【TechWeb】12 month 12 Daily news , Sponsored by the National Engineering Laboratory of deep learning technology and application WAVE SUMMIT+ 2021 The deep learning developer summit was held today . This summit , Flying oar open source framework v2.2 Make a big show .
Ma Yanjun, senior director of Baidu deep learning technology platform department, and Baidu AI Xin Zhou, director of product R & D department , Just the new version of the propeller This feature is explained in detail with the application of ground deployment .
How will the new version of the propeller solve AI Provide a new answer to the landing industry ？ Look at it together. .
Practice integration and innovation , Continuous accumulation and breakthrough of core technology of flying propeller
Summit site , Ma Yanjun said , As the first self-developed propeller in China 、 Rich in functions 、 Open source and open industrial deep learning platform , Continuous accumulation and breakthrough of core technologies , Newly released open source framework 2.2 edition , Involving deep learning and development 、 Training 、 Text tasks are extremely optimized 、 Hardware efficient adaptation 、 Low threshold reasoning deployment and other innovative technologies , Enabling developers , Provide technical source power for industrial application and frontier exploration .
In terms of development , The paddle provides rich API, Support developer convenience 、 Develop deep learning models efficiently . Propeller frame v2.2 Of API More abundant 、 Efficient and maintain good compatibility , Targeted enrichment 100 Multiple API, It can support a wider range of model development , Especially for the model application of Scientific Computing , Fourier transform is added 、Jacobian/Hessian/VJP/JVP A series of API, Support quantum computing 、 Life science 、 Computational fluid dynamics 、 Molecular dynamics and other applications , Help explore cutting-edge technologies .
In terms of training , New release of end-to-end adaptive large-scale distributed training technology . For different models and hardware , Abstract into a unified distributed computing view and resource view , Through hardware aware segmentation and mapping function and end-to-end cost model , The optimal model segmentation and hardware combination strategy are searched , Set the model parameters 、 gradient 、 The optimizer state is allocated to different computing cards according to the optimal policy , Save storage 、 Load balancing 、 The purpose of improving training performance .
Based on a new end-to-end adaptive large-scale distributed training technology , Baidu PaddlePaddle in Pengcheng cloud brain II Adaptive optimization on Cluster , Before training speed reaches optimization 2.1 times . And Pengcheng, the world's first knowledge enhancement model recently released - Baidu · Literary heart , It is also based on end-to-end adaptive large-scale distributed training technology .
Text task , From text processing 、 Training 、 Decoding to deployment for full acceleration . Upgrade support for string tensors , Provide end-to-end text task development experience for developers . In terms of the pre training model , in the light of Transformer Encoder The network structure realizes the ultimate performance optimization , And through the user-defined operator function , Integrated NVIDIA FasterTransformer High performance operator . Based on these optimizations , The framework forms a whole process development experience integrating training and promotion for the pre training model , Save deployment code 94%.
Hardware access , multi-level 、 The low-cost hardware adaptation scheme reduces the adaptation cost between the framework and the chip . Baidu self research Kernel Primitive API、NNAdapter、 compiler CINN（ pre-release ） Three optimization schemes , Respectively for AI Operator Library 、 chart 、 Deep integration optimization of software and hardware at the back end of the compiler , The cost of hardware adaptation is greatly reduced , Enabling hardware ecosystem .
Continuously reduce the application threshold , Propeller model library 、 New upgrade of Enterprise Edition
In addition to the leading release of propeller deep learning framework technology , The summit also brought new upgrades to the industrial open source model library and enterprise version of the propeller .
Ma Yanjun said at the meeting , at present , Baidu PaddlePaddle's official support for the industrial open source algorithm model exceeds 400 individual , And publish 13 individual PP Series model , Achieve a balance in accuracy and performance , Completely open up the reasoning deployment tool chain .
While the application capacity of the propeller industry is comprehensively upgraded , The enterprise version of flying propeller is also focusing on improving the model deployment capability . At the meeting , Xinzhou has brought the deployment and upgrading of the model of the flying oar enterprise version and the flying oar EsayDL New release of desktop version .
Propeller enterprise edition includes EasyDL and BML Dual platform development mode , Committed to improving AI Development efficiency and resource utilization efficiency , At present, it has become the most widely used AI Development platform . The new upgrade of the model deployment is based on the propeller reasoning deployment tool chain , Deep integration with platform , Create automated and efficient enterprise deployment capabilities .
The first is automatic model combination compression , Significantly improve reasoning performance . be based on PaddleSlim, According to the characteristics of different models and hardware , Several fully automatic combined compression pipelines are designed , Can automatically select the best compression path . For common models , Accuracy loss control in 1% Next , The acceleration ratio can reach 3-5 times .
Secondly, based on the propeller reasoning engine , Widely adaptive reasoning chip . The new version uses a propeller inference engine , It is widely adapted to reasoning chips and has excellent performance . at present , Platform completed 9345 Real test and tuning of a combination of model chips , You can override 95% Demand scenarios for , Savings compared to self adaptation 97% Development time of .
Finally, model service and intelligent edge console , Significantly improve deployment efficiency . Especially the newly released intelligent edge console , It provides a fully visual operation interface , The efficiency of model and business integration is significantly improved , Model deployment time has been reduced from days to 5 minute . Xinzhou demonstrated how to 5 Minutes to let the robot dog learn new skills of gesture recognition .
Last , Flying propeller EasyDL The desktop version is newly released . Developers don't have to configure all kinds of environments , High efficiency local modeling can be achieved by one click high-speed installation on the desktop ,1 Minutes to install ,15 Model development can be completed in minutes , Local data management 、 Computing power scheduling 、 Deploy the application , Give Way AI“ You can get ”.
Baidu PaddlePaddle's industrial deep learning open source open platform originated from industrial practice , It is Baidu's practice of integration and innovation 、 Reducing the threshold of industry development AI Mass production platform . Iterative updates from generation to generation of the propeller , It is also a step-by-step upward climb of China's artificial intelligence industry .AI Promote China's industrial prosperity , The oars have been on the road .