def get_full_actor_cls(
self, trial: "Trial", actor_cls: ActorClass
) -> Optional[ActorClass]:
"""Get a fully configured actor class.
Returns the actor handle if the placement group is ready. In this case,
the placement group is moved to `self._in_use_pgs` and removed from
`self._ready`.
Args:
trial: "Trial" object to start
actor_cls: Ray actor class.
Returns:
Configured ActorClass or None
"""
pgf = trial.placement_group_factory
if not self._ready[pgf]:
return None
pg = self._ready[pgf].pop()
self._in_use_pgs[pg] = trial
self._in_use_trials[trial] = pg
logger.debug(f"For trial {trial} use pg {pg.id}")
# We still have to pass resource specs
if not pgf.head_bundle_is_empty:
# Pass the full resource specs of the first bundle per default
head_bundle = pg.bundle_specs[0].copy()
num_cpus = head_bundle.pop("CPU", 0)
num_gpus = head_bundle.pop("GPU", 0)
memory = head_bundle.pop("memory", None)
object_store_memory = head_bundle.pop("object_store_memory", None)
# Only custom resources remain in `head_bundle`
resources = head_bundle
return actor_cls.options(
placement_group=pg,
placement_group_bundle_index=0,
placement_group_capture_child_tasks=True,
num_cpus=num_cpus,
num_gpus=num_gpus,
memory=memory,
object_store_memory=object_store_memory,
resources=resources,
)
else:
return actor_cls.options(
placement_group=pg,
placement_group_capture_child_tasks=True,
num_cpus=0,
num_gpus=0,
resources={},
)
def java_actor_class(class_name):
from ray.actor import ActorClass
return ActorClass._ray_from_function_descriptor(
Language.JAVA,
JavaFunctionDescriptor(class_name, "<init>", ""),
0, # max_reconstructions,
None, # num_cpus,
None, # num_gpus,
None, # memory,
None, # object_store_memory,
None) # resources,
def java_actor_class(class_name: str):
"""Define a Java actor class.
Args:
class_name: Java class name.
"""
from ray.actor import ActorClass
return ActorClass._ray_from_function_descriptor(
Language.JAVA,
JavaFunctionDescriptor(class_name, "<init>", ""),
{},
)
def java_actor_class(class_name):
from ray.actor import ActorClass
return ActorClass._ray_from_function_descriptor(
Language.JAVA,
JavaFunctionDescriptor(class_name, "<init>", ""),
max_restarts=0,
max_task_retries=0,
num_cpus=None,
num_gpus=None,
memory=None,
object_store_memory=None,
resources=None,
accelerator_type=None,
)
def cpp_actor_class(create_function_name: str, class_name: str):
"""Define a Cpp actor class.
Args:
create_function_name: Create cpp class function name.
class_name: Cpp class name.
"""
from ray.actor import ActorClass
print("create func=", create_function_name, "class_name=", class_name)
return ActorClass._ray_from_function_descriptor(
Language.CPP,
CppFunctionDescriptor(create_function_name, "PYTHON", class_name),
{},
)
def get_full_actor_cls(self, trial: "Trial",
actor_cls: ActorClass) -> Optional[ActorClass]:
"""Get a fully configured actor class.
Returns the actor handle if the placement group is ready. In this case,
the placement group is moved to `self._in_use_pgs` and removed from
`self._ready`.
Args:
trial ("Trial"): "Trial" object to start
actor_cls: Ray actor class.
Returns:
Configured ActorClass or None
"""
pgf = trial.placement_group_factory
if not self._ready[pgf]:
return None
pg = self._ready[pgf].pop()
self._in_use_pgs[pg] = trial
self._in_use_trials[trial] = pg
# We still have to pass resource specs
# Pass the full resource specs of the first bundle per default
first_bundle = pg.bundle_specs[0].copy()
num_cpus = first_bundle.pop("CPU", None)
num_gpus = first_bundle.pop("GPU", None)
# Only custom resources remain in `first_bundle`
resources = first_bundle or None
if num_cpus is None:
# If the placement group specifically set the number
# of CPUs to 0, use this.
num_cpus = pgf.head_cpus
logger.debug(f"For trial {trial} use pg {pg.id}")
return actor_cls.options(
placement_group=pg,
placement_group_bundle_index=0,
placement_group_capture_child_tasks=True,
num_cpus=num_cpus,
num_gpus=num_gpus,
resources=resources)
def java_actor_class(class_name):
"""Define a Java actor class.
Args:
class_name (str): Java class name.
"""
from ray.actor import ActorClass
return ActorClass._ray_from_function_descriptor(
Language.JAVA,
JavaFunctionDescriptor(class_name, "<init>", ""),
max_restarts=0,
max_task_retries=0,
num_cpus=None,
num_gpus=None,
memory=None,
object_store_memory=None,
resources=None,
accelerator_type=None,
runtime_env=None)
def get_full_actor_cls(self, trial: Trial,
actor_cls: ActorClass) -> Optional[ActorClass]:
"""Get a fully configured actor class.
Returns the actor handle if the placement group is ready. In this case,
the placement group is moved to `self._in_use_pgs` and removed from
`self._ready`.
Args:
trial (Trial): Trial object to start
actor_cls: Ray actor class.
Returns:
Configured ActorClass or None
"""
pgf = trial.placement_group_factory
if not self._ready[pgf]:
return None
pg = self._ready[pgf].pop()
self._in_use_pgs[pg] = trial
self._in_use_trials[trial] = pg
# We still have to pass resource specs
# Pass the full resource specs of the first bundle per default
first_bundle = pg.bundle_specs[0].copy()
num_cpus = first_bundle.pop("CPU", None)
num_gpus = first_bundle.get("GPU", None)
# Only custom resources remain in `first_bundle`
resources = first_bundle or None
return actor_cls.options(placement_group=pg,
placement_group_bundle_index=0,
num_cpus=num_cpus,
num_gpus=num_gpus,
resources=resources)
def cpp_actor_class(create_function_name, class_name):
"""Define a Cpp actor class.
Args:
create_function_name (str): Create cpp class function name.
class_name (str): Cpp class name.
"""
from ray.actor import ActorClass
print("create func=", create_function_name, "class_name=", class_name)
return ActorClass._ray_from_function_descriptor(
Language.CPP,
CppFunctionDescriptor(create_function_name, "PYTHON", class_name),
max_restarts=0,
max_task_retries=0,
num_cpus=None,
num_gpus=None,
memory=None,
object_store_memory=None,
resources=None,
accelerator_type=None,
runtime_env=None,
)
