【Ray】ray.remote和option

https://docs.ray.io/en/latest/ray-core/package-ref.html?highlight=ray.remote#ray-remote

1 ray.remote

定义义远程函数或 actor 类。
remote 支持重启、分配资源等功能。

用法1：装饰器的方式

用作装饰器，来修饰函数或者类。
比如：

>>> import ray
>>>
>>> @ray.remote
... def f(a, b, c):
...     return a + b + c
>>>
>>> object_ref = f.remote(1, 2, 3)
>>> result = ray.get(object_ref)
>>> assert result == (1 + 2 + 3)
>>>
>>> @ray.remote
... class Foo:
...     def __init__(self, arg):
...         self.x = arg
...
...     def method(self, a):
...         return self.x + a
>>>
>>> actor_handle = Foo.remote(123)
>>> object_ref = actor_handle.method.remote(321)
>>> result = ray.get(object_ref)
>>> assert result == (123 + 321)

用法2：作为函数使用

使用函数调用来创建远程函数或actor。

>>> def g(a, b, c):
...     return a + b + c
>>>
>>> remote_g = ray.remote(g)
>>> object_ref = remote_g.remote(1, 2, 3)
>>> assert ray.get(object_ref) == (1 + 2 + 3)

>>> class Bar:
...     def __init__(self, arg):
...         self.x = arg
...
...     def method(self, a):
...         return self.x + a
>>>
>>> RemoteBar = ray.remote(Bar)
>>> actor_handle = RemoteBar.remote(123)
>>> object_ref = actor_handle.method.remote(321)
>>> result = ray.get(object_ref)
>>> assert result == (123 + 321)

2 option

用来改变动态修改remote定义的参数。配置并覆盖任务调用参数。 参数与可以传递给 ray.remote 的参数相同。不支持覆盖 max_calls。

>>> @ray.remote(num_gpus=1, max_calls=1, num_returns=2)
... def f():
...     return 1, 2
>>>
>>> f_with_2_gpus = f.options(num_gpus=2) 
>>> object_ref = f_with_2_gpus.remote() 
>>> assert ray.get(object_ref) == (1, 2) 

>>> @ray.remote(num_cpus=2, resources={"CustomResource": 1})
... class Foo:
...     def method(self):
...         return 1
>>>
>>> Foo_with_no_resources = Foo.options(num_cpus=1, resources=None)
>>> foo_actor = Foo_with_no_resources.remote()
>>> assert ray.get(foo_actor.method.remote()) == 1

3 remote参数

1. num_returns – This is only for remote functions. It specifies the
   number of object refs returned by the remote function invocation.
   Pass “dynamic” to allow the task to decide how many return values to
   return during execution, and the caller will receive an
   ObjectRef[ObjectRefGenerator] (note, this setting is experimental).

   num_cpus – The quantity of CPU cores to reserve for this task or for
   the lifetime of the actor.

   num_gpus – The quantity of GPUs to reserve for this task or for the
   lifetime of the actor.

   resources (Dict[str, float]) – The quantity of various custom
   resources to reserve for this task or for the lifetime of the actor.
   This is a dictionary mapping strings (resource names) to floats.

   accelerator_type – If specified, requires that the task or actor run
   on a node with the specified type of accelerator. See
   ray.accelerators for accelerator types.

   memory – The heap memory request for this task/actor.

   max_calls – Only for remote functions. This specifies the maximum
   number of times that a given worker can execute the given remote
   function before it must exit (this can be used to address memory
   leaks in third-party libraries or to reclaim resources that cannot
   easily be released, e.g., GPU memory that was acquired by
   TensorFlow). By default this is infinite.

   max_restarts – Only for actors. This specifies the maximum number of
   times that the actor should be restarted when it dies unexpectedly.
   The minimum valid value is 0 (default), which indicates that the
   actor doesn’t need to be restarted. A value of -1 indicates that an
   actor should be restarted indefinitely.

   max_task_retries – Only for actors. How many times to retry an actor
   task if the task fails due to a system error, e.g., the actor has
   died. If set to -1, the system will retry the failed task until the
   task succeeds, or the actor has reached its max_restarts limit. If
   set to n > 0, the system will retry the failed task up to n times,
   after which the task will throw a RayActorError exception upon
   ray.get. Note that Python exceptions are not considered system
   errors and will not trigger retries.

   max_retries – Only for remote functions. This specifies the maximum
   number of times that the remote function should be rerun when the
   worker process executing it crashes unexpectedly. The minimum valid
   value is 0, the default is 4 (default), and a value of -1 indicates
   infinite retries.

   runtime_env (Dict[str, Any]) – Specifies the runtime environment for
   this actor or task and its children. See Runtime environments for
   detailed documentation. This API is in beta and may change before
   becoming stable.

   retry_exceptions – Only for remote functions. This specifies whether
   application-level errors should be retried up to max_retries times.
   This can be a boolean or a list of exceptions that should be
   retried.

   scheduling_strategy – Strategy about how to schedule a remote
   function or actor. Possible values are None: ray will figure out the
   scheduling strategy to use, it will either be the
   PlacementGroupSchedulingStrategy using parent’s placement group if
   parent has one and has placement_group_capture_child_tasks set to
   true, or “DEFAULT”; “DEFAULT”: default hybrid scheduling; “SPREAD”:
   best effort spread scheduling; PlacementGroupSchedulingStrategy:
   placement group based scheduling.

   _metadata – Extended options for Ray libraries. For example, _metadata={“workflows.io/options”: } for Ray workflows.