arXiv subjects

Xinming Wei

Publications and source records attributed to Xinming Wei.

9 recordsLinked to original sources

arXiv2026

ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels

LLMs scale Mixture-of-Experts (MoE) parameters for superior intelligence, but massive weights and dynamic computation impede efficient serving. Existing instance-level prefill-decode disaggregation isolates the phases on separate full-model replicas. As MoE weights grow, each instance may span tens to hundreds of GPUs, making resource allocation increasingly coarse. Configured prefill-to-decode ratios thus often mismatch demand, overprovisioning one phase while overloading the other. Prefill-decode colocation avoids this duplication, but existing Green Context solutions partition each GPU by phase and fix phase resources during a kernel. They cannot track resource changes across operations or layerwise variation in routed expert load, causing head-of-line blocking or idle reserved resources. Partitioning every GPU also leaves each phase with fewer local resources, forces wider parallelism and more communication, and lets prefill and decode traffic interfere on the shared network. We present ExpertPlex, which shares massive MoE experts across phases while disaggregating lightweight attention modules. Expert sharing eliminates over 95% of duplicate model weights and multiplexes dynamically sparse computation, while attention disaggregation reduces attention communication cost. ExpertPlex further uses (1) adaptive persistent kernels to schedule dynamic expert computation at tile granularity for efficient, isolated execution; (2) attention-initiated MoE communication to avoid network interference and enable cross-phase communication-computation overlap; and (3) a tile-to-cluster model to optimize these mechanisms for maximum goodput. Experiments serving MiniMax-M2.7 and GLM-5.1-FP8 show that ExpertPlex improves goodput by up to 2.01$\times$ over instance-level prefill-decode disaggregation and 1.66$\times$ over prefill-decode colocation.

cs.DC↗

arXiv2026

UltraEP: Unleash MoE Training and Inference on Rack-Scale Nodes with Near-Optimal Load Balancing

Large-scale expert parallelism (EP) is becoming pivotal for training and serving frontier MoE models, but it also amplifies device-level expert load imbalance into compute stragglers, token all-to-all bottlenecks, and activation-memory spikes. Existing balancers redistribute experts periodically based on historical load, which becomes unreliable for production deployments with non-stationary load patterns. We present UltraEP, the first exact-load, real-time balancer for large-EP MoE training and serving prefill on rack-scale nodes (RSNs). Leveraging the extended scale-up connectivity among dozens of GPUs within RSNs, UltraEP rebalances every microbatch and layer on critical paths, which requires nontrivial co-design of plan solving and expert replication communication to minimize exposed overhead. To this end, UltraEP eagerly reacts to post-gating load with an efficient quota-driven planner, and executes the resulting irregular expert-state transfers with RSN-native persistent tile streaming and relay-based fan-out mitigation. We evaluate UltraEP in a multi-RSN deployment of up to 256 GPUs, using cutting-edge MoE models from 106B to 671B parameters. Averaged across training and serving, UltraEP achieves 94.3% of the force-balanced ideal throughput, delivering 1.49$\times$ improvement over no-balancing, while reducing the final inter-rank imbalance from 1.30$-$4.01 to 1.01$-$1.04.

cs.DC↗

arXiv2026

ReLibra: Routing-Replay-Guided Load Balancing for MoE Training in Reinforcement Learning

Load imbalance is a long-standing challenge in Mixture-of-Experts (MoE) training and is exacerbated in reinforcement learning (RL) for LLMs, where hot experts can shift frequently across micro-batches. Existing MoE training systems rely on historical loads to predict future expert demand, making them less effective under sharp fluctuations. We propose ReLibra, an MoE RL training system that exploits a unique opportunity in RL's rollout-training workflow, routing replay, to enable fine-grained load balancing at micro-batch granularity. Because rollout and training process the same tokens with the same MoE parameters, the token-to-expert routing decisions are known before training starts. Leveraging this information, ReLibra places two MoE load-balancing mechanisms at inter- and intra-batch timescales, matching their communication patterns to hierarchical network bandwidths. At the inter-batch timescale, ReLibra performs expert reordering to redistribute experts for batch-level cross-node balancing; at the intra-batch timescale, it dynamically performs expert replication within a node to absorb micro-batch-level load fluctuations. Experiments on diverse MoE LLMs and RL workloads show that ReLibra improves training throughput by up to 1.6$\times$ over Megatron-LM and by up to 1.2$\times$ over EPLB, even when EPLB is given oracle loads. Moreover, ReLibra remains within 6%-10% of the throughput of an idealized balanced baseline.

cs.LG↗

arXiv2026

Heddle: A Distributed Orchestration System for Agentic RL Rollout

Agentic Reinforcement Learning (RL) enables LLMs to solve complex tasks by alternating between a data-collection rollout phase and a policy training phase. During rollout, the agent generates trajectories, i.e., multi-step interactions between LLMs and external tools. Yet, frequent tool calls induce long-tailed trajectory generation that bottlenecks rollouts. This stems from step-centric designs that ignore trajectory context, triggering three system problems for long-tail trajectory generation: queueing delays, interference overhead, and inflated per-token time. We propose Heddle, a trajectory-centric system to optimize the when, where, and how of agentic rollout execution. Heddle integrates three core mechanisms: trajectory-level scheduling using runtime prediction and progressive priority to minimize cumulative queueing; trajectory-aware placement via presorted dynamic programming and opportunistic migration during idle tool call intervals to minimize interference; and trajectory-adaptive resource manager that dynamically tunes model parallelism to accelerate the per-token time of long-tail trajectories while maintaining high throughput for short trajectories. Evaluations across diverse agentic RL workloads demonstrate that Heddle effectively neutralizes the long-tail bottleneck, achieving up to 2.5$\times$ higher end-to-end rollout throughput compared to state-of-the-art baselines.

cs.LG↗

arXiv2025

ACE-GNN: Adaptive GNN Co-Inference with System-Aware Scheduling in Dynamic Edge Environments

The device-edge co-inference paradigm effectively bridges the gap between the high resource demands of Graph Neural Networks (GNNs) and limited device resources, making it a promising solution for advancing edge GNN applications. Existing research enhances GNN co-inference by leveraging offline model splitting and pipeline parallelism (PP), which enables more efficient computation and resource utilization during inference. However, the performance of these static deployment methods is significantly affected by environmental dynamics such as network fluctuations and multi-device access, which remain unaddressed. We present ACE-GNN, the first Adaptive GNN Co-inference framework tailored for dynamic Edge environments, to boost system performance and stability. ACE-GNN achieves performance awareness for complex multi-device access edge systems via system-level abstraction and two novel prediction methods, enabling rapid runtime scheme optimization. Moreover, we introduce a data parallelism (DP) mechanism in the runtime optimization space, enabling adaptive scheduling between PP and DP to leverage their distinct advantages and maintain stable system performance. Also, an efficient batch inference strategy and specialized communication middleware are implemented to further improve performance. Extensive experiments across diverse applications and edge settings demonstrate that ACE-GNN achieves a speedup of up to 12.7x and an energy savings of 82.3% compared to GCoDE, as well as 11.7 better energy efficiency than Fograph.

cs.DC↗

arXiv2025

Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving

AI agent frameworks operate in isolation, forcing agents to rediscover solutions and repeat mistakes across different systems. Despite valuable problem-solving experiences accumulated by frameworks like smolagents, OpenHands, and OWL, this knowledge remains trapped within individual systems, preventing the emergence of collective intelligence. Current memory systems focus on individual agents or framework-specific demonstrations, failing to enable cross-architecture knowledge transfer. We introduce AGENT KB, a universal memory infrastructure enabling seamless experience sharing across heterogeneous agent frameworks without retraining. AGENT KB aggregates trajectories into a structured knowledge base and serves lightweight APIs. At inference time, hybrid retrieval operates through two stages: planning seeds agents with cross-domain workflows, while feedback applies targeted diagnostic fixes. A disagreement gate ensures retrieved knowledge enhances rather than disrupts reasoning, addressing knowledge interference in cross-framework transfer. We validate AGENT KB across major frameworks on GAIA, Humanity's Last Exam, GPQA, and SWE-bench. Results show substantial improvements across diverse model families: compared to baseline pass@1, smolagents with AGENT KB achieve up to 18.7pp gains at pass@3 (55.2% -> 73.9%), while OpenHands improves 4.0pp on SWE-bench pass@1 (24.3% -> 28.3%). Similar improvements are observed across all base model families. Ablations confirm that hybrid retrieval and feedback stages are essential, with automatically generated experiences matching manual curation. This establishes the foundation for collective agent intelligence through shared memory infrastructures.

cs.CL↗

arXiv2025

Agent.xpu: Efficient Scheduling of Agentic LLM Workloads on Heterogeneous SoC

Personal LLM agents increasingly combine foreground reactive interactions with background proactive monitoring, forming long-lived, stateful LLM flows that interleave prefill and token-by-token decode. While modern heterogeneous SoCs integrate CPUs, iGPUs, and NPUs to support on-device intelligence, existing LLM engines assume static, single-shot inference and lack mechanisms for flow-level concurrency, prioritization, and efficient accelerator coordination. As a result, commodity SoCs remain poorly matched to the dynamic, mixed-criticality execution patterns of personal agents. This paper presents Agent$.$xpu, the first LLM engine that orchestrates concurrent reactive and proactive LLM flows on commodity SoCs. Extensive profiling uncovers unique SoC characteristics of operator-accelerator affinity, asymmetric DDR contention, and stage-divergent batching behaviors distinct from cloud-serving assumptions. Agent$.$xpu introduces three key techniques: a heterogeneous execution graph (HEG) capturing NPU/iGPU affinity and elastic operator binding; flow-aware NPU-iGPU coordination with stage elasticity, decoupling prefill and decode to reduce bandwidth contention and enforce priorities; and fine-grained preemption with slack-aware piggybacking to guarantee reactive responsiveness without starving proactive work. Across realistic personal-agent workloads, Agent$.$xpu delivers 1.2-4.9$\times$ proactive throughput and reduces reactive latency by at least 91%, compared with both industrial iGPU-only serving engine and NPU-iGPU static inference with optimal tensor-partitioning schemes. Agent$.$xpu also minimizes energy consumption and graphics interference via controlled iGPU usage.

cs.DC↗

arXiv2024

Autoformalize Mathematical Statements by Symbolic Equivalence and Semantic Consistency

Autoformalization, the task of automatically translating natural language descriptions into a formal language, poses a significant challenge across various domains, especially in mathematics. Recent advancements in large language models (LLMs) have unveiled their promising capabilities to formalize even competition-level math problems. However, we observe a considerable discrepancy between pass@1 and pass@k accuracies in LLM-generated formalizations. To address this gap, we introduce a novel framework that scores and selects the best result from k autoformalization candidates based on two complementary self-consistency methods: symbolic equivalence and semantic consistency. Elaborately, symbolic equivalence identifies the logical homogeneity among autoformalization candidates using automated theorem provers, and semantic consistency evaluates the preservation of the original meaning by informalizing the candidates and computing the similarity between the embeddings of the original and informalized texts. Our extensive experiments on the MATH and miniF2F datasets demonstrate that our approach significantly enhances autoformalization accuracy, achieving up to 0.22-1.35x relative improvements across various LLMs and baseline methods.

cs.CL↗

arXiv2024

The Dawn of AI-Native EDA: Opportunities and Challenges of Large Circuit Models

Within the Electronic Design Automation (EDA) domain, AI-driven solutions have emerged as formidable tools, yet they typically augment rather than redefine existing methodologies. These solutions often repurpose deep learning models from other domains, such as vision, text, and graph analytics, applying them to circuit design without tailoring to the unique complexities of electronic circuits. Such an AI4EDA approach falls short of achieving a holistic design synthesis and understanding, overlooking the intricate interplay of electrical, logical, and physical facets of circuit data. This paper argues for a paradigm shift from AI4EDA towards AI-native EDA, integrating AI at the core of the design process. Pivotal to this vision is the development of a multimodal circuit representation learning technique, poised to provide a comprehensive understanding by harmonizing and extracting insights from varied data sources, such as functional specifications, RTL designs, circuit netlists, and physical layouts. We champion the creation of large circuit models (LCMs) that are inherently multimodal, crafted to decode and express the rich semantics and structures of circuit data, thus fostering more resilient, efficient, and inventive design methodologies. Embracing this AI-native philosophy, we foresee a trajectory that transcends the current innovation plateau in EDA, igniting a profound shift-left in electronic design methodology. The envisioned advancements herald not just an evolution of existing EDA tools but a revolution, giving rise to novel instruments of design tools that promise to radically enhance design productivity and inaugurate a new epoch where the optimization of circuit performance, power, and area (PPA) is achieved not incrementally, but through leaps that redefine the benchmarks of electronic systems' capabilities.

cs.AR↗