WeHub Mirror
92 lines
3.3 KiB
Python
92 lines
3.3 KiB
Python
from typing import TYPE_CHECKING, Dict
|
|
|
|
from ray.train.v2._internal.exceptions import RayTrainError
|
|
from ray.util.annotations import PublicAPI
|
|
|
|
if TYPE_CHECKING:
|
|
from ray.train.v2.api.preemption import PreemptionInfo
|
|
|
|
|
|
@PublicAPI(stability="alpha")
|
|
class TrainingFailedError(RayTrainError):
|
|
"""Exception raised when training fails from a `trainer.fit()` call.
|
|
This is either :class:`ray.train.WorkerGroupError` or :class:`ray.train.ControllerError`.
|
|
"""
|
|
|
|
|
|
@PublicAPI(stability="alpha")
|
|
class WorkerGroupError(TrainingFailedError):
|
|
"""Exception raised from the worker group during training.
|
|
|
|
Args:
|
|
error_message: A human-readable error message describing the training worker failures.
|
|
worker_failures: A mapping from worker rank to the exception that
|
|
occurred on that worker during training.
|
|
"""
|
|
|
|
def __init__(self, error_message: str, worker_failures: Dict[int, Exception]):
|
|
super().__init__("Training failed due to worker errors:\n" + error_message)
|
|
self._error_message = error_message
|
|
self.worker_failures = worker_failures
|
|
|
|
def __reduce__(self):
|
|
return (self.__class__, (self._error_message, self.worker_failures))
|
|
|
|
|
|
@PublicAPI(stability="alpha")
|
|
class ControllerError(TrainingFailedError):
|
|
"""Exception raised when training fails due to a controller error.
|
|
|
|
Args:
|
|
controller_failure: The exception that occurred on the controller.
|
|
"""
|
|
|
|
def __init__(self, controller_failure: Exception):
|
|
super().__init__(
|
|
"Training failed due to controller error:\n" + str(controller_failure)
|
|
)
|
|
self.controller_failure = controller_failure
|
|
self.with_traceback(controller_failure.__traceback__)
|
|
|
|
def __reduce__(self):
|
|
return (self.__class__, (self.controller_failure,))
|
|
|
|
|
|
@PublicAPI(stability="alpha")
|
|
class PreemptionError(TrainingFailedError):
|
|
"""Exception raised when training is interrupted by node preemption.
|
|
|
|
Distinct from :class:`WorkerGroupError` so that a planned preemption
|
|
consumes a separate retry budget (``FailureConfig.max_preemption_failures``,
|
|
default -1 = unlimited) rather than ``max_failures``, which is reserved for
|
|
real failures (OOM, hardware faults, user-code bugs).
|
|
|
|
Args:
|
|
preemption_info: Which nodes / world ranks were preempted and the
|
|
reclaim deadline, for logging and debugging which preemption caused
|
|
the restart.
|
|
drain_timed_out: True when Ray Train stopped waiting because the reclaim
|
|
deadline passed while workers were still running (and tore them down
|
|
itself); False when every worker had already exited. Distinguishes
|
|
a forced teardown from an observed one when debugging.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
preemption_info: "PreemptionInfo",
|
|
drain_timed_out: bool = False,
|
|
):
|
|
self.preemption_info = preemption_info
|
|
self.drain_timed_out = drain_timed_out
|
|
super().__init__(
|
|
"Training was interrupted by node preemption "
|
|
f"(preempted_ranks={preemption_info.preempted_ranks}, "
|
|
f"drain_timed_out={drain_timed_out})."
|
|
)
|
|
|
|
def __reduce__(self):
|
|
return (
|
|
self.__class__,
|
|
(self.preemption_info, self.drain_timed_out),
|
|
)
|