【摘要】 面向6G“内生智能”愿景，无线大模型的边缘部署面临着纯数据驱动模型物理一致性差、泛化能力弱，以及端侧高昂算力与通信开销的双重挑战。基于此，提出一种双向赋能的智能通算融合无线大模型架构。在大模型赋能通信维度，提出物理启发的空口设计，在模型底层嵌入电磁机理与常微分方程以隐式逼近空间信道梯度，突破跨域泛化瓶颈；同时，构建基于可微分投影的跨层联合优化框架，高效求解物理层-MAC层混合整数非线性规划问题。在通信服务大模型维度，设计了端边协同分层架构，依托混合专家机制实现全局知识与端侧算力的按需解耦；进一步融合微批次通算流水线并行、硬件零拷贝调度与空中计算技术，将参数同步轮次降维至常数级，可较大限度地消除通信等待气泡。所提架构在理论和工程层面实现了物理先验、数据驱动、异构算力与空口传输的协同，为未来6G泛在算力网络和内生智能通信系统的标准演进提供了理论支撑与宏观蓝图。更多还原

【Abstract】 Driven by the 6G vision of endogenous intelligence, deploying large AI models(LAMs) in wireless networks is highly promising yet challenging, where purely data-driven models suffer from poor physical consistency, weak generalization and expensive computational and communication overheads on resource-constrained edge devices. To address these issues, a mutually empowering wireless LAM(WLAM) architecture is proposed to intelligently integrate communication and computing. To empower communications via LAMs, a physics-inspired air-interface design is introduced. By structurally embedding electromagnetic priors and ordinary differential equations at the model’s lower layers to implicitly approximate spatial channel gradients, this design overcomes cross-domain generalization limits. Furthermore, a cross-layer joint optimization framework based on differentiable projection is developed to efficiently solve the mixed-integer nonlinear programming problems spanning the physical and MAC layers. Conversely, in terms of communication facilitating LAM deployment, a device-edge collaborative layered architecture is designed to utilize the mixture-of-experts mechanism for on-demand decoupling of global knowledge and device-side computing power. Furthermore, by integrating micro-batch pipeline parallelism, hardware zero-copy scheduling, and over-the-air computation, parameter synchronization rounds are reduced to a constant order, virtually eliminating communication pipeline bubbles. Ultimately, the proposed architecture synthesizes physical priors, data-driven learning, heterogeneous computing and air-interface transmission at both theoretical and engineering levels, providing theoretical support and a macroscopic blueprint for the standards evolution of future 6G ubiquitous computing power networks and endogenous intelligent communication systems.更多还原