TorchX

TorchX is an application spec SDK for PyTorch related apps. It defines standard structs called specs that represent the job definition of an application. The application spec is the common language between TorchX runners and pipeline adapters. Once an application’s spec is created, the application can be run as a standalone job on a cluster or as a stage in an ML pipeline/workflow. TorchX works with several mainstream job schedulers and ML pipeline platforms so chances are you are already familiar with one that works with TorchX.

Just getting started? First learn the basic concepts and take a look at the builtin components library.

Not finding the component or adapter you are looking for? Write a custom one that fits your needs by using our SDK.

Documentation

Usage

• Basics
• Quickstart
• CLI
• Configuring
• App Best Practices
• Component Best Practices

Examples

• Application Examples
• Pipelines Examples

API

• torchx.specs
• torchx.components
• torchx.runner
• torchx.schedulers
• torchx.pipelines

Components Library

Components

• Train
• Serve
• Interpret
• Metrics
• Hyperparameter Optimization
• Base
• Distributed
• Utils

Runtime Library

Runtime

• Overview
• Hyper Parameter Optimization
• Tracking

Works With

Schedulers

• Localhost
• Kubernetes
• Slurm

Pipeline Adapters

• Kubeflow Pipelines

Experimental

Beta Features

• Coming Soon