Operator fusion, as a key performance optimization technique in the deployment of AI models, significantly improves execution efficiency and has been widely adopted in modern AI compilers. However, for cascaded reduction operations involving multiple loops with inter-loop data dependencies, such as the safe softmax followed by GEMM within attention mechanisms, existing compilers lack effective automated fusion and kernel generation capabilities. Although some works have addressed specific instances through hand-crafted fusion strategies, their solutions are limited in generality and difficult to extend to other similar structures. Given the prevalence of such computational patterns in deep learning models, there remains significant untapped potential in achieving general and automated fusion optimization. In this paper, we present a formal theoretical methodology for analyzing cascaded reductions which can fuse them into a single loop and introduce an incremental computation form. Based on this methodology, we design Red uction Fuser (RedFuser), a framework that automatically identifies supported cascaded reduction patterns and generates optimized fused kernels. Experiments show that RedFuser successfully fuses diverse workloads, achieving up to 2× to 5× speedup over state-of-the-art AI compilers and matching the performance of highly optimized hand-written kernels.

The rapid adoption of Large Language Models (LLMs) in cloud environments has intensified the demand for high-performance AI training and inference, where Remote Direct Memory Access (RDMA) plays a critical role. However, existing RDMA virtualization solutions, such as Single-Root Input/Output Virtualization (SR-IOV), face significant limitations in scalability, performance, and stability. These issues include lengthy container initialization times, hardware resource constraints, and inefficient traffic steering. To address these challenges, we propose Stellar, a new generation RDMA network for cloud AI. Stellar introduces three key innovations: Para-Virtualized Direct Memory Access (PVDMA) for on-demand memory pinning, extended Memory Translation Table (eMTT) for optimized GPU Direct RDMA (GDR) performance, and RDMA Packet Spray for efficient multi-path utilization. Deployed in our large-scale AI clusters, Stellar spins up virtual devices in seconds, reduces container initialization time by 15 times, and improves LLM training speed by up to 14%. Our evaluations demonstrate that Stellar significantly outperforms existing solutions, offering a scalable, stable, and high-performance RDMA network for cloud AI.

The rapid adoption of AI and convenience offered by cloud services have resulted in the growing demands for GPUs in the cloud. Generally, GPUs are physically attached to host servers as PCIe devices. However, the fixed assembly combination of host servers and GPUs is extremely inefficient in resource utilization, upgrade, and maintenance. Due to these issues, the GPU disaggregation technique has been proposed to decouple GPUs from host servers. It aggregates GPUs into a pool and allocates GPU node(s) according to user demands. However, existing GPU disaggregation systems have flaws in software-hardware compatibility, disaggregation scope, and capacity. In this article, we present a new implementation of datacenter-scale GPU disaggregation, named DxPU. DxPU efficiently solves the above problems and can flexibly allocate as many GPU node(s) as users demand. To understand the performance overhead incurred by DxPU, we build up a performance model for AI specific workloads. With the guidance of modeling results, we develop a prototype system, which has been deployed into the datacenter of a leading cloud provider for a test run. We also conduct detailed experiments to evaluate the performance overhead caused by our system. The results show that the overhead of DxPU is less than 10%, compared with native GPU servers, in most of user scenarios.