = None try: with self.named_temporary_file( mode="w", prefix="requirements_", suffix=".txt", delete=False ) as temp_file: temp_file.write(new_requirements) temp_file.flush() # close the file before reading in install_from_file for Windows compatibility package_api.install_from_file(temp_file.name) except Exception as e: print('ERROR: Failed installing requirements.txt:\n{}'.format(requirements_text)) raise e finally: if self._session.debug_mode and temp_file: rm_file(temp_file.name) # call post installation callback requirements_manager.post_install(self._session) # mark as successful installation repo_requirements_installed = True # if we reached here without installing anything, and # we failed installing from cached requirements, them this is an error if cached_requirements_failed and not repo_requirements_installed: raise ValueError("Failed installing task requirements and repository requirements") def named_temporary_file(self, *args, **kwargs): kwargs.setdefault("delete", not self._session.debug_mode) return named_temporary_file(*args, **kwargs) def list(self, **_): # display all registered working machines workers = self.get("get_all", last_seen=0) print_parameters(workers) def _register(self, queues=()): self.queues = queues try: self.get("register", worker=self.worker_id, queues=queues) # If we got here - we've registered self._worker_registered = True # except APIError as error: # if error.codes != WORKER_ALREADY_REGISTERED: # raise # except ValueError: # raise ValueError("Worker cannot register itself with backend service") except Exception as e: self.log("Worker failed registering itself with backend service: {}".format(e)) def _unregister(self, queues=()): self.queues = () try: self.get("unregister", worker=self.worker_id, queues=queues) self._worker_registered = False except Exception as e: self.log("Worker failed unregistering itself with backend service: {}".format(e)) def log_traceback(self, err): if isinstance(err, APIError): tb = err.get_traceback() if tb: print("Server traceback:\n{}".format(tb)) if self._session.debug_mode: self.log(traceback.format_exc()) def debug(self, message): if self._session.debug_mode: print("clearml_agent: {}".format(message)) def find_python_executable_for_version(self, config_version): # type: (Text) -> Tuple[Text, Text, Text] """ Find python executable with version ``config_version``. The search starts with the executable named ``python``. If not found in the path, increasingly major components of the semantic version are dropped with the last executable to be searched being a plain ``python``. A result is returned only if ``python --version``'s output matches config_version. For example: if config_version=3.6.5, the search order is: 1. python3.6.5 2. python3.6 3. python3 4. python if config_version=3.6, the search starts with ``python3.6`` and so on. :return: 3-tuple: 1. Python executable version as reported by itself with ``--version`` 2. The version suffix of the executable name (e.g. ``python3`` -> ``3``) 3. The executable name itself (e.g. ``python3``) """ def suffixes(it): it = list(it) for i in range(len(it) + 1): yield it[:i] def rreplace(s, old, new, count): return (s[::-1].replace(old[::-1], new[::-1], count))[::-1] if is_windows_platform(): python_executables = [ (version, config_version if os.path.sep in config_version else 'python{}'.format(version)) for version in map( ".".join, reversed(list(suffixes( rreplace( rreplace(config_version.split(os.path.sep)[-1].lower(), 'python', '', 1), '.exe', '', 1).split(".")))) ) ] else: python_executables = [ (version, config_version if os.path.sep in config_version else 'python{}'.format(version)) for version in map( ".".join, reversed(list(suffixes( rreplace(config_version.split(os.path.sep)[-1], 'python', '', 1).split(".")))) ) ] for version, executable in python_executables: self.log.debug("Searching for {}".format(executable)) if find_executable(executable): try: output = Argv(executable, "--version").get_output( stderr=subprocess.STDOUT ) except subprocess.CalledProcessError as ex: self.log.warning("error getting %s version: %s", executable, ex) continue match = re.search( r"Python ({}(?:\.\d+)*)".format( r"\d+" if not config_version or os.path.sep in config_version else config_version), output ) if match: self.log.debug("Found: {}".format(executable)) return match.group(1), version or '.'.join(match.group(1).split('.')[:2]), executable raise CommandFailedError( "Python executable with version {!r} defined in configuration file, " "key 'agent.default_python', not found in path, tried: {}".format( config_version, list(zip(*python_executables))[1] ) ) def install_virtualenv( self, venv_dir=None, requested_python_version=None, standalone_mode=False, execution_info=None, cached_requirements=None): # type: (str, str, bool, ExecutionInfo, dict) -> Tuple[Path, RequirementsManager, bool] """ Install a new python virtual environment, removing the old one if exists :return: virtualenv directory, requirements manager to use with task, True if there is a cached venv entry """ if self._session.config.get("agent.ignore_requested_python_version", None): requested_python_version = '' requested_python_version = \ requested_python_version or \ Text(self._session.config.get("agent.python_binary", None)) or \ Text(self._session.config.get("agent.default_python", None)) if self.is_conda: executable_version_suffix = \ requested_python_version[max(requested_python_version.find('python'), 0):].replace('python', '') executable_name = 'python' else: try: executable_version, executable_version_suffix, executable_name = \ self.find_python_executable_for_version(requested_python_version) except Exception: def_python_version = Text(self._session.config.get("agent.python_binary", None)) or \ Text(self._session.config.get("agent.default_python", None)) print('Warning: could not locate requested Python version {}, reverting to version {}'.format( requested_python_version, def_python_version)) executable_version, executable_version_suffix, executable_name = \ self.find_python_executable_for_version(def_python_version) self._session.config.put("agent.default_python", executable_version) self._session.config.put("agent.python_binary", executable_name) venv_dir = Path(venv_dir) if venv_dir else \ Path(self._session.config["agent.venvs_dir"], executable_version_suffix) first_time = not standalone_mode and ( is_windows_platform() or self.is_conda or not venv_dir.is_dir() or not self.is_venv_update ) requirements_manager = self._get_requirements_manager( base_interpreter=executable_name ) if not standalone_mode: rm_tree(normalize_path(venv_dir, WORKING_REPOSITORY_DIR)) package_manager_params = dict( session=self._session, python=executable_version_suffix if self.is_conda else executable_name, path=venv_dir, requirements_manager=requirements_manager, execution_info=execution_info, ) global_package_manager_params = dict( interpreter=executable_name, session=self._session, ) call_package_manager_create = False if not self.is_conda and standalone_mode: # pip with standalone mode get_pip = partial(VirtualenvPip, **package_manager_params) self.package_api = get_pip() self.global_package_api = SystemPip(**global_package_manager_params) elif not self.is_conda: if self.is_venv_update: self.package_api = VenvUpdateAPI( url=self._session.config["agent.venv_update.url"] or DEFAULT_VENV_UPDATE_URL, **package_manager_params ) else: self.package_api = VirtualenvPip(**package_manager_params) if first_time: self.package_api.remove() call_package_manager_create = True self.global_package_api = SystemPip(**global_package_manager_params) elif standalone_mode: # conda with standalone mode get_conda = partial(CondaAPI, **package_manager_params) self.package_api = get_conda() else: get_conda = partial(CondaAPI, **package_manager_params) self.package_api = get_conda() # no support for reusing Conda environments self.package_api.remove() call_package_manager_create = True if venv_dir.exists(): timestamp = datetime.utcnow().strftime("%Y-%m-%d-%H-%M-%S") new_venv_folder = venv_dir.with_name( "{}_{}".format(venv_dir.name, timestamp) ) self.warning( 'Path "{}" exists, using "{}" instead'.format( venv_dir, new_venv_folder ) ) venv_dir = new_venv_folder self.package_api = get_conda(path=venv_dir) # check if we have a cached folder if cached_requirements and self.package_api.get_cached_venv( requirements=cached_requirements, docker_cmd=execution_info.docker_cmd if execution_info else None, python_version=package_manager_params['python'], cuda_version=self._session.config.get("agent.cuda_version"), destination_folder=Path(venv_dir) ): print('::: Using Cached environment {} :::'.format(self.package_api.get_last_used_entry_cache())) return venv_dir, requirements_manager, True # create the initial venv if call_package_manager_create: self.package_api.create() return venv_dir, requirements_manager, False def parse_requirements(self, reqs_file=None, overrides=None): os = None session = self._session if overrides: overrides = ConfigFactory.parse_string("\n".join(overrides)) os = overrides.pop("os", None) ConfigTree.merge_configs(session.config, overrides) if reqs_file: contents = Path(reqs_file).read_text() else: contents = ensure_text(sys.stdin.read()) session.finalize_config(session.config) requirements_manager = self._get_requirements_manager(os_override=os) requirements_manager.translator.enabled = False print(requirements_manager.replace(contents)) def get_docker_config_cmd(self, docker_args): docker_image = str(ENV_DOCKER_IMAGE.get() or self._session.config.get("agent.default_docker.image", "nvidia/cuda")) \ if not docker_args else docker_args[0] docker_arguments = docker_image.split(' ') if docker_image else [] if len(docker_arguments) > 1: docker_image = docker_arguments[0] docker_arguments = docker_arguments[1:] else: docker_arguments = self._session.config.get("agent.default_docker.arguments", None) or [] if isinstance(docker_arguments, six.string_types): docker_arguments = [docker_arguments] # store docker arguments self._docker_image = docker_image self._docker_arguments = docker_arguments print("Running in Docker {} mode (v19.03 and above) - using default docker image: {} {}\n".format( '*standalone*' if self._standalone_mode else '', self._docker_image, self._docker_arguments or '')) temp_config = deepcopy(self._session.config) mounted_cache_dir = self._docker_fixed_user_cache # '/root/.clearml/cache' mounted_pip_dl_dir = '/root/.clearml/pip-download-cache' mounted_vcs_cache = '/root/.clearml/vcs-cache' mounted_venv_dir = '/root/.clearml/venvs-builds' temp_config.put("sdk.storage.cache.default_base_dir", mounted_cache_dir) temp_config.put("agent.pip_download_cache.path", mounted_pip_dl_dir) temp_config.put("agent.vcs_cache.path", mounted_vcs_cache) temp_config.put("agent.package_manager.system_site_packages", True) temp_config.put("agent.package_manager.conda_env_as_base_docker", False) temp_config.put("agent.default_python", "") temp_config.put("agent.python_binary", "") temp_config.put("agent.cuda_version", "") temp_config.put("agent.cudnn_version", "") temp_config.put("agent.venvs_dir", mounted_venv_dir) temp_config.put("agent.git_user", (ENV_AGENT_GIT_USER.get() or self._session.config.get("agent.git_user", None))) temp_config.put("agent.git_pass", (ENV_AGENT_GIT_PASS.get() or self._session.config.get("agent.git_pass", None))) if temp_config.get("agent.venvs_cache.path", None): temp_config.put("agent.venvs_cache.path", '/root/.clearml/venvs-cache') self._host_ssh_cache = mkdtemp(prefix='clearml_agent.ssh.') self._temp_cleanup_list.append(self._host_ssh_cache) return temp_config, partial(self._get_docker_config_cmd, temp_config=temp_config) def _get_docker_config_cmd(self, temp_config, **kwargs): host_cache = Path(os.path.expandvars( self._session.config["sdk.storage.cache.default_base_dir"])).expanduser().as_posix() host_pip_dl = Path(os.path.expandvars( self._session.config["agent.pip_download_cache.path"])).expanduser().as_posix() host_vcs_cache = Path(os.path.expandvars( self._session.config["agent.vcs_cache.path"])).expanduser().as_posix() host_venvs_cache = Path(os.path.expandvars( self._session.config["agent.venvs_cache.path"])).expanduser().as_posix() \ if self._session.config.get("agent.venvs_cache.path", None) else None host_ssh_cache = self._host_ssh_cache host_apt_cache = Path(os.path.expandvars(self._session.config.get( "agent.docker_apt_cache", '~/.clearml/apt-cache'))).expanduser().as_posix() host_pip_cache = Path(os.path.expandvars(self._session.config.get( "agent.docker_pip_cache", '~/.clearml/pip-cache'))).expanduser().as_posix() # make sure all folders are valid Path(host_apt_cache).mkdir(parents=True, exist_ok=True) Path(host_pip_cache).mkdir(parents=True, exist_ok=True) Path(host_cache).mkdir(parents=True, exist_ok=True) Path(host_pip_dl).mkdir(parents=True, exist_ok=True) Path(host_vcs_cache).mkdir(parents=True, exist_ok=True) Path(host_ssh_cache).mkdir(parents=True, exist_ok=True) if host_venvs_cache: Path(host_venvs_cache).mkdir(parents=True, exist_ok=True) # copy the .ssh folder to a temp folder, to be mapped into docker # noinspection PyBroadException try: if Path(host_ssh_cache).is_dir(): shutil.rmtree(host_ssh_cache, ignore_errors=True) shutil.copytree(Path('~/.ssh').expanduser().as_posix(), host_ssh_cache) except Exception: host_ssh_cache = None self.log.warning('Failed creating temporary copy of ~/.ssh for git credential') # check if the .git credentials exist: try: host_git_credentials = [ f.as_posix() for f in [Path('~/.git-credentials').expanduser(), Path('~/.gitconfig').expanduser()] if f.is_file()] except Exception: host_git_credentials = None extra_shell_script_str = "" if self._extra_shell_script: cmds = self._extra_shell_script if not isinstance(cmds, (list, tuple)): cmds = [cmds] extra_shell_script_str = " ; ".join(map(str, cmds)) + " ; " bash_script = self._session.config.get("agent.docker_init_bash_script", None) preprocess_bash_script = self._session.config.get("agent.docker_preprocess_bash_script", None) install_opencv_libs = self._session.config.get("agent.docker_install_opencv_libs", True) self.temp_config_path = self.temp_config_path or safe_mkstemp( suffix=".cfg", prefix=".clearml_agent.", text=True, name_only=True ) mounted_cache_dir = temp_config.get("sdk.storage.cache.default_base_dir") mounted_pip_dl_dir = temp_config.get("agent.pip_download_cache.path") mounted_vcs_cache = temp_config.get("agent.vcs_cache.path") mounted_venvs_cache = temp_config.get("agent.venvs_cache.path", "") # Make sure we have created the configuration file for the executor if not self.dump_config(self.temp_config_path, config=temp_config): self.log.warning('Could not update docker configuration file {}'.format(self.temp_config_path)) docker_cmd = dict( worker_id=self.worker_id, # docker_image=docker_image, # docker_arguments=docker_arguments, extra_docker_arguments=self._extra_docker_arguments, extra_shell_script=extra_shell_script_str, python_version='python3', conf_file=self.temp_config_path, host_apt_cache=host_apt_cache, host_pip_cache=host_pip_cache, host_ssh_cache=host_ssh_cache, host_git_credentials=host_git_credentials, host_cache=host_cache, mounted_cache=mounted_cache_dir, host_pip_dl=host_pip_dl, mounted_pip_dl=mounted_pip_dl_dir, host_vcs_cache=host_vcs_cache, mounted_vcs_cache=mounted_vcs_cache, host_venvs_cache=host_venvs_cache, mounted_venvs_cache=mounted_venvs_cache, standalone_mode=self._standalone_mode, force_current_version=self._force_current_version, bash_script=bash_script, preprocess_bash_script=preprocess_bash_script, install_opencv_libs=install_opencv_libs, ) docker_cmd.update(kwargs) return self._get_docker_cmd(**docker_cmd) @staticmethod def _get_docker_cmd( worker_id, docker_image, docker_arguments, python_version, conf_file, host_apt_cache, host_pip_cache, host_ssh_cache, host_cache, mounted_cache, host_pip_dl, mounted_pip_dl, host_vcs_cache, mounted_vcs_cache, host_venvs_cache, mounted_venvs_cache, standalone_mode=False, extra_docker_arguments=None, extra_shell_script=None, force_current_version=None, host_git_credentials=None, bash_script=None, preprocess_bash_script=None, install_opencv_libs=None, docker_bash_setup_script=None, ): docker = 'docker' base_cmd = [docker, 'run', '-t'] update_scheme = "" dockers_nvidia_visible_devices = 'all' gpu_devices = os.environ.get('NVIDIA_VISIBLE_DEVICES', None) if gpu_devices is None or gpu_devices.lower().strip() == 'all': if ENV_DOCKER_SKIP_GPUS_FLAG.get(): dockers_nvidia_visible_devices = os.environ.get('NVIDIA_VISIBLE_DEVICES') or \ dockers_nvidia_visible_devices else: base_cmd += ['--gpus', 'all', ] elif gpu_devices.strip() and gpu_devices.strip() != 'none': if ENV_DOCKER_SKIP_GPUS_FLAG.get(): dockers_nvidia_visible_devices = gpu_devices else: base_cmd += ['--gpus', '\"device={}\"'.format(gpu_devices), ] # We are using --gpu, so we should not pass NVIDIA_VISIBLE_DEVICES, I think. # base_cmd += ['-e', 'NVIDIA_VISIBLE_DEVICES=' + gpu_devices, ] elif gpu_devices.strip() == 'none': dockers_nvidia_visible_devices = gpu_devices if docker_arguments: docker_arguments = list(docker_arguments) \ if isinstance(docker_arguments, (list, tuple)) else [docker_arguments] base_cmd += [a for a in docker_arguments if a] if extra_docker_arguments: extra_docker_arguments = [extra_docker_arguments] \ if isinstance(extra_docker_arguments, six.string_types) else extra_docker_arguments base_cmd += [str(a) for a in extra_docker_arguments if a] # check if running inside a kubernetes if ENV_DOCKER_HOST_MOUNT.get() or (os.environ.get('KUBERNETES_SERVICE_HOST') and os.environ.get('KUBERNETES_PORT')): # map network to sibling docker, unless we have other network argument if not any(a.strip().startswith('--network') for a in base_cmd): # noinspection PyBroadException try: network_mode = get_bash_output( 'docker inspect --format=\'{{.HostConfig.NetworkMode}}\' $(basename $(cat /proc/1/cpuset))') if network_mode: base_cmd += ['--network', network_mode] except Exception: pass base_cmd += ['-e', 'NVIDIA_VISIBLE_DEVICES={}'.format(dockers_nvidia_visible_devices)] # check if we need to map host folders if ENV_DOCKER_HOST_MOUNT.get(): # expect CLEARML_AGENT_K8S_HOST_MOUNT = '/mnt/host/data:/root/.clearml' k8s_node_mnt, _, k8s_pod_mnt = ENV_DOCKER_HOST_MOUNT.get().partition(':') # search and replace all the host folders with the k8s host_mounts = [host_apt_cache, host_pip_cache, host_pip_dl, host_cache, host_vcs_cache, host_venvs_cache] for i, m in enumerate(host_mounts): if not m: continue if k8s_pod_mnt not in m: print('Warning: K8S mount missing, ignoring cached folder {}'.format(m)) host_mounts[i] = None else: host_mounts[i] = m.replace(k8s_pod_mnt, k8s_node_mnt, 1) host_apt_cache, host_pip_cache, host_pip_dl, host_cache, host_vcs_cache, host_venvs_cache = host_mounts # copy the configuration file into the mounted folder new_conf_file = os.path.join(k8s_pod_mnt, '.clearml_agent.{}.cfg'.format(quote(worker_id, safe=""))) try: rm_tree(new_conf_file) rm_file(new_conf_file) shutil.copy(conf_file, new_conf_file) conf_file = new_conf_file.replace(k8s_pod_mnt, k8s_node_mnt) except Exception: raise ValueError('Error: could not copy configuration file into: {}'.format(new_conf_file)) if host_ssh_cache: new_ssh_cache = os.path.join(k8s_pod_mnt, '.clearml_agent.{}.ssh'.format(quote(worker_id, safe=""))) try: rm_tree(new_ssh_cache) shutil.copytree(host_ssh_cache, new_ssh_cache) host_ssh_cache = new_ssh_cache.replace(k8s_pod_mnt, k8s_node_mnt) except Exception: raise ValueError('Error: could not copy .ssh directory into: {}'.format(new_ssh_cache)) base_cmd += ['-e', 'CLEARML_WORKER_ID='+worker_id, ] # update the docker image, so the system knows where it runs base_cmd += ['-e', 'CLEARML_DOCKER_IMAGE={} {}'.format(docker_image, ' '.join(docker_arguments)).strip()] # if we are running a RC version, install the same version in the docker # because the default latest, will be a release version (not RC) specify_version = '' try: from clearml_agent.version import __version__ _version_parts = __version__.split('.') if force_current_version or 'rc' in _version_parts[-1].lower() or 'rc' in _version_parts[-2].lower(): specify_version = '=={}'.format(__version__) except: pass if os.environ.get('FORCE_LOCAL_CLEARML_AGENT_WHEEL'): local_wheel = os.path.expanduser(os.environ.get('FORCE_LOCAL_CLEARML_AGENT_WHEEL')) docker_wheel = str(Path('/tmp') / Path(local_wheel).name) base_cmd += ['-v', local_wheel + ':' + docker_wheel] clearml_agent_wheel = '\"{}\"'.format(docker_wheel) else: # clearml-agent{specify_version} clearml_agent_wheel = 'clearml-agent{specify_version}'.format(specify_version=specify_version) if not standalone_mode: if not bash_script: # Find the highest python version installed, or install from apt-get # python+pip is the requirement to match bash_script = [ "echo 'Binary::apt::APT::Keep-Downloaded-Packages \"true\";' > /etc/apt/apt.conf.d/docker-clean", "chown -R root /root/.cache/pip", "export DEBIAN_FRONTEND=noninteractive", "export CLEARML_APT_INSTALL=\"$CLEARML_APT_INSTALL{}\"".format( ' libsm6 libxext6 libxrender-dev libglib2.0-0' if install_opencv_libs else ""), "[ ! -z $(which git) ] || export CLEARML_APT_INSTALL=\"$CLEARML_APT_INSTALL git\"", "declare LOCAL_PYTHON", "for i in {{10..5}}; do which {python_single_digit}.$i && " + "{python_single_digit}.$i -m pip --version && " + "export LOCAL_PYTHON=$(which {python_single_digit}.$i) && break ; done", "[ ! -z $LOCAL_PYTHON ] || export CLEARML_APT_INSTALL=\"$CLEARML_APT_INSTALL {python_single_digit}-pip\"", # noqa "[ -z \"$CLEARML_APT_INSTALL\" ] || (apt-get update && apt-get install -y $CLEARML_APT_INSTALL)", ] if preprocess_bash_script: bash_script = preprocess_bash_script + bash_script docker_bash_script = " ; ".join([line for line in bash_script if line]) \ if not isinstance(bash_script, str) else bash_script # make sure that if we do not have $LOCAL_PYTHON defined # we set it to python3 update_scheme += ( docker_bash_script + " ; " + "[ ! -z $LOCAL_PYTHON ] || export LOCAL_PYTHON={python} ; " + "$LOCAL_PYTHON -m pip install -U \"pip{pip_version}\" ; " + "$LOCAL_PYTHON -m pip install -U {clearml_agent_wheel} ; ").format( python_single_digit=python_version.split('.')[0], python=python_version, pip_version=PackageManager.get_pip_version(), clearml_agent_wheel=clearml_agent_wheel) if host_git_credentials: for git_credentials in host_git_credentials: base_cmd += ['-v', '{}:/root/{}'.format(git_credentials, Path(git_credentials).name)] if docker_bash_setup_script and docker_bash_setup_script.strip('\n '): extra_shell_script = (extra_shell_script or '') + \ ' ; '.join(line.strip().replace('\"', '\\\"') for line in docker_bash_setup_script.split('\n') if line.strip()) + \ ' ; ' base_cmd += ( ['-v', conf_file+':'+DOCKER_ROOT_CONF_FILE] + (['-v', host_ssh_cache+':/root/.ssh'] if host_ssh_cache else []) + (['-v', host_apt_cache+':/var/cache/apt/archives'] if host_apt_cache else []) + (['-v', host_pip_cache+':/root/.cache/pip'] if host_pip_cache else []) + (['-v', host_pip_dl+':'+mounted_pip_dl] if host_pip_dl else []) + (['-v', host_cache+':'+mounted_cache] if host_cache else []) + (['-v', host_vcs_cache+':'+mounted_vcs_cache] if host_vcs_cache else []) + (['-v', host_venvs_cache + ':' + mounted_venvs_cache] if host_venvs_cache else []) + ['--rm', docker_image, 'bash', '-c', update_scheme + extra_shell_script + "cp {} {} ; ".format(DOCKER_ROOT_CONF_FILE, DOCKER_DEFAULT_CONF_FILE) + "NVIDIA_VISIBLE_DEVICES={nv_visible} $LOCAL_PYTHON -u -m clearml_agent ".format( nv_visible=dockers_nvidia_visible_devices, python=python_version) ]) return base_cmd def _run_as_user_patch(self, command, trains_conf, script_dir, venv_folder, sdk_cache_folder, user_uid): class RunasArgv(Argv): def __init__(self, *args): super(RunasArgv, self).__init__(*args) self.uid = 0 self.gid = 0 def call_subprocess(self, func, censor_password=False, *args, **kwargs): self._log.debug("running: %s: %s", func.__name__, list(self)) with self.normalize_exception(censor_password): return func(list(self), *args, preexec_fn=self._change_uid, **kwargs) def set_uid(self, user_uid, user_gid): from pwd import getpwnam try: self.uid = getpwnam(user_uid).pw_uid self.gid = getpwnam(user_gid).pw_gid except Exception: raise ValueError("Could not find requested user uid={} gid={}".format(user_uid, user_gid)) def _change_uid(self): os.setgid(self.gid) os.setuid(self.uid) # create a home folder for our user clearml_agent_home = self._run_as_user_home + '{}'.format( '.'+str(Singleton.get_slot()) if Singleton.get_slot() else '') try: home_folder = self._run_as_user_home rm_tree(home_folder) Path(home_folder).mkdir(parents=True, exist_ok=True) except: home_folder = os.path.join('/home', self._run_as_user_home) rm_tree(home_folder) Path(home_folder).mkdir(parents=True, exist_ok=True) # move our entire venv into the new home venv_folder = venv_folder.as_posix() if not venv_folder.endswith(os.path.sep): venv_folder += os.path.sep new_venv_folder = os.path.join(home_folder, 'venv/') shutil.move(venv_folder, new_venv_folder) # allow everyone to access it for f in Path(new_venv_folder).rglob('*'): try: f.chmod(0o0777) except: pass # make sure we will be able to access the cache folder (we assume we have the ability change mod) if sdk_cache_folder: sdk_cache_folder = Path(os.path.expandvars(sdk_cache_folder)).expanduser().absolute() try: sdk_cache_folder.chmod(0o0777) except: pass for f in sdk_cache_folder.rglob('*'): try: f.chmod(0o0777) except: pass # make sure we could access the trains_conf file try: user_trains_conf = os.path.join(home_folder, 'clearml.conf') shutil.copy(trains_conf, user_trains_conf) Path(user_trains_conf).chmod(0o0777) except: user_trains_conf = trains_conf # patch venv folder to new location script_dir = script_dir.replace(venv_folder, new_venv_folder) # New command line execution command = RunasArgv('bash', '-c', 'HOME=\"{}\" PATH=\"{}\" PYTHONPATH=\"{}\" TRAINS_CONFIG_FILE={} {}'.format( home_folder, os.environ.get('PATH', '').replace(venv_folder, new_venv_folder), os.environ.get('PYTHONPATH', '').replace(venv_folder, new_venv_folder), user_trains_conf, command.serialize().replace(venv_folder, new_venv_folder))) command.set_uid(user_uid=user_uid, user_gid=user_uid) return command, script_dir def _kill_daemon(self, dynamic_gpus=False): worker_id, worker_name = self._generate_worker_id_name(dynamic_gpus=dynamic_gpus) # Iterate over all running process for pid, uid, slot, file in sorted(Singleton.get_running_pids(), key=lambda x: x[1] or ''): if pid < 0 or uid is None: continue # if dynamic gpus kill all children if dynamic_gpus and uid == worker_id: print('Terminating clearml-agent worker_id={} pid={}'.format(uid, pid)) if not terminate_process(pid, timeout=120): warning('Could not terminate process pid={}'.format(pid)) return True # wither we have a match for the worker_id or we just pick the first one, and kill it. if (worker_id and uid == worker_id) or (not worker_id and uid.startswith('{}:'.format(worker_name))): # this is us kill it print('Terminating clearml-agent worker_id={} pid={}'.format(uid, pid)) timeout = 120 if uid.startswith('{}:dgpu'.format(worker_name)) else 10 if not terminate_process(pid, timeout=timeout): error('Could not terminate process pid={}'.format(pid)) return True print('Could not find a running clearml-agent instance with worker_name={} worker_id={}'.format( worker_name, worker_id)) return False def _singleton(self, dynamic_gpus=False): # ensure singleton worker_id, worker_name = self._generate_worker_id_name(dynamic_gpus=dynamic_gpus) # if we are running in services mode, we allow double register since # docker-compose will kill instances before they cleanup self.worker_id, worker_slot = Singleton.register_instance( unique_worker_id=worker_id, worker_name=worker_name, api_client=self._session.api_client, allow_double=bool(ENV_DOCKER_HOST_MOUNT.get()) # and bool(self._services_mode), ) if self.worker_id is None: error('Instance with the same WORKER_ID [{}] is already running'.format(worker_id)) exit(1) # update folders based on free slot self._session.create_cache_folders(slot_index=worker_slot) def _generate_worker_id_name(self, dynamic_gpus=False): worker_id = self._session.config["agent.worker_id"] worker_name = self._session.config["agent.worker_name"] if not worker_id and os.environ.get('NVIDIA_VISIBLE_DEVICES') is not None: nvidia_visible_devices = os.environ.get('NVIDIA_VISIBLE_DEVICES') if nvidia_visible_devices and nvidia_visible_devices.lower() != 'none': worker_id = '{}:{}gpu{}'.format( worker_name, 'd' if dynamic_gpus else '', nvidia_visible_devices) elif nvidia_visible_devices == '': pass else: worker_name = '{}:cpu'.format(worker_name) return worker_id, worker_name def _resolve_queue_names(self, queues, create_if_missing=False): if not queues: default_queue = self._session.send_api(queues_api.GetDefaultRequest()) return [default_queue.id] queues = return_list(queues) if not create_if_missing: return [self._resolve_name(q if isinstance(q, str) else q.name, "queues") for q in queues] queue_ids = [] for q in queues: try: q_id = self._resolve_name(q if isinstance(q, str) else q.name, "queues") except: self._session.send_api(queues_api.CreateRequest(name=q if isinstance(q, str) else q.name)) q_id = self._resolve_name(q if isinstance(q, str) else q.name, "queues") queue_ids.append(q_id) return queue_ids if __name__ == "__main__": pass