ONEFLOW V0.8.0 officially released

It's today OneFlow Open source 717 God ,OneFlow v0.8.0 Official release . This update contains 523 individual commit, Please update the list completely see link ：https://github.com/Oneflow-Inc/oneflow/releases/tag/v0.8.0, Welcome to download and experience the new version , Looking forward to your feedback .

 

OneFlow v0.8.0  It mainly includes the following new highlights, functions and optimization ：

 

1. Compatibility

 

PyTorch compatible API More perfect . Add a series of and PyTorch 1.10.0 Compatible functions and interfaces , Including new and PyTorch The alignment of the 68 individual API, Repair the 84 Operator and interface compatibility bug, Support users will be more PyTorch The model is migrated to OneFlow.

 

2.Global Operators support

 

All operator pairs Global Tensor The support of is more complete and efficient . Repair the 28 individual Global Tensor dependent bug, newly added 180 individual Global Operator single test , Users can use Global Tensor Develop distributed models .

 

3. performance

 

Graph The performance of advanced features is more mature ：

 

• In addition to the original ZeRO-DP outside ,ZeRO The zero redundancy optimizer can work with MP,2-D,3-D Use in parallel , Further save the cost of video memory .
• Graph Proposed a new pipelined parallel API, While simplifying pipelined parallel configuration, accelerate pipelined parallel and 3-D Parallel performance .
• In order to further improve Graph.debug Debugging efficiency , Add about logic diagram 、light plan Physical diagram 、 Memory analysis 、Python Multi dimensional debugging functions such as stack information .

 

 

 

be based on OneFlow v0.8.0 and LiBai v0.2.0 , GPT and BERT 3-D The speed optimization under parallel is obvious , The speed performance in multiple dimensions exceeds that in the same configuration Megatron-LM（ See details for data ：https://libai.readthedocs.io/en/latest/tutorials/get_started/Benchmark.html）.

 

4.OneEmbedding Components

 

OneEmbedding It is an expansion component specially designed for large-scale recommendation system , High performance 、 Scalable 、 High flexibility . It has the following characteristics ：

 

• Support for tiered storage , Dynamic capacity expansion Embedding, Users can expand at a lower cost Embedding Capacity
• Hybrid parallel strategy , It can easily expand the model horizontally to the scene of multiple machines and multiple cards
• Communication quantization compression function , In the parallel scenario , Quantize and compress the communication data , To reduce traffic , Improve your training speed
• Efficient data pipeline , Execute the parts of the model without data dependency in advance , Overlap in time
• Support automatic hybrid accuracy training , Part of the calculation will be converted into FP16 Data type calculation , Improve training speed while reducing the occupation of video memory , And it can ensure the convergence accuracy of the model
• Provide a series of high-performance for common operations of recommended system models CUDA operator
• Support flexible model building

 

API file ：

https://docs.oneflow.org/master/cookies/one_embedding.html

 

5. Multi device adaptation

 

OneFlow Provide a simple, efficient and extensible hardware abstraction layer EP（Execution Provider）, To cope with the complexity of adapting to different hardware . After introducing the hardware abstraction layer , Users do not need to pay attention to the specific implementation details of the underlying hardware and framework , Each module of the framework can be adapted to new hardware devices without modification , meanwhile , Users only need to follow the Convention of hardware abstract interface and the actual situation of hardware equipment , Implement a series of interfaces , You can complete the hardware adaptation .

 

EP It also defines a set of basic computing interfaces Primitive, be based on Primitive The interface is re implemented Kernel. comparison EP Runtime interface provided ,Primitive The interface provided is more flexible , Different interfaces are independent of each other , Each interface represents the specific computing power that a certain hardware device can provide .

 

6. Debugging tool stack

 

Add debugging tool stack OneFlow-Profiler And AutoProf.

 

OneFlow-Profiler A tool used to collect various performance related information in the framework execution process . This tool can count the execution time of operators or system components , Allocation of memory and video memory , At the same time, the input and parameter information corresponding to the operator can be recorded , This information can be used by developers to analyze the most expensive part of the framework implementation , To achieve targeted optimization （https://github.com/Oneflow-Inc/oneflow/pull/8047）.

 

 

AutoProf It's a test OneFlow and PyTorch Framework for operator performance , by OneFlow Provide elegance 、 Detect efficiently OneFlow API And PyTorch API Whether the method of alignment , Allow users to automatically compare OneFlow API And PyTorch API Performance results of （https://github.com/Oneflow-Inc/oneflow/pull/8207）.

 

7. Handling of abnormal error reports

 

Greatly improve OneFlow API Exception handling flow in , perfect API Error reporting prompt information under abnormal conditions .

 

8.API file

 

perfect OneFlow API The content of the document is 20 The rest , According to the function API Document module reconfiguration . Except for general operators API Outside , Detailed instructions OneFlow oneflow.nn.graph、oneflow.embedding、oneflow.autograd  And other modules and environment variables .

 

see OneFlow v0.8.0 Complete update , Please click on “ Read the original ”.

Everyone else is watching

• OneFlow Source list ：GDB Compile debugging

• Reading Pathways： The next step forward is OneFlow

• OneFlow The source code parsing ： Automatic inference of operator signature

• Hinton： My 50 years of in-depth study career and Research on mental skills

• LLVM The father of ： Why should we rebuild AI Infrastructure software

• Quantitative model of parallel computing and its application in deep learning engine

• Large model training is difficult ？ Superior efficiency 、 Easy-to-use “ Li Bai ” Here comes the model library

---

This article is from WeChat official account. - OneFlow（OneFlowTechnology）.
If there is any infringement , Please contact the [email protected] Delete .
Participation of this paper “OSC Source creation plan ”, You are welcome to join us , share .