Use flying paddles to handle large model production in one stop. The PaddleFleetX large model development kit is the first in the industry.

On November 30, the WAVE SUMMIT 2022 Deep Learning Developer Summit, hosted by the National Engineering Research Center for Deep Learning Technology and Applications and hosted by Baidu Flying Paddle, was held as scheduled. At the summit, Baidu AI Technology Ecosystem General Manager Ma Yanjun released the latest technology and ecological progress of the Paddle Deep Learning Platform. The newly released Paddle Open Source Framework version 2.4 brought the industry's first end-to-end large model development kit PaddleFleetX, uniting 12 companies. Hardware ecological partners released the Flying Paddle ecological release version, and the AI ​​Studio learning and training community upgraded to launch enterprise training and ecological heterogeneous computing centers, etc. This series of measures marks that Feipiao continues to consolidate its AI technology base and continues to deepen industrial integration and innovation.

Picture: Baidu AI Technology Ecosystem General Manager Ma Yanjun shares the latest release of the Flying Paddle platform

The Flying Paddle open source framework version 2.4 is released, continuing to optimize with underlying technology innovation Performance

The 2.4 version of the Flying Paddle upgrade makes framework development more flexible and convenient, continues to lead the way in large-scale model distributed training, and enables high-performance inference deployment in all scenarios.

In terms of development, version 2.4 of the Flying Paddle open source framework has added more than 160 APIs for important scenarios such as sparse computing and graph learning, and the API development threshold and cost have been significantly reduced, making One-third of the new APIs this time come from contributions from ecological developers. In response to the needs of AI for Science scenarios, version 2.4 implements a general high-order automatic differentiation function to better support scientific computing-related applications. At the same time, Flying Paddle has comprehensively improved the scalability and deployment flexibility of the core dynamic-to-static technology. The new model's dynamic-to-static success rate reaches 92%, giving full play to the respective advantages of dynamic and static images.

In terms of training, version 2.4 has been newly upgraded and launched the GPU-based ultra-large-scale graph model training engine PGLBox, which is the first in the industry to implement an integrated graph learning solution that can simultaneously support complex algorithms, ultra-large graphs, and ultra-large discrete models. . In addition, Fei Paddle's collective communication distributed training performance has also been optimized to the extreme, providing a comprehensive and rich distributed training performance optimization system for large model training. Based on this, Fei Paddle has won the internationally authoritative AI training benchmark test twice in a row this year. No. 1 on the MLPerf Training list.

As the "last mile" of AI implementation in the industry, the reasoning and deployment process of the model is very critical. First of all, for the reasoning of large models, version 2.4 of the Flying Paddle open source framework supports functions such as adaptive model segmentation and distributed reasoning. Relying on the dynamic and static capabilities of the Flying Paddle framework, it can achieve automatic deep fusion and high-performance optimization, fully supporting large model applications. Landed. At the same time, in order to fundamentally solve the three major problems faced by AI application implementation: scene fragmentation, high development costs, and slow inference speed, Feipiao has launched a new full-scenario high-performance AI deployment tool FastDeploy, which is a one-stop solution. To meet the deployment requirements of multiple scenarios, multiple frameworks, and multiple hardware on devices, edges, and clouds, the API design is not only unified and easy to use, but also supports deep linkage between automated compression and high-performance inference engines, giving full play to the advantages of integrated software and hardware, and has the industry's leading edge Leading reasoning performance provides optimal solutions for AI industry applications.

Continue to lower the application threshold and accelerate the implementation of AI applications

The application of large models can lower the threshold for AI applications, but the development, training, inference and deployment processes of large models still pose great challenges. In order to better support the implementation of large model applications, PaddleFleetX has released a new end-to-end large model development kit PaddleFleetX. PaddleFleet Training and debugging of hardware platforms in multi-cloud environments. At the same time, Paddle Fleet Quantitative deployment.

At this summit, the number of open source algorithms in the Flying Paddle industry-level open source model library has been added to more than 600, covering mainstream task scenarios such as vision, natural language, and timing modeling; The number of PP series featured models with balanced precision and performance that have been polished by industrial practice has increased to 42; the number of practical examples of flying paddle industry has increased to 68, covering ten key industry scenarios such as finance, industry, transportation, Internet, security, and education; the number of flying paddle industry models has been released The one-stop entrance to the paddle industry-level model library aggregates model knowledge and toolsets, and connects the entire process of model selection, rapid experience, model development and use, and model deployment. Fei Paddle has accumulated strength from the aspects of open source industrial models and reference examples to provide convenient support for enterprises to apply AI technology and jointly solve the pain points and difficulties of the industry.

Join hands with ecological partners in various fields to create and share, and jointly build a prosperous Flying Paddle AI ecosystem

Promoting the application of artificial intelligence in the industry cannot be separated from the cooperation with hardware ecological partners. In May of this year, Flying Paddle jointly launched the "Hardware Ecosystem Co-Creation Plan" with its hardware ecosystem partners to fully cooperate in multiple dimensions such as joint research and development, resource sharing, joint licensing, and training empowerment. As of November, the number of members of Fei Paddle’s “Hardware Ecosystem Co-Creation Plan” has increased from 13 to 28. At the same time, Flying Paddle has joined hands with 12 manufacturers including Nvidia, Arm, Cambrian, Kunlun Core, Tianshu Intelligent Core, Graphcore, and Suiyuan to jointly release the Flying Paddle ecological release to provide developers with a better integrated software and hardware experience. At the same time, we will work together to build a prosperous hardware ecosystem.

In order to further accelerate the intelligent upgrading of the AI ​​industry, Fei Paddle has gone deep into the industrial scene and joined forces with China Energy Group, Industrial and Commercial Bank of China, China Unicom, PetroChina, and China Academy of Railway Sciences. , China Mobile, China FAW and other leading companies in the industry have released a collection plan for industrial practice examples to link companies and developers to solve key problems in actual industrial scenarios in various industries and help the AI ​​industry to implement.

Feipiao Construction’s AI Studio learning and training community is committed to making AI learning and application easier, and has now become the largest AI developer community in China. At this summit, AI Studio upgraded and added two new sections: corporate training and ecological heterogeneous computing power center.

Enterprise training provides a practical training platform for enterprises to apply deep learning. Up to now, Feipiao has jointly held events with more than 20 well-known companies to jointly explore AI application solutions and cultivate AI talents. For a long time, AI Studio has provided developers with rich computing resources including CPU and GPU, and has been connected to Sugon DCU computing power to provide developers with an online experience center and a convenient experience environment for hardware infrastructure and ecology. The heterogeneous computing power center welcomes more hardware manufacturers to settle in.

The Baidu Flying Paddle deep learning open source open platform derived from industrial practice is an AI large-scale production platform for Baidu to practice integrated innovation and lower the threshold. In the future, as the core technology of deep learning continues to be developed, Fei Paddle will better support the application of AI and large models, continue to build a more prosperous AI ecosystem, promote the industry to accelerate intelligent upgrading, and let AI benefit thousands of industries.

The above is the detailed content of Use flying paddles to handle large model production in one stop. The PaddleFleetX large model development kit is the first in the industry.. For more information, please follow other related articles on the PHP Chinese website!