class RayTrialExecutor(TrialExecutor):
"""An implementation of TrialExecutor based on Ray."""
def __init__(self,
queue_trials: bool = False,
reuse_actors: bool = False,
result_buffer_length: Optional[int] = None,
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
# The maxlen will be updated when `set_max_pending_trials()` is called
self._cached_actor_pg = deque(maxlen=1)
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.last_pg_recon = 0
self.pg_recon_interval = float(
os.environ.get("TUNE_PLACEMENT_GROUP_RECON_INTERVAL", "5"))
self._default_buffer_length = result_buffer_length or int(
os.getenv("TUNE_RESULT_BUFFER_LENGTH", 1000))
self._buffer_length = result_buffer_length
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) -> None:
if len(self._cached_actor_pg) > 0:
logger.warning(
"Cannot update maximum number of queued actors for reuse "
"during a run.")
else:
self._cached_actor_pg = deque(maxlen=max_pending)
self._pg_manager.set_max_staging(max_pending)
def stage_and_update_status(self, trials: Iterable[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 len(self._cached_actor_pg) > 0:
existing_runner, pg = self._cached_actor_pg.popleft()
logger.debug(f"Trial {trial}: Reusing cached runner "
f"{existing_runner}")
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 len(self._cached_actor_pg) > 0:
existing_runner, pg = self._cached_actor_pg.popleft()
logger.debug(
f"Cannot reuse cached runner {existing_runner} for new trial")
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)
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
kwargs["sync_function_tpl"] = trial.sync_to_cloud
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):
buffer_length = self._buffer_length
# If buffer length has not been explicitly set, we determine
# it automatically
if buffer_length is None:
if trial.checkpoint_at_end:
# If a trial checkpoint can be triggered externally,
# it is not safe to buffer results.
buffer_length = 1
else:
# Else, use the default buffer length
buffer_length = self._default_buffer_length
else:
if trial.checkpoint_at_end:
if log_once("trial_executor_buffer_checkpoint"):
logger.warning(
"You passed `checkpoint_at_end` to `tune.run()`, "
"but still requested buffered training. "
"If used with a custom stopper or early stopping, "
"checkpoints may be created later than desired.")
if buffer_length > 1:
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._notify_trainable_of_new_resources_if_needed(trial)
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 _notify_trainable_of_new_resources_if_needed(self, trial: Trial):
if trial.has_new_resources:
trainable = trial.runner
trial.has_new_resources = False
with self._change_working_directory(trial):
with warn_if_slow("update_resources"):
try:
ray.get(trainable._update_resources.remote(
trial.placement_group_factory if trial.
uses_placement_groups else trial.resources),
timeout=DEFAULT_GET_TIMEOUT)
except GetTimeoutError:
logger.exception(
"Trial %s: updating resources timed out.", trial)
def _stop_trial(self,
trial: Trial,
error=False,
error_msg=None,
destroy_pg_if_cannot_replace=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.
If the placement group will be used right away
(destroy_pg_if_cannot_replace=False), 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 (len(self._cached_actor_pg) <
(self._cached_actor_pg.maxlen or float("inf")))):
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.append((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,
destroy_pg_if_cannot_replace)
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: Trial,
checkpoint: Optional[Checkpoint] = None,
train: bool = 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: Trial,
error: bool = False,
error_msg: Optional[str] = None,
destroy_pg_if_cannot_replace: bool = True) -> None:
"""Only returns resources if resources allocated.
If destroy_pg_if_cannot_replace is False, the Trial placement group
will not be removed if it can't replace any staging ones."""
prior_status = trial.status
self._stop_trial(
trial,
error=error,
error_msg=error_msg,
destroy_pg_if_cannot_replace=destroy_pg_if_cannot_replace)
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: Trial) -> None:
"""Continues the training of this trial."""
self._train(trial)
def pause_trial(self, trial: Trial) -> None:
"""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: Trial,
new_config: Dict,
new_experiment_tag: str,
logger_creator: Optional[Callable[[Dict], "ray.tune.Logger"]] = None
) -> bool:
"""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) -> List[Trial]:
"""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) -> Optional[Trial]:
"""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
) -> Optional[Trial]:
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) -> List[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) -> bool:
"""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: Resources) -> bool:
"""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) -> str:
"""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(NODE_ID_PREFIX) and (
total_resources.get(name, 0.) > 0 or "_group_" not in name)
])
if customs:
status += " ({})".format(customs)
return status
else:
return "Resources requested: ?"
def resource_string(self) -> str:
"""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, trials: List[Trial]) -> None:
"""Before step() is 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, trials: List[Trial]) -> None:
self._just_staged_trials.clear()
if time.time() > self.last_pg_recon + self.pg_recon_interval:
# Only do this every now and then - usually the placement groups
# should not get out of sync, and calling this often is inefficient
self._pg_manager.reconcile_placement_groups(trials)
self.last_pg_recon = time.time()
self._pg_manager.cleanup()
def force_reconcilation_on_next_step_end(self) -> None:
self.last_pg_recon = -float("inf")
def save(self,
trial,
storage=Checkpoint.PERSISTENT,
result: Optional[Dict] = None) -> Checkpoint:
"""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) -> None:
"""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: Trial) -> Dict:
"""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) -> bool:
if self._resources_initialized:
self._update_avail_resources()
return self._avail_resources.gpu > 0
def cleanup(self, trials: List[Trial]) -> None:
self._trial_cleanup.cleanup(partial=False)
self._pg_manager.reconcile_placement_groups(trials)
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
class RayTrialExecutor(TrialExecutor):
"""An implementation of TrialExecutor based on Ray."""
def __init__(
self,
reuse_actors: bool = False,
result_buffer_length: Optional[int] = None,
refresh_period: Optional[float] = None,
wait_for_placement_group: Optional[float] = None,
):
super(RayTrialExecutor, self).__init__()
# future --> (type, trial/pg)
self._futures = {}
force_trial_cleanup = int(
os.environ.get("TUNE_FORCE_TRIAL_CLEANUP_S", "0"))
self._get_next_event_wait = int(
os.environ.get("TUNE_GET_EXECUTOR_EVENT_WAIT_S", "5"))
if force_trial_cleanup:
self._trial_cleanup = _TrialCleanup(force_trial_cleanup)
else:
self._trial_cleanup = None
self._has_cleaned_up_pgs = False
self._reuse_actors = reuse_actors
# The maxlen will be updated when `set_max_pending_trials()` is called
self._cached_actor_pg = deque(maxlen=1)
self._avail_resources = Resources(cpu=0, gpu=0)
self._pg_manager = PlacementGroupManager(prefix=get_tune_pg_prefix())
self._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.last_pg_recon = 0
self.pg_recon_interval = float(
os.environ.get("TUNE_PLACEMENT_GROUP_RECON_INTERVAL", "5"))
self._buffer_length = result_buffer_length or int(
os.getenv("TUNE_RESULT_BUFFER_LENGTH", 1))
self._buffer_min_time_s = float(
os.getenv("TUNE_RESULT_BUFFER_MIN_TIME_S", 0.0))
self._buffer_max_time_s = float(
os.getenv("TUNE_RESULT_BUFFER_MAX_TIME_S", 100.0))
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 set_max_pending_trials(self, max_pending: int) -> None:
if len(self._cached_actor_pg) > 0:
logger.warning(
"Cannot update maximum number of queued actors for reuse "
"during a run.")
else:
self._cached_actor_pg = deque(maxlen=max_pending)
self._pg_manager.set_max_staging(max_pending)
def _stage_and_update_status(self, trials: Iterable[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 not in (Trial.PENDING, Trial.PAUSED):
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._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.
"""
# TODO(xwjiang): This method should consider `self._cached_actor_pg`.
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._trials_to_cache.add(trial)
logger_creator = partial(noop_logger_creator, logdir=trial.logdir)
if len(self._cached_actor_pg) > 0:
assert self._reuse_actors
existing_runner, pg = self._cached_actor_pg.popleft()
logger.debug(f"Trial {trial}: Reusing cached runner "
f"{existing_runner}")
trial.set_runner(existing_runner)
if pg:
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
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 not self._pg_manager.has_ready(trial):
return None
full_actor_class = self._pg_manager.get_full_actor_cls(
trial, _actor_cls)
# 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 trial.uses_cloud_checkpointing:
# We keep these kwargs separate for backwards compatibility
# with trainables that don't provide these keyword arguments
kwargs["remote_checkpoint_dir"] = trial.remote_checkpoint_dir
kwargs["sync_function_tpl"] = trial.sync_function_tpl
# Throw a meaningful error if trainable does not use the
# new API
sig = inspect.signature(trial.get_trainable_cls())
try:
sig.bind_partial(**kwargs)
except Exception as e:
raise RuntimeError(
"Your trainable class does not accept a "
"`remote_checkpoint_dir` or `sync_function_tpl` argument "
"in its constructor, but you've passed a "
"`upload_dir` to your SyncConfig. Without accepting "
"these parameters and passing them to the base trainable "
"constructor in the init call, cloud checkpointing is "
"effectively disabled. To resolve this issue, add the "
"parameters to your trainable class constructor or "
"disable cloud checkpointing by setting `upload_dir=None`."
) from e
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_future(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._futures) // 10),
)
with self._change_working_directory(trial):
buffer_length = self._buffer_length
if buffer_length > 1 and trial.checkpoint_at_end:
# If a trial checkpoint can be triggered externally,
# it is not safe to buffer results.
if log_once("trial_executor_buffer_checkpoint"):
logger.warning("Disabling buffered training as you passed "
"`checkpoint_at_end` to `tune.run()`.")
buffer_length = 1
if buffer_length > 1:
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._futures[remote] = (ExecutorEventType.TRAINING_RESULT, trial)
trial_item = self._find_future(trial)
assert len(trial_item) < 2, trial_item
def _start_trial(self, trial) -> bool:
"""Starts trial and restores last result if trial was paused.
Args:
trial (Trial): The trial to start.
Returns:
True if trial was started successfully, False otherwise.
See `RayTrialExecutor.restore` for possible errors raised.
"""
self.set_status(trial, Trial.PENDING)
runner = self._setup_remote_runner(trial)
if not runner:
return False
trial.set_runner(runner)
self._notify_trainable_of_new_resources_if_needed(trial)
self.restore(trial)
self.set_status(trial, Trial.RUNNING)
if trial in self._staged_trials:
self._staged_trials.remove(trial)
if not trial.is_restoring:
self._train(trial)
return True
def _notify_trainable_of_new_resources_if_needed(self, trial: Trial):
if trial.has_new_resources:
trainable = trial.runner
trial.has_new_resources = False
with self._change_working_directory(trial):
with warn_if_slow("update_resources"):
try:
ray.get(
trainable._update_resources.remote(
trial.placement_group_factory),
timeout=DEFAULT_GET_TIMEOUT,
)
except GetTimeoutError:
logger.exception(
"Trial %s: updating resources timed out.", trial)
def _stop_trial(self, trial: 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.
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 (len(self._cached_actor_pg) <
(self._cached_actor_pg.maxlen or float("inf")))):
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:
# True if a placement group was replaced
self._cached_actor_pg.append((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)
with self._change_working_directory(trial):
future = trial.runner.stop.remote()
pg = self._pg_manager.remove_from_in_use(trial)
self._futures[future] = (ExecutorEventType.STOP_RESULT, pg)
if self._trial_cleanup: # force trial cleanup within a deadline
self._trial_cleanup.add(future)
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: Trial) -> bool:
"""Starts the trial.
Will not return resources if trial repeatedly fails on start.
Args:
trial (Trial): Trial to be started.
Returns:
True if the remote runner has been started. False if trial was
not started (e.g. because of lacking resources/pending PG).
"""
try:
return self._start_trial(trial)
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_future(self, trial):
out = [rid for rid, t in self._futures.items() if t[1] is trial]
assert (
len(out) <=
1), "Expecting one future for any given trial at any given time."
return out
def stop_trial(self,
trial: Trial,
error: bool = False,
error_msg: Optional[str] = None) -> None:
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)
out = self._find_future(trial)
for result_id in out:
self._futures.pop(result_id)
def continue_training(self, trial: Trial) -> None:
"""Continues the training of this trial."""
self._train(trial)
def reset_trial(
self,
trial: Trial,
new_config: Dict,
new_experiment_tag: str,
logger_creator: Optional[Callable[[Dict], "ray.tune.Logger"]] = None,
) -> bool:
"""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 _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)
resources = ray.cluster_resources()
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) -> bool:
"""Returns whether there are resources available for this trial.
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
"""
return (trial in self._staged_trials or self._pg_manager.can_stage()
or self._pg_manager.has_ready(trial, update=True))
def debug_string(self) -> str:
"""Returns a human readable message for printing to the console."""
total_resources = self._pg_manager.occupied_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", 0.0)),
_to_gb(self._avail_resources.memory),
_to_gb(
total_resources.pop("object_store_memory", 0.0)),
_to_gb(self._avail_resources.object_store_memory),
))
customs = ", ".join([
"{}/{} {}".format(
total_resources.get(name, 0.0),
self._avail_resources.get_res_total(name),
name,
) for name in self._avail_resources.custom_resources
if not name.startswith(NODE_ID_PREFIX) and
(total_resources.get(name, 0.0) > 0 or "_group_" not in name)
])
if customs:
status += " ({})".format(customs)
return status
else:
return "Resources requested: ?"
def on_step_begin(self, trials: List[Trial]) -> None:
"""Before step() is 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, trials: List[Trial]) -> None:
self._do_force_trial_cleanup()
if time.time() > self.last_pg_recon + self.pg_recon_interval:
# Only do this every now and then - usually the placement groups
# should not get out of sync, and calling this often is inefficient
self._pg_manager.reconcile_placement_groups(trials)
self.last_pg_recon = time.time()
self._pg_manager.cleanup()
def _do_force_trial_cleanup(self) -> None:
if self._trial_cleanup:
while True:
next_future_to_clean = self._trial_cleanup.get_next()
if not next_future_to_clean:
break
if next_future_to_clean in self._futures.keys():
_, pg = self._futures.pop(next_future_to_clean)
post_stop_cleanup(next_future_to_clean, pg)
else:
# This just means that before the deadline reaches,
# the future is already cleaned up.
pass
def force_reconcilation_on_next_step_end(self) -> None:
self.last_pg_recon = -float("inf")
def save(self,
trial,
storage=Checkpoint.PERSISTENT,
result: Optional[Dict] = None) -> Checkpoint:
"""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.
"""
logger.info(f"saving trial {trial}")
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._futures[value] = (ExecutorEventType.SAVING_RESULT, trial)
return checkpoint
def restore(self, trial) -> None:
"""Restores training state from a given model checkpoint.
Args:
trial (Trial): The trial to be restored.
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.
"""
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 trial.uses_cloud_checkpointing 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 no cloud checkpoints are 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` for remote "
"storage-based restoration")
self._futures[remote] = (ExecutorEventType.RESTORING_RESULT, trial)
trial.restoring_from = checkpoint
def export_trial_if_needed(self, trial: Trial) -> Dict:
"""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) -> bool:
if self._resources_initialized:
self._update_avail_resources()
return self._avail_resources.gpu > 0
def cleanup(self, trials: List[Trial]) -> None:
while True:
if self._trial_cleanup and self._trial_cleanup.is_empty():
break
elif not self._trial_cleanup and len(self._futures) == 0:
break
self._do_force_trial_cleanup()
ready, _ = ray.wait(list(self._futures.keys()), timeout=0)
if not ready:
continue
event_type, trial_or_pg = self._futures.pop(ready[0])
if event_type == ExecutorEventType.STOP_RESULT:
post_stop_cleanup(ready[0], trial_or_pg)
self._pg_manager.reconcile_placement_groups(trials)
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 get_next_executor_event(self, live_trials: Set[Trial],
next_trial_exists: bool) -> ExecutorEvent:
"""Get the next executor event to be processed in TrialRunner.
In case there are multiple events available for handling, the next
event is determined by the following priority:
1. if there is `next_trial_exists`, and if there is cached resources
to use, PG_READY is emitted.
2. if there is `next_trial_exists` and there is no cached resources
to use, wait on pg future and randomized other futures. If multiple
futures are ready, pg future will take priority to be handled first.
3. if there is no `next_trial_exists`, wait on just randomized other
futures.
An example of #3 would be synchronous hyperband. Although there are pgs
ready, the scheduler is holding back scheduling new trials since the
whole band of trials is waiting for the slowest trial to finish. In
this case, we prioritize handling training result to avoid deadlock
situation.
This is a blocking wait with a timeout (specified with env var).
The reason for the timeout is
we still want to print status info periodically in TrialRunner for
better user experience.
The handle of `ExecutorEvent.STOP_RESULT` is purely internal to
RayTrialExecutor itself. All the other future results are handled by
TrialRunner.
In the future we may want to do most of the handle of
`ExecutorEvent.RESTORE_RESULT` and `SAVING_RESULT` in
RayTrialExecutor itself and only notify TrialRunner to invoke
corresponding callbacks. This view is more consistent with our goal
of TrialRunner responsible for external facing Trial state transition,
while RayTrialExecutor responsible for internal facing transitions,
namely, `is_saving`, `is_restoring` etc.
Also you may notice that the boundary between RayTrialExecutor and
PlacementGroupManager right now is really blurry. This will be
improved once we move to an ActorPool abstraction.
`next_trial_exists` means that there is a trial to run - prioritize
returning PG_READY in this case.
"""
# First update status of staged placement groups
self._stage_and_update_status(live_trials)
while True:
###################################################################
# when next_trial_exists and there are cached resources
###################################################################
# There could be existing PGs from either `self._cached_actor_pg`
# or from `self._pg_manager._ready`. If so and if there is indeed
# a next trial to run, we return `PG_READY` future for trial
# runner. The next trial can then be scheduled on this PG.
if next_trial_exists:
if len(self._cached_actor_pg) > 0:
return ExecutorEvent(ExecutorEventType.PG_READY)
# TODO(xwjiang): Expose proper API when we decide to do
# ActorPool abstraction.
if any(len(r) > 0 for r in self._pg_manager._ready.values()):
return ExecutorEvent(ExecutorEventType.PG_READY)
###################################################################
# Prepare for futures to wait
###################################################################
futures_to_wait = list(self._futures.keys())
random.shuffle(futures_to_wait)
if next_trial_exists:
# Only wait for pg explicitly if there is next trial to run.
# In which case, handling PG_READY triumphs handling other events.
# Since we want to place pending trial ASAP.
futures_to_wait = (self._pg_manager.get_staging_future_list() +
futures_to_wait)
logger.debug(f"get_next_executor_event before wait with futures "
f"{futures_to_wait} and "
f"next_trial_exists={next_trial_exists}")
ready_futures, _ = ray.wait(futures_to_wait,
num_returns=1,
timeout=self._get_next_event_wait)
###################################################################
# Dealing with no future returned case.
###################################################################
if len(ready_futures) == 0:
if len(self._futures) == 0:
# No running trial and timing out with wait, could be we may
# have insufficient cluster resources that makes tune run
# infeasible.
# TODO: Move InsufficientResourceManager's logic
# to TrialExecutor. It is not Runner's responsibility!
return ExecutorEvent(
ExecutorEventType.NO_RUNNING_TRIAL_TIMEOUT)
else:
# Training simply takes long time, yield the control back to main
# event loop to print progress info etc.
return ExecutorEvent(ExecutorEventType.YIELD)
###################################################################
# If there is future returned.
###################################################################
assert len(ready_futures) == 1
ready_future = ready_futures[0]
###################################################################
# If it is a PG_READY event.
###################################################################
if ready_future not in self._futures.keys():
# This is a ready future.
self._pg_manager.handle_ready_future(ready_future)
return ExecutorEvent(ExecutorEventType.PG_READY)
###################################################################
# non PG_READY event
###################################################################
result_type, trial_or_pg = self._futures.pop(ready_future)
if result_type == ExecutorEventType.STOP_RESULT:
pg = trial_or_pg
post_stop_cleanup(ready_future, pg)
else:
trial = trial_or_pg
assert isinstance(trial, Trial)
try:
future_result = ray.get(ready_future)
# For local mode
if isinstance(future_result, _LocalWrapper):
future_result = future_result.unwrap()
if result_type in (
ExecutorEventType.TRAINING_RESULT,
ExecutorEventType.SAVING_RESULT,
ExecutorEventType.RESTORING_RESULT,
):
logger.debug(
f"Returning [{result_type}] for trial {trial}")
return ExecutorEvent(result_type,
trial,
result=future_result)
else:
raise TuneError(
f"Unexpected future type - [{result_type}]")
except Exception:
return ExecutorEvent(ExecutorEventType.ERROR, trial,
traceback.format_exc())
class RayTrialExecutor(TrialExecutor):
"""An implemention of TrialExecutor based on Ray."""
def __init__(self,
queue_trials=False,
reuse_actors=False,
ray_auto_init=False,
refresh_period=RESOURCE_REFRESH_PERIOD):
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 id generated.
# We use self._paused to store paused trials here.
self._paused = {}
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_runner(self, trial, reuse_allowed):
if (self._reuse_actors and reuse_allowed
and self._cached_actor is not None):
logger.debug("Reusing cached runner {} for {}".format(
self._cached_actor, trial.trial_id))
existing_runner = self._cached_actor
self._cached_actor = None
else:
if self._cached_actor:
logger.debug(
"Cannot reuse cached runner {} for new trial".format(
self._cached_actor))
self._cached_actor.stop.remote()
self._cached_actor.__ray_terminate__.remote()
self._cached_actor = None
existing_runner = None
cls = ray.remote(
num_cpus=trial.resources.cpu,
num_gpus=trial.resources.gpu,
memory=trial.resources.memory,
object_store_memory=trial.resources.object_store_memory,
resources=trial.resources.custom_resources)(
trial._get_trainable_cls())
trial.init_logger()
# We checkpoint metadata here to try mitigating logdir duplication
self.try_checkpoint_metadata(trial)
remote_logdir = trial.logdir
if existing_runner:
trial.runner = existing_runner
if not self.reset_trial(trial, trial.config, trial.experiment_tag):
raise AbortTrialExecution(
"Trainable runner reuse requires reset_config() to be "
"implemented and return True.")
return existing_runner
def logger_creator(config):
# Set the working dir in the remote process, for user file writes
if not os.path.exists(remote_logdir):
os.makedirs(remote_logdir)
if not ray.worker._mode() == ray.worker.LOCAL_MODE:
os.chdir(remote_logdir)
return NoopLogger(config, remote_logdir)
# Logging for trials is handled centrally by TrialRunner, so
# configure the remote runner to use a noop-logger.
return cls.remote(config=trial.config, logger_creator=logger_creator)
def _train(self, trial):
"""Start one iteration of training and save remote id."""
assert trial.status == Trial.RUNNING, trial.status
remote = trial.runner.train.remote()
# Local Mode
if isinstance(remote, dict):
remote = _LocalWrapper(remote)
self._running[remote] = trial
def _start_trial(self, trial, checkpoint=None):
"""Starts trial and restores last result if trial was paused.
Raises:
ValueError if restoring from checkpoint fails.
"""
prior_status = trial.status
self.set_status(trial, Trial.RUNNING)
trial.runner = self._setup_runner(
trial,
reuse_allowed=checkpoint is not None
or trial._checkpoint.value is not None)
if not self.restore(trial, checkpoint):
if trial.status == Trial.ERROR:
raise RuntimeError(
"Restore from checkpoint failed for Trial {}.".format(
str(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
else:
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.
"""
if stop_logger:
trial.close_logger()
if error:
self.set_status(trial, Trial.ERROR)
else:
self.set_status(trial, Trial.TERMINATED)
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 {}".format(trial.runner))
self._cached_actor = trial.runner
else:
logger.debug(
"Destroying actor for trial {}.".format(trial))
trial.runner.stop.remote()
trial.runner.__ray_terminate__.remote()
except Exception:
logger.exception("Error stopping runner for Trial %s", str(trial))
self.set_status(trial, Trial.ERROR)
finally:
trial.runner = None
def start_trial(self, trial, checkpoint=None):
"""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.
"""
self._commit_resources(trial.resources)
try:
self._start_trial(trial, checkpoint)
except Exception as e:
logger.exception("Error starting runner for Trial %s", str(trial))
error_msg = traceback.format_exc()
time.sleep(2)
self._stop_trial(trial, error=True, error_msg=error_msg)
if isinstance(e, AbortTrialExecution):
return # don't retry fatal Tune errors
try:
# This forces the trial to not start from checkpoint.
trial.clear_checkpoint()
logger.info(
"Trying to start runner for Trial %s without checkpoint.",
str(trial))
self._start_trial(trial)
except Exception:
logger.exception(
"Error starting runner for Trial %s, aborting!",
str(trial))
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
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("Returning resources for Trial %s.", str(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):
"""Tries to invoke `Trainable.reset_config()` 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.
Returns:
True if `reset_config` is successful else False.
"""
trial.experiment_tag = new_experiment_tag
trial.config = new_config
trainable = trial.runner
with warn_if_slow("reset_config"):
reset_val = ray.get(trainable.reset_config.remote(new_config))
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])
# 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):
for i in range(num_retries):
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 not resources:
logger.warning(
"Cluster resources not detected or are 0. Retrying...")
time.sleep(0.5)
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 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.DISK):
"""Saves the trial's state to a checkpoint."""
trial._checkpoint.storage = storage
trial._checkpoint.last_result = trial.last_result
if storage == Checkpoint.MEMORY:
trial._checkpoint.value = trial.runner.save_to_object.remote()
else:
# Keeps only highest performing checkpoints if enabled
if trial.keep_checkpoints_num:
try:
last_attr_val = trial.last_result[
trial.checkpoint_score_attr]
if (trial.compare_checkpoints(last_attr_val)
and not math.isnan(last_attr_val)):
trial.best_checkpoint_attr_value = last_attr_val
self._checkpoint_and_erase(trial)
except KeyError:
logger.warning(
"Result dict has no key: {}. keep"
"_checkpoints_num flag will not work".format(
trial.checkpoint_score_attr))
else:
with warn_if_slow("save_to_disk"):
trial._checkpoint.value = ray.get(
trial.runner.save.remote())
return trial._checkpoint.value
def _checkpoint_and_erase(self, trial):
"""Checkpoints the model and erases old checkpoints
if needed.
Parameters
----------
trial : trial to save
"""
with warn_if_slow("save_to_disk"):
trial._checkpoint.value = ray.get(trial.runner.save.remote())
if len(trial.history) >= trial.keep_checkpoints_num:
ray.get(trial.runner.delete_checkpoint.remote(trial.history[-1]))
trial.history.pop()
trial.history.insert(0, trial._checkpoint.value)
def restore(self, trial, checkpoint=None):
"""Restores training state from a given model checkpoint.
This will also sync the trial results to a new location
if restoring on a different node.
"""
if checkpoint is None or checkpoint.value is None:
checkpoint = trial._checkpoint
if checkpoint is None or checkpoint.value is None:
return True
if trial.runner is None:
logger.error("Unable to restore - no runner.")
self.set_status(trial, Trial.ERROR)
return False
try:
value = checkpoint.value
if checkpoint.storage == Checkpoint.MEMORY:
assert type(value) != Checkpoint, type(value)
trial.runner.restore_from_object.remote(value)
else:
# TODO: Somehow, the call to get the current IP on the
# remote actor can be very slow - a better fix would
# be to use an actor table to detect the IP of the Trainable
# and rsync the files there.
# See https://github.com/ray-project/ray/issues/5168
with warn_if_slow("get_current_ip"):
worker_ip = ray.get(trial.runner.current_ip.remote())
with warn_if_slow("sync_to_new_location"):
trial.sync_logger_to_new_location(worker_ip)
with warn_if_slow("restore_from_disk"):
ray.get(trial.runner.restore.remote(value))
trial.last_result = checkpoint.last_result
return True
except Exception:
logger.exception("Error restoring runner for Trial %s.", trial)
self.set_status(trial, Trial.ERROR)
return False
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:
return ray.get(
trial.runner.export_model.remote(trial.export_formats))
return {}
class RayTrialExecutor(TrialExecutor):
"""An implementation of TrialExecutor based on Ray."""
def __init__(self,
queue_trials=False,
reuse_actors=False,
ray_auto_init=False,
refresh_period=RESOURCE_REFRESH_PERIOD):
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 id generated.
# We use self._paused to store paused trials here.
self._paused = {}
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)
remote_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):
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._cached_actor.stop.remote()
self._cached_actor.__ray_terminate__.remote()
self._cached_actor = None
cls = ray.remote(
num_cpus=trial.resources.cpu,
num_gpus=trial.resources.gpu,
memory=trial.resources.memory,
object_store_memory=trial.resources.object_store_memory,
resources=trial.resources.custom_resources)(
trial.get_trainable_cls())
def logger_creator(config):
# Set the working dir in the remote process, for user file writes
os.makedirs(remote_logdir, exist_ok=True)
if not ray.worker._mode() == ray.worker.LOCAL_MODE:
os.chdir(remote_logdir)
return NoopLogger(config, remote_logdir)
# 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.
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 cls.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):
"""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.
See `RayTrialExecutor.restore` for possible errors raised.
"""
prior_status = trial.status
self.set_status(trial, Trial.RUNNING)
trial.set_runner(
runner or self._setup_remote_runner(
trial,
reuse_allowed=checkpoint is not None
or trial.has_checkpoint()))
self.restore(trial, checkpoint)
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 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.
"""
if stop_logger:
trial.close_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):
trial.runner.stop.remote()
trial.runner.__ray_terminate__.remote()
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):
"""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.
"""
self._commit_resources(trial.resources)
try:
self._start_trial(trial, checkpoint)
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):
"""Tries to invoke `Trainable.reset_config()` 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.
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_config"):
try:
reset_val = ray.get(
trainable.reset_config.remote(new_config),
DEFAULT_GET_TIMEOUT)
except RayTimeoutError:
logger.exception("Trial %s: reset_config 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], 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):
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 not resources:
logger.warning(
"Cluster resources not detected or are 0. Retrying...")
time.sleep(0.5)
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.
Args:
trial (Trial): The state of this 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 future, 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)
else:
with warn_if_slow("save_checkpoint_to_storage"):
# TODO(ujvl): Make this asynchronous.
value = ray.get(trial.runner.save.remote())
checkpoint = Checkpoint(storage, value, result)
with warn_if_slow("on_checkpoint", DEFAULT_GET_TIMEOUT) as profile:
try:
trial.on_checkpoint(checkpoint)
except Exception:
logger.exception("Trial %s: Error handling checkpoint %s",
trial, checkpoint.value)
return None
if profile.too_slow and trial.sync_on_checkpoint:
logger.warning(
"Consider turning off forced head-worker trial checkpoint "
"syncs by setting sync_on_checkpoint=False. Note that this "
"might result in faulty trial restoration for some worker "
"failure modes.")
return checkpoint.value
def restore(self, trial, checkpoint=None):
"""Restores training state from a given model checkpoint.
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.warning("Trial %s: Reading checkpoint into memory.",
trial)
data_dict = TrainableUtil.pickle_checkpoint(value)
with self._change_working_directory(trial):
remote = trial.runner.restore_from_object.remote(data_dict)
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")
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),
DEFAULT_GET_TIMEOUT)
return {}
def has_gpus(self):
if self._resources_initialized:
self._update_avail_resources()
return self._avail_resources.gpu > 0
@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
class _ResourceUpdater:
"""Periodic Resource updater for Tune.
Initially, all resources are set to 0. The updater will try to update resources
when (1) init ResourceUpdater (2) call "update_avail_resources", "num_cpus"
or "num_gpus".
The update takes effect when (1) Ray is initialized (2) the interval between
this and last update is larger than "refresh_period"
"""
def __init__(self, refresh_period: Optional[float] = None):
self._avail_resources = Resources(cpu=0, gpu=0)
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._last_resource_refresh = float("-inf")
self.update_avail_resources()
def update_avail_resources(self, num_retries=5):
if not ray.is_initialized():
return
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(
f"Cluster resources not detected or are 0. Attempt #{i + 1}...",
)
time.sleep(0.5)
resources = ray.cluster_resources()
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()
def debug_string(self, total_resources: Dict[str, Any]) -> str:
"""Returns a human readable message for printing to the console."""
if self._last_resource_refresh > 0:
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", 0.0)),
_to_gb(self._avail_resources.memory),
_to_gb(total_resources.pop("object_store_memory", 0.0)),
_to_gb(self._avail_resources.object_store_memory),
)
)
customs = ", ".join(
[
"{}/{} {}".format(
total_resources.get(name, 0.0),
self._avail_resources.get_res_total(name),
name,
)
for name in self._avail_resources.custom_resources
if not name.startswith(NODE_ID_PREFIX)
and (total_resources.get(name, 0.0) > 0 or "_group_" not in name)
]
)
if customs:
status += f" ({customs})"
return status
else:
return "Resources requested: ?"
def get_num_cpus(self) -> int:
self.update_avail_resources()
return self._avail_resources.cpu
def get_num_gpus(self) -> int:
self.update_avail_resources()
return self._avail_resources.gpu
def __reduce__(self):
# Do not need to serialize resources, because we can always
# update it again. This also prevents keeping outdated resources
# when deserialized.
return _ResourceUpdater, (self._refresh_period,)