Fan Lai

Haizhong Zheng, Rui Liu, Fan Lai et al.

One-shot coreset selection aims to select a representative subset of the training data, given a pruning rate, that can later be used to train future models while retaining high accuracy. State-of-the-art coreset selection methods pick the highest importance examples based on an importance metric and are found to perform well at low pruning rates. However, at high pruning rates, they suffer from a catastrophic accuracy drop, performing worse than even random sampling. This paper explores the reasons behind this accuracy drop both theoretically and empirically. We first propose a novel metric to measure the coverage of a dataset on a specific distribution by extending the classical geometric set cover problem to a distribution cover problem. This metric helps explain why coresets selected by SOTA methods at high pruning rates perform poorly compared to random sampling because of worse data coverage. We then propose a novel one-shot coreset selection method, Coverage-centric Coreset Selection (CCS), that jointly considers overall data coverage upon a distribution as well as the importance of each example. We evaluate CCS on five datasets and show that, at high pruning rates (e.g., 90%), it achieves significantly better accuracy than previous SOTA methods (e.g., at least 19.56% higher on CIFAR10) as well as random selection (e.g., 7.04% higher on CIFAR10) and comparable accuracy at low pruning rates. We make our code publicly available at https://github.com/haizhongzheng/Coverage-centric-coreset-selection.

Jiachen Liu, Fan Lai, Yinwei Dai et al.

Federated learning (FL) is an emerging machine learning (ML) paradigm that enables heterogeneous edge devices to collaboratively train ML models without revealing their raw data to a logically centralized server. However, beyond the heterogeneous device capacity, FL participants often exhibit differences in their data distributions, which are not independent and identically distributed (Non-IID). Many existing works present point solutions to address issues like slow convergence, low final accuracy, and bias in FL, all stemming from client heterogeneity. In this paper, we explore an additional layer of complexity to mitigate such heterogeneity by grouping clients with statistically similar data distributions (cohorts). We propose Auxo to gradually identify such cohorts in large-scale, low-availability, and resource-constrained FL populations. Auxo then adaptively determines how to train cohort-specific models in order to achieve better model performance and ensure resource efficiency. Our extensive evaluations show that, by identifying cohorts with smaller heterogeneity and performing efficient cohort-based training, Auxo boosts various existing FL solutions in terms of final accuracy (2.1% - 8.2%), convergence time (up to 2.2x), and model bias (4.8% - 53.8%).

Sanjay Sri Vallabh Singapuram, Fan Lai, Chuheng Hu et al.

The need to train DNN models on end-user devices (e.g., smartphones) is increasing with the need to improve data privacy and reduce communication overheads. Unlike datacenter servers with powerful CPUs and GPUs, modern smartphones consist of a diverse collection of specialized cores following a system-on-a-chip (SoC) architecture that together perform a variety of tasks. We observe that training DNNs on a smartphone SoC without carefully considering its resource constraints can not only lead to suboptimal training performance but significantly affect user experience as well. In this paper, we present Swan, a neural engine to optimize DNN training on smartphone SoCs without hurting user experience. Extensive large-scale evaluations show that Swan can improve performance by 1.2 - 23.3x over the state-of-the-art.

Yifan Yu, Yuqing Jian, Junxiong Wang et al.

Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR models agree with their own generations, while DLMs often do not. We define the introspective acceptance rate, which measures whether a model accepts its previously generated tokens. This reveals why AR training has a structural advantage: causal masking and logit shifting implicitly enforce introspective consistency. Motivated by this observation, we introduce Introspective Diffusion Language Model (I-DLM), a paradigm that retains diffusion-style parallel decoding while inheriting the introspective consistency of AR training. I-DLM uses a novel introspective strided decoding (ISD) algorithm, which enables the model to verify previously generated tokens while advancing new ones in the same forward pass. From a systems standpoint, we build I-DLM inference engine on AR-inherited optimizations and further customize it with a stationary-batch scheduler. To the best of our knowledge, I-DLM is the first DLM to match the quality of its same-scale AR counterpart while outperforming prior DLMs in both model quality and practical serving efficiency across 15 benchmarks. It reaches 69.6 on AIME-24 and 45.7 on LiveCodeBench-v6, exceeding LLaDA-2.1-mini (16B) by more than 26 and 15 points, respectively. Beyond quality, I-DLM is designed for the growing demand of large-concurrency serving, delivering about 3x higher throughput than prior state-of-the-art DLMs.

Sean Nian, Jiahao Fang, Qilong Feng et al.

KV cache restoration has emerged as a dominant bottleneck in serving long-context LLM workloads, including multi-turn conversations, retrieval-augmented generation, and agentic pipelines. Existing approaches treat restoration as a per-request tradeoff between recomputation and I/O transfer, recomputing KV states from scratch or offloading them from external storage (e.g., CPU memory or remote machines). However, existing advances fail to exploit parallelism across tokens, layers, and distributed deployments, and critically ignore resource contention under batched serving. We present CacheFlow, a KV cache restoration framework that rethinks cache restoration as a multi-dimensional parallel execution problem. CacheFlow introduces a unified 3D parallelism abstraction across tokens, layers, and GPUs, enabling fine-grained overlap of recomputation and I/O along the structural dependencies of transformer inference. At the core of CacheFlow is a batch-aware two-pointer scheduler that jointly optimizes compute and I/O allocation across requests by prioritizing operations with the highest marginal reduction in recomputation cost. Our evaluations show that CacheFlow reduces Time-To-First-Token (TTFT) by 10%-62% over existing advances across diverse models, workloads, and hardware.

Yibo Huang, Zhenning Yang, Jiarong Xing et al.

Efficiently serving embedding-based recommendation (EMR) models remains a significant challenge due to their increasingly large memory requirements. Today's practice splits the model across many monolithic servers, where a mix of GPUs, CPUs, and DRAM is provisioned in fixed proportions. This approach leads to suboptimal resource utilization and increased costs. Disaggregating embedding operations from neural network inference is a promising solution but raises novel networking challenges. In this paper, we discuss the design of FlexEMR for optimized EMR disaggregation. FlexEMR proposes two sets of techniques to tackle the networking challenges: Leveraging the temporal and spatial locality of embedding lookups to reduce data movement over the network, and designing an optimized multi-threaded RDMA engine for concurrent lookup subrequests. We outline the design space for each technique and present initial results from our early prototype.

Ting Sun, Penghan Wang, Fan Lai

Large language models (LLMs) have facilitated a wide range of applications with distinct service-level objectives (SLOs), from latency-sensitive online tasks like interactive chatbots to throughput-oriented offline workloads like data synthesis. The existing deployment model, which dedicates machines to each workload, simplifies SLO management but often leads to poor resource utilization. This paper introduces HyGen, an interference-aware LLM serving system that enables efficient co-location of online and offline workloads while preserving SLOs. HyGen incorporates two key innovations: (1) performance control mechanisms, including a latency predictor to estimate batch execution time and an SLO-aware profiler to quantify latency interference, and (2) SLO-aware offline scheduling policies that maximize serving throughput and prevent starvation. Our evaluation on production workloads shows that HyGen achieves up to 3.9-5.8x throughput gains over online and hybrid serving baselines, while ensuring latency SLOs. The code of HyGen is publicly available at https://github.com/UIUC-MLSys/HyGen.

Hoai-Chau Tran, Chi H. Nguyen, Duy M. H. Nguyen et al.

The Segment Anything Model (SAM) achieves strong open-vocabulary segmentation, but its ViT-based image encoders dominate inference latency and memory. Existing activation compression methods, such as token merging, reduce the token length to process, yet introduce non-trivial runtime overhead and encounter catastrophic quality drop under high compression. Other methods applying Sparse Attention focus on attention alone, leaving the MLP fully dense and capping achievable speedup. We propose SparseSAM, a (i) training-free structured sparsification framework that jointly accelerates attention and MLP layers while preserving token identity. SparseSAM introduces (ii) Stripe-Sort Attention, which uses a deterministic Z-order permutation to transform dense attention into static hardware-friendly sparse patterns, eliminating dynamic masking overhead. SparseSAM further introduces a (iii) Residual-Consistency MLP that routes only informative tokens through the MLP while propagating remaining tokens through the residual pathway. Across four segmentation benchmarks, SparseSAM loses only 0.004 mIoU at a 0.4 density and 0.021 mIoU at 0.3, a 2.10x reduction in accuracy loss versus token merging advances, while achieving 2x faster inference and 2.8x memory reduction.

Ting Sun, Penghan Wang, Fan Lai

The rapid rise of large language models (LLMs) in text streaming services has introduced significant cost and Quality of Experience (QoE) challenges in serving millions of daily requests, especially in meeting Time-To-First-Token (TTFT) and Time-Between-Token (TBT) requirements for real-time interactions. Our real-world measurements show that both server-based and on-device deployments struggle to meet diverse QoE demands: server deployments face high costs and last-hop issues (e.g., Internet latency and dynamics), while on-device LLM inference is constrained by resources. We introduce DiSCo, a device-server cooperative scheduler designed to optimize users' QoE by adaptively routing requests and migrating response generation between endpoints while maintaining cost constraints. DiSCo employs cost-aware scheduling, leveraging the predictable speed of on-device LLM inference with the flexible capacity of server-based inference to dispatch requests on the fly, while introducing a token-level migration mechanism to ensure consistent token delivery during migration. Evaluations on real-world workloads -- including commercial services like OpenAI GPT and DeepSeek, and open-source deployments such as LLaMA3 -- show that DiSCo can improve users' QoE by reducing tail TTFT (11-52\%) and mean TTFT (6-78\%) across different model-device configurations, while dramatically reducing serving costs by up to 84\% through its migration mechanism while maintaining comparable QoE levels.

Jiachen Liu, Fan Lai, Ding Ding et al.

In recent years, collaborative learning (CL) has emerged as a promising approach for machine learning (ML) and data science across distributed edge devices. As the deployment of CL jobs increases, they inevitably contend for limited resources. However, efficient resource scheduling in this context is challenging because of the ephemeral nature and resource heterogeneity of devices, coupled with the overlapping resource requirements of diverse CL jobs. Existing resource managers often assign devices to CL jobs randomly for simplicity and scalability, but this approach compromises job efficiency. In this paper, we present Venn, a CL resource manager that efficiently schedules ephemeral, heterogeneous devices among multiple CL jobs to reduce the average job completion time (JCT). Venn formulates the Intersection Resource Scheduling (IRS) problem to identify complex resource contention among multiple CL jobs. It then proposes a contention-aware scheduling heuristic to minimize the average scheduling delay. Furthermore, it proposes a resource-aware device-to-job matching heuristic to optimize response collection time by mitigating stragglers. Our evaluation shows that, compared to the state-of-the-art CL resource managers, Venn improves the average JCT by up to 1.88x. The code is available at https://github.com/SymbioticLab/Venn.

Daiyang Yu, Xinyu Chen, Yihan Zhang et al.

Machine learning (ML) is increasingly applied to optimize system performance in tasks such as resource management and network simulation. Unlike traditional ML tasks (e.g., image classification), networked systems often operate in heterogeneous, long-running, and dynamic environment states, where input conditions (e.g., network loads) and operational objectives can shift over time and across settings. Existing learning-based systems offer little support for adaptation, resulting in costly model training, extensive data collection, degraded system performance, and slow responsiveness. This paper presents EMA, the first model adaptation system supporting learning-based systems to adapt to evolving environments with minimal operational overhead. EMA takes a system-driven, data-centric approach that accommodates diverse system and model designs while addressing two key deployment challenges. First, it reduces expensive model training by introducing state transformers that align the input state of a new environment with previously similar states, allowing models to warm-start adaptation. Second, it addresses the often-overlooked yet costly process of data labeling--collecting ground truth for exploring and training on various system decisions--by prioritizing labeling high-utility data while balancing the tradeoff between training and labeling cost. Evaluations on eight representative learning-based systems show that EMA reduces adaptation costs (e.g., GPU training time) by 14.9-42.4% while improving system performance (e.g., network throughput) by 6.9-31.3%.

Haoyi Song, Ruihan Ji, Naichen Shi et al.

Large language models (LLMs) have transformed natural language processing, but their reliable deployment requires effective uncertainty quantification (UQ). Existing UQ methods are often heuristic and lack a probabilistic interpretation. This paper begins by providing a theoretical justification for the role of perturbations in UQ for LLMs. We then introduce a dual random walk perspective, modeling input-output pairs as two Markov chains with transition probabilities defined by semantic similarity. Building on this, we propose a fully probabilistic framework based on an inverse model, which quantifies uncertainty by evaluating the diversity of the input space conditioned on a given output through systematic perturbations. Within this framework, we define a new uncertainty measure, Inv-Entropy. A key strength of our framework is its flexibility: it supports various definitions of uncertainty measures, embeddings, perturbation strategies, and similarity metrics. We also propose GAAP, a perturbation algorithm based on genetic algorithms, which enhances the diversity of sampled inputs. In addition, we introduce a new evaluation metric, Temperature Sensitivity of Uncertainty (TSU), which directly assesses uncertainty without relying on correctness as a proxy. Extensive experiments demonstrate that Inv-Entropy outperforms existing semantic UQ methods. The code to reproduce the results can be found at https://github.com/UMDataScienceLab/Uncertainty-Quantification-for-LLMs.

Fan Lai, Yinwei Dai, Sanjay S. Singapuram et al.

We present FedScale, a federated learning (FL) benchmarking suite with realistic datasets and a scalable runtime to enable reproducible FL research. FedScale datasets encompass a wide range of critical FL tasks, ranging from image classification and object detection to language modeling and speech recognition. Each dataset comes with a unified evaluation protocol using real-world data splits and evaluation metrics. To reproduce realistic FL behavior, FedScale contains a scalable and extensible runtime. It provides high-level APIs to implement FL algorithms, deploy them at scale across diverse hardware and software backends, and evaluate them at scale, all with minimal developer efforts. We combine the two to perform systematic benchmarking experiments and highlight potential opportunities for heterogeneity-aware co-optimizations in FL. FedScale is open-source and actively maintained by contributors from different institutions at http://fedscale.ai. We welcome feedback and contributions from the community.

Jimmy Shong, Yuhan Ding, Yihan Jiang et al.

Machine learning (ML) systems expose a rapidly expanding configuration space spanning model-parallelism strategies, communication optimizations, and low-level runtime parameters. End-to-end system efficiency is highly sensitive to these choices, yet identifying high-performance configurations is challenging due to heterogeneous feature types (e.g., sparse and dense parameters), conditional dependencies (e.g., valid execution parameters only under specific upstream decisions), and the high search (profiling) cost. Existing approaches either optimize a narrow subset of configuration dimensions or rely on ad-hoc heuristics that fail to generalize as configuration spaces continue to grow. We present AutoScout, a general-purpose systems configurator for ML training, fine-tuning, and inference. It formulates the system configuration as a mixed-discrete/continuous optimization problem with hierarchical dependencies and introduces a hybrid optimization framework that jointly refines sparse structural decisions and dense execution parameters. To reduce profiling cost, AutoScout adaptively prioritizes high-impact configuration features and ensembles simulators with varying fidelity. Across diverse models, hardware platforms, and deployment objectives, AutoScout consistently identifies high-performance configurations, achieving 2.7-3.0$\times$ training speedup over expert-tuned settings.

Jiachen Liu, Jae-Won Chung, Zhiyu Wu et al.

Large language models (LLMs) are now at the core of conversational AI services such as real-time translation and chatbots, which provide live user interaction by incrementally streaming text to the user. However, existing LLM serving systems fail to provide good user experience because their optimization metrics are not always aligned with user experience. In this paper, we first introduce and define the notion of Quality-of-Experience (QoE) for text streaming services by considering each user's end-to-end interaction timeline. Based on this, we propose Andes, a QoE-aware LLM serving system that enhances user experience by ensuring that users receive the first token promptly and subsequent tokens at a smooth, digestible pace, even during surge periods. This is enabled by Andes's preemptive request scheduler that dynamically prioritizes requests at the token granularity based on each request's expected QoE gain and GPU resource usage. Our evaluations demonstrate that, compared to state-of-the-art LLM serving systems, Andes improves the average QoE by up to $4.7\times$ given the same GPU resource, or saves up to 61% GPU resources while maintaining the same high QoE.

Kevin Li, Dibyadeep Saha, Avni Kanodia et al.

As Low-Rank Adaptation (LoRA) becomes the standard approach for efficiently fine-tuning large language models (LLMs), shared clusters increasingly execute many concurrent LoRA training jobs over the same frozen backbone. While recent advances enable batching (co-locating) multiple adapters during serving, efficient training-time co-location of heterogeneous LoRA adapters presents unique challenges. Jobs often differ in adapter rank, batch size, and resource allocation, and naïve batching can introduce synchronization stalls, communication overheads, and per-job slowdowns that are worse than executing independently. We introduce tLoRA, a framework that enables efficient batch training of multiple LoRA jobs. tLoRA fuses adapters that share the same base model into an elastic shared super-model, exploiting existing distributed training frameworks to derive parallelism plans that share resources effectively. At the kernel level, tLoRA employs a fused LoRA kernel that adaptively reconstructs low-rank computation tiles and schedules rank-aware nano-batches to maximize overlap between computation and communication across adapters. At the scheduling layer, tLoRA incorporates an online, residual-capacity-aware scheduler that adaptively groups jobs to maximize collective throughput. Evaluations using real-world cluster traces demonstrate that tLoRA improves training throughput by 1.2--1.8x, job training completion time by 2.3--5.4x, and GPU utilization by 37%.

Haizhong Zheng, Yang Zhou, Brian R. Bartoldson et al.

Reinforcement learning, such as PPO and GRPO, has powered recent breakthroughs in LLM reasoning. Scaling rollout to sample more prompts enables models to selectively use higher-quality data for training, which can stabilize RL training and improve model performance. However, this comes at the cost of significant computational overhead. In this paper, we show that a substantial portion of this overhead can be avoided by skipping uninformative prompts before rollout. Our analysis of reward dynamics reveals a strong temporal consistency in prompt value: prompts that are uninformative in one epoch of training are likely to remain uninformative in future epochs. Based on these insights, we propose GRESO (GRPO with Efficient Selective Rollout), an online, lightweight pre-rollout filtering algorithm that predicts and skips uninformative prompts using reward training dynamics. By evaluating GRESO on a broad range of math reasoning benchmarks and models, such as Qwen2.5-Math-1.5B, DeepSeek-R1-Distill-Qwen-1.5B, and Qwen2.5-Math-7B, we show that GRESO achieves up to 2.4x wall-clock time speedup in rollout and up to 2.0x speedup in total training time without accuracy degradation.

Haizhong Zheng, Xiaoyan Bai, Xueshen Liu et al.

Large Language Models (LLMs) have achieved remarkable success with their billion-level parameters, yet they incur high inference overheads. The emergence of activation sparsity in LLMs provides a natural approach to reduce this cost by involving only parts of the parameters for inference. However, existing methods only focus on utilizing this naturally formed activation sparsity in a post-training setting, overlooking the potential for further amplifying this inherent sparsity. In this paper, we hypothesize that LLMs can learn to be efficient by achieving more structured activation sparsity. To achieve this, we introduce a novel training algorithm, Learn-To-be-Efficient (LTE), designed to train efficiency-aware LLMs to learn to activate fewer neurons and achieve a better trade-off between sparsity and performance. Furthermore, unlike SOTA MoEfication methods, which mainly focus on ReLU-based models, LTE can also be applied to LLMs like LLaMA using non-ReLU activations. Extensive evaluation on language understanding, language generation, and instruction tuning tasks show that LTE consistently outperforms SOTA baselines. Along with our hardware-aware custom kernel implementation, LTE reduces LLaMA2-7B inference latency by 25% at 50% sparsity.

Jiachen Liu, Jiaxin Pei, Jintao Huang et al.

Scientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way. This compilation imposes two structural costs: a Storytelling Tax, where failed experiments, rejected hypotheses, and the branching exploration process are discarded to fit a linear narrative; and an Engineering Tax, where the gap between reviewer-sufficient prose and agent-sufficient specification leaves critical implementation details unwritten. Tolerable for human readers, these costs become critical when AI agents must understand, reproduce, and extend published work. We introduce the Agent-Native Research Artifact (Ara), a protocol that replaces the narrative paper with a machine-executable research package structured around four layers: scientific logic, executable code with full specifications, an exploration graph that preserves the failures compilation discards, and evidence grounding every claim in raw outputs. Three mechanisms support the ecosystem: a Live Research Manager that captures decisions and dead ends during ordinary development; an Ara Compiler that translates legacy PDFs and repos into Aras; and an Ara-native review system that automates objective checks so human reviewers can focus on significance, novelty, and taste. On PaperBench and RE-Bench, Ara raises question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. On RE-Bench's five open-ended extension tasks, preserved failure traces in Ara accelerate progress, but can also constrain a capable agent from stepping outside the prior-run box depending on the agent's capabilities.

Mingyan Gao, Yanzi Li, Banruo Liu et al.

Multi-agent systems (MAS) decompose complex tasks and delegate subtasks to different large language model (LLM) agents and tools. Prior studies have reported the superior accuracy performance of MAS across diverse domains, enabled by long-horizon context tracking and error correction through role-specific agents. However, the design and deployment of MAS incur higher complexity and runtime cost compared to single-agent systems (SAS). Meanwhile, frontier LLMs, such as OpenAI-o3 and Gemini-2.5-Pro, have rapidly advanced in long-context reasoning, memory retention, and tool usage, mitigating many limitations that originally motivated MAS designs. In this paper, we conduct an extensive empirical study comparing MAS and SAS across various popular agentic applications. We find that the benefits of MAS over SAS diminish as LLM capabilities improve, and we propose efficient mechanisms to pinpoint the error-prone agent in MAS. Furthermore, the performance discrepancy between MAS and SAS motivates our design of a hybrid agentic paradigm, request cascading between MAS and SAS, to improve both efficiency and capability. Our design improves accuracy by 1.1-12% while reducing deployment costs by up to 20% across various agentic applications.

Wei Zhang, Zhiyu Wu, Yi Mu et al.

The integration of Large Language Models (LLMs) into diverse applications, ranging from interactive chatbots and cloud AIOps to intelligent agents, has introduced a wide spectrum of Service Level Objectives (SLOs) for responsiveness. These workloads include latency-sensitive requests focused on per-token latency in streaming chat, throughput-intensive requests that require rapid full responses to invoke tools, and collective requests with dynamic dependencies arising from self-reflection or agent-based reasoning. This workload diversity, amplified by unpredictable request information such as response lengths and runtime dependencies, makes existing schedulers inadequate even within their design envelopes. In this paper, we define service gain as the useful service delivered by completing requests. We observe that as SLO directly reflects the actual performance needs of requests, completing a request much faster than its SLO (e.g., deadline) yields limited additional service gain. Based on this insight, we introduce Tempo, the first systematic SLO-aware scheduler designed to maximize service gain across diverse LLM workloads. Tempo allocates just enough serving bandwidth to meet each SLO, maximizing residual capacity for others best-effort workloads. Instead of assuming request information or none at all, it adopts a hybrid scheduling strategy: using quantile-based response upper bounds and dependency-graph matching for conservative initial estimates, prioritizing requests by service gain density, and refining decisions online as generation progresses. Our evaluation across diverse workloads, including chat, reasoning, and agentic pipelines, shows that Tempo improves end-to-end service gain by up to 8.3$\times$ and achieves up to 10.3$\times$ SLO goodput compared to state-of-the-art designs

Yifan Yu, Yu Gan, Nikhil Sarda et al.

Large language models (LLMs) have excelled in various applications, yet serving them at scale is challenging due to their substantial resource demands and high latency. Our real-world studies reveal that over 70% of user requests to LLMs have semantically similar counterparts, suggesting the potential for knowledge transfer among requests. However, naively caching and reusing past responses leads to a big quality drop. In this paper, we introduce IC-Cache, a caching system that enables live LLM capability augmentation to improve serving efficiency: by leveraging historical request-response pairs from larger models as in-context examples, IC-Cache empowers small LLMs to imitate and even exceed the compositional abilities (e.g., reasoning) of their larger counterparts, enabling selective offloading of requests to reduce cost and latency. Achieving this live augmentation at scale introduces intricate trade-offs between response quality, latency, and system throughput. For a new request, IC-Cache efficiently selects similar, high-utility examples to prepend them to the new request's input. At scale, it adaptively routes requests across LLMs of varying capabilities, accounting for response quality and serving loads. IC-Cache employs a cost-aware cache replay mechanism that refines example quality offline to maximize online cache utility and efficiency. Evaluations on millions of realistic requests demonstrate that IC-Cache improves LLM serving throughput by 1.4-5.9x and reduces latency by 28-71% without hurting response quality.

Yuxuan Zhu, Jiachen Liu, Mosharaf Chowdhury et al.

Federated learning (FL) aims to train machine learning (ML) models across potentially millions of edge client devices. Yet, training and customizing models for FL clients is notoriously challenging due to the heterogeneity of client data, device capabilities, and the massive scale of clients, making individualized model exploration prohibitively expensive. State-of-the-art FL solutions personalize a globally trained model or concurrently train multiple models, but they often incur suboptimal model accuracy and huge training costs. In this paper, we introduce FedTrans, a multi-model FL training framework that automatically produces and trains high-accuracy, hardware-compatible models for individual clients at scale. FedTrans begins with a basic global model, identifies accuracy bottlenecks in model architectures during training, and then employs model transformation to derive new models for heterogeneous clients on the fly. It judiciously assigns models to individual clients while performing soft aggregation on multi-model updates to minimize total training costs. Our evaluations using realistic settings show that FedTrans improves individual client model accuracy by 14% - 72% while slashing training costs by 1.6X - 20X over state-of-the-art solutions.

Qinyi Luo, Penghan Wang, Wei Zhang et al.

Huge embedding tables in modern deep learning recommender models (DLRM) require prohibitively large memory during training and inference. This paper proposes FIITED, a system to automatically reduce the memory footprint via FIne-grained In-Training Embedding Dimension pruning. By leveraging the key insight that embedding vectors are not equally important, FIITED adaptively adjusts the dimension of each individual embedding vector during model training, assigning larger dimensions to more important embeddings while adapting to dynamic changes in data. We prioritize embedding dimensions with higher frequencies and gradients as more important. To enable efficient pruning of embeddings and their dimensions during model training, we propose an embedding storage system based on virtually-hashed physically-indexed hash tables. Experiments on two industry models and months of realistic datasets show that FIITED can reduce DLRM embedding size by more than 65% while preserving model quality, outperforming state-of-the-art in-training embedding pruning methods. On public datasets, FIITED can reduce the size of embedding tables by 2.1x to 800x with negligible accuracy drop, while improving model throughput.

Banruo Liu, Wei-Yu Lin, Minghao Fang et al.

The rise of compound AI serving -- integrating multiple operators in a pipeline that may span edge and cloud tiers -- enables end-user applications such as autonomous driving, generative AI-powered meeting companions, and immersive gaming. Achieving high service goodput -- i.e., meeting service level objectives (SLOs) for pipeline latency, accuracy, and costs -- requires effective planning of operator placement, configuration, and resource allocation across infrastructure tiers. However, the diverse SLO requirements, varying edge capabilities, and high query volumes create an enormous planning search space, rendering current solutions fundamentally limited for real-time serving and cost-efficient deployments. This paper presents Circinus, an SLO-aware query planner for large-scale compound AI workloads. Circinus novelly decomposes multi-query planning and multi-dimensional SLO objectives while preserving global decision quality. By exploiting plan similarities within and across queries, it significantly reduces search steps. It further improves per-step efficiency with a precision-aware plan profiler that incrementally profiles and strategically applies early stopping based on imprecise estimates of plan performance. At scale, Circinus selects query-plan combinations to maximize global SLO goodput. Evaluations in real-world settings show that Circinus improves service goodput by 3.2-5.0$\times$, accelerates query planning by 4.2-5.8$\times$, achieving query response in seconds, while reducing deployment costs by 3.2-4.0$\times$ over state of the arts even in their intended single-tier deployments.

Ayush Barik, Sofia Stoica, Nikhil Sarda et al.

Text-to-audio diffusion models produce high-fidelity audio but require tens of function evaluations (NFEs), incurring multi-second latency and limited throughput. We present SoundWeaver, the first training-free, model-agnostic serving system that accelerates text-to-audio diffusion by warm-starting from semantically similar cached audio. SoundWeaver introduces three components: a Reference Selector that retrieves and temporally aligns cached candidates via semantic and duration-aware gating; a Skip Gater that dynamically determines the percentage of NFEs to skip; and a lightweight Cache Manager that maintains cache utility through quality-aware eviction and refinement. On real-world audio traces, SoundWeaver achieves 1.8--3.0$ \times $ latency reduction with a cache of only ${\sim}$1K entries while preserving or improving perceptual quality.

Udbhav Bamba, Minghao Fang, Yifan Yu et al.

Reinforcement learning algorithms such as GRPO have driven recent advances in large language model (LLM) reasoning. While scaling the number of rollouts stabilizes training, existing approaches suffer from limited exploration on challenging prompts and leave informative feedback signals underexploited, due to context-independent rollout allocation across prompts (e.g., generating 16 rollouts per prompt) and relying heavily on sparse rewards. This paper presents XRPO(eXplore - eXploit GRPO), a unified framework that recasts policy optimization through the principled lens of rollout exploration-exploitation. To enhance exploration, XRPO introduces a mathematically grounded rollout allocator that adaptively prioritizes prompts with higher potential for uncertainty reduction. It further addresses stagnation on zero-reward prompts through an in-context seeding strategy that injects curated exemplars, steering the model into more difficult reasoning trajectories. To strengthen exploitation, XRPO develops a group-relative, novelty-aware advantage sharpening mechanism that leverages sequence likelihoods to amplify low-probability yet correct responses, thereby extending the policy's reach beyond sparse rewards. Experiments across diverse math and coding benchmarks on both reasoning and non-reasoning models demonstrate that XRPO outperforms existing advances (e.g., GRPO and GSPO) up to 4% pass@1 and 6% cons@32, while accelerating training convergence by up to 2.7X.

Kaizhuo Yan, Yingjie Yu, Yifan Yu et al.

Proximal Policy Optimization (PPO)-based reinforcement learning from human feedback (RLHF) is a widely adopted paradigm for aligning large language models (LLMs) with human preferences. However, its training pipeline suffers from substantial inefficiencies due to sequential multi-model dependencies (e.g., reward model depends on actor outputs) and long-tail response lengths, where a few long responses straggle the stage completion. We present OPPO, a novel, lightweight, and model-agnostic PPO-based RLHF framework that improves training efficiency by overlapping pipeline execution. OPPO introduces two novel techniques: (1) Intra-step overlap, which streams upstream model outputs (e.g., actor model) in right-sized chunks, enabling the downstream model (e.g., reward) to begin prefill while the upstream continues decoding; and (2) Inter-step overlap, which adaptively overcommits a few prompts and defers long generations to future steps, mitigating tail latency without discarding partial work. OPPO integrates easily with existing PPO implementations with a few lines of code change. Extensive evaluations show that OPPO accelerates PPO-based RLHF training by $1.8 \times-2.8 \times$ and improves GPU utilization by $1.4 \times-2.1 \times$ without compromising training convergence.

Yiding Wang, Decang Sun, Kai Chen et al.

Training deep neural networks (DNNs) is time-consuming. While most existing solutions try to overlap/schedule computation and communication for efficient training, this paper goes one step further by skipping computing and communication through DNN layer freezing. Our key insight is that the training progress of internal DNN layers differs significantly, and front layers often become well-trained much earlier than deep layers. To explore this, we first introduce the notion of training plasticity to quantify the training progress of internal DNN layers. Then we design Egeria, a knowledge-guided DNN training system that employs semantic knowledge from a reference model to accurately evaluate individual layers' training plasticity and safely freeze the converged ones, saving their corresponding backward computation and communication. Our reference model is generated on the fly using quantization techniques and runs forward operations asynchronously on available CPUs to minimize the overhead. In addition, Egeria caches the intermediate outputs of the frozen layers with prefetching to further skip the forward computation. Our implementation and testbed experiments with popular vision and language models show that Egeria achieves 19%-43% training speedup w.r.t. the state-of-the-art without sacrificing accuracy.

Naichen Shi, Fan Lai, Raed Al Kontar et al.

In this paper we propose Fed-ensemble: a simple approach that bringsmodel ensembling to federated learning (FL). Instead of aggregating localmodels to update a single global model, Fed-ensemble uses random permutations to update a group of K models and then obtains predictions through model averaging. Fed-ensemble can be readily utilized within established FL methods and does not impose a computational overhead as it only requires one of the K models to be sent to a client in each communication round. Theoretically, we show that predictions on newdata from all K models belong to the same predictive posterior distribution under a neural tangent kernel regime. This result in turn sheds light onthe generalization advantages of model averaging. We also illustrate thatFed-ensemble has an elegant Bayesian interpretation. Empirical results show that our model has superior performance over several FL algorithms,on a wide range of data sets, and excels in heterogeneous settings often encountered in FL applications.

Fan Lai, Xiangfeng Zhu, Harsha V. Madhyastha et al.

Federated Learning (FL) is an emerging direction in distributed machine learning (ML) that enables in-situ model training and testing on edge data. Despite having the same end goals as traditional ML, FL executions differ significantly in scale, spanning thousands to millions of participating devices. As a result, data characteristics and device capabilities vary widely across clients. Yet, existing efforts randomly select FL participants, which leads to poor model and system efficiency. In this paper, we propose Oort to improve the performance of federated training and testing with guided participant selection. With an aim to improve time-to-accuracy performance in model training, Oort prioritizes the use of those clients who have both data that offers the greatest utility in improving model accuracy and the capability to run training quickly. To enable FL developers to interpret their results in model testing, Oort enforces their requirements on the distribution of participant data while improving the duration of federated testing by cherry-picking clients. Our evaluation shows that, compared to existing participant selection mechanisms, Oort improves time-to-accuracy performance by 1.2x-14.1x and final model accuracy by 1.3%-9.8%, while efficiently enforcing developer-specified model testing criteria at the scale of millions of clients.