Shulou(Shulou.com)12/24 Report--

On December 16, at the "2023 Open Atomic developers Conference" sponsored by the Open Atomic Open Source Foundation, Ant Group led the open source graph database TuGraph, time series database CeresDB, privacy computing framework SecretFlow, front-end framework OpenSumi, and data domain large model open source framework DB-GPT was selected into the "2023 Fast-growing Open Source Project".

▲ figure: TuGraph, CeresDB, hidden language SecretFlow, OpenSumi, DB-GPT were selected as "2023 fast-growing open source projects". It is understood that Ant takes open source key basic software as its core strategy. The five selected projects show that the open source investment of ants in the key basic software field has been recognized, and the cross-border open source project of large model has been paid attention to, which also brings inspiration for the industry to explore the application of large model.

Take CeresDB as an example, this is a distributed database for storing and managing time series data, which provides high-performance read and write, high compression ratio and low-cost storage, visual query and other features for time series data, and is suitable for scenarios such as IoT of the Internet of things, operation and maintenance monitoring, financial analysis, and so on. Ren Chunshao, director of the CeresDB open source project, said, "We hope that open source will help developers solve the horizontal expansion of time series data storage and the pain points of high availability, and help the development efficiency of massive data scenarios." At present, the kernel HoraeDB of the project has been donated to the Apache Open Source Foundation.

AI and the wave of large models have brought opportunities and challenges to scientific and technological innovation. For open source, it also contains new vitality. For example, in the field of basic software development, the combination of large models and databases has produced a new spark.

In June this year, Ant Group launched the database domain large model framework DB-GPT. Combined with large model and database, DB-GPT can systematically build enterprise knowledge base, generative report analysis system (GBI), daily data processing and report forms and other applications. Chen Faqiang, director of the DB-GPT open source project, said, "based on large models and databases, enterprises and developers can build their own applications with less code. We hope that DB-GPT will build the infrastructure in the field of large models, making it easier and more convenient to build large model applications around databases." It is understood that starting from a personal project, DB-GPT has grown into an open source community of nearly Wanxing in half a year, and has been recognized by the industry and developers.

The conference also discussed some pain points and difficulties of hard core technology. The outbreak of AI and large models brings arithmetic anxiety and puts forward more stringent requirements for the efficiency of the system. Around the application development of the large model, Zhang Ke, head of AI infrastructure of Ant Group, shared Ant's AI engineering practice. He said that through the intelligence of the system, ants have solved the problem of efficiency optimization of the system, comprehensively covering the efficiency improvement in many fields, such as training reasoning engine framework, online service, GPU cluster scheduling and virtualization, engineering intelligence and so on. At present, the distributed training system service DLRover, video memory and transmission optimization library GLake in this system have been opened.

Zhang Ke also said that ants have built heterogeneous clusters with more than 10,000 calories, and the arithmetic efficiency (HFU) of training above kcal is more than 60%. The effective training time accounts for more than 95%. Under the same model effect, the throughput performance of RLHF training is 3.59 times higher than that of the industry solution, and the reasoning performance is about 2 times higher than that of the industry solution, which is at the advanced level of the industry. " At present, this set of technical capabilities have been applied to the ant lark model, which will bring more green and efficient computing power, "stresses Zhang Ke.

▲ Picture: ant Group Zhang Ke shared AI Engineering practice at this meeting, Ant Group also hosted a sub-forum on "Frontier Theory on Information and data Security". Domestic and foreign experts discussed large model privacy protection, data security, quantum cryptography and so on. The open source "Copper Lock" project, led by Ant Group, has set up a project management committee, which will focus on cryptography research, industrial application promotion and ecological co-construction of commercial passwords.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.