DeepSeek has forged a collaborative partnership with Peking University and Tsinghua University, culminating in the publication of a paper on the ArXiv platform. This paper introduces an innovative large-model reasoning framework, named DualPath, tailored specifically for intelligent agent systems. The primary objective of this framework is to tackle the Input/Output (I/O) bottleneck that arises due to Key-Value (KV)-Cache loading in scenarios involving lengthy contexts.

DualPath achieves this by introducing a secondary pathway, structured as 'storage → decoding engine → prefilling'. This novel approach capitalizes on the underutilized storage network bandwidth of the decoding engine, in conjunction with Remote Direct Memory Access (RDMA) high-speed networks. The result is a seamless global pooling and load balancing of cluster storage bandwidth.

In rigorous testing conducted with a production-grade model boasting 660 billion parameters, DualPath demonstrated remarkable performance enhancements. Specifically, it elevated offline reasoning throughput by 1.87 times and online service throughput by an average of 1.96 times. Notably, these improvements were achieved without compromising the speed of token-by-token generation, while simultaneously optimizing the latency associated with the first token.