/elastic

PyTorch elastic training

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

LicenseCircleCI

PyTorch Elastic

PyTorch Elastic (torchelastic) is a framework that enables distributed training jobs to be executed in a fault tolerant and elastic manner. It provides the primitives and interfaces for you to write your distributed PyTorch job in such a way that it can be run on multiple machines with elasticity; that is, your distributed job is able to start as soon as min number of workers are present and allowed to grow up to max number of workers without being stopped or restarted.

Use cases

Fault Tolerant Jobs

Jobs that run on infrastructure where nodes get replaced frequently, either due to flaky hardware or by design. Or mission critical production grade jobs that need to be run with resilience to failures.

Dynamic Capacity Management

Jobs that run on leased capacity that can be taken away at any time (e.g. AWS spot instances) or shared pools where the pool size can change dynamically based on demand.

Quickstart

Use one of the included examples and get a job running by following our Quickstart guide.

Requirements

torchelastic requires

  • python3
  • torch
  • etcd

Installation

pip install torchelastic

How torchelastic works

torchelastic induces users to think about their PyTorch jobs as a train_step and state. It provides the basic control loop that repetitively executes the user provided train_step being aware of faults, exceptions, and cluster membership changes.

Train Step

The train_step is a unit of work, typically (although not necessarily) mapping to the processing of a mini-batch of training data. All k workers in a distributed torchelastic job run the train_step(), each contributing to the final output. What each worker does in a train_step and what input data it consumes is a user-defined.

In the simplest use-case, torchelastic drives the execution of train_step until either:

  1. the input data is exhausted
  2. an unrecoverable failure condition is encountered
  3. some other user-defined end of job criteria is met

Each train_step may not be fully independent since the computations performed in the previous train_step can be used and/or updated in the next one. Information is carried across train_step calls using the state object.

State

The state, as the name implies, is an object that carries persistent information throughout the lifetime of the job and is expected to be updated on each train_step. For example, in a training job, one of the information that the state carries is the model weights. In practice it contains other (meta)data that must be persisted between train_steps, for instance, the offset or index of the data stream. The state object is the only input parameter to the train_step.

Train Loop

The train_step is executed in a train_loop by torchelastic. The train_loop is a fancy while-loop that enables the execution of the job with fault tolerance and elasticity. torchelastic works at train_step granularity, hence when a fault occurs during a train_step the computations performed during the failed train_step are considered lost and the state is restored to the previously succeeded train_step.

Rendezvous

Torchelastic jobs define a [min, max] range of number of workers that it can run with. For instance, [2, 10] means that the job can start when at least two workers are present, and can be scaled up to ten workers at runtime.

Each time there is a change in membership in the set of workers, torchelastic runs arendezvous, which serves the following purposes:

  1. barrier - all nodes will block until rendezvous is complete before resuming execution.
  2. role assignment - on each rendezvous each node is assigned a unique integer valued rank between [0, n) where n is the world size (total number of workers).
  3. world size broadcast - on each rendezvous all nodes receive the new world_size.

The resource manager is free to add/remove instances from a torchelastic job as long as the total number of workers remain within [2, 10]. This is what we refer to as elasticity. Additionally, in the event of a worker node failure, as long as the failed node is replaced, torchelastic will detect this event as a membership change and admit the new worker into the group, making the job fault-tolerant.

For additional details refer to the README in the rendezvous module.

Usage

Please refer to the usage documentation for details on how to write and configure a torchelastic job.

See the CONTRIBUTING file for how to help out.

License

torchelastic is BSD licensed, as found in the LICENSE file.