当前位置：网站首页>AI chief architect 8-aica-gao Xiang, in-depth understanding and practice of propeller 2.0

AI chief architect 8-aica-gao Xiang, in-depth understanding and practice of propeller 2.0

2022-06-28 08:15:00 【The mountain of ignorance, the valley of despair, the slope of 】

1、 Introduce

Insert picture description here
Catalog
Insert picture description here
The longitudinal , Hardware and under development , The transverse , Academia and Industry
Insert picture description here
history
Insert picture description here
Industry landing and developers to optimize
Insert picture description here

2、 In depth understanding of

Most concerned issues
Insert picture description here
Corresponding
Insert picture description here

2.0 Usability

Insert picture description here
A combination of motion and stillness , Foundation and high level API
Insert picture description here
tensor Calculation
Insert picture description here
Networking is convenient
Insert picture description here
High level networking
Insert picture description here
Integration of foundation and high level
Insert picture description here
Dynamic graph mode
Insert picture description here
Programming paradigm evolution ,cuda、 Static diagram 、 Dynamic graph
Insert picture description here
Dynamic and static conversion
Insert picture description here
Control flow comparison
Insert picture description here
High development efficiency , Low deployment efficiency , Static diagram networking in advance , Optimize network structure
Insert picture description here
Dynamic and static conversion
Insert picture description here
Model saving and loading , Dynamic graph code to build network structure , Static graph can directly save the network structure
Insert picture description here
Develop reasoning processes
Insert picture description here

2.1 Distributed training

Insert picture description here
Challenge
Insert picture description here
Large scale distributed training
Insert picture description here
Data parallelism , Tensor model parallelism - Distribute the parameters of the tensor , Pipeline parallel - Network structure splitting , Grouping parameter parallelism is more detailed on the basis of model parallelism
Insert picture description here
Set communication and parameter server , Go to center and have center
Insert picture description here
How to configure distributed training
Insert picture description here
Collective communication is simpler
Insert picture description here
Combined parallel training , Wenxin big model
Insert picture description here

2.2 Hardware matching and porting are convenient

Insert picture description here
Hardware ecology ,CPU,GPU Training ,GPU Reasoning （ Server side , Mobile , Web side ）,FPGU
Insert picture description here
Hardware test adaptation architecture diagram
Insert picture description here
Deployment advantages
Insert picture description here

2.4 It's easy to update iterations , The algorithm model base is complete

Insert picture description here

3、 Operator development

Insert picture description here
Operator writing , Forward and reverse
Insert picture description here
After writing the operator alignment , Compile
Insert picture description here
Use custom operators
Insert picture description here

QA:

Define the underlying principles and ideas of the framework

原网站

版权声明
本文为[The mountain of ignorance, the valley of despair, the slope of ]所创，转载请带上原文链接，感谢
https://yzsam.com/2022/179/202206280805465993.html

边栏推荐

猜你喜欢

随机推荐