Home

Welcome to the oneflow wiki!

OneFlow roadmap

April 2021

version: OneFlow v0.4.0

OneFlow 1.0 interface project

A new, easy-to-use Python user interface, fully compatible with PyTorch, support OOP modeling and distributed Eager training !

• oneflow.tensor
• oneflow.autograd
• oneflow.nn.Module
• high level API (oneflow.Model) support one code switch between Lazy and Eager
• oneflow.ComputeGraph
• oneflow.autograd.Function
• OpBuilder
• Multi-Client

2-D SBP

Support hierarchical N-D SBP Parallel for super large scale distributed training with data/model parallel at the same time.

• parallel distribution
• hierarchical boxing sub graph builder
• hierarchical parallel cast
• GPT-3 model release and benchmark

Device

Support more general computing chips and AI computing chips.

• extend device type
• support Cambrian