def testStopTrial(self):
ray.init(num_cpus=4, num_gpus=2)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 5
},
"resources": Resources(cpu=1, gpu=1),
}
trials = [
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs)
]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[1].status, Trial.PENDING)
# Stop trial while running
runner.stop_trial(trials[0])
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.PENDING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.PENDING)
# Stop trial while pending
runner.stop_trial(trials[-1])
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.TERMINATED)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[2].status, Trial.RUNNING)
self.assertEqual(trials[-1].status, Trial.TERMINATED)
def has_resources(self, resources):
"""Returns whether this runner has at least the specified resources.
This refreshes the Ray cluster resources if the time since last update
has exceeded self._refresh_period. This also assumes that the
cluster is not resizing very frequently.
"""
if resources.has_placement_group:
return self._pg_manager.can_stage()
self._update_avail_resources()
currently_available = Resources.subtract(self._avail_resources,
self._committed_resources)
have_space = (
resources.cpu_total() <= currently_available.cpu
and resources.gpu_total() <= currently_available.gpu
and resources.memory_total() <= currently_available.memory
and resources.object_store_memory_total() <=
currently_available.object_store_memory and all(
resources.get_res_total(res) <= currently_available.get(res)
for res in resources.custom_resources))
if have_space:
# The assumption right now is that we block all trials if one
# trial is queued.
self._trial_queued = False
return True
can_overcommit = self._queue_trials and not self._trial_queued
if can_overcommit:
self._trial_queued = True
logger.warning(
"Allowing trial to start even though the "
"cluster does not have enough free resources. Trial actors "
"may appear to hang until enough resources are added to the "
"cluster (e.g., via autoscaling). You can disable this "
"behavior by specifying `queue_trials=False` in "
"ray.tune.run().")
return True
return False
def has_resources(self, resources):
"""Returns whether this runner has at least the specified resources.
This refreshes the Ray cluster resources if the time since last update
has exceeded self._refresh_period. This also assumes that the
cluster is not resizing very frequently.
"""
if time.time() - self._last_resource_refresh > self._refresh_period:
self._update_avail_resources()
currently_available = Resources.subtract(self._avail_resources,
self._committed_resources)
have_space = (
resources.cpu_total() <= currently_available.cpu
and resources.gpu_total() <= currently_available.gpu and all(
resources.get_res_total(res) <= currently_available.get(res)
for res in resources.custom_resources))
if have_space:
return True
can_overcommit = self._queue_trials
if (resources.cpu_total() > 0 and currently_available.cpu <= 0) or \
(resources.gpu_total() > 0 and currently_available.gpu <= 0) or \
any((resources.get_res_total(res_name) > 0
and currently_available.get(res_name) <= 0)
for res_name in resources.custom_resources):
can_overcommit = False # requested resource is already saturated
if can_overcommit:
logger.warning(
"Allowing trial to start even though the "
"cluster does not have enough free resources. Trial actors "
"may appear to hang until enough resources are added to the "
"cluster (e.g., via autoscaling). You can disable this "
"behavior by specifying `queue_trials=False` in "
"ray.tune.run().")
return True
return False
def testCheckpointingAtEnd(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"checkpoint_at_end": True,
"resources": Resources(cpu=1, gpu=1),
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process result
runner.step() # Process result, dispatch save
self.assertEqual(trials[0].last_result[DONE], True)
runner.step() # Process save
self.assertEqual(trials[0].has_checkpoint(), True)
def testResourceScheduler(self):
ray.init(num_cpus=4, num_gpus=1)
kwargs = {
"stopping_criterion": {
"training_iteration": 1
},
"resources": Resources(cpu=1, gpu=1),
}
trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)]
snapshot = TrialStatusSnapshot()
runner = TrialRunner(callbacks=[TrialStatusSnapshotTaker(snapshot)])
for t in trials:
runner.add_trial(t)
while not runner.is_finished():
runner.step()
self.assertLess(snapshot.max_running_trials(), 2)
self.assertTrue(snapshot.all_trials_are_terminated())
def testCustomResources(self):
ray.init(num_cpus=4, num_gpus=2, resources={"a": 2})
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 1
},
"resources": Resources(cpu=1, gpu=0, custom_resources={"a": 2}),
}
trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[1].status, Trial.PENDING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.PENDING)
def update_resources(self, resources: Union[Dict, Callable,
PlacementGroupFactory]):
"""EXPERIMENTAL: Updates the resource requirements.
Should only be called when the trial is not running.
Raises:
ValueError if trial status is running.
"""
if self.status is Trial.RUNNING:
raise ValueError("Cannot update resources while Trial is running.")
if isinstance(resources, PlacementGroupFactory):
self.placement_group_factory = resources
else:
self.resources = Resources(**resources)
self._setup_resources()
self.invalidate_json_state()
def testTrialErrorResumeTrue(self):
ray.init(num_cpus=3, local_mode=True, include_dashboard=False)
runner = TrialRunner(local_checkpoint_dir=self.tmpdir)
kwargs = {
"stopping_criterion": {
"training_iteration": 4
},
"resources": Resources(cpu=1, gpu=0),
}
trials = [
Trial("__fake", config={"mock_error": True}, **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
]
for t in trials:
runner.add_trial(t)
while not runner.is_finished():
runner.step()
runner.checkpoint(force=True)
assert trials[0].status == Trial.ERROR
del runner
new_runner = TrialRunner(run_errored_only=True,
resume=True,
local_checkpoint_dir=self.tmpdir)
assert len(new_runner.get_trials()) == 3
assert Trial.ERROR not in (t.status for t in new_runner.get_trials())
# The below is just a check for standard behavior.
disable_error = False
for t in new_runner.get_trials():
if t.config.get("mock_error"):
t.config["mock_error"] = False
disable_error = True
assert disable_error
while not new_runner.is_finished():
new_runner.step()
assert Trial.ERROR not in (t.status for t in new_runner.get_trials())
def testErrorHandling(self):
ray.init(num_cpus=4, num_gpus=2)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 1
},
"resources": Resources(cpu=1, gpu=1),
}
_global_registry.register(TRAINABLE_CLASS, "asdf", None)
trials = [Trial("asdf", **kwargs), Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.ERROR)
self.assertEqual(trials[1].status, Trial.PENDING)
runner.step()
self.assertEqual(trials[0].status, Trial.ERROR)
self.assertEqual(trials[1].status, Trial.RUNNING)
def testPauseThenResume(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"resources": Resources(cpu=1, gpu=1),
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial
runner.step() # Process result
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(ray.get(trials[0].runner.get_info.remote()), None)
self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1)
runner.trial_executor.pause_trial(trials[0])
self.assertEqual(trials[0].status, Trial.PAUSED)
def testRestoreMetricsAfterCheckpointing(self):
ray.init(num_cpus=1, num_gpus=1)
observer = TrialResultObserver()
runner = TrialRunner(callbacks=[observer])
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
while not runner.is_finished():
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
kwargs["restore_path"] = trials[0].checkpoint.dir_or_data
kwargs.pop("stopping_criterion")
kwargs.pop("checkpoint_freq") # No checkpointing for next trial
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
observer.reset()
while not observer.just_received_a_result():
runner.step()
self.assertEqual(trials[1].last_result["timesteps_since_restore"], 10)
self.assertEqual(trials[1].last_result["iterations_since_restore"], 1)
self.assertGreater(trials[1].last_result["time_since_restore"], 0)
while not observer.just_received_a_result():
runner.step()
self.assertEqual(trials[1].last_result["timesteps_since_restore"], 20)
self.assertEqual(trials[1].last_result["iterations_since_restore"], 2)
self.assertGreater(trials[1].last_result["time_since_restore"], 0)
self.addCleanup(shutil.rmtree, trials[0].checkpoint.dir_or_data)
def testFractionalGpus(self):
ray.init(num_cpus=4, num_gpus=1)
runner = TrialRunner()
kwargs = {
"resources": Resources(cpu=1, gpu=0.5),
}
trials = [
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
Trial("__fake", **kwargs),
]
for t in trials:
runner.add_trial(t)
for _ in range(10):
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[1].status, Trial.RUNNING)
self.assertEqual(trials[2].status, Trial.PENDING)
self.assertEqual(trials[3].status, Trial.PENDING)
def testFailFast(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner(fail_fast=True)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 0,
"config": {
"mock_error": True,
"persistent_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
while not runner.is_finished():
runner.step()
self.assertEqual(trials[0].status, Trial.ERROR)
# Somehow with `fail_fast=True`, if one errors out, the others are
# then stopped with `TERMINATED` status.
self.assertEqual(trials[1].status, Trial.TERMINATED)
self.assertRaises(TuneError, lambda: runner.step())
def testFailureRecoveryDisabled(self):
ray.init(num_cpus=1, num_gpus=1)
searchalg, scheduler = create_mock_components()
runner = TrialRunner(searchalg, scheduler=scheduler)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 0,
"config": {
"mock_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
while not runner.is_finished():
runner.step()
self.assertEqual(trials[0].status, Trial.ERROR)
self.assertEqual(trials[0].num_failures, 1)
self.assertEqual(len(searchalg.errored_trials), 1)
self.assertEqual(len(scheduler.errored_trials), 1)
def testFailFastRaise(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner(fail_fast=TrialRunner.RAISE)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 0,
"config": {
"mock_error": True,
"persistent_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process result, dispatch save
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process save
with self.assertRaises(Exception):
runner.step() # Error
def testMultiStepRun2(self):
"""Checks that runner.step throws when overstepping."""
ray.init(num_cpus=1)
runner = TrialRunner()
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"resources": Resources(cpu=1, gpu=0),
}
trials = [Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step()
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertRaises(TuneError, runner.step)
def testChangeResources(self):
"""Checks that resource requirements can be changed on fly."""
ray.init(num_cpus=2)
class ChangingScheduler(FIFOScheduler):
def on_trial_result(self, trial_runner, trial, result):
if result["training_iteration"] == 1:
executor = trial_runner.trial_executor
executor.stop_trial(trial)
trial.update_resources(dict(cpu=2, gpu=0))
executor.start_trial(trial)
return TrialScheduler.CONTINUE
runner = TrialRunner(scheduler=ChangingScheduler())
kwargs = {
"stopping_criterion": {
"training_iteration": 2
},
"resources": Resources(cpu=1, gpu=0),
}
trials = [Trial("__fake", **kwargs)]
for t in trials:
runner.add_trial(t)
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(
runner.trial_executor._pg_manager.occupied_resources().get("CPU"),
1)
self.assertRaises(
ValueError, lambda: trials[0].update_resources(dict(cpu=2, gpu=0)))
runner.step()
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(
runner.trial_executor._pg_manager.occupied_resources().get("CPU"),
2)
def default_resource_request(cls, config):
num_workers = config.get("num_workers",
kwargs.get("num_workers", 1))
num_cpus_per_worker = config.get(
"num_cpus_per_worker", kwargs.get("num_cpus_per_worker",
1))
use_gpu = config.get("use_gpu", kwargs.get("use_gpu"))
use_local = config.get("use_local",
kwargs.get("use_local", False))
if use_local:
remote_worker_count = num_workers - 1
local_cpus = 1
local_gpus = int(use_gpu)
else:
remote_worker_count = num_workers
local_cpus = 0
local_gpus = 0
return Resources(
cpu=int(local_cpus * num_cpus_per_worker),
gpu=int(local_gpus),
extra_cpu=int(remote_worker_count * num_cpus_per_worker),
extra_gpu=int(int(use_gpu) * remote_worker_count))
def testFailureRecoveryNodeRemoval(self):
# Node removal simulation only works with resource requests
os.environ["TUNE_PLACEMENT_GROUP_AUTO_DISABLED"] = "1"
ray.init(num_cpus=1, num_gpus=1)
searchalg, scheduler = create_mock_components()
runner = TrialRunner(searchalg, scheduler=scheduler)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 1,
"config": {
"mock_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
with patch("ray.cluster_resources") as resource_mock:
resource_mock.return_value = {"CPU": 1, "GPU": 1}
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
runner.step() # Process result, dispatch save
runner.step() # Process save
self.assertEqual(trials[0].status, Trial.RUNNING)
# Mimic a node failure
resource_mock.return_value = {"CPU": 0, "GPU": 0}
runner.step() # Detect node failure
self.assertEqual(trials[0].status, Trial.PENDING)
self.assertEqual(trials[0].num_failures, 1)
self.assertEqual(len(searchalg.errored_trials), 0)
self.assertEqual(len(scheduler.errored_trials), 1)
def testFailFastRaise(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner(fail_fast=TrialRunner.RAISE)
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
"max_failures": 0,
"config": {
"mock_error": True,
"persistent_error": True,
},
}
runner.add_trial(Trial("__fake", **kwargs))
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
with self.assertRaises(Exception):
while not runner.is_finished():
runner.step()
# Not critical checks. Only to showcase the difference
# with none raise type FailFast.
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(trials[1].status, Trial.PENDING)
def testRestoreMetricsAfterCheckpointing(self):
ray.init(num_cpus=1, num_gpus=1)
runner = TrialRunner()
kwargs = {
"resources": Resources(cpu=1, gpu=1),
"checkpoint_freq": 1,
}
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial
self.assertEqual(trials[0].status, Trial.RUNNING)
self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1)
# checkpoint = runner.trial_executor.save(trials[0])
runner.step() # Process result, dispatch save
runner.step() # Process save
runner.trial_executor.stop_trial(trials[0])
kwargs["restore_path"] = trials[0].checkpoint.value
kwargs.pop("checkpoint_freq") # No checkpointing for next trial
runner.add_trial(Trial("__fake", **kwargs))
trials = runner.get_trials()
runner.step() # Start trial, dispatch restore
self.assertEqual(trials[0].status, Trial.TERMINATED)
self.assertEqual(trials[1].status, Trial.RUNNING)
runner.step() # Process restore
runner.step() # Process result
self.assertEqual(trials[1].last_result["timesteps_since_restore"], 10)
self.assertEqual(trials[1].last_result["iterations_since_restore"], 1)
self.assertGreater(trials[1].last_result["time_since_restore"], 0)
runner.step() # Process restore
self.assertEqual(trials[1].last_result["timesteps_since_restore"], 20)
self.assertEqual(trials[1].last_result["iterations_since_restore"], 2)
self.assertGreater(trials[1].last_result["time_since_restore"], 0)
self.addCleanup(os.remove, trials[0].checkpoint.value)
class RayTrialExecutor(TrialExecutor):
"""An implementation of TrialExecutor based on Ray."""
def __init__(self,
queue_trials: bool = False,
reuse_actors: bool = False,
refresh_period: Optional[float] = None,
wait_for_placement_group: Optional[float] = None):
super(RayTrialExecutor, self).__init__(queue_trials)
# Check for if we are launching a trial without resources in kick off
# autoscaler.
self._trial_queued = False
self._running = {}
# Since trial resume after paused should not run
# trial.train.remote(), thus no more new remote object ref generated.
# We use self._paused to store paused trials here.
self._paused = {}
self._trial_cleanup = _TrialCleanup()
self._has_cleaned_up_pgs = False
self._reuse_actors = reuse_actors
self._cached_actor_pg = (None, None)
self._avail_resources = Resources(cpu=0, gpu=0)
self._committed_resources = Resources(cpu=0, gpu=0)
self._pg_manager = PlacementGroupManager(prefix=get_tune_pg_prefix())
self._staged_trials = set()
self._just_staged_trials = set()
self._trial_just_finished = False
self._trial_just_finished_before = False
self._resources_initialized = False
if refresh_period is None:
refresh_period = float(
os.environ.get("TUNE_STATE_REFRESH_PERIOD",
TUNE_STATE_REFRESH_PERIOD))
self._refresh_period = refresh_period
self._wait_for_pg = wait_for_placement_group or float(
os.environ.get("TUNE_PLACEMENT_GROUP_WAIT_S", "-1"))
if self._wait_for_pg < 0:
self._wait_for_pg = None
self._buffer_length = int(os.getenv("TUNE_RESULT_BUFFER_LENGTH", 1000))
self._buffer_min_time_s = float(
os.getenv("TUNE_RESULT_BUFFER_MIN_TIME_S", 0.))
self._buffer_max_time_s = float(
os.getenv("TUNE_RESULT_BUFFER_MAX_TIME_S", 100.))
self._last_resource_refresh = float("-inf")
self._last_ip_refresh = float("-inf")
self._last_ip_addresses = set()
self._last_nontrivial_wait = time.time()
if ray.is_initialized():
self._update_avail_resources()
def in_staging_grace_period(self) -> bool:
"""Returns True if trials have recently been staged."""
return self._pg_manager.in_staging_grace_period()
def set_max_pending_trials(self, max_pending: int):
self._pg_manager.set_max_staging(max_pending)
def stage_and_update_status(self, trials: List[Trial]):
"""Check and update statuses of scheduled placement groups.
Stages placement groups of all trials.
"""
if not self._has_cleaned_up_pgs:
# Clean up existing placement groups after trigger the tuning
# run step() method for the first time
self._pg_manager.cleanup_existing_pg()
self._has_cleaned_up_pgs = True
for trial in trials:
if trial.status != Trial.PENDING:
continue
if not trial.uses_placement_groups:
continue
if trial in self._staged_trials:
continue
if self._pg_manager.trial_in_use(trial):
continue
if not self._pg_manager.stage_trial_pg(trial):
# Break if we reached the limit of pending placement groups.
break
self._staged_trials.add(trial)
self._just_staged_trials.add(trial)
self._pg_manager.update_status()
def get_staged_trial(self):
"""Get a trial whose placement group was successfully staged.
Can also return None if no trial is available.
Returns:
Trial object or None.
"""
for trial in self._staged_trials:
if self._pg_manager.has_ready(trial):
return trial
return None
def _setup_remote_runner(self, trial):
trial.init_logdir()
# We checkpoint metadata here to try mitigating logdir duplication
self.try_checkpoint_metadata(trial)
logger_creator = partial(noop_logger_creator, logdir=trial.logdir)
if self._reuse_actors and self._cached_actor_pg[0] is not None:
logger.debug(f"Trial {trial}: Reusing cached runner "
f"{self._cached_actor_pg[0]}")
existing_runner, pg = self._cached_actor_pg
self._cached_actor_pg = (None, None)
trial.set_runner(existing_runner)
if pg and trial.uses_placement_groups:
self._pg_manager.assign_cached_pg(pg, trial)
if not self.reset_trial(trial, trial.config, trial.experiment_tag,
logger_creator):
raise AbortTrialExecution(
"Trainable runner reuse requires reset_config() to be "
"implemented and return True.")
return existing_runner
if self._cached_actor_pg[0]:
logger.debug("Cannot reuse cached runner {} for new trial".format(
self._cached_actor_pg[0]))
existing_runner, pg = self._cached_actor_pg
if pg:
self._pg_manager.return_or_clean_cached_pg(pg)
with self._change_working_directory(trial):
self._trial_cleanup.add(trial, actor=existing_runner)
self._cached_actor_pg = (None, None)
trainable_cls = trial.get_trainable_cls()
if not trainable_cls:
raise AbortTrialExecution(
f"Invalid trainable: {trial.trainable_name}. If you passed "
f"a string, make sure the trainable was registered before.")
_actor_cls = _class_cache.get(trainable_cls)
if trial.uses_placement_groups:
if not self._pg_manager.has_ready(trial, update=True):
if trial not in self._staged_trials:
if self._pg_manager.stage_trial_pg(trial):
self._staged_trials.add(trial)
self._just_staged_trials.add(trial)
just_staged = trial in self._just_staged_trials
# This part of the code is mostly here for testing
# purposes. If self._wait_for_pg is set, we will wait here
# for that many seconds until the placement group is ready.
# This ensures that the trial can be started right away and
# not just in the next step() of the trial runner.
# We only do this if we have reason to believe that resources
# will be ready, soon, i.e. when a) we just staged the PG,
# b) another trial just exited, freeing resources, or c)
# when there are no currently running trials.
if self._wait_for_pg is not None and (
just_staged or self._trial_just_finished_before
or not self.get_running_trials()):
logger.debug(
f"Waiting up to {self._wait_for_pg} seconds for "
f"placement group of trial {trial} to become ready.")
wait_end = time.monotonic() + self._wait_for_pg
while time.monotonic() < wait_end:
self._pg_manager.update_status()
if self._pg_manager.has_ready(trial):
break
time.sleep(0.1)
else:
return None
if not self._pg_manager.has_ready(trial):
# PG may have become ready during waiting period
return None
full_actor_class = self._pg_manager.get_full_actor_cls(
trial, _actor_cls)
else:
full_actor_class = _actor_cls.options(
num_cpus=trial.resources.cpu,
num_gpus=trial.resources.gpu,
memory=trial.resources.memory or None,
object_store_memory=trial.resources.object_store_memory
or None,
resources=trial.resources.custom_resources)
# Clear the Trial's location (to be updated later on result)
# since we don't know where the remote runner is placed.
trial.set_location(Location())
logger.debug("Trial %s: Setting up new remote runner.", trial)
# Logging for trials is handled centrally by TrialRunner, so
# configure the remote runner to use a noop-logger.
trial_config = copy.deepcopy(trial.config)
trial_config[TRIAL_INFO] = TrialInfo(trial)
stdout_file, stderr_file = trial.log_to_file
trial_config[STDOUT_FILE] = stdout_file
trial_config[STDERR_FILE] = stderr_file
kwargs = {
"config": trial_config,
"logger_creator": logger_creator,
}
if issubclass(trial.get_trainable_cls(), DurableTrainable):
kwargs["remote_checkpoint_dir"] = trial.remote_checkpoint_dir
with self._change_working_directory(trial):
return full_actor_class.remote(**kwargs)
def _train(self, trial):
"""Start one iteration of training and save remote id."""
if self._find_item(self._paused, trial):
raise TuneError(
"Should not call `train` on PAUSED trial {}. "
"This is an internal error - please file an issue "
"on https://github.com/ray-project/ray/issues/.".format(
str(trial)))
if self._find_item(self._running, trial):
logging.debug(
"Trial {} already has a queued future. Skipping this "
"`train` call. This may occur if a trial has "
"been unpaused within a scheduler callback.".format(
str(trial)))
return
assert trial.status == Trial.RUNNING, trial.status
buffer_time_s = max(
self._buffer_min_time_s,
min(self._buffer_max_time_s,
len(self._running) // 10))
with self._change_working_directory(trial):
if self._buffer_length > 1:
buffer_length = self._buffer_length
if trial.checkpoint_freq > 0:
buffer_length = min(buffer_length, trial.checkpoint_freq)
remote = trial.runner.train_buffered.remote(
buffer_time_s, buffer_length)
else:
remote = trial.runner.train.remote()
# Local Mode
if isinstance(remote, dict):
remote = _LocalWrapper(remote)
self._running[remote] = trial
trial_item = self._find_item(self._running, trial)
assert len(trial_item) < 2, trial_item
def _start_trial(self,
trial,
checkpoint=None,
runner=None,
train=True) -> bool:
"""Starts trial and restores last result if trial was paused.
Args:
trial (Trial): The trial to start.
checkpoint (Optional[Checkpoint]): The checkpoint to restore from.
If None, and no trial checkpoint exists, the trial is started
from the beginning.
runner (Trainable): The remote runner to use. This can be the
cached actor. If None, a new runner is created.
train (bool): Whether or not to start training.
Returns:
True if trial was started successfully, False otherwise.
See `RayTrialExecutor.restore` for possible errors raised.
"""
prior_status = trial.status
self.set_status(trial, Trial.PENDING)
if runner is None:
runner = self._setup_remote_runner(trial)
if not runner:
return False
trial.set_runner(runner)
self.restore(trial, checkpoint)
self.set_status(trial, Trial.RUNNING)
if trial in self._staged_trials:
self._staged_trials.remove(trial)
previous_run = self._find_item(self._paused, trial)
if prior_status == Trial.PAUSED and previous_run:
# If Trial was in flight when paused, self._paused stores result.
self._paused.pop(previous_run[0])
self._running[previous_run[0]] = trial
elif train and not trial.is_restoring:
self._train(trial)
return True
def _stop_trial(self, trial, error=False, error_msg=None):
"""Stops this trial.
Stops this trial, releasing all allocating resources. If stopping the
trial fails, the run will be marked as terminated in error, but no
exception will be thrown.
If the trial should be paused (``pause=True``), we do not remove
its placement group (or a surrogate placement group).
Args:
error (bool): Whether to mark this trial as terminated in error.
error_msg (str): Optional error message.
"""
self.set_status(trial, Trial.ERROR if error else Trial.TERMINATED)
self._trial_just_finished = True
trial.set_location(Location())
try:
trial.write_error_log(error_msg)
if hasattr(trial, "runner") and trial.runner:
if (not error and self._reuse_actors
and self._cached_actor_pg[0] is None):
logger.debug("Reusing actor for %s", trial.runner)
# Move PG into cache (disassociate from trial)
pg = self._pg_manager.cache_trial_pg(trial)
if pg or not trial.uses_placement_groups:
# True if a placement group was replaced
self._cached_actor_pg = (trial.runner, pg)
should_destroy_actor = False
else:
# False if no placement group was replaced. This should
# only be the case if there are no more trials with
# this placement group factory to run
logger.debug(
"Could not cache of trial {trial} actor for "
"reuse, as there are no pending trials "
"requiring its resources.")
should_destroy_actor = True
else:
should_destroy_actor = True
if should_destroy_actor:
logger.debug("Trial %s: Destroying actor.", trial)
# Try to return the placement group for other trials to use
self._pg_manager.return_pg(trial)
with self._change_working_directory(trial):
self._trial_cleanup.add(trial, actor=trial.runner)
if trial in self._staged_trials:
self._staged_trials.remove(trial)
except Exception:
logger.exception("Trial %s: Error stopping runner.", trial)
self.set_status(trial, Trial.ERROR)
finally:
trial.set_runner(None)
def start_trial(self, trial, checkpoint=None, train=True) -> bool:
"""Starts the trial.
Will not return resources if trial repeatedly fails on start.
Args:
trial (Trial): Trial to be started.
checkpoint (Checkpoint): A Python object or path storing the state
of trial.
train (bool): Whether or not to start training.
Returns:
True if the remote runner has been started. False if trial was
not started (e.g. because of lacking resources/pending PG).
"""
if not trial.uses_placement_groups:
self._commit_resources(trial.resources)
try:
return self._start_trial(trial, checkpoint, train=train)
except AbortTrialExecution:
logger.exception("Trial %s: Error starting runner, aborting!",
trial)
time.sleep(2)
error_msg = traceback.format_exc()
self._stop_trial(trial, error=True, error_msg=error_msg)
return False
except Exception:
logger.exception("Trial %s: Unexpected error starting runner.",
trial)
time.sleep(2)
error_msg = traceback.format_exc()
self._stop_trial(trial, error=True, error_msg=error_msg)
# Note that we don't return the resources, since they may
# have been lost. TODO(ujvl): is this the right thing to do?
return False
def _find_item(self, dictionary, item):
out = [rid for rid, t in dictionary.items() if t is item]
return out
def stop_trial(self, trial, error=False, error_msg=None):
"""Only returns resources if resources allocated."""
prior_status = trial.status
self._stop_trial(trial, error=error, error_msg=error_msg)
if prior_status == Trial.RUNNING:
logger.debug("Trial %s: Returning resources.", trial)
if not trial.uses_placement_groups:
self._return_resources(trial.resources)
out = self._find_item(self._running, trial)
for result_id in out:
self._running.pop(result_id)
def continue_training(self, trial):
"""Continues the training of this trial."""
self._train(trial)
def pause_trial(self, trial):
"""Pauses the trial.
If trial is in-flight, preserves return value in separate queue
before pausing, which is restored when Trial is resumed.
"""
trial_future = self._find_item(self._running, trial)
if trial_future:
self._paused[trial_future[0]] = trial
super(RayTrialExecutor, self).pause_trial(trial)
def reset_trial(self,
trial,
new_config,
new_experiment_tag,
logger_creator=None):
"""Tries to invoke `Trainable.reset()` to reset trial.
Args:
trial (Trial): Trial to be reset.
new_config (dict): New configuration for Trial trainable.
new_experiment_tag (str): New experiment name for trial.
logger_creator (Optional[Callable[[Dict], Logger]]): Function
that instantiates a logger on the actor process.
Returns:
True if `reset_config` is successful else False.
"""
trial.set_experiment_tag(new_experiment_tag)
trial.set_config(new_config)
trainable = trial.runner
# Pass magic variables
extra_config = copy.deepcopy(new_config)
extra_config[TRIAL_INFO] = TrialInfo(trial)
stdout_file, stderr_file = trial.log_to_file
extra_config[STDOUT_FILE] = stdout_file
extra_config[STDERR_FILE] = stderr_file
with self._change_working_directory(trial):
with warn_if_slow("reset"):
try:
reset_val = ray.get(trainable.reset.remote(
extra_config, logger_creator),
timeout=DEFAULT_GET_TIMEOUT)
except GetTimeoutError:
logger.exception("Trial %s: reset timed out.", trial)
return False
return reset_val
def get_running_trials(self):
"""Returns the running trials."""
return list(self._running.values())
def get_alive_node_ips(self):
now = time.time()
if now - self._last_ip_refresh < self._refresh_period:
return self._last_ip_addresses
logger.debug("Checking ips from Ray state.")
self._last_ip_refresh = now
nodes = ray.state.nodes()
ip_addresses = set()
for node in nodes:
if node["alive"]:
ip_addresses.add(node["NodeManagerAddress"])
self._last_ip_addresses = ip_addresses
return ip_addresses
def get_current_trial_ips(self):
return {t.node_ip for t in self.get_running_trials()}
def get_next_failed_trial(self):
"""Gets the first trial found to be running on a node presumed dead.
Returns:
A Trial object that is ready for failure processing. None if
no failure detected.
"""
if ray.worker._mode() != ray.worker.LOCAL_MODE:
live_cluster_ips = self.get_alive_node_ips()
if live_cluster_ips - self.get_current_trial_ips():
for trial in self.get_running_trials():
if trial.node_ip and trial.node_ip not in live_cluster_ips:
return trial
return None
def get_next_available_trial(self, timeout: Optional[float] = None):
if not self._running:
return None
shuffled_results = list(self._running.keys())
random.shuffle(shuffled_results)
# Note: We shuffle the results because `ray.wait` by default returns
# the first available result, and we want to guarantee that slower
# trials (i.e. trials that run remotely) also get fairly reported.
# See https://github.com/ray-project/ray/issues/4211 for details.
start = time.time()
ready, _ = ray.wait(shuffled_results, timeout=timeout)
if not ready:
return None
result_id = ready[0]
wait_time = time.time() - start
if wait_time > NONTRIVIAL_WAIT_TIME_THRESHOLD_S:
self._last_nontrivial_wait = time.time()
if time.time() - self._last_nontrivial_wait > BOTTLENECK_WARN_PERIOD_S:
logger.warning(
"Over the last {} seconds, the Tune event loop has been "
"backlogged processing new results. Consider increasing your "
"period of result reporting to improve performance.".format(
BOTTLENECK_WARN_PERIOD_S))
self._last_nontrivial_wait = time.time()
return self._running[result_id]
def fetch_result(self, trial):
"""Fetches result list of the running trials.
Returns:
Result of the most recent trial training run.
"""
trial_future = self._find_item(self._running, trial)
if not trial_future:
raise ValueError("Trial was not running.")
self._running.pop(trial_future[0])
with warn_if_slow("fetch_result"):
result = ray.get(trial_future[0], timeout=DEFAULT_GET_TIMEOUT)
# For local mode
if isinstance(result, _LocalWrapper):
result = result.unwrap()
if not isinstance(result, list):
return [result]
return result
def _commit_resources(self, resources):
committed = self._committed_resources
all_keys = set(resources.custom_resources).union(
set(committed.custom_resources))
custom_resources = {
k: committed.get(k) + resources.get_res_total(k)
for k in all_keys
}
self._committed_resources = Resources(
committed.cpu + resources.cpu_total(),
committed.gpu + resources.gpu_total(),
committed.memory + resources.memory_total(),
committed.object_store_memory +
resources.object_store_memory_total(),
custom_resources=custom_resources)
def _return_resources(self, resources):
committed = self._committed_resources
all_keys = set(resources.custom_resources).union(
set(committed.custom_resources))
custom_resources = {
k: committed.get(k) - resources.get_res_total(k)
for k in all_keys
}
self._committed_resources = Resources(
committed.cpu - resources.cpu_total(),
committed.gpu - resources.gpu_total(),
custom_resources=custom_resources)
assert self._committed_resources.is_nonnegative(), (
"Resource invalid: {}".format(resources))
def _update_avail_resources(self, num_retries=5):
if time.time() - self._last_resource_refresh < self._refresh_period:
return
logger.debug("Checking Ray cluster resources.")
resources = None
for i in range(num_retries):
if i > 0:
logger.warning(
"Cluster resources not detected or are 0. Attempt #"
"%s...", i + 1)
time.sleep(0.5)
try:
resources = ray.cluster_resources()
except Exception as exc:
# TODO(rliaw): Remove this when local mode is fixed.
# https://github.com/ray-project/ray/issues/4147
logger.debug(f"{exc}: Using resources for local machine.")
resources = ResourceSpec().resolve(True).to_resource_dict()
if resources:
break
if not resources:
# NOTE: This hides the possibility that Ray may be waiting for
# clients to connect.
resources.setdefault("CPU", 0)
resources.setdefault("GPU", 0)
logger.warning("Cluster resources cannot be detected or are 0. "
"You can resume this experiment by passing in "
"`resume=True` to `run`.")
resources = resources.copy()
num_cpus = resources.pop("CPU", 0)
num_gpus = resources.pop("GPU", 0)
memory = ray_constants.from_memory_units(resources.pop("memory", 0))
object_store_memory = ray_constants.from_memory_units(
resources.pop("object_store_memory", 0))
custom_resources = resources
self._avail_resources = Resources(
int(num_cpus),
int(num_gpus),
memory=int(memory),
object_store_memory=int(object_store_memory),
custom_resources=custom_resources)
self._last_resource_refresh = time.time()
self._resources_initialized = True
def has_resources_for_trial(self, trial: Trial):
"""Returns whether this runner has resources available for this trial.
If using placement groups, this will return True as long as we
didn't reach the maximum number of pending trials. It will also return
True if the trial placement group is already staged.
Args:
trial: Trial object which should be scheduled.
Returns:
boolean
"""
if trial.uses_placement_groups:
return trial in self._staged_trials or self._pg_manager.can_stage(
) or self._pg_manager.has_ready(trial, update=True)
return self.has_resources(trial.resources)
def has_resources(self, resources):
"""Returns whether this runner has at least the specified resources.
This refreshes the Ray cluster resources if the time since last update
has exceeded self._refresh_period. This also assumes that the
cluster is not resizing very frequently.
"""
if resources.has_placement_group:
return self._pg_manager.can_stage()
self._update_avail_resources()
currently_available = Resources.subtract(self._avail_resources,
self._committed_resources)
have_space = (
resources.cpu_total() <= currently_available.cpu
and resources.gpu_total() <= currently_available.gpu
and resources.memory_total() <= currently_available.memory
and resources.object_store_memory_total() <=
currently_available.object_store_memory and all(
resources.get_res_total(res) <= currently_available.get(res)
for res in resources.custom_resources))
if have_space:
# The assumption right now is that we block all trials if one
# trial is queued.
self._trial_queued = False
return True
can_overcommit = self._queue_trials and not self._trial_queued
if can_overcommit:
self._trial_queued = True
logger.warning(
"Allowing trial to start even though the "
"cluster does not have enough free resources. Trial actors "
"may appear to hang until enough resources are added to the "
"cluster (e.g., via autoscaling). You can disable this "
"behavior by specifying `queue_trials=False` in "
"ray.tune.run().")
return True
return False
def debug_string(self):
"""Returns a human readable message for printing to the console."""
total_resources = self._pg_manager.total_used_resources(
self._committed_resources)
if self._resources_initialized:
status = ("Resources requested: {}/{} CPUs, {}/{} GPUs, "
"{}/{} GiB heap, {}/{} GiB objects".format(
total_resources.pop("CPU",
0), self._avail_resources.cpu,
total_resources.pop("GPU", 0),
self._avail_resources.gpu,
_to_gb(total_resources.pop("memory", 0.)),
_to_gb(self._avail_resources.memory),
_to_gb(total_resources.pop("object_store_memory",
0.)),
_to_gb(self._avail_resources.object_store_memory)))
customs = ", ".join([
"{}/{} {}".format(total_resources.get(name, 0.),
self._avail_resources.get_res_total(name),
name)
for name in self._avail_resources.custom_resources
if not name.startswith(ray.resource_spec.NODE_ID_PREFIX)
])
if customs:
status += " ({})".format(customs)
return status
else:
return "Resources requested: ?"
def resource_string(self):
"""Returns a string describing the total resources available."""
if self._resources_initialized:
res_str = ("{} CPUs, {} GPUs, "
"{} GiB heap, {} GiB objects".format(
self._avail_resources.cpu,
self._avail_resources.gpu,
_to_gb(self._avail_resources.memory),
_to_gb(self._avail_resources.object_store_memory)))
if self._avail_resources.custom_resources:
custom = ", ".join(
"{} {}".format(self._avail_resources.get_res_total(name),
name)
for name in self._avail_resources.custom_resources)
res_str += " ({})".format(custom)
return res_str
else:
return "? CPUs, ? GPUs"
def on_step_begin(self, trial_runner):
"""Before step() called, update the available resources."""
self._update_avail_resources()
self._trial_just_finished_before = self._trial_just_finished
self._trial_just_finished = False
def on_step_end(self, trial_runner):
self._just_staged_trials.clear()
self._pg_manager.reconcile_placement_groups(trial_runner.get_trials())
self._pg_manager.cleanup()
def save(self, trial, storage=Checkpoint.PERSISTENT, result=None):
"""Saves the trial's state to a checkpoint asynchronously.
Args:
trial (Trial): The trial to be saved.
storage (str): Where to store the checkpoint. Defaults to
PERSISTENT.
result (dict): The state of this trial as a dictionary to be saved.
If result is None, the trial's last result will be used.
Returns:
Checkpoint object, or None if an Exception occurs.
"""
result = result or trial.last_result
with self._change_working_directory(trial):
if storage == Checkpoint.MEMORY:
value = trial.runner.save_to_object.remote()
checkpoint = Checkpoint(storage, value, result)
trial.on_checkpoint(checkpoint)
else:
value = trial.runner.save.remote()
checkpoint = Checkpoint(storage, value, result)
trial.saving_to = checkpoint
self._running[value] = trial
return checkpoint
def restore(self, trial, checkpoint=None, block=False):
"""Restores training state from a given model checkpoint.
Args:
trial (Trial): The trial to be restored.
checkpoint (Checkpoint): The checkpoint to restore from. If None,
the most recent PERSISTENT checkpoint is used. Defaults to
None.
block (bool): Whether or not to block on restore before returning.
Raises:
RuntimeError: This error is raised if no runner is found.
AbortTrialExecution: This error is raised if the trial is
ineligible for restoration, given the Tune input arguments.
"""
if checkpoint is None or checkpoint.value is None:
checkpoint = trial.checkpoint
if checkpoint.value is None:
return
if trial.runner is None:
raise RuntimeError(
"Trial {}: Unable to restore - no runner found.".format(trial))
value = checkpoint.value
if checkpoint.storage == Checkpoint.MEMORY:
logger.debug("Trial %s: Attempting restore from object", trial)
# Note that we don't store the remote since in-memory checkpoints
# don't guarantee fault tolerance and don't need to be waited on.
with self._change_working_directory(trial):
trial.runner.restore_from_object.remote(value)
else:
logger.debug("Trial %s: Attempting restore from %s", trial, value)
if issubclass(trial.get_trainable_cls(),
DurableTrainable) or not trial.sync_on_checkpoint:
with self._change_working_directory(trial):
remote = trial.runner.restore.remote(value)
elif trial.sync_on_checkpoint:
# This provides FT backwards compatibility in the
# case where a DurableTrainable is not provided.
logger.debug("Trial %s: Reading checkpoint into memory", trial)
obj = TrainableUtil.checkpoint_to_object(value)
with self._change_working_directory(trial):
remote = trial.runner.restore_from_object.remote(obj)
else:
raise AbortTrialExecution(
"Pass in `sync_on_checkpoint=True` for driver-based trial"
"restoration. Pass in an `upload_dir` and a Trainable "
"extending `DurableTrainable` for remote storage-based "
"restoration")
if block:
ray.get(remote)
else:
self._running[remote] = trial
trial.restoring_from = checkpoint
def export_trial_if_needed(self, trial):
"""Exports model of this trial based on trial.export_formats.
Return:
A dict that maps ExportFormats to successfully exported models.
"""
if trial.export_formats and len(trial.export_formats) > 0:
with self._change_working_directory(trial):
return ray.get(trial.runner.export_model.remote(
trial.export_formats),
timeout=DEFAULT_GET_TIMEOUT)
return {}
def has_gpus(self):
if self._resources_initialized:
self._update_avail_resources()
return self._avail_resources.gpu > 0
def cleanup(self):
self._trial_cleanup.cleanup(partial=False)
self._pg_manager.cleanup(force=True)
self._pg_manager.cleanup_existing_pg(block=True)
@contextmanager
def _change_working_directory(self, trial):
"""Context manager changing working directory to trial logdir.
Used in local mode.
For non-local mode it is no-op.
"""
if ray.worker._mode() == ray.worker.LOCAL_MODE:
old_dir = os.getcwd()
try:
os.chdir(trial.logdir)
yield
finally:
os.chdir(old_dir)
else:
yield
def __init__(self,
trainable_name,
config=None,
trial_id=None,
local_dir=DEFAULT_RESULTS_DIR,
evaluated_params=None,
experiment_tag="",
resources=None,
stopping_criterion=None,
remote_checkpoint_dir=None,
checkpoint_freq=0,
checkpoint_at_end=False,
sync_on_checkpoint=True,
keep_checkpoints_num=None,
checkpoint_score_attr=TRAINING_ITERATION,
export_formats=None,
restore_path=None,
trial_name_creator=None,
loggers=None,
sync_to_driver_fn=None,
max_failures=0):
"""Initialize a new trial.
The args here take the same meaning as the command line flags defined
in ray.tune.config_parser.
"""
validate_trainable(trainable_name)
# Trial config
self.trainable_name = trainable_name
self.trial_id = Trial.generate_id() if trial_id is None else trial_id
self.config = config or {}
self.local_dir = local_dir # This remains unexpanded for syncing.
#: Parameters that Tune varies across searches.
self.evaluated_params = evaluated_params or {}
self.experiment_tag = experiment_tag
trainable_cls = self.get_trainable_cls()
if trainable_cls and hasattr(trainable_cls,
"default_resource_request"):
default_resources = trainable_cls.default_resource_request(
self.config)
if default_resources:
if resources:
raise ValueError(
"Resources for {} have been automatically set to {} "
"by its `default_resource_request()` method. Please "
"clear the `resources_per_trial` option.".format(
trainable_cls, default_resources))
resources = default_resources
self.location = Location()
self.resources = resources or Resources(cpu=1, gpu=0)
self.stopping_criterion = stopping_criterion or {}
self.loggers = loggers
self.sync_to_driver_fn = sync_to_driver_fn
self.verbose = True
self.max_failures = max_failures
# Local trial state that is updated during the run
self.last_result = {}
self.last_update_time = -float("inf")
# stores in memory max/min/last result for each metric by trial
self.metric_analysis = {}
self.export_formats = export_formats
self.status = Trial.PENDING
self.start_time = None
self.logdir = None
self.runner = None
self.result_logger = None
self.last_debug = 0
self.error_file = None
self.error_msg = None
self.custom_trial_name = None
# Checkpointing fields
if remote_checkpoint_dir:
self.remote_checkpoint_dir_prefix = remote_checkpoint_dir
else:
self.remote_checkpoint_dir_prefix = None
self.checkpoint_freq = checkpoint_freq
self.checkpoint_at_end = checkpoint_at_end
self.sync_on_checkpoint = sync_on_checkpoint
newest_checkpoint = Checkpoint(Checkpoint.PERSISTENT, restore_path)
self.checkpoint_manager = CheckpointManager(
keep_checkpoints_num, checkpoint_score_attr,
checkpoint_deleter(str(self), self.runner))
self.checkpoint_manager.newest_checkpoint = newest_checkpoint
# Restoration fields
self.restoring_from = None
self.num_failures = 0
self.num_consecutive_start_attempts = 0
# AutoML fields
self.results = None
self.best_result = None
self.param_config = None
self.extra_arg = None
self._nonjson_fields = [
"loggers",
"sync_to_driver_fn",
"results",
"best_result",
"param_config",
"extra_arg",
]
if trial_name_creator:
self.custom_trial_name = trial_name_creator(self)
def testSerialization(self):
original = Resources(1, 0, 0, 1, custom_resources={"a": 1, "b": 2})
jsoned = resources_to_json(original)
new_resource = json_to_resources(jsoned)
self.assertEqual(original, new_resource)
def default_resource_request(cls, config: Dict) -> Resources:
return Resources(cpu=0,
gpu=0,
extra_cpu=num_cpus_per_worker * num_workers,
extra_gpu=num_gpus_per_worker * num_workers)
def testResourceNumericalError(self):
resource = Resources(cpu=0.99, gpu=0.99, custom_resources={"a": 0.99})
small_resource = Resources(cpu=0.33, gpu=0.33, custom_resources={"a": 0.33})
for i in range(3):
resource = Resources.subtract(resource, small_resource)
self.assertTrue(resource.is_nonnegative())
def default_resource_request(cls, config):
return Resources(cpu=config["cpu"], gpu=config["gpu"])
def create_trial(cpu, gpu=0):
return Trial("__fake", resources=Resources(cpu=cpu, gpu=gpu))
def testStartFailure(self):
_global_registry.register(TRAINABLE_CLASS, "asdf", None)
trial = Trial("asdf", resources=Resources(1, 0))
self.trial_executor.start_trial(trial)
self.assertEqual(Trial.ERROR, trial.status)
class RayTrialExecutor(TrialExecutor):
"""An implementation of TrialExecutor based on Ray."""
def __init__(self,
queue_trials=False,
reuse_actors=False,
ray_auto_init=None,
refresh_period=RESOURCE_REFRESH_PERIOD):
if ray_auto_init is None:
if os.environ.get("TUNE_DISABLE_AUTO_INIT") == "1":
logger.info("'TUNE_DISABLE_AUTO_INIT=1' detected.")
ray_auto_init = False
else:
ray_auto_init = True
super(RayTrialExecutor, self).__init__(queue_trials)
# Check for if we are launching a trial without resources in kick off
# autoscaler.
self._trial_queued = False
self._running = {}
# Since trial resume after paused should not run
# trial.train.remote(), thus no more new remote object ref generated.
# We use self._paused to store paused trials here.
self._paused = {}
self._trial_cleanup = _TrialCleanup()
self._reuse_actors = reuse_actors
self._cached_actor = None
self._avail_resources = Resources(cpu=0, gpu=0)
self._committed_resources = Resources(cpu=0, gpu=0)
self._resources_initialized = False
self._refresh_period = refresh_period
self._last_resource_refresh = float("-inf")
self._last_nontrivial_wait = time.time()
if not ray.is_initialized() and ray_auto_init:
logger.info("Initializing Ray automatically."
"For cluster usage or custom Ray initialization, "
"call `ray.init(...)` before `tune.run`.")
ray.init()
if ray.is_initialized():
self._update_avail_resources()
def _setup_remote_runner(self, trial, reuse_allowed):
trial.init_logger()
# We checkpoint metadata here to try mitigating logdir duplication
self.try_checkpoint_metadata(trial)
logger_creator = partial(noop_logger_creator, logdir=trial.logdir)
if (self._reuse_actors and reuse_allowed
and self._cached_actor is not None):
logger.debug("Trial %s: Reusing cached runner %s", trial,
self._cached_actor)
existing_runner = self._cached_actor
self._cached_actor = None
trial.set_runner(existing_runner)
if not self.reset_trial(trial, trial.config, trial.experiment_tag,
logger_creator):
raise AbortTrialExecution(
"Trainable runner reuse requires reset_config() to be "
"implemented and return True.")
return existing_runner
if self._cached_actor:
logger.debug("Cannot reuse cached runner {} for new trial".format(
self._cached_actor))
with self._change_working_directory(trial):
self._trial_cleanup.add(trial, actor=self._cached_actor)
self._cached_actor = None
_actor_cls = _class_cache.get(trial.get_trainable_cls())
full_actor_class = _actor_cls.options(
num_cpus=trial.resources.cpu,
num_gpus=trial.resources.gpu,
memory=trial.resources.memory or None,
object_store_memory=trial.resources.object_store_memory or None,
resources=trial.resources.custom_resources)
# Clear the Trial's location (to be updated later on result)
# since we don't know where the remote runner is placed.
trial.set_location(Location())
logger.debug("Trial %s: Setting up new remote runner.", trial)
# Logging for trials is handled centrally by TrialRunner, so
# configure the remote runner to use a noop-logger.
trial_config = copy.deepcopy(trial.config)
trial_config[TRIAL_INFO] = TrialInfo(trial)
stdout_file, stderr_file = trial.log_to_file
trial_config[STDOUT_FILE] = stdout_file
trial_config[STDERR_FILE] = stderr_file
kwargs = {
"config": trial_config,
"logger_creator": logger_creator,
}
if issubclass(trial.get_trainable_cls(), DurableTrainable):
kwargs["remote_checkpoint_dir"] = trial.remote_checkpoint_dir
with self._change_working_directory(trial):
return full_actor_class.remote(**kwargs)
def _train(self, trial):
"""Start one iteration of training and save remote id."""
if self._find_item(self._paused, trial):
raise TuneError(
"Should not call `train` on PAUSED trial {}. "
"This is an internal error - please file an issue "
"on https://github.com/ray-project/ray/issues/.".format(
str(trial)))
if self._find_item(self._running, trial):
logging.debug(
"Trial {} already has a queued future. Skipping this "
"`train` call. This may occur if a trial has "
"been unpaused within a scheduler callback.".format(
str(trial)))
return
assert trial.status == Trial.RUNNING, trial.status
with self._change_working_directory(trial):
remote = trial.runner.train.remote()
# Local Mode
if isinstance(remote, dict):
remote = _LocalWrapper(remote)
self._running[remote] = trial
trial_item = self._find_item(self._running, trial)
assert len(trial_item) < 2, trial_item
def _start_trial(self, trial, checkpoint=None, runner=None, train=True):
"""Starts trial and restores last result if trial was paused.
Args:
trial (Trial): The trial to start.
checkpoint (Optional[Checkpoint]): The checkpoint to restore from.
If None, and no trial checkpoint exists, the trial is started
from the beginning.
runner (Trainable): The remote runner to use. This can be the
cached actor. If None, a new runner is created.
train (bool): Whether or not to start training.
See `RayTrialExecutor.restore` for possible errors raised.
"""
prior_status = trial.status
if runner is None:
# TODO: Right now, we only support reuse if there has been
# previously instantiated state on the worker. However,
# we should consider the case where function evaluations
# can be very fast - thereby extending the need to support
# reuse to cases where there has not been previously
# instantiated state before.
reuse_allowed = checkpoint is not None or trial.has_checkpoint()
runner = self._setup_remote_runner(trial, reuse_allowed)
trial.set_runner(runner)
self.restore(trial, checkpoint)
self.set_status(trial, Trial.RUNNING)
previous_run = self._find_item(self._paused, trial)
if prior_status == Trial.PAUSED and previous_run:
# If Trial was in flight when paused, self._paused stores result.
self._paused.pop(previous_run[0])
self._running[previous_run[0]] = trial
elif train and not trial.is_restoring:
self._train(trial)
def _stop_trial(self, trial, error=False, error_msg=None,
stop_logger=True):
"""Stops this trial.
Stops this trial, releasing all allocating resources. If stopping the
trial fails, the run will be marked as terminated in error, but no
exception will be thrown.
Args:
error (bool): Whether to mark this trial as terminated in error.
error_msg (str): Optional error message.
stop_logger (bool): Whether to shut down the trial logger.
"""
self.set_status(trial, Trial.ERROR if error else Trial.TERMINATED)
trial.set_location(Location())
try:
trial.write_error_log(error_msg)
if hasattr(trial, "runner") and trial.runner:
if (not error and self._reuse_actors
and self._cached_actor is None):
logger.debug("Reusing actor for %s", trial.runner)
self._cached_actor = trial.runner
else:
logger.debug("Trial %s: Destroying actor.", trial)
with self._change_working_directory(trial):
self._trial_cleanup.add(trial, actor=trial.runner)
except Exception:
logger.exception("Trial %s: Error stopping runner.", trial)
self.set_status(trial, Trial.ERROR)
finally:
trial.set_runner(None)
if stop_logger:
trial.close_logger()
def start_trial(self, trial, checkpoint=None, train=True):
"""Starts the trial.
Will not return resources if trial repeatedly fails on start.
Args:
trial (Trial): Trial to be started.
checkpoint (Checkpoint): A Python object or path storing the state
of trial.
train (bool): Whether or not to start training.
"""
self._commit_resources(trial.resources)
try:
self._start_trial(trial, checkpoint, train=train)
except AbortTrialExecution:
logger.exception("Trial %s: Error starting runner, aborting!",
trial)
time.sleep(2)
error_msg = traceback.format_exc()
self._stop_trial(trial, error=True, error_msg=error_msg)
except Exception:
logger.exception("Trial %s: Unexpected error starting runner.",
trial)
time.sleep(2)
error_msg = traceback.format_exc()
self._stop_trial(trial, error=True, error_msg=error_msg)
# Note that we don't return the resources, since they may
# have been lost. TODO(ujvl): is this the right thing to do?
def _find_item(self, dictionary, item):
out = [rid for rid, t in dictionary.items() if t is item]
return out
def stop_trial(self, trial, error=False, error_msg=None, stop_logger=True):
"""Only returns resources if resources allocated."""
prior_status = trial.status
self._stop_trial(
trial, error=error, error_msg=error_msg, stop_logger=stop_logger)
if prior_status == Trial.RUNNING:
logger.debug("Trial %s: Returning resources.", trial)
self._return_resources(trial.resources)
out = self._find_item(self._running, trial)
for result_id in out:
self._running.pop(result_id)
def continue_training(self, trial):
"""Continues the training of this trial."""
self._train(trial)
def pause_trial(self, trial):
"""Pauses the trial.
If trial is in-flight, preserves return value in separate queue
before pausing, which is restored when Trial is resumed.
"""
trial_future = self._find_item(self._running, trial)
if trial_future:
self._paused[trial_future[0]] = trial
super(RayTrialExecutor, self).pause_trial(trial)
def reset_trial(self,
trial,
new_config,
new_experiment_tag,
logger_creator=None):
"""Tries to invoke `Trainable.reset()` to reset trial.
Args:
trial (Trial): Trial to be reset.
new_config (dict): New configuration for Trial trainable.
new_experiment_tag (str): New experiment name for trial.
logger_creator (Callable[[Dict], Logger]): A function that
instantiates a logger on the actor process.
Returns:
True if `reset_config` is successful else False.
"""
trial.experiment_tag = new_experiment_tag
trial.config = new_config
trainable = trial.runner
with self._change_working_directory(trial):
with warn_if_slow("reset"):
try:
reset_val = ray.get(
trainable.reset.remote(new_config, logger_creator),
timeout=DEFAULT_GET_TIMEOUT)
except GetTimeoutError:
logger.exception("Trial %s: reset timed out.", trial)
return False
return reset_val
def get_running_trials(self):
"""Returns the running trials."""
return list(self._running.values())
def get_alive_node_ips(self):
nodes = ray.state.nodes()
ip_addresses = set()
for node in nodes:
if node["alive"]:
ip_addresses.add(node["NodeManagerAddress"])
return ip_addresses
def get_current_trial_ips(self):
return {t.node_ip for t in self.get_running_trials()}
def get_next_failed_trial(self):
"""Gets the first trial found to be running on a node presumed dead.
Returns:
A Trial object that is ready for failure processing. None if
no failure detected.
"""
if ray.worker._mode() != ray.worker.LOCAL_MODE:
live_cluster_ips = self.get_alive_node_ips()
if live_cluster_ips - self.get_current_trial_ips():
for trial in self.get_running_trials():
if trial.node_ip and trial.node_ip not in live_cluster_ips:
return trial
return None
def get_next_available_trial(self):
shuffled_results = list(self._running.keys())
random.shuffle(shuffled_results)
# Note: We shuffle the results because `ray.wait` by default returns
# the first available result, and we want to guarantee that slower
# trials (i.e. trials that run remotely) also get fairly reported.
# See https://github.com/ray-project/ray/issues/4211 for details.
start = time.time()
[result_id], _ = ray.wait(shuffled_results)
wait_time = time.time() - start
if wait_time > NONTRIVIAL_WAIT_TIME_THRESHOLD_S:
self._last_nontrivial_wait = time.time()
if time.time() - self._last_nontrivial_wait > BOTTLENECK_WARN_PERIOD_S:
logger.warning(
"Over the last {} seconds, the Tune event loop has been "
"backlogged processing new results. Consider increasing your "
"period of result reporting to improve performance.".format(
BOTTLENECK_WARN_PERIOD_S))
self._last_nontrivial_wait = time.time()
return self._running[result_id]
def fetch_result(self, trial):
"""Fetches one result of the running trials.
Returns:
Result of the most recent trial training run.
"""
trial_future = self._find_item(self._running, trial)
if not trial_future:
raise ValueError("Trial was not running.")
self._running.pop(trial_future[0])
with warn_if_slow("fetch_result"):
result = ray.get(trial_future[0], timeout=DEFAULT_GET_TIMEOUT)
# For local mode
if isinstance(result, _LocalWrapper):
result = result.unwrap()
return result
def _commit_resources(self, resources):
committed = self._committed_resources
all_keys = set(resources.custom_resources).union(
set(committed.custom_resources))
custom_resources = {
k: committed.get(k) + resources.get_res_total(k)
for k in all_keys
}
self._committed_resources = Resources(
committed.cpu + resources.cpu_total(),
committed.gpu + resources.gpu_total(),
committed.memory + resources.memory_total(),
committed.object_store_memory +
resources.object_store_memory_total(),
custom_resources=custom_resources)
def _return_resources(self, resources):
committed = self._committed_resources
all_keys = set(resources.custom_resources).union(
set(committed.custom_resources))
custom_resources = {
k: committed.get(k) - resources.get_res_total(k)
for k in all_keys
}
self._committed_resources = Resources(
committed.cpu - resources.cpu_total(),
committed.gpu - resources.gpu_total(),
custom_resources=custom_resources)
assert self._committed_resources.is_nonnegative(), (
"Resource invalid: {}".format(resources))
def _update_avail_resources(self, num_retries=5):
resources = None
for i in range(num_retries):
if i > 0:
logger.warning(
"Cluster resources not detected or are 0. Attempt #"
"%s...", i + 1)
time.sleep(0.5)
try:
resources = ray.cluster_resources()
except Exception:
# TODO(rliaw): Remove this when local mode is fixed.
# https://github.com/ray-project/ray/issues/4147
logger.debug("Using resources for local machine.")
resources = ResourceSpec().resolve(True).to_resource_dict()
if resources:
break
if not resources:
# NOTE: This hides the possibility that Ray may be waiting for
# clients to connect.
resources.setdefault("CPU", 0)
resources.setdefault("GPU", 0)
logger.warning("Cluster resources cannot be detected or are 0. "
"You can resume this experiment by passing in "
"`resume=True` to `run`.")
resources = resources.copy()
num_cpus = resources.pop("CPU", 0)
num_gpus = resources.pop("GPU", 0)
memory = ray_constants.from_memory_units(resources.pop("memory", 0))
object_store_memory = ray_constants.from_memory_units(
resources.pop("object_store_memory", 0))
custom_resources = resources
self._avail_resources = Resources(
int(num_cpus),
int(num_gpus),
memory=int(memory),
object_store_memory=int(object_store_memory),
custom_resources=custom_resources)
self._last_resource_refresh = time.time()
self._resources_initialized = True
def has_resources(self, resources):
"""Returns whether this runner has at least the specified resources.
This refreshes the Ray cluster resources if the time since last update
has exceeded self._refresh_period. This also assumes that the
cluster is not resizing very frequently.
"""
if time.time() - self._last_resource_refresh > self._refresh_period:
self._update_avail_resources()
currently_available = Resources.subtract(self._avail_resources,
self._committed_resources)
have_space = (
resources.cpu_total() <= currently_available.cpu
and resources.gpu_total() <= currently_available.gpu
and resources.memory_total() <= currently_available.memory
and resources.object_store_memory_total() <=
currently_available.object_store_memory and all(
resources.get_res_total(res) <= currently_available.get(res)
for res in resources.custom_resources))
if have_space:
# The assumption right now is that we block all trials if one
# trial is queued.
self._trial_queued = False
return True
can_overcommit = self._queue_trials and not self._trial_queued
if can_overcommit:
self._trial_queued = True
logger.warning(
"Allowing trial to start even though the "
"cluster does not have enough free resources. Trial actors "
"may appear to hang until enough resources are added to the "
"cluster (e.g., via autoscaling). You can disable this "
"behavior by specifying `queue_trials=False` in "
"ray.tune.run().")
return True
return False
def debug_string(self):
"""Returns a human readable message for printing to the console."""
if self._resources_initialized:
status = ("Resources requested: {}/{} CPUs, {}/{} GPUs, "
"{}/{} GiB heap, {}/{} GiB objects".format(
self._committed_resources.cpu,
self._avail_resources.cpu,
self._committed_resources.gpu,
self._avail_resources.gpu,
_to_gb(self._committed_resources.memory),
_to_gb(self._avail_resources.memory),
_to_gb(
self._committed_resources.object_store_memory),
_to_gb(self._avail_resources.object_store_memory)))
customs = ", ".join([
"{}/{} {}".format(
self._committed_resources.get_res_total(name),
self._avail_resources.get_res_total(name), name)
for name in self._avail_resources.custom_resources
if not name.startswith(ray.resource_spec.NODE_ID_PREFIX)
])
if customs:
status += " ({})".format(customs)
return status
else:
return "Resources requested: ?"
def resource_string(self):
"""Returns a string describing the total resources available."""
if self._resources_initialized:
res_str = ("{} CPUs, {} GPUs, "
"{} GiB heap, {} GiB objects".format(
self._avail_resources.cpu,
self._avail_resources.gpu,
_to_gb(self._avail_resources.memory),
_to_gb(self._avail_resources.object_store_memory)))
if self._avail_resources.custom_resources:
custom = ", ".join(
"{} {}".format(
self._avail_resources.get_res_total(name), name)
for name in self._avail_resources.custom_resources)
res_str += " ({})".format(custom)
return res_str
else:
return "? CPUs, ? GPUs"
def on_step_begin(self, trial_runner):
"""Before step() called, update the available resources."""
self._update_avail_resources()
def save(self, trial, storage=Checkpoint.PERSISTENT, result=None):
"""Saves the trial's state to a checkpoint asynchronously.
Args:
trial (Trial): The trial to be saved.
storage (str): Where to store the checkpoint. Defaults to
PERSISTENT.
result (dict): The state of this trial as a dictionary to be saved.
If result is None, the trial's last result will be used.
Returns:
Checkpoint object, or None if an Exception occurs.
"""
result = result or trial.last_result
with self._change_working_directory(trial):
if storage == Checkpoint.MEMORY:
value = trial.runner.save_to_object.remote()
checkpoint = Checkpoint(storage, value, result)
trial.on_checkpoint(checkpoint)
else:
value = trial.runner.save.remote()
checkpoint = Checkpoint(storage, value, result)
trial.saving_to = checkpoint
self._running[value] = trial
return checkpoint
def restore(self, trial, checkpoint=None, block=False):
"""Restores training state from a given model checkpoint.
Args:
trial (Trial): The trial to be restored.
checkpoint (Checkpoint): The checkpoint to restore from. If None,
the most recent PERSISTENT checkpoint is used. Defaults to
None.
block (bool): Whether or not to block on restore before returning.
Raises:
RuntimeError: This error is raised if no runner is found.
AbortTrialExecution: This error is raised if the trial is
ineligible for restoration, given the Tune input arguments.
"""
if checkpoint is None or checkpoint.value is None:
checkpoint = trial.checkpoint
if checkpoint.value is None:
return
if trial.runner is None:
raise RuntimeError(
"Trial {}: Unable to restore - no runner found.".format(trial))
value = checkpoint.value
if checkpoint.storage == Checkpoint.MEMORY:
logger.debug("Trial %s: Attempting restore from object", trial)
# Note that we don't store the remote since in-memory checkpoints
# don't guarantee fault tolerance and don't need to be waited on.
with self._change_working_directory(trial):
trial.runner.restore_from_object.remote(value)
else:
logger.debug("Trial %s: Attempting restore from %s", trial, value)
if issubclass(trial.get_trainable_cls(), DurableTrainable):
with self._change_working_directory(trial):
remote = trial.runner.restore.remote(value)
elif trial.sync_on_checkpoint:
# This provides FT backwards compatibility in the
# case where a DurableTrainable is not provided.
logger.debug("Trial %s: Reading checkpoint into memory", trial)
obj = TrainableUtil.checkpoint_to_object(value)
with self._change_working_directory(trial):
remote = trial.runner.restore_from_object.remote(obj)
else:
raise AbortTrialExecution(
"Pass in `sync_on_checkpoint=True` for driver-based trial"
"restoration. Pass in an `upload_dir` and a Trainable "
"extending `DurableTrainable` for remote storage-based "
"restoration")
if block:
ray.get(remote)
else:
self._running[remote] = trial
trial.restoring_from = checkpoint
def export_trial_if_needed(self, trial):
"""Exports model of this trial based on trial.export_formats.
Return:
A dict that maps ExportFormats to successfully exported models.
"""
if trial.export_formats and len(trial.export_formats) > 0:
with self._change_working_directory(trial):
return ray.get(
trial.runner.export_model.remote(trial.export_formats),
timeout=DEFAULT_GET_TIMEOUT)
return {}
def has_gpus(self):
if self._resources_initialized:
self._update_avail_resources()
return self._avail_resources.gpu > 0
def cleanup(self):
self._trial_cleanup.cleanup(partial=False)
@contextmanager
def _change_working_directory(self, trial):
"""Context manager changing working directory to trial logdir.
Used in local mode.
For non-local mode it is no-op.
"""
if ray.worker._mode() == ray.worker.LOCAL_MODE:
old_dir = os.getcwd()
try:
os.chdir(trial.logdir)
yield
finally:
os.chdir(old_dir)
else:
yield
