Beijing, China – Peking University (PKU) and Huawei jointly announced the release of a fully open-source inference solution for the DeepSeek large language model, marking a significant step towards democratizing access to advanced AI technologies. The announcement, made on March 9, 2025, by PKU’s High-Performance Computing Platform, signals a commitment to open-source collaboration and technological self-reliance in China’s burgeoning AI landscape.
The solution leverages PKU’s in-house developed SCOW computing platform system and the HeSi scheduling system. It seamlessly integrates DeepSeek, openEuler, MindSpore, and other open-source components like vLLM/RAY, enabling highly efficient DeepSeek inference on Huawei’s Ascend AI processors. This integration supports unified training and inference deployment across large-scale computing clusters.
This open-source solution represents a paradigm shift in AI development, stated a representative from PKU’s High-Performance Computing Platform. By embracing open collaboration, we are breaking down technological barriers and fostering a global community of innovators.
Key Features and Performance:
The fully open-source nature of the solution allows developers to easily access the source code and customize it to meet specific needs. Importantly, the performance of this open-source solution reportedly rivals that of closed-source alternatives. Specifically, performance data released by PKU indicates the open-source solution achieves a system output throughput of 1198, compared to 1288 for a closed-source alternative, using a DeepSeek-R1-w8a8 model on a 2*Atlas 800I A2 hardware configuration, with an input length of 4096 and output length of 1024, supporting 128 concurrent users.
Deployment on Weiming Excellence No. 1 Cluster:
The DeepSeek inference solution has been successfully deployed on the Weiming Excellence No. 1 cluster, a cutting-edge intelligent computing platform developed and maintained by PKU’s Computing Center. This cluster provides robust computing power to the PKU Kunpeng Ascend Science and Education Innovation Excellence Center.
Launched on November 18, 2024, the Weiming Excellence No. 1 cluster is the first domestic intelligent computing platform based on independently developed basic software from a Chinese university. The initial phase integrates 20 Ascend AI servers and 10 Kunpeng general-purpose servers, providing an AI computing power of 30.64 PFlops (half-precision).
A Fully Independent Technology Stack:
The cluster boasts a fully domestically produced and independent technology stack, covering the entire ecosystem from chip instruction sets to scheduling systems. The HeSi and SCOW systems, developed in-house, are used for critical software components such as the scheduling system and portal, offering strong scalability and adaptability. This allows for broad support of both domestic and international mainstream processors, including Kunpeng and Ascend, as well as open-source frameworks and models like vLLM, MindSpore, and DeepSeek.
Open Source as the Future of AI:
The release of the DeepSeek inference solution underscores the growing importance of open source in driving technological innovation. By fostering collaboration and knowledge sharing, open source accelerates technology adoption, promotes cross-disciplinary partnerships, and creates a virtuous cycle of development.
The solution incorporates deep optimizations at the openEuler operating system level, including heterogeneous scheduling with load-aware MoE (Mixture of Experts) for fine-grained task allocation. Furthermore, it employs heterogeneous fusion for efficient memory management, reducing system memory fragmentation. The Bisheng compiler is also utilized to further optimize performance.
As global demand for technological transparency and customization increases, the open-source model is poised to become the preferred approach for enterprises, research institutions, and developers alike. This initiative by Peking University and Huawei is a testament to the power of open collaboration in shaping the future of artificial intelligence.
References:
- 北京大学高性能计算校级公共平台. (2025, March 9). 北京大学联合华为发布全栈开源DeepSeek推理方案. [Peking University and Huawei Release Fully Open-Source DeepSeek Inference Solution]. Retrieved from [Insert Official Source Link Here if Available].
Views: 0