arXiv subjects

Yifan Qiao

Publications and source records attributed to Yifan Qiao.

At least 19 recordsLinked to original sources

arXiv2026

Lizard: Bandwidth-Adaptive Real-Time Video Analytics through Content-Aware Packet Discarding at Last-Mile Edge Routers

The timeliness and accuracy of edge-based video analytics can be hindered by drastic reductions in available bandwidth (ABW) at last-mile edge routers, causing prolonged queuing delays. This work proposes Lizard, a system that leverages video-content-aware packet discarding to mitigate the negative effects of drastic ABW degradation that may frequently occur at a last-mile edge router by judiciously discarding packets that contain frame blocks less important to the analytics at the destination. To achieve this, we first devise a frame-block-aware RTP header extension to effectively decouple packet dependencies to encode frame blocks. Second, Lizard uses a priority-based feedback mechanism that dynamically evaluates packet priorities based on relative accuracy impacts. Third, we develop an adaptive phase-transition-based packet discarding strategy at the router to discard packets that represent unimportant blocks. Our evaluation of Lizard shows improvements over existing methods are substantial: 53.2% reduction in latency and 27.1% increase in analysis accuracy.

cs.NI↗

arXiv2026

A Deep Neural Network based Level-1 Trigger for the Magnetic Monopole Search in the SCEP Experiment

The Search for Cosmic Exotic Particles (SCEP) experiment proposes a multilayer magnetic monopole (MM) detection array with a target exposure of O(10^4) m^2 year. Continuously recording the raw waveforms at this scale would generate approximately O(10) EB of data per year, making full waveform storage impractical. A Level-1 trigger is therefore required to reduce the stored-data volume while retaining rare MM signals. MM waveforms vary with particle velocity and trajectory, so a single filter kernel cannot match all possible signals. A mismatch between the kernel and the signal waveform can reduce the detection probability. This work therefore develops a compact template-bank construction method for MM detection. Each incoming waveform is filtered with all kernels in the bank, and the best-matched kernel is identified from their responses. A multilayer perceptron (MLP) trigger is further developed to combine all template responses and improve signal detection. With a 31-sample window and a stored-data fraction of 10^(-3), the MLP increases the mean net acceptance P_net from 6.72% for the conventional template-bank trigger to 8.79%, a relative improvement of 30.8%. The network is then implemented on a Xilinx AC701 board equipped with a field-programmable gate array (FPGA) to process a 20-bit input waveform sampled at 1 MHz using 200 MHz trigger logic. At the target stored-data fraction, the fixed-point implementation reduces P_net by only 3/(3.2 x 10^6) relative to the floating-point reference. On-board tests reproduce the bit-accurate reference trigger outputs and sustain continuous 1 MHz processing without dropped input samples, verifying the functional correctness and real-time operation of the FPGA-based Level-1 trigger.

physics.ins-det↗

arXiv2026

Demystifying Numerical Instability in LLM Inference: Achieving Reproducible Inference for Mission-Critical Tasks with HEAL

As Large Language Models (LLMs) deploy into mission-critical domains (e.g., finance, medicine, and law), output reproducibility has become a strict system requirement. While practitioners use greedy decoding to eliminate algorithmic stochasticity, empirical deployments with 16-bit precisions still exhibit catastrophic output divergence across heterogeneous GPUs. Through SASS-level profiling, we reveal that this inconsistency is fundamentally driven by truncation errors introduced during downcasting at kernel boundaries. However, achieving reproducibility via a global FP32 pipeline incurs prohibitive system penalties: bypassing 16-bit hardware accelerators hurts compute efficiency, while upcasting the KV cache doubles memory overhead. To bridge this gap, we propose Hybrid Error ALleviation (HEAL), a targeted intervention that approximates FP32 precision while resolving hardware constraints through two targeted mechanisms. First, recognizing that floating-point formats underutilize their bit-width for Q, K, V tensors, HEAL applies INT16 quantization that preserves numerical stability without expanding the KV cache footprint. Second, HEAL synthesizes high-precision matrix multiplications via an algebraic error compensation strategy, executing entirely on high-throughput 16-bit Tensor Cores. To evaluate our approach practically, we introduce MCR-Bench, a benchmark targeting reproducibility in mission-critical tasks. HEAL achieves the same level of reproducibility on downstream tasks as the FP32 baseline while reducing the performance overhead by up to 7.1x.

cs.LG↗

arXiv2026

Prism: Cost-Efficient Multi-LLM Serving via GPU Memory Ballooning

Inference providers must maintain availability for many LLMs, including low-volume but essential models, making resource efficiency increasingly important as token prices fall. Analysis of production traces reveals a dynamic bursty-group pattern in which sets of models become active together and shift over time; existing space- and time-sharing approaches lack principled mechanisms to adapt to this variability, forcing trade-offs between SLO adherence and efficiency. We observe that elastic memory allocation can unify spatial and temporal sharing. Based on this insight, we have developed Prism, a memory-centric LLM co-serving framework that applies memory ballooning to reclaim memory across models and support both forms of sharing under a single scheme. Prism's balloon driver, referred to as kvcached, has been open-sourced at https://github.com/ovg-project/kvcached, and deployed in production environments across 10K+ GPUs.

cs.DC↗

arXiv2026

Idleness is Relative: Exploiting Tool-Call Idle Windows for Offloading in Agentic Systems with MORI

Modern LLM serving systems increasingly host agentic workloads, whose sessions issue tens of model invocations interleaved with tool calls, accumulating KV cache that can be reused across steps. As requests' total KV cache size easily exceeds GPU HBM capacity, researchers offload them to CPU DRAM. However, tool-call durations span orders of magnitude, and the cost of transferring KV cache between tiers makes it impractical to re-place entries on every call. We observe that agentic programs exhibit a two-phase structure: busy phases of rapid short tool calls and idle phases dominated by long-running calls. Current eviction policies such as LRU fail to capture this property. A binary busy/idle label also falls short because the ratio of busy to idle programs may not match the hardware's GPU-to-CPU capacity ratio. When it does not, one tier sits underutilized while the other is oversubscribed, wasting memory or forcing unnecessary evictions. We present MORI, an agent serving system that solves the above problem. Our key insight is that idleness is a continuous, relative spectrum. MORI ranks all active programs by idleness, assigns the busiest to GPU HBM and the most idle to CPU DRAM, dynamically shifts the partition boundary to match hardware capacity, and enforces admission control at each memory tier. Evaluated on real coding agent workloads collected from Claude Code across four GPU and model pairs, MORI delivers 20--71% higher throughput and 18--43% lower TTFT than the best baseline with offloading.

cs.OS↗

arXiv2026

SERL: Self-Examining Reinforcement Learning on Open-Domain

Reinforcement Learning (RL) has been shown to improve the capabilities of large language models (LLMs). However, applying RL to open-domain tasks faces two key challenges: (1) the inherent subjectivity of these tasks prevents the verifiable rewards as required by Reinforcement Learning with Verifiable Rewards (RLVR); (2) Reinforcement Learning from Human Feedback (RLHF) relies on external reward mechanisms. To overcome these limitations, we propose Self-Examining Reinforcement Learning (SERL), a novel self-improving framework where the LLM serves as both Actor and Judge. SERL introduces two synergistic reward mechanisms without any external signals. On the one hand, to improve the Actor's capability, we derive rewards from Copeland-style pairwise comparison judgments across a group of generated responses. On the other hand, a self-consistency reward that encourages coherent judgments is proposed to improve the Judge's reliability. This process refines the Judge's capability, which in turn provides a more robust reward for Actor. Experiments show that our method outperforms existing self-improvement training methods. SERL improves the LC win rate of Qwen3-8B on AlpacaEval 2 from 52.37% to 59.90%. To the best of our knowledge, our method achieves state-of-the-art performance among self-improving approaches. Furthermore, it achieves a performance comparable to significantly larger models like Qwen3-32B, demonstrating superior effectiveness and robustness on open-domain tasks.

cs.LG↗

arXiv2026

HALO: A Fine-Grained Resource Sharing Quantum Operating System

As quantum computing enters the cloud era, thousands of users must share access to a small number of quantum processors. Users need to wait minutes to days to start their jobs, which only takes a few seconds for execution. Current quantum cloud platforms employ a fair-share scheduler, as there is no way to multiplex a quantum computer among multiple programs at the same time, leaving many qubits idle and significantly under-utilizing the hardware. This imbalance between high user demand and scarce quantum resources has become a key barrier to scalable and cost-effective quantum computing. We present HALO, the first quantum operating system design that supports fine-grained resource-sharing. HALO introduces two complementary mechanisms. First, a hardware-aware qubit-sharing algorithm that places shared helper qubits on regions of the quantum computer that minimize routing overhead and avoid cross-talk noise between different users' processes. Second, a shot-adaptive scheduler that allocates execution windows according to each job's sampling requirements, improving throughput and reducing latency. Together, these mechanisms transform the way quantum hardware is scheduled and achieve more fine-grained parallelism. We evaluate HALO on the IBM Torino quantum computer on helper qubit intense benchmarks. Compared to state-of-the-art systems such as HyperQ, HALO improves overall hardware utilization by up to 2.44x, increasing throughput by 4.44x, and maintains fidelity loss within 33%, demonstrating the practicality of resource-sharing in quantum computing.

cs.OS↗

arXiv2026

SkyNomad: On Using Multi-Region Spot Instances to Minimize AI Batch Job Cost

AI batch jobs such as model training, inference pipelines, and data analytics require substantial GPU resources and often need to finish before a deadline. Spot instances offer 3-10x lower cost than on-demand instances, but their unpredictable availability makes meeting deadlines difficult. Existing systems either rely solely on spot instances and risk deadline violations, or operate in simplified single-region settings. These approaches overlook substantial spatial and temporal heterogeneity in spot availability, lifetimes, and prices. We show that exploiting such heterogeneity to access more spot capacity is the key to reduce the job execution cost. We present SkyNomad, a multi-region scheduling system that maximizes spot usage and minimizes cost while guaranteeing deadlines. SkyNomad uses lightweight probing to estimate availability, predicts spot lifetimes, accounts for migration cost, and unifies regional characteristics and deadline pressure into a monetary cost model that guides scheduling decisions. Our evaluation shows that SkyNomad achieves 1.25-3.96x cost savings in real cloud deployments and performs within 10% cost differences of an optimal policy in simulation, while consistently meeting deadlines.

cs.DC↗

arXiv2025

Lazarus: Resilient and Elastic Training of Mixture-of-Experts Models

Sparsely-activated Mixture-of-Experts (MoE) architecture has increasingly been adopted to further scale large language models (LLMs). However, frequent failures still pose significant challenges as training scales. The cost of even a single failure is significant, as all GPUs need to idle wait until the failure is resolved, potentially losing considerable training progress as training has to restart from checkpoints. This problem is exacerbated by the growing use of spot instances on public clouds for model training, which despite offering substantial cost savings, introduce frequent preemptions-essentially failures that regularly occur throughout the training process. Existing solutions for efficient fault-tolerant training either lack elasticity or rely on building resiliency into pipeline parallelism, which cannot be applied to MoE models due to the expert parallelism strategy adopted by the MoE architecture. We present Lazarus, a system for resilient and elastic training of MoE models. Lazarus adaptively allocates expert replicas to address the inherent imbalance in expert workload and speeds up training, while a provably optimal expert placement algorithm is developed to maximize the probability of recovery upon failures. Through adaptive expert placement and a flexible token dispatcher, Lazarus can also fully utilize all available nodes after failures, leaving no GPU idle. Our evaluation shows that Lazarus outperforms existing MoE training systems by up to 5.7x under frequent node failures and 3.4x on a real spot instance trace.

cs.DC↗

arXiv2025

Balancing Signal and Variance: Adaptive Offline RL Post-Training for VLA Flow Models

Vision-Language-Action (VLA) models based on flow matching have shown excellent performance in general-purpose robotic manipulation tasks. However, the action accuracy of these models on complex downstream tasks is unsatisfactory. One important reason is that these models rely solely on the post-training paradigm of imitation learning, which makes it difficult to have a deeper understanding of the distribution properties of data quality, which is exactly what Reinforcement Learning (RL) excels at. In this paper, we theoretically propose an offline RL post-training objective for VLA flow models and induce an efficient and feasible offline RL fine-tuning algorithm -- Adaptive Reinforced Flow Matching (ARFM). By introducing an adaptively adjusted scaling factor in the VLA flow model loss, we construct a principled bias-variance trade-off objective function to optimally control the impact of RL signal on flow loss. ARFM adaptively balances RL advantage preservation and flow loss gradient variance control, resulting in a more stable and efficient fine-tuning process. Extensive simulation and real-world experimental results show that ARFM exhibits excellent generalization, robustness, few-shot learning, and continuous learning performance.

cs.RO↗

arXiv2025

ConServe: Fine-Grained GPU Harvesting for LLM Online and Offline Co-Serving

Large language model (LLM) serving demands low latency and high throughput, but high load variability makes it challenging to achieve high GPU utilization. In this paper, we identify a synergetic but overlooked opportunity to co-serve latency-critical online requests alongside latency-tolerant offline tasks such as model benchmarking. While promising, existing serving systems fail to co-serve them efficiently, as their coarse-grained resource management at the request or iteration level cannot harvest millisecond-level GPU idle cycles without introducing interference that violates online latency objectives. ConServe is a new LLM co-serving system that achieves high throughput and strong online latency guarantees by managing resources at finer granularities. ConServe introduces three techniques: (1) a latency-aware token-level scheduler that precisely sizes offline batches and tokens to fit within online latency objectives; (2) sub-iteration, layer-wise preemption that allows offline tasks to yield to online load spikes; and (3) incremental KV cache management that enables preempting and resuming offline requests at near-zero cost. Evaluations with Llama-3.1 and Qwen-2.5 models on real-world workloads show that ConServe delivers an average of 2.2$\times$ higher throughput and reduces online serving tail latency by 2.9$\times$ on average compared to state-of-the-art systems.

cs.DC↗

arXiv2025

Towards Efficient and Practical GPU Multitasking in the Era of LLM

GPU singletasking is becoming increasingly inefficient and unsustainable as hardware capabilities grow and workloads diversify. We are now at an inflection point where GPUs must embrace multitasking, much like CPUs did decades ago, to meet the demands of modern AI workloads. In this work, we highlight the key requirements for GPU multitasking, examine prior efforts, and discuss why they fall short. To advance toward efficient and practical GPU multitasking, we envision a resource management layer, analogous to a CPU operating system, to handle various aspects of GPU resource management and sharing. We outline the challenges and potential solutions, and hope this paper inspires broader community efforts to build the next-generation GPU compute paradigm grounded in multitasking.

cs.OS↗

arXiv2025

PrefillOnly: An Inference Engine for Prefill-only Workloads in Large Language Model Applications

Besides typical generative applications, like ChatGPT, GitHub Copilot, and Cursor, we observe an emerging trend that LLMs are increasingly used in traditional discriminative tasks, such as recommendation, credit verification, and data labeling. The key characteristic of these emerging use cases is that the LLM generates only a single output token, rather than an arbitrarily long sequence of tokens. We call this prefill-only workload. However, since existing LLM engines assume arbitrary output lengths, they fail to leverage the unique properties of prefill-only workloads. In this paper, we present PrefillOnly, the first LLM inference engine that improves the inference throughput and latency by fully embracing the properties of prefill-only workloads. First, since it generates only one token, PrefillOnly only needs to store the KV cache of only the last computed layer, rather than of all layers. This drastically reduces the GPU memory footprint of LLM inference and allows handling long inputs without using solutions that reduces throughput, such as cross-GPU KV cache parallelization. Second, because the output length is fixed, rather than arbitrary, PrefillOnly can precisely determine the job completion time (JCT) of each prefill-only request before it starts. This enables efficient JCT-aware scheduling policies such as shortest remaining job first. PrefillOnly can process upto 4x larger queries per second without inflating average and P99 latency.

cs.DC↗

arXiv2025

LSTM-based Selective Dense Text Retrieval Guided by Sparse Lexical Retrieval

This paper studies fast fusion of dense retrieval and sparse lexical retrieval, and proposes a cluster-based selective dense retrieval method called CluSD guided by sparse lexical retrieval. CluSD takes a lightweight cluster-based approach and exploits the overlap of sparse retrieval results and embedding clusters in a two-stage selection process with an LSTM model to quickly identify relevant clusters while incurring limited extra memory space overhead. CluSD triggers partial dense retrieval and performs cluster-based block disk I/O if needed. This paper evaluates CluSD and compares it with several baselines for searching in-memory and on-disk MS MARCO and BEIR datasets.

cs.IR↗

arXiv2024

DRust: Language-Guided Distributed Shared Memory with Fine Granularity, Full Transparency, and Ultra Efficiency

Despite being a powerful concept, distributed shared memory (DSM) has not been made practical due to the extensive synchronization needed between servers to implement memory coherence. This paper shows a practical DSM implementation based on the insight that the ownership model embedded in programming languages such as Rust automatically constrains the order of read and write, providing opportunities for significantly simplifying the coherence implementation if the ownership semantics can be exposed to and leveraged by the runtime. This paper discusses the design and implementation of DistR, a Rust-based DSM system that outperforms the two state-of-the-art DSM systems GAM and Grappa by up to 2.64x and 29.16x in throughput, and scales much better with the number of servers.

cs.DC↗

arXiv2024

A Tale of Two Paths: Toward a Hybrid Data Plane for Efficient Far-Memory Applications

With rapid advances in network hardware, far memory has gained a great deal of traction due to its ability to break the memory capacity wall. Existing far memory systems fall into one of two data paths: one that uses the kernel's paging system to transparently access far memory at the page granularity, and a second that bypasses the kernel, fetching data at the object granularity. While it is generally believed that object fetching outperforms paging due to its fine-grained access, it requires significantly more compute resources to run object-level LRU and eviction. We built Atlas, a hybrid data plane enabled by a runtime-kernel co-design that simultaneously enables accesses via these two data paths to provide high efficiency for real-world applications. Atlas uses always-on profiling to continuously measure page locality. For workloads already with good locality, paging is used to fetch data, whereas for those without, object fetching is employed. Object fetching moves objects that are accessed close in time to contiguous local space, dynamically improving locality and making the execution increasingly amenable to paging, which is much more resource-efficient. Our evaluation shows that Atlas improves the throughput (e.g., by 1.5x and 3.2x) and reduces the tail latency (e.g., by one and two orders of magnitude) when using remote memory, compared with AIFM and Fastswap, the state-of-the-art techniques respectively in the two categories.

cs.DC↗

arXiv2024

Weighted KL-Divergence for Document Ranking Model Refinement

Transformer-based retrieval and reranking models for text document search are often refined through knowledge distillation together with contrastive learning. A tight distribution matching between the teacher and student models can be hard as over-calibration may degrade training effectiveness when a teacher does not perform well. This paper contrastively reweights KL divergence terms to prioritize the alignment between a student and a teacher model for proper separation of positive and negative documents. This paper analyzes and evaluates the proposed loss function on the MS MARCO and BEIR datasets to demonstrate its effectiveness in improving the relevance of tested student models.

cs.IR↗

arXiv2024

Approximate Cluster-Based Sparse Document Retrieval with Segmented Maximum Term Weights

This paper revisits cluster-based retrieval that partitions the inverted index into multiple groups and skips the index partially at cluster and document levels during online inference using a learned sparse representation. It proposes an approximate search scheme with two parameters to control the rank-safeness competitiveness of pruning with segmented maximum term weights within each cluster. Cluster-level maximum weight segmentation allows an improvement in the rank score bound estimation and threshold-based pruning to be approximately adaptive to bound estimation tightness, resulting in better relevance and efficiency. The experiments with MS MARCO passage ranking and BEIR datasets demonstrate the usefulness of the proposed scheme with a comparison to the baselines. This paper presents the design of this approximate retrieval scheme with rank-safeness analysis, compares clustering and segmentation options, and reports evaluation results.

cs.IR↗