Rong Gu

Zhibin Wang, Ziyu Zhong, Nuo Shen et al.

Speculative decoding and dynamic sparse attention are two complementary approaches for accelerating long-context LLM inference: the former amortizes target-model execution across multiple verifier queries, while the latter reduces each query's KV-cache working set. Directly combining them, however, exposes a structural mismatch: speculative verification relies on cross-query commonality, whereas dynamic sparse attention assigns query-specific sparse layouts. This mismatch limits KV-block reuse, amplifies NSA's branch-wise overheads, and makes verification strategy selection input- and regime-dependent. We present SpecSA, a sparse speculative-verification framework that turns dynamic sparse attention into a verification-oriented workload. SpecSA combines overlap-aware grouped-query execution, refresh/reuse-based NSA kernel fusion, and profile-guided prompt-adaptive orchestration to improve cross-query reuse, reduce selected-index and branch-fusion overheads, and select effective draft-verification strategies under user-specified precision classes. Experiments on NVIDIA H100 GPUs show that SpecSA achieves up to 3.49x end-to-end throughput over autoregressive NSA decoding and up to 6.86x kernel speedups for sparse speculative verification.

Yuhang Zhou, Zhibin Wang, Peng Jiang et al.

Training large language models faces frequent interruptions due to various faults, demanding robust fault-tolerance. Existing backup-free methods, such as redundant computation, dynamic parallelism, and data rerouting, each incur performance penalties, whether from ongoing overhead, lengthy reconfigurations, or post-recovery inefficiencies. We propose Chameleon, an adaptive fault-tolerant system that intelligently selects optimal recovery strategies when a failure occurs. Chameleon achieves this through a unified performance model, expedient execution plan search, accurate performance estimation, and efficient communication optimizations. Experiments on a 32-card cluster show that Chameleon maintains a performance gap of within 11.00% between post-recovery and failure-free training, while preserving model convergence and efficient memory usage. Compared to state-of-the-art methods, Chameleon achieves up to 1.229x and 1.355x higher average throughput than Oobleck and Recycle, respectively.

Chengying Huan, Ziheng Meng, Yongchao Liu et al.

Graph Chain-of-Thought (Graph-CoT) enables large language models (LLMs) to perform step-by-step reasoning over graph-structured knowledge, but existing pipelines suffer from low accuracy, excessive token usage, high latency, and low throughput due to single-agent monolithic prompts, repeated context re-encoding, and inefficient serving execution. We present GLM, the first multi-agent Graph-CoT system co-designed with an optimized LLM serving architecture. GLM decomposes reasoning into specialized agents for classification, reasoning, action generation, and graph retrieval, enabling branching and selective context sharing to reduce prompt length and reasoning iterations while preserving reasoning quality, thereby improving accuracy and reducing overall token consumption. To scale inference, we introduce a Graph-CoT-aware LLM inference mechanism with graph-specific KV-cache management, priority-based eviction, and pipelined execution to improve serving efficiency. Experiments demonstrate that GLM improves answer accuracy by up to 38%, reduces token cost by up to 95.7%, lowers inference latency by 90.3%, and achieves up to 15.1x higher throughput compared to state-of-the-art Graph-CoT baselines, enabling efficient adoption for complex real-world reasoning at scale.

Zhibin Wang, Shipeng Li, Yuhang Zhou et al.

User experience is a critical factor Large Language Model (LLM) serving systems must consider, where service level objectives (SLOs) considering the experience of individual requests and system level metrics (SLMs) considering the overall system performance are two key performance measures. However, we observe two notable issues in existing metrics: 1) manually delaying the delivery of some tokens can improve SLOs, and 2) actively abandoning requests that do not meet SLOs can improve SLMs, both of which are counterintuitive. In this paper, we revisit SLOs and SLMs in LLM serving, and propose a new SLO that aligns with user experience. Based on the SLO, we propose a comprehensive metric framework called smooth goodput, which integrates SLOs and SLMs to reflect the nature of user experience in LLM serving. Through this unified framework, we reassess the performance of different LLM serving systems under multiple workloads. Evaluation results show that our metric framework provides a more comprehensive view of token delivery and request processing, and effectively captures the optimal point of user experience and system performance with different serving strategies.

Zhibin Wang, Rui Ning, Chao Fang et al.

Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely FlashForge. FlashForge delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that FlashForge achieves an average 1.9x speedup and 120.9x memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and 3.8x end-to-end time per output token compared to the vLLM.

Zhibin Wang, Shipeng Li, Xue Li et al.

Large language models have been widely deployed in various applications, encompassing both interactive online tasks and batched offline tasks. Given the burstiness and latency sensitivity of online tasks, over-provisioning resources is common practice. This allows for the integration of latency-insensitive offline tasks during periods of low online load, enhancing resource utilization. However, strategically serving online and offline tasks through a preemption mechanism fails to fully leverage the flexibility of offline tasks and suffers from KV cache recomputation and irregular workloads. In this paper, we introduce Echo, a collaborative online-offline task serving system, including a scheduler, a KV cache manager, and estimation toolkits. The scheduler and KV cache manager work tightly to maximize the throughput of offline tasks, while the estimator further predicts execution time to ensure online task SLOs. The scheduler leverages the batch information of last iteration to reduce the search space for finding the optimal schedule. The KV cache manager sets the priority of the KV cache based on the type of tasks and the opportunity of prefix sharing to reduce the recomputation. Finally, the estimation toolkits predict the execution time, future memory consumption, and the throughput of offline tasks to guide the scheduler, KV cache manager, and the system deployer. Evaluation based on real-world workloads demonstrates that Echo can increase offline task throughput by up to $3.3\times$, while satisfying online task SLOs.

Zhibin Wang, Zetao Hong, Xue Li et al.

Large Language Model (LLM) inference has emerged as a fundamental paradigm. In real-world scenarios, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose ARES, an adaptive decoding rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with : A dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 74.77% and achieving up to 2.24 times higher goodput.

Hongxu Pan, Shuxian Hu, Mo Zhou et al.

Researchers have proposed various methods of incorporating more structured information into the design of Graph Neural Networks (GNNs) to enhance their expressiveness. However, these methods are either computationally expensive or lacking in provable expressiveness. In this paper, we observe that the chords increase the complexity of the graph structure while contributing little useful information in many cases. In contrast, chordless structures are more efficient and effective for representing the graph. Therefore, when leveraging the information of cycles, we choose to omit the chords. Accordingly, we propose a Chordless Structure-based Graph Neural Network (CSGNN) and prove that its expressiveness is strictly more powerful than the k-hop GNN (KPGNN) with polynomial complexity. Experimental results on real-world datasets demonstrate that CSGNN outperforms existing GNNs across various graph tasks while incurring lower computational costs and achieving better performance than the GNNs of 3-WL expressiveness.

Jianzhou Chen, Jinyang Sun, Xiumei Wang et al.

Heart failure is one of the leading causes of death worldwide, with millons of deaths each year, according to data from the World Health Organization (WHO) and other public health agencies. While significant progress has been made in the field of heart failure, leading to improved survival rates and improvement of ejection fraction, there remains substantial unmet needs, due to the complexity and multifactorial characteristics. Therefore, we propose a composable strategy framework for assessment and treatment optimization in heart failure. This framework simulates the doctor-patient consultation process and leverages multi-modal algorithms to analyze a range of data, including video, physical examination, text results as well as medical history. By integrating these various data sources, our framework offers a more holistic evaluation and optimized treatment plan for patients. Our results demonstrate that this multi-modal approach outperforms single-modal artificial intelligence (AI) algorithms in terms of accuracy in heart failure (HF) prognosis prediction. Through this method, we can further evaluate the impact of various pathological indicators on HF prognosis,providing a more comprehensive evaluation.

Rong Gu

Most reinforcement learning (RL) platforms use high-level programming languages, such as OpenAI Gymnasium using Python. These frameworks provide various API and benchmarks for testing RL algorithms in different domains, such as autonomous driving (AD) and robotics. These platforms often emphasise the design of RL algorithms and the training performance but neglect the correctness of models and reward functions, which can be crucial for the successful application of RL. This paper proposes using formal methods to model AD systems and demonstrates how model checking (MC) can be used in RL for AD. Most studies combining MC and RL focus on safety, such as safety shields. However, this paper shows different facets where MC can strengthen RL. First, an MC-based model pre-analysis can reveal bugs with respect to sensor accuracy and learning step size. This step serves as a preparation of RL, which saves time if bugs exist and deepens users' understanding of the target system. Second, reward automata can benefit the design of reward functions and greatly improve learning performance especially when the learning objectives are multiple. All these findings are supported by experiments.