ray/doc/source/train/architecture.rst
2021-10-18 22:27:46 -07:00

44 lines
1.3 KiB
ReStructuredText

.. _train-arch:
Ray Train Architecture
======================
A diagram of the Ray Train architecture is provided below.
.. image:: train-arch.svg
:width: 70%
:align: center
Trainer
-------
The Trainer is the main class that is exposed in the Ray Train API that users will interact with.
* The user will pass in a *function* which defines the training logic.
* The Trainer will create an :ref:`Executor <train-arch-executor>` to run the distributed training.
* The Trainer will handle callbacks based on the results from the BackendExecutor.
.. _train-arch-executor:
Executor
--------
The executor is an interface which handles execution of distributed training.
* The executor will handle the creation of an actor group and will be initialized in conjunction with a backend.
* Worker resources, number of workers, and placement strategy will be passed to the Worker Group.
Backend
-------
A backend is used in conjunction with the executor to initialize and manage framework-specific communication protocols.
Each communication library (Torch, Horovod, TensorFlow, etc.) will have a separate backend and will take a specific configuration value.
WorkerGroup
-----------
The WorkerGroup is a generic utility class for managing a group of Ray Actors.
* This is similar in concept to Fiber's `Ring <https://uber.github.io/fiber/experimental/ring/>`_.