riidZddlZddlZddlZddlmZmZddlmZm Z m Z ddl m Z m Z mZmZdedefd ZGd d ZdS) zTracker for XGBoost collective.N)IntEnumunique)DictOptionalUnion)_LIB _check_call_deprecate_positional_args make_jcargsaddrreturncDtj|dddS)z Get network family from address.Nr)socket getaddrinfo)r s U/var/www/html/bet.cuttalo.com/ml/venv/lib/python3.11/site-packages/xgboost/tracker.py get_familyr s  dD ) )! ,Q //ceZdZdZeGddeZe dddddede e d ed e d ed d f dZ ddZ ddZ ddZdd e ed d fdZd ee ee efffdZd S) RabitTrackeraYTracker for the collective used in XGBoost, acting as a coordinator between workers. Parameters ---------- n_workers: The total number of workers in the communication group. host_ip: The IP address of the tracker node. XGBoost can try to guess one by probing with sockets. But it's best to explicitly pass an address. port: The port this tracker should listen to. XGBoost can query an available port from the OS, this configuration is useful for restricted network environments. sortby: How to sort the workers for rank assignment. The default is host, but users can set the `DMLC_TASK_ID` via arguments of :py:meth:`~xgboost.collective.init` and obtain deterministic rank assignment through sorting by task name. Available options are: - host - task timeout : Timeout for constructing (bootstrap) and shutting down the communication group, doesn't apply to communication when the group is up and running. The timeout value should take the time of data loading and pre-processing into account, due to potential lazy execution. By default the Tracker doesn't have any timeout to avoid pre-mature aborting. The :py:meth:`.wait_for` method has a different timeout parameter that can stop the tracker even if the tracker is still being used. A value error is raised when timeout is reached. Examples -------- .. code-block:: python from xgboost.tracker import RabitTracker from xgboost import collective as coll tracker = RabitTracker(host_ip="127.0.0.1", n_workers=2) tracker.start() with coll.CommunicatorContext(**tracker.worker_args()): ret = coll.broadcast("msg", 0) assert str(ret) == "msg" ceZdZdZdZdS)RabitTracker._SortByrrN)__name__ __module__ __qualname__HOSTTASKrr_SortByrNsrrrhost)sortbytimeout n_workershost_ipportr!r"rNc ftj}|dvrtd|t|t |||d|dkr |jjn |jjt|}ttj |tj |||_ dS)N)r taskz-Expecting either 'host' or 'task' for sortby.rabitr )r r#r%dmlc_communicatorr!r")ctypesc_void_p ValueErrorrr rrrintr r XGTrackerCreatebyrefhandle)selfr#r$r%r!r"r0argss r__init__zRabitTracker.__init__Ss"" ) ) )LMM M   w   %(.&(8(84<$$dl>OLL     D(v|F/C/CDDEEE rc~t|dr,|j}|`ttj|dSdS)zInternal function for testing.r0N)hasattrr0r r XGTrackerFree)r1r0s rfreezRabitTracker.freensI 4 " " 4[F *622 3 3 3 3 3 4 4rc.|dSN)r7r1s r__del__zRabitTracker.__del__us rclttj|jt dS)zStart the tracker. Once started, the client still need to call the :py:meth:`wait_for` method in order to wait for it to finish (think of it as a thread). N)r r XGTrackerRunr0r r:s rstartzRabitTracker.startxs+ D%dk;==AABBBBBrcpttj|jt |dS)zWait for the tracker to finish all the work and shutdown. When timeout is reached, a value error is raised. By default we don't have timeout since we don't know how long it takes for the model to finish training. )r"N)r r XGTrackerWaitForr0r )r1r"s rwait_forzRabitTracker.wait_fors2 D)$+{77S7S7STTUUUUUrctj}ttj|jtj||jJtj |j}|S)zGet arguments for workers.) r*c_char_pr r XGTrackerWorkerArgsr0r/valuejsonloads)r1c_envenvs r worker_argszRabitTracker.worker_argssZ!!D,T[&,u:M:MNNOOO{&&&j%% r)r)rNr9)rrr__doc__rrrr r-rstrr3r7r;r>rArrrJrrrrrso::x ' V   #     44444CCCCVV VVVVVT#uS#X"67rr)rKr*rFrenumrrtypingrrrcorer r r r rLr-rrrrrrPs%% ((((((((((LLLLLLLLLLLL0S0S0000 }}}}}}}}}}r