Mingxing Zhang

Ruoyu Qin, Weiran He, Yaoyu Wang et al.

Prefill-decode (PD) disaggregation has become the standard architecture for large-scale LLM serving, but in practice its deployment boundary is still determined by KVCache transfer. In conventional dense-attention models, prefill generates huge KVCache traffics that keep prefill and decode tightly coupled within a single high-bandwidth network domain, limiting heterogeneous deployment and resource elasticity. Recent hybrid-attention architectures substantially reduce KVCache size, making cross-cluster KVCache transport increasingly plausible. However, smaller KVCache alone does not make heterogeneous cross-datacenter PD serving practical: real workloads remain bursty, request lengths are highly skewed, prefix caches are unevenly distributed, and inter-cluster bandwidth fluctuates. A naive design that fully externalizes prefill can therefore still suffer from congestion, unstable queueing, and poor utilization. We present Prefill-as-a-Service (PrfaaS), a cross-datacenter serving architecture that selectively offloads long-context prefill to standalone, compute-dense prefill clusters and transfers the resulting KVCache over commodity Ethernet to local PD clusters for decode. Rather than treating reduced KVCache as sufficient, PrfaaS combines model-side KV efficiency with system-side selective offloading, bandwidth-aware scheduling, and cache-aware request placement. This design removes the requirement that heterogeneous accelerators share the same low-latency RDMA fabric, enabling independent scaling of prefill and decode capacity across loosely coupled clusters. In a case study using an internal 1T-parameter hybrid model, a PrfaaS-augmented heterogeneous deployment achieves 54% and 32% higher serving throughput than homogeneous PD and naive heterogeneous baselines, respectively, while consuming only modest cross-datacenter bandwidth.

Enzhe Lu, Zhejun Jiang, Jingyuan Liu et al. · pku

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

Tiancheng Hu, Jin Qin, Zheng Wang et al.

Disaggregation maps parts of an AI workload to different types of GPUs, offering a path to utilize modern heterogeneous GPU clusters. However, existing solutions operate at a coarse granularity and are tightly coupled to specific model architectures, leaving much room for performance improvement. This paper presents Tessera, the first kernel disaggregation system to improve performance and cost efficiency on heterogeneous GPUs for large model inference. Our key insight is that kernels within a single application exhibit diverse resource demands, making them the most suitable granularity for aligning computation with hardware capabilities. Tessera integrates offline analysis with online adaptation by extracting precise inter-kernel dependencies from PTX to ensure correctness, overlapping communication with computation through a pipelined execution model, and employing workload-aware scheduling with lightweight runtime adaptation. Extensive evaluations across five heterogeneous GPUs and four model architectures, scaling up to 16 GPUs, show that Tessera improves serving throughput and cost efficiency by up to 2.3x and 1.6x, respectively, compared to existing disaggregation methods, while generalizing to model architectures where prior approaches do not apply. Surprisingly, a heterogeneous GPU pair under Tessera can even exceed the throughput of two homogeneous high-end GPUs at a lower cost.

Mingxing Zhang, Nicola Rossberg, Simone Innocente et al.

In recent years, machine learning models have been increasingly applied to spectroscopic datasets for chemical and biomedical analysis. For their successful adoption, particularly in clinical and safety-critical settings, professionals and researchers must be able to understand and trust the reasoning behind model predictions. However, the inherently high dimensionality and strong collinearity of spectroscopy data pose a fundamental challenge to model explainability. These properties not only complicate model training but also undermine the stability and consistency of explanations, leading to fluctuations in feature importance across repeated training runs. Feature extraction techniques have been used to reduce the input dimensionality; these new features hinder the connection between the prediction and the original signal. This study proposes SHAPCA, an explainable machine learning pipeline that combines Principal Component Analysis (for dimensionality reduction) and Shapely Additive exPlanations (for post hoc explanation) to provide explanations in the original input space, which a practitioner can interpret and link back to the biological components. The proposed framework enables analysis from both global and local perspectives, revealing the spectral bands that drive overall model behaviour as well as the instance-specific features that influence individual predictions. Numerical analysis demonstrated the interpretability of the results and greater consistency across different runs.

Feng Ren, Ruoyu Qin, Teng Ma et al.

Modern GPU clusters are built upon a complex hierarchy of heterogeneous interconnects, ranging from multi-rail RDMA to proprietary fabrics such as Multi-Node NVLink and Ascend UB. Orchestrating these diverse links effectively remains a critical challenge in disaggregated LLM serving. Operating Mooncake TE on thousands of GPUs exposed a critical limitation shared by existing frameworks: imperative, statically bound path selection. This rigidity forces engines to rely on state-blind striping that ignores congestion signals, creating communication silos, wasting multi-rail bandwidth due to head-of-line blocking, and leading to operational fragility where routine faults require manual intervention. We present TENT, a data-movement engine that decouples transfer intent from physical execution. Instead of locking workloads to fixed backends, TENT unifies heterogeneous interconnects into a single dynamic resource pool. Applications simply declare transfer intents, while TENT dynamically decomposes elephant flows into fine-grained slices and "sprays" them across links based on instantaneous link quality. This telemetry-driven orchestration eliminates head-of-line blocking and enables transparent, sub-50 ms self-healing by rerouting slices around failures without application logic. TENT serves as the production data plane for LLM inference and RL pipelines at multiple industrial sites. Our evaluation on H800 HGX clusters shows that TENT outperforms state-of-the-art baselines, including Mooncake TE, NIXL, and UCCL. In LLM inference with SGLang HiCache, TENT achieves up to 1.36x higher throughput and 26% lower P90 TTFT than Mooncake TE. In RL pipelines, TENT accelerates parameter updates in Moonshot Checkpoint Engine by 20-26%.

Jialiang Huang, Teng Ma, Zheng Liu et al.

Serverless computing is renowned for its computation elasticity, yet its full potential is often constrained by the requirement for functions to operate within local and dedicated background environments, resulting in limited memory elasticity. To address this limitation, this paper introduces TrEnv-X, a co-designed integration of the serverless platform with the operating system and CXL/RDMA-based remote memory pools. TrEnv-X's core innovations are repurposable sandboxes, which can be shared across different functions to decrease the associated creation overhead, and OS-level memory templates, which enable rapid state restoration from CXL/RDMA-based remote memory pools. To further demonstrate TrEnv-X's versatility, we generalize its design from traditional containers for microVM-based agent workloads and introduce new optimizations, including browser sharing and a page cache bypassing mechanism. Our evaluation shows that TrEnv-X achieves up to 7x reduction in P99 latency and 48% memory savings for container-based functions. When applied to LLM agents, it reduces the P99 latency by up to 58% and memory usage by 61% compared to state-of-the-art systems like E2B.

Huipeng Ma, Luan Zhang, Dandan Song et al.

In multi-hop reasoning, multi-round retrieval-augmented generation (RAG) methods typically rely on LLM-generated content as the retrieval query. However, these approaches are inherently vulnerable to knowledge overshadowing - a phenomenon where critical information is overshadowed during generation. As a result, the LLM-generated content may be incomplete or inaccurate, leading to irrelevant retrieval and causing error accumulation during the iteration process. To address this challenge, we propose ActiShade, which detects and activates overshadowed knowledge to guide large language models (LLMs) in multi-hop reasoning. Specifically, ActiShade iteratively detects the overshadowed keyphrase in the given query, retrieves documents relevant to both the query and the overshadowed keyphrase, and generates a new query based on the retrieved documents to guide the next-round iteration. By supplementing the overshadowed knowledge during the formulation of next-round queries while minimizing the introduction of irrelevant noise, ActiShade reduces the error accumulation caused by knowledge overshadowing. Extensive experiments show that ActiShade outperforms existing methods across multiple datasets and LLMs.

Xun Sun, Shaoyuan Chen, Pingchuan Ma et al.

Mixture-of-Experts (MoE) serving relies on wide expert parallelism (EP) to aggregate the memory capacity and bandwidth of many GPUs within one inference instance. This efficiency comes with a systems cost: every decoding step depends on token dispatch and combination across all active EP ranks, so even one rank failure can disrupt the entire service. Existing EP stacks handle such failures poorly because they treat membership as a fixed configuration established at initialization. The same rank set determines communicator state, expert placement, and the routing metadata baked into CUDA execution graphs, leaving the system with no way to shrink around a failure while keeping the instance valid. This paper argues that partial-failure tolerance should instead be formulated as a live EP validity problem. We present EEP, a communication and runtime substrate that represents membership as explicit, mutable runtime state. EEP repairs the specific state invalidated by a fault: it restores peer reachability without rebuilding the communication substrate, repairs lost expert coverage through a bandwidth-aware hierarchy, and reintegrates repaired ranks without forcing healthy ranks to recapture their CUDA graphs. We implement EEP in an EP serving stack integrated with SGLang and evaluate it under steady-state serving, failure recovery, and rank reintegration. The results show that explicit mutable membership preserves the steady-state fast path, staying within 4.4% of a fixed-membership DeepEP baseline under static serving, while turning a local rank fault from whole-instance downtime into two bounded interruptions. On a single-rank failure workload, EEP incurs an 11s recovery pause and an 8s reintegration pause, and restores throughput to within 95% of the pre-fault level within 52s, whereas a fixed-membership full-restart baseline remains unavailable until 348s.

Shaoyuan Chen, Wencong Xiao, Yutong Lin et al.

Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but also introduce significant challenges in real-world serving due to inefficient use of the expensive, computation-optimized accelerators. Although disaggregated serving architectures have been proposed to split different phases of LLM inference, the efficiency of decoding phase is still low. This is caused by the varying resource demands of different operators in the transformer-based LLMs. Specifically, the attention operator is memory-intensive, exhibiting a memory access pattern that clashes with the strengths of modern accelerators, especially for long context requests. To enhance the efficiency of LLM decoding, we introduce model-attention disaggregation. This approach leverages a collection of cheap, memory-optimized devices for the attention operator while still utilizing high-end accelerators for other parts of the model. This heterogeneous setup ensures that each component is tailored to its specific workload, maximizing overall performance and cost efficiency. Our comprehensive analysis and experiments confirm the viability of splitting the attention computation over multiple devices. Also, the communication bandwidth required between heterogeneous devices proves to be manageable with prevalent networking technologies. To further validate our theory, we develop and deploy Lamina, an LLM inference system that incorporates model-attention disaggregation in a distributed heterogeneous cluster. Experimental results indicate that Lamina can provide 16.1 ~ 90.1% higher estimated throughput than existing solutions with similar costs.

Yaoqi Chen, Jinkai Zhang, Baotong Lu et al. · microsoft-research

The growing context lengths of large language models (LLMs) pose significant challenges for efficient inference, primarily due to GPU memory and bandwidth constraints. We present RetroInfer, a novel system that reconceptualizes the key-value (KV) cache as a vector storage system which exploits the inherent attention sparsity to accelerate long-context LLM inference. At its core is the wave index, an Attention-aWare VEctor index that enables efficient and accurate retrieval of critical tokens through techniques such as tripartite attention approximation, accuracy-bounded attention estimation, and segmented clustering. Complementing this is the wave buffer, which coordinates KV cache placement and overlaps computation and data transfer across GPU and CPU to sustain high throughput. Unlike prior sparsity-based methods that struggle with token selection and hardware coordination, RetroInfer delivers robust performance without compromising model accuracy. Experiments on long-context benchmarks show up to 4.5X speedup over full attention within GPU memory limits and up to 10.5X over sparse attention baselines when KV cache is extended to CPU memory, all while preserving full-attention-level accuracy.

Jiahao Wang, Weiyu Xie, Mingxing Zhang et al.

Retrieval-Augmented Generation enhances Large Language Models by integrating external knowledge, which reduces hallucinations but increases prompt length. This increase leads to higher computational costs and longer Time to First Token (TTFT). To mitigate this issue, existing solutions aim to reuse the preprocessed KV cache of each retrieved chunk to accelerate RAG. However, the lack of cross-chunk contextual information leads to a significant drop in generation quality, leaving the potential benefits of KV cache reuse largely unfulfilled. The challenge lies in how to reuse the precomputed KV cache of chunks while preserving generation quality. We propose FusionRAG, a novel inference framework that optimizes both the preprocessing and reprocessing stages of RAG. In the offline preprocessing stage, we embed information from other related text chunks into each chunk, while in the online reprocessing stage, we recompute the KV cache for tokens that the model focuses on. As a result, we achieve a better trade-off between generation quality and efficiency. According to our experiments, FusionRAG significantly improves generation quality at the same recomputation ratio compared to previous state-of-the-art solutions. By recomputing fewer than 15% of the tokens, FusionRAG achieves up to 70% higher normalized F1 scores than baselines and reduces TTFT by 2.66x-9.39x compared to Full Attention.

Ruoyu Qin, Weiran He, Weixiao Huang et al.

Reinforcement Learning (RL) has become critical for advancing modern Large Language Models (LLMs), yet existing synchronous RL systems face severe performance bottlenecks. The rollout phase, which dominates end-to-end iteration time, suffers from substantial long-tail latency and poor resource utilization due to inherent workload imbalance. We present Seer, a novel online context learning system that addresses these challenges by exploiting previously overlooked similarities in output lengths and generation patterns among requests sharing the same prompt. Seer introduces three key techniques: divided rollout for dynamic load balancing, context-aware scheduling, and adaptive grouped speculative decoding. Together, these mechanisms substantially reduce long-tail latency and improve resource efficiency during rollout. Evaluations on production-grade RL workloads demonstrate that Seer improves end-to-end rollout throughput by 74% to 97% and reduces long-tail latency by 75% to 93% compared to state-of-the-art synchronous RL systems, significantly accelerating RL training iterations.

Ruoyu Qin, Zheming Li, Weiran He et al.

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. It features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache. The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs). Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake's innovative architecture enables Kimi to handle 75% more requests.

Mingxing Zhang, Zhengchun Zhou, Lanping Li et al.

Sequences play an important role in many engineering applications and systems. Searching sequences with desired properties has long been an interesting but also challenging research topic. This article proposes a novel method, called HpGAN, to search desired sequences algorithmically using generative adversarial networks (GAN). HpGAN is based on the idea of zero-sum game to train a generative model, which can generate sequences with characteristics similar to the training sequences. In HpGAN, we design the Hopfield network as an encoder to avoid the limitations of GAN in generating discrete data. Compared with traditional sequence construction by algebraic tools, HpGAN is particularly suitable for intractable problems with complex objectives which prevent mathematical analysis. We demonstrate the search capabilities of HpGAN in two applications: 1) HpGAN successfully found many different mutually orthogonal complementary code sets (MOCCS) and optimal odd-length Z-complementary pairs (OB-ZCPs) which are not part of the training set. In the literature, both MOCSSs and OB-ZCPs have found wide applications in wireless communications. 2) HpGAN found new sequences which achieve four-times increase of signal-to-interference ratio--benchmarked against the well-known Legendre sequence--of a mismatched filter (MMF) estimator in pulse compression radar systems. These sequences outperform those found by AlphaSeq.