Skip to content

AI-ML-Projects/ftlib

 
 

Repository files navigation

FTLib

Build Status License

FTLib (Fault-Tolerant Library) is a framework to keep data-parallel distributed training continue regardless worker loss or join. It exposes collective communication APIs with fault-tolerance support by gluing a consensus to a communication library, both of which can be user-specific. A distributed training using FTLib is able to continue as long as at least one single worker is alive and when new workers join the training.

Status

Prototyping

Design

Develop Guide

TODO Please refer to the design docs.

See also

Getting started

Where to use FTLib

  • Less reliable infrastructure/script

Distributed training jobs running on less reliable infrastructure risks more as any worker or communication failure will leads to the termination of the entire job.

  • Dynamic workload system

A system may reduce the total workload of distributed training jobs to release resources so that resource can be squeezed out for jobs with higher priority. Without such jobs with higher-priority, the system can increase the workload to avoid resource idling.

Requirements

The requirements for using FTLib differs with choices of consensus and communication library. Please refer the requirements.txt under each consensus and communication library(Not available, still in todo list).

Usage

Please refer test for details on how to use FTLib in distributed training.

Layout

.
├── CHANGELOG.md
├── deploy
├── docs
│   ├── design
│   └── imgs
├── ftlib
│   ├── consensus
│   ├── commlib
│   ├── ftlib_status.py
│   ├── __init__.py
│   └── rank_assign_scheme.py
├── LICENSE
├── OWNERS
├── README.md
├── requirements.txt
├── ROADMAP
├── scripts
└── test

License

FTLib is Apache license. Implementations of consensus and communication library may come with different licenses.

About

Fault-tolerant for DL frameworks

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 79.4%
  • C++ 11.5%
  • Dockerfile 3.2%
  • Go 2.9%
  • CMake 1.5%
  • Shell 1.5%