DeepSeek联合北大、清华研究团队发布DualPath推理系统，打破大模型存储带宽瓶颈
北京大学、清华大学与DeepSeek-AI的联合研究团队发布了一项针对大语言模型推理架构优化的最新研究成果。该团队成功研发了名为DualPath的全新推理系统，专门解决智能体工作负载下KV-Cache存储I/O带宽受限的技术难题。https://cdn3.ldstatic.com/original/4X/1/a/b/1ab57fcd4a05f491bbf03e45eae5d29996b00da1.png
https://cdn3.ldstatic.com/original/4X/9/8/e/98e2c9dffd6367462dd4539a5e8d89f9d3cceb79.png
https://cdn3.ldstatic.com/original/4X/4/4/9/449880f0999763479effd19620cc00c89b54810f.png
依然27B小模型
https://cdn3.ldstatic.com/original/4X/7/7/3/7737f9c766957e34da6871902e1e7a9d2aca40f3.png
arXiv.org (https://arxiv.org/abs/2602.21548)
https://cdn3.ldstatic.com/optimized/4X/c/d/4/cd49b65780faf86c14ed9761c9c522acfb73adde_2_500x500.png
DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference (https://arxiv.org/abs/2602.21548)
The performance of multi-turn, agentic LLM inference is increasingly dominated by KV-Cache storage I/O rather than computation. In prevalent disaggregated architectures, loading the massive KV-Cache from external storage creates a fundamental...
页:
[1]