As the root technology of software, artificial intelligence framework has become the core force to accelerate the development of artificial intelligence model and promote the development of industrial intelligence. The Shengsi Artificial Intelligence Framework Summit 2024, with the theme of "Rising for Wisdom, the Source of Innovation", was held at the Beijing National Convention Center on the 22nd, aiming at bringing together the innovative forces of the AI industry, promoting the continuous innovation of root technology, and building a new ecology of artificial intelligence open source.
As the mainstream open source framework in China, Shengsi MindSpore builds a vibrant developer community and is committed to promoting the progress of China’s artificial intelligence industry. In the past two years, open source framework products including Shengsi MindSpore have been developing continuously. IDC research shows that Shengsi MindSpore has made outstanding achievements in ecological construction.
He You, an academician of China Academy of Engineering and vice chairman of China Artificial Intelligence Society, said in his speech: China Artificial Intelligence Society has worked closely with Shengsi MindSpore, joined hands with Pengcheng Laboratory and Shengsi MindSpore to carry out the Academic Fund 2.0 project, accelerated the original academic innovation based on Shengsi and the publication of papers, brought the original force of academic talents into play, and combined with the technical originality of Shengsi, supported the prosperity of artificial intelligence industry with academic innovation.
Zhou Jun, president of Huawei ICTMarketing, said: The parallel computing ability, concise programming ability and convenient deployment ability of the AI framework have gradually become the key success factors of large-scale model training. Shengsi will continue to innovate and explore, focus on root technology, and use a more open technology system to enable partners and developers to achieve large-scale model innovation more flexibly and accelerate the application scale of large models.
JackDongarra, winner of the Turing Prize in 2021 and an outstanding professor of computer science at the University of Tennessee, shared the development trend of computing and framework. He pointed out that open source has many advantages in essence and can provide a better software foundation for the future. The development of MindSpore open source framework has promoted the improvement of software performance.
AIforScience new paradigm, generative AI-enabled aerodynamic shape design
Tang Zhigong, academician of China Academy of Sciences and chairman of China Aerodynamics Society, said: Based on MindSpore, the large model platform of generative aerodynamic design breaks the traditional design paradigm, and the design time is shortened from month level to minute level, which meets the requirements of conceptual design. In the future, the platform will be extended to aviation, aerospace, shipbuilding, high-speed rail, energy, automobiles and other industries, leading the leap-forward development of industrial equipment design and manufacturing capabilities.
Independent innovation and development of generative aerodynamic design large model platform based on Shengsi AI framework. In the model development stage, MindSporeFlow, a lifting framework and fluid mechanics suite, provides a comprehensive library of scientific computing algorithms and a general interface for the model, which improves the efficiency of model development. In the model training stage, using MindSpore multi-dimensional parallel interface, based on the computing power support of Chengdu Intelligent Computing Center, the model and data can be expanded efficiently; In the model deployment stage, the professional knowledge is integrated into the large model platform of aerodynamic design by using the Shengsi large model suite. Non-AI tools such as large language model, aerodynamic shape design model, aerodynamic prediction model and storm software are connected in series, and the aerodynamic design large model platform can support a variety of aerodynamic shape design scenarios.
Shengsi MindSpore2.3 is newly released, and the large-scale model development training is simpler, more stable and more efficient.
Shengsi MindSpore actively explores cutting-edge technologies and supports the original and efficient training of large models. Through eight parallel technologies, such as original multi-copy and multi-pipeline interleaving, the cluster linearity reaches 90% (less than 60% in the industry), and the computing power utilization rate reaches 55% (less than 40% in the industry) through the whole map optimization and sinking execution. Aiming at the common problems of high cluster failure rate and long recovery time, the failure recovery can be completed in 20 minutes by compiling snapshots and deterministic CKPT technology.
In the deployment of large-scale model, Shengsi realizes the unification of scripts, distributed strategies and runtime by upgrading the architecture of training and pushing, and the reasoning deployment of Baichuan2-13B takes only one day. In the large model reasoning, the reasoning throughput is improved by more than 2 times through LLMServing; Upgrade the model compression tool "Golden Hoop 2.0" to achieve 100 billion model compression to ten times.
In order to lower the development threshold, Shengsi continuously upgraded the MindSporeTransFormers large model suite, and provided MindSporeOne production suite. The whole process can be developed out of the box in one week.
Innovate the AI+ scientific computing paradigm and incubate the basic big model in the scientific field. MindSpore has been exploring AI bio-computing for many years, and has built an AI bio-computing suite with top scientific research institutions and partners, including more than 20 SOTA models such as protein structure prediction and generation, to accelerate innovation in related fields.
Facing the long-term planning, we will improve MindSpore’s deep-rooted technology and continue to evolve to help the big model industry land.
Four actions empower academics and ecology
Ding Cheng, chairman of Shengsi MindSpore Open Source Community, announced four actions of Shengsi Empowering Academics and Ecology at the conference. The Academic Thesis Fund 2.0, which cooperates with China Artificial Intelligence Society and Pengcheng Laboratory, will join hands with more than 50 global AI scholars in the next three years to explore new academic peaks and achieve double improvement in quantity and quality; Shengsi development board application innovation action, based on the orange development board, provides systematic cases, tutorials and support to help developers get started quickly, get started quickly and quickly create personalized applications; Shengsi accelerates the incubation of native big models, and supports more partners to move from Shengsi migration and adaptation to native development through incentives, special technical support and joint market promotion; Shengsi open source community internship activities, through code practice, let developers really practice and grow.
Awarded the outstanding contribution mentor and developer of Shengsi MindSpore.
As Gitee- the largest open source code hosting platform in China, open source China selected MindSpore as the "Best Open Source Contribution Project of Giee China". According to Giee Index 2.0, MindSpore performed well in various indicators and became the whole category index NO.1 under the Giee-AI field classification. Ma Yue, chairman of Open Source China, said: The ecological prosperity cannot be separated from the contribution of every project, and MindSpore is a treasure cherished by Gitee.
In March, 2020, Huawei launched the MindSporeAI framework. After opening the source, it received positive response from domestic and foreign developers, with tens of millions of visits, more than 6.87 million downloads and installations, ranking first in the code cloud open source project, with more than 5,500 service enterprises and 360 cooperative universities. Since 2023, more than 1,200 top conference papers have been published based on Shengsi framework, ranking first in China and second in the world among all AI frameworks, and it has become the most innovative and dynamic AI open source community in China.
This conference selected 10 "Outstanding Contribution Mentors of Shengsi MindSpore" and 15 "Outstanding Contribution Developers of Shengsi MindSpore" for their outstanding contributions to the community.
关于作者