Local certification is the task of verifying graph properties using certificates assigned to vertices and checked using only local information. A central question is which properties admit constant-size certificates, independently of network size. A fundamental test case is the parity of the network order. Bousquet, Feuilloley, Valenzuela, and Zeitoun (DISC 2026) showed that the complexity of certifying even order depends dramatically on the verification model. On arbitrary graphs, three bits per vertex suffice with unique identifiers and verification radius two. In the anonymous radius-one model, they established an $Ω(\log\log^* n)$ lower bound and an $O(\log n)$ upper bound. They also showed that constant-size certificates suffice in this model on bounded-expansion classes. In this paper, we settle the complexity of anonymous parity certification on general graphs and further investigate the role of graph sparsity. Our main contributions are: * We close the gap left by Bousquet et al., proving that the local complexity of certifying even order in the anonymous radius-one model is $Θ(\log\log n)$ bits. * We show that on graphs of degeneracy at most $d$, parity can be certified with $O(\log d)$ bits, while $Ω(\log\log d)$ bits are necessary. In particular, every class of bounded degeneracy admits constant-size certificates, strictly generalizing the result for bounded-expansion classes. * We show that constant-size certification cannot be guaranteed on nowhere-dense classes, a broader notion of sparsity that includes bounded expansion. We construct a fixed, subgraph-closed nowhere-dense class on which parity certification requires $Θ(\log\log n)$ bits. Together, our results provide a more complete picture of anonymous parity certification on sparse graph classes, showing that its complexity depends on the particular structural notion of sparsity.
Long Dang, Ali Alfatemi, Kaiqi Xiong, Mohammed Aledhari, Abdellah Chehri, Shawqi Al-Maliki, Mohamed Rahouti
Distributed machine learning (DML), including conventional distributed training and federated learning (FL), enables collaborative model development across cloud-edge-IoT systems without fully centralizing raw data. However, distributing data, computation, updates, and orchestration increases the attack surface and introduces coupled challenges across privacy, robustness, reliability, fairness, and governance. This survey presents a security-centered, heterogeneity-aware roadmap for trustworthy DML, with FL treated as a prominent high-risk case. We first characterize the assets, adversaries, trust assumptions, and coordination patterns that define DML/FL security boundaries. We then analyze how statistical, system, and trust heterogeneity, including non-IID data, resource asymmetry, intermittent participation, and cross-silo/cross-device trust differences, affect convergence, fairness, attack detectability, and defense reliability. Next, we synthesize attack-defense interactions across poisoning, backdoors, Byzantine manipulation, privacy leakage, model extraction, network/orchestration attacks, Sybil behavior, and strategic participation. We further compare layered defenses, including robust aggregation, drift-aware optimization, differential privacy, secure aggregation, homomorphic encryption, secure multiparty computation, trusted execution environments, watermarking, provenance, authenticated communication, and monitoring. Finally, we discuss deployment, evaluation, governance, and open research guidelines, emphasizing reproducible, non-independent and identically distributed (non-IID) partitions, attack-aware benchmarks, privacy budgets, attack success rate, client-level disparity, latency, communication, memory, energy/carbon overhead, and auditable lifecycle artifacts.
Microservice architectures rely on load balancers to distribute requests across service replicas, yet existing designs largely optimize for balanced routing without considering how routing shapes autoscaling signals and resource efficiency. We present Slant, a resource-aware load balancing framework that intentionally creates safe traffic imbalance to improve resource utilization. Slant profiles service-specific critical points, computes conservative multi-dimensional capacity, derives imbalance-aware routing weights that maintain hot and cold pods, and falls back to safer behavior under anomalies. By exposing clearer utilization signals, Slant works with existing vertical pod autoscalers (VPAs) without modification. Experiments on three microservice benchmarks show that Slant reduces allocated CPU cores by up to 26\% while lowering P99 latency by up to 18\% compared to Round-Robin. Under dynamic workloads, Slant triggers VPA scale-up 45\% faster and completes scale-in 40\% faster. Under anomalies, it recovers 40\% faster from pod failures and limits latency amplification to 1.8$\times$ versus 2.6$\times$ during downstream slowdowns.
The empirical success of matrix-aware optimizers, such as Muon (Jordan et al., 2024) and Shampoo (arXiv:1802.09568), highlights the potential of non-Euclidean geometry for deep learning, including large language model (LLM) training. Despite this progress, the theoretical understanding of stochastic optimization in general normed spaces remains limited, particularly in distributed settings. We analyze minibatch stochastic gradient descent (SGD) for smooth nonconvex optimization with respect to an arbitrary norm on a finite-dimensional space. Using the framework of $κ$-regular normed spaces (arXiv:0809.0813), we derive high-probability oracle complexity bounds under a light-tail condition stated directly in the dual norm induced by the update norm. The bounds make the dependence on geometry explicit through the regularity constant $κ$ of the dual space. In the Euclidean case, our bound recovers the standard minibatch SGD rate up to logarithmic factors. For some non-Euclidean geometries and noise models, it is sharper than bounds obtained by converting a Euclidean noise assumption, reducing a linear dimension dependence to a logarithmic one. Building on this framework, we introduce Rennala NSGD, a semi-asynchronous distributed algorithm, and establish wall-clock time complexity guarantees under fixed, heterogeneous worker computation times. In the Euclidean case ($κ=1$), our bound matches the parameter dependence of classical Rennala SGD up to logarithmic factors.
Conventional vision pipelines route every pixel to a host processor, which must then wait for a complete frame before it can act and devote part of each frame period to routine pre-processing. We examine an alternative organisation in which this work is distributed across hardwired distributed processing units (HDPUs): small, fixed datapaths attached directly to the sensor, each committed to a single always-on task, operating truly in parallel on the same pixel stream and delivering compact results to the host. From one library of 3x3 operators sharing a common two-clock latency, we construct three HDPUs (Canny edges, Harris corners, and a line tracker that issues a steering decision) and run them side by side on a single camera stream in a low-cost Artix-7 FPGA, at a cost of 2,568 LUTs and an estimated 0.318 W. Every stage of every unit is bit-exact with its software model. In a full-system simulation with a datasheet camera model, each decision leaves its HDPU within 11.03 camera lines of the pixel that produced it, 22-67x sooner than the earliest moment at which a frame-based program on the host could respond; the HDPU latency varies by no more than 17 ns, whereas the host's own run times reached 12.5x their median. Offloading the three tasks returns 18-30% of a core to the host in every frame period at 640x480, and 63% of the frame period, with all threads in use, at 1080p. The cost of the units grows with frame width only through their line buffers, and in a closed-loop model of a line-following vehicle the HDPU raises the highest speed at which the line is held from 4.75 to 10.00 m/s.
Mariana do Carmo Barata Gonçalves, Valderi Reis Quietinho Leithardt
This study comparatively evaluates post-quantum cryptography (PQC) in automotive communication scenarios, considering the threat posed by quantum algorithms such as Shor's and Grover's algorithms. The simulations use OMNeT++, Veins, INET, and SUMO on two hardware platforms and examine ML-KEM/CRYSTALS-Kyber, ML-DSA/CRYSTALS-Dilithium, and FALCON. ML-KEM and ML-DSA are finalized NIST standards, whereas FALCON is a NIST-selected candidate pending final standardization. The evaluation considers a rural V2V scenario and an urban V2X scenario supported by a roadside unit. Kyber exhibits low {key encapsulation mechanism (KEM)}\color{black} operation latency, Dilithium provides balanced signature performance, and FALCON has high key-generation and signing costs but efficient verification. The results are interpreted as a comparative assessment among the evaluated PQC schemes under the reported simulation conditions; {a conventional ECDSA/ECDH baseline measured in the same simulation setup is also reported for reference.}\color{black} The resulting dataset can support future studies of cryptographic selection in automotive systems.
Michal Orzechowski, Lukasz Opiola, Ignacio Lamata Martinez, Marinos Ioannides, Panayiotis N. Panayiotou, Lukasz Dutka, Renata G. Slota, Jacek Kitowski
The complexity of high-quality 3D digitised cultural heritage objects creates challenges for existing data management systems as they need to develop metadata management and processing capabilities to provide semantic insight into the interconnectivity of data that constitutes cultural heritage objects. To address these challenges, we propose a global federated authentication and authorisation mechanism, a data and metadata management system, and an integrated engine for designing and executing automated workflows that facilitate the processing of both data and metadata. The solution is evaluated with three distinct 3D digitised cultural objects and presents the complete process from data upload to cultural heritage object publication
Po-Hung Chen, Bo-Hung Chen, Dah-Wei Chiou, Jie-Hong Roland Jiang
The recently developed quantum circuit cutting technique greatly extends the capabilities of current noisy intermediate-scale quantum (NISQ) hardware. However, it introduces substantial overhead in both classical postprocessing and quantum resources, as the postprocessing complexity and sampling cost scale exponentially with the number of circuit cuts. In this work, we propose an enhanced circuit cutting framework, ShotQC, which effectively reduces the sampling overhead through two key optimizations: shot distribution and cut parameterization. The former employs an adaptive Monte Carlo strategy to dynamically allocate more quantum shots to subcircuit configurations that contribute more to the variance in the final outcome. The latter exploits additional degrees of freedom in postprocessing to further suppress variance. Integrating these optimizations, ShotQC significantly reduces the sampling overhead without increasing classical postprocessing complexity, as demonstrated across a range of benchmark circuits.
Serverless computing, particularly Function-as-a-Service (FaaS), has revolutionized cloud computing by abstracting infrastructure management and enabling dynamic resource allocation. This paper examines the performance and compatibility of OpenFaaS, an open-source serverless platform, when deployed on various Kubernetes distributions, including Kubeadm, K3s, MicroK8s, and K0s. Moreover, leveraging the CloudLab infrastructure, this study examines the impact of Python, Go, and Node$.$js programming languages on the performance of Kubernetes-enabled OpenFaaS, specifically when these languages are used to develop functions deployed on the platform. The performance is evaluated and analyzed under various levels of concurrent invocations using several usage-level metrics, such as throughput and CPU usage, as well as responsiveness metrics, such as delay. According to our findings, Go consistently outperforms Python and Node$.$js in terms of throughput and CPU usage, making it the ideal runtime for serverless applications. Among the Kubernetes distributions, K3s and Kubeadm exhibit superior performance, with Kubeadm maintaining low latency and efficient CPU usage, and K3s demonstrating high throughput. This study provides valuable insights into optimizing the Kubernetes-enabled OpenFaaS platform, highlighting the strengths and trade-offs of different Kubernetes distributions and language runtimes.
Xin Huang, Teng Ma, Zhuoran Li, Yibo Zhao, Zhaoshi Li, Qianchao Zhu, Wei Yang, Chenxi Wang 等 10 位作者
KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We present FlatCake, a flat KV storage architecture that exposes host DRAM and NVMe SSD as peer media in a unified 64-bit address space. Exploiting the fact that KV blocks are write-once, immutable, and reused at session scope, FlatCake places each block directly onto its target medium in session-contiguous, layer-major order at creation time and never migrates it. FlatCake streams coalesced SSD reads directly into a bounded GPU window via GPUDirect Storage and decouples DRAM and SSD request queues to prevent head-of-line blocking. On a production Claude Code trace with DeepSeek-V4-Flash, FlatCake sustains a 94% cache hit rate as load increases (versus 64% for Mooncake), eliminates Mooncake's $6.5\times$--$6.7\times$ SSD read amplification, and reduces peak-load mean TTFT by $1.6\times$ while improving total token throughput by $1.6\times$. Across 8K--128K contexts on Llama-3.1-70B, FlatCake further reduces mean TTFT by up to $2.4\times$ over Mooncake and $1.8\times$ over Mooncake-GDS.
Trust and reputation mechanisms can reduce the influence of malicious clients in federated learning (FL), but strict update filtering is difficult under non-independent and identically distributed (non-IID) data because honest local updates may temporarily reduce performance on a shared validation proxy. Extending an earlier trust-based semi-decentralized FL (SDFL) framework, this paper operationalizes the trust signals and introduces peer-relative update screening that combines a robust validation-utility score with direction- and norm-based update quality. The method is evaluated with 50 MNIST clients, 20% persistent label-flip attackers, Dirichlet heterogeneity levels of 0.5 and 0.1, 100 communication rounds, and five paired random seeds. Under moderate heterogeneity, the proposed method reaches 98.37% final accuracy while rejecting 95.82% of malicious updates and retaining 90.30% of benign updates; the final benign and malicious trust distributions are completely separated (KS=1.00). Under severe heterogeneity, it reaches 95.25% accuracy, rejects 88.55% of malicious updates, and retains 76.90% of benign updates (KS=0.96). Trust weighting alone preserves slightly higher utility in the most heterogeneous condition, but accepts all malicious submissions. The results therefore distinguish influence mitigation from explicit containment and quantify the utility-selectivity trade-off introduced by screening. The study provides an empirically validated off-chain governance core that can feed blockchain-based audit and incentive layers in SDFL.
Sparsh Gupta, Kamalavasan Kamalakkannan, Maxim Moraru, Galen Shipman, Patrick Diehl
Scientific applications still rely heavily on legacy Fortran codebases built for homogeneous, CPU-based systems. As High-Performance Computing (HPC) shifts toward GPU-accelerated architectures, many accelerators lack native Fortran bindings, creating a need to translate and optimize legacy code for portability. Kokkos offers performance portability through a single-source C++ abstraction, but manual porting from Fortran demands significant expertise and time. While large language models (LLMs) show promise in source-to-source code generation, their use in autonomous workflows for translating and optimizing parallel code is largely unexplored. This paper presents an agentic AI workflow in which specialized LLM agents collaborate to translate, validate, compile, run, test, debug, and optimize Fortran kernels into portable Kokkos C++ programs without human intervention, bounded by configurable retry and optimization thresholds. We evaluate the pipeline over five runs per configuration on an NVIDIA A100, characterizing its variability in success rates, token cost, and optimization behavior, on benchmark kernels including a real-world multi-file kernel from a production climate model. Proprietary models such as GPT-5 and o4-mini-high completed the full workflow most reliably, producing optimized code that exceeded the Fortran baselines for only a few U.S. dollars, whereas open-weights models such as gpt-oss-120b and Qwen3.5-397B-A17B handled simpler kernels but struggled with complex Kokkos abstractions such as parallel reductions and hierarchical parallelism. This work demonstrates, at the kernel level, the feasibility of agentic AI for Fortran-to-Kokkos transformation and, more broadly, for structured software-engineering tasks such as automated code modernization and refactoring.
Zejian Wang, Yifan Zhao, Kaihui Ren, Xiaoling Zhou, Yubin Wang, Chenhao Guan, Lin Zhang, Yufeng Li 等 14 位作者
Fully Sharded Data Parallelism (FSDP2) maintains parameters and gradients as per-parameter tensors, whereas its collective path operates on flat, rank-major buffers. Converting between these layouts copies all-gather outputs into parameter storage and packs gradients into full-group fp32 buffers. Layout conversion alone accounts for 6.4% of step time in Llama-3.1 70B training at 8K tokens per GPU. Communication kernels impose a second cost: a default NCCL reduce-scatter can slow a concurrent GEMM that would otherwise complete in a single wave by 69%. We present TIDE, a set of zero-copy collectives built on public NCCL 2.31+ APIs. TIDE preserves FSDP2's per-parameter storage and compute kernels while folding layout placement into collective addressing. TIDE-AG delivers shards directly to parameter storage using NVSwitch multicast and paced recursive-doubling RDMA, with no SMs used for transport or placement. TIDE-RS consumes per-parameter bf16 gradients directly and reduces them into fp32 shards (a mixed-precision mode that NCCL does not provide) using copy engines and short-lived thread blocks (CTAs) for node-local accumulation and NVSwitch for cross-node reduction. A two-phase pipeline separates gradient reads from network progress; Slot-Reuse optionally stages gradients in vacated all-gather slots. On 64 GPUs, TIDE reduces step time by up to 27.1% relative to Stock FSDP2 and 6.0% relative to FSDP2+SymmMem across Llama-3.1 70B, Wan2.2, and Qwen3-30B-A3B. It reduces reduce-scatter SM time by factors of 12.5-14.5 and cuts interference with concurrent compute by up to an order of magnitude. Its reduce-scatter buffer remains nearly constant from 16 to 64 GPUs at a fixed per-GPU output size, and its peak memory is up to 4.7 GiB below FSDP2+SymmMem's and, with Slot-Reuse on Wan2.2, below Stock FSDP2's.
Serverless platforms increasingly host machine learning (ML) inference workflows, modeled as directed acyclic graphs (DAGs) of interdependent functions. Beyond container startup, each ML node must load libraries, model weights, and CUDA kernels before it can run. These costs accumulate along dependent DAG paths and place initialization on the end-to-end critical path. Existing systems optimize one half of this problem. DAG-agnostic pre-loading prepares nodes the request may never reach, wasting host and GPU memory. DAG-aware serving routes and scales whole functions, leaving the loading stages inside a node untouched. This paper presents GearUp, a serverless DAG inference serving system that removes ML-node loading from the end-to-end critical path at low resource cost. GearUp profiles the loading stages of each node and schedules them inside the slack time the DAG exposes, hiding initialization behind upstream execution. It reuses initialized instances within and across requests to avoid redundant cold starts. A probabilistic, traffic-adaptive keep-alive policy sets how long idle instances are retained, balancing memory cost against cold-start risk under bursty load. On the AMBER Alert and FINRA-ML workflows with NVIDIA L40S GPUs, GearUp reduces median end-to-end latency by 64.5% and 75.9% against SMIless, for median speedups of 3.1x and 10.0x over cold execution. Integrated GPU and host memory allocation fall by up to 83.9% and 94.6% against eager pre-loading in InstaInfer. GearUp achieves 4.2x strong scaling on four workers.
The short-time Fourier transform (STFT) computes the Fourier transform of a short-duration window of a longer signal, which characterizes how a frequency changes over time. While the Fourier transform algorithms, particularly the Fast Fourier Transform (FFT) algorithm, has been well developed to take advantage of parallel computing, there is limited work on understanding the performance of parallel short-time Fourier transform algorithms. In this study, two parallel short-time Fourier transform algorithms, one is FFT-based and the other a quasi-parallel feedforward algorithm, are implemented and analyzed in comparison with each other. It is found that while the quasi-parallel algorithm outperforms in runtime under low concurrency settings, the FFT-based algorithm scales better and has lower memory usage. For each algorithm, three parallel approaches, shared-memory, distributed-memory and hybrid distributed-shared-memory, are tested. Profiling results show that all implementations are compute bound except for the hybrid feedforward algorithm, and that the distributed-memory variants incur significantly higher communication cost. Overall, this study presents a deeper and more comprehensive analysis and demonstrates the defining characteristics of different families of parallel short-time Fourier transform algorithms, in which the FFT-based approach uses less memory while the quasi-parallel feedforward algorithm has lower runtime but higher serial fraction.
It was recently shown (arXiv:2610.09618) that the infimum of the approximation ratios of deterministic constant-round LOCAL algorithms for minimum dominating set on planar graphs is 7. The $(7+\varepsilon)$-approximation given there runs in $T_0+O((c_0/\varepsilon)^κ)$ rounds with $κ=\ln 3/\ln(24/23)\approx 25.81$, and it is asked there whether this exponent can be substantially reduced. We replace the heaviest-edge star selection in its clustering step by a random-centre rule: every part becomes a centre independently with probability $p$, and every non-centre joins its heaviest centre neighbour. For planar graphs this gives $κ<5.89$ with $p=3/8$ (three fair coins per part and level) and $κ<6.7$ with $p=1/2$. An analysis that counts only the join edges cannot go below $5.74$. For every proper minor-closed class of density $a$, the new exponent is never worse, and with about $\frac12\log_2 a$ fair coins per part and level it is smaller by a factor that tends to $8$ as the density grows. This is optimal for the rule when $a$ is an integer and the class contains all $K_{a,n}$. The planar bounds hold verbatim on the torus, the projective plane and the Klein bottle. We then turn to the dependence on $\varepsilon$. On paths, every deterministic $r$-round algorithm for fractional domination, and every randomized one that always outputs a dominating set, has (expected) ratio at least $1+1/(r+4)$. So these approximations need $Ω(1/\varepsilon)$ rounds on every minor-closed class that contains all paths, and on paths this is tight up to an additive constant. On forests and on outerplanar graphs, with no bound on the degree, $O(1/\varepsilon)$ rounds suffice, so there the optimal number of rounds is $Θ(1/\varepsilon)$. The planar level bounds, the bounds on paths and the deterministic bound on forests are machine-checked in Lean 4 with Mathlib.
Yaqi Qiao, Ping He, Songrun Xie, Ayush Barik, Chensong Zhang, Zhengzhong Tu, Fan Lai
Diffusion models have become the central backbone for modern image, video, and audio generation, but their efficient service remains a challenge. Unlike autoregressive decoding, diffusion inference repeatedly updates high-dimensional spatial or temporal latents over many denoising steps. This all-region execution pattern makes generation latency high and limits serving throughput. Existing multi-GPU parallelization methods can reduce per-step computation, but often introduce substantial activation exchange overhead, causing communication to offset or even outweigh the benefits of parallel execution. This paper presents FlashDiff, a diffusion serving system that improves inference efficiency through adaptive regional execution and scheduling. FlashDiff is based on the observation that diffusion refinement is not uniform across latent regions or denoising steps: different regions often stabilize at different rates, while neighboring steps exhibit strong temporal correlation. FlashDiff leverages these properties to selectively execute only regions that require further refinement and to reallocate the resulting compute slack across concurrent serving requests. FlashDiff consists of three mechanisms. First, it decomposes the latent representation into coherent execution regions using early-stage attention signals, preserving semantic structure while exposing fine-grained parallelism. Second, it uses a lightweight runtime controller to estimate region activity and bypass low-impact updates when further refinement is unlikely to affect output quality. Third, it applies an affinity-aware online scheduler that co-locates dependent regions, balances residual load across GPUs, and reuses reclaimed compute capacity to improve serving efficiency. Across real-world image, video, and audio workloads, FlashDiff reduces end-to-end serving latency by 30-97% and improves throughput by 1.2-2.2x.
This paper presents a federated learning (FL) framework for semantic communication (SemCom) over the multiple access channel (MAC), where multiple distributed transmitters simultaneously convey semantic information to a common receiver using end-to-end learned semantic representations. Semantic feature extraction, compression, wireless transmission, and semantic inference are jointly optimized through FL-based SemCom. Unlike conventional SemCom systems that train users independently, simultaneous semantic transmissions over the MAC enable concurrent training of multiple transmitters within the same transmission interval, thereby reducing training overhead and channel usage. Federated semantic aggregation is incorporated with shared and personalized encoder components, allowing transmitters with limited local training data to benefit from semantic representations learned from transmitters possessing larger datasets while preserving local adaptation. Evaluated using the CIFAR-10 dataset over fading channels, the proposed approach maintains high semantic inference accuracy under shared wireless access and significantly improves performance under severe data imbalance. In the meantime, simultaneous semantic transmissions can significantly reduce training overhead while preserving robust semantic performance across distributed transmitters with heterogeneous local datasets. These results highlight the potential of combining FL and SemCom to support scalable and communication-efficient edge intelligence in future wireless networks.
Prefill and decode are complementary in resource usage, but often divergent in parallelism preferences. Executing the two phases separately preserves their preferred parallelism but leaves complementary resources underutilized, while co-batching improves utilization but often requires a uniform parallel configuration that hurts execution efficiency. To resolve this utilization-efficiency dilemma, we propose Conflux Parallelism (XP), a parallel execution abstraction that co-executes prefill and decode with different parallelism within the same batch. Conflux Parallelism exploits operator-level Transformer structure to construct a heterogeneous execution path with different batching and parallelism strategies across operators, combining the best of co-batching and phase-specific execution. Building on Conflux Parallelism, we design ConfluxServe, an SLO-aware serving system that maximizes cluster goodput under TTFT and TBT constraints. Evaluation on production workloads shows that ConfluxServe improves serving goodput by 1.26-1.62x over the strongest baselines.
Hybrid large language models (LLMs) have emerged as a promising way to combine the modeling capability of the Attention mechanism with the efficiency of state space models (SSMs). However, most existing hybrid LLMs still inherit a static design philosophy: once the Attention-to-SSM layout is chosen, every request must follow the same hybrid execution. This one-layout-fits-all design is inefficient for real-world serving workloads, where different requests rely on contextual information in different ways. This paper presents Morph, a serving system that co-designs the model architecture and runtime for request-adaptive hybrid LLM inference. Its model architecture pairs Attention and SSM layers within each block and uses prefill-oriented gating to select a request-specific layout that remains fixed throughout decoding. Its serving runtime preserves batching efficiency under divergent layer choices through selective batching and enables correct prefix reuse through hybrid-aware context caching. Experiments on six datasets show that Morph achieves a better quality-efficiency trade-off than existing static hybrid baselines. System-level evaluation further shows that Morph improves the serving throughput by up to 3.87x over baselines while substantially reducing inference latency.
Shuhuai Li, Shengdong Ye, Dongdong Ge, Yinyu Ye, Jianghao Lin
Agentic services alternate between model inference on clustered GPUs and external tools on distributed CPUs. In this scenario, existing systems (1) overlook tool-return readiness in service evaluation, wasting limited resources, and (2) use heuristic strategies to decide admission and placement on GPUs, resulting in suboptimal scheduling decisions. To bridge this gap, we present TAO, a tool-aware joint scheduling framework for LLM-based agentic serving. Its value interface combines near-term tool readiness, node-specific cache credit with time-aware decay, and token footprint; a dynamic program jointly selects and places waiting sessions under capacity constraints. Across the 12 reported model-workload-concurrency settings, TAO reduces latency by up to 26.8%, reduces mean time to first token (TTFT) by up to 36.3%, and increases decode throughput by up to 29.5% relative to the best evaluated state-of-the-art baseline for each metric.
We analyze the differential privacy of gossip averaging on a graph with independent Gaussian noise at every node, as seen by a subset of observing nodes. We characterize the privacy loss of any sensitive node over multiple gossip steps as that of a Gaussian mechanism whose sensitivity is the solution to a binary quadratic problem with a coefficient matrix determined by the gossip matrix. We recognize this matrix as a Toeplitz matrix plus a small correction, and show that its Toeplitz symbol controls how the DP guarantees grow with the number of steps. Under a computable condition on the Toeplitz symbol, the privacy guarantee matches that of a trusted central aggregator, and we give a utility bound for strongly convex learning problems that similarly matches the centralized case. Finally, we give numerical recipes for upper and lower bounds on the DP guarantees. On realistic test cases, the two bounds agree to within a few percent.
The fastest machine-learning kernels are compositions of components, such as loaders, matrix-multiply mainloops, epilogues and communication stages, that run concurrently on one GPU streaming multiprocessor and share its on-chip memory, barriers, warps and registers. Because these components hard-code where they sit in this shared space, they cannot be compiled, shipped or replaced on their own: combining components from different sources means forking them by hand or keeping them in separate kernels. We present PipeLink, a contract-guided linker for GPU kernel components. Each component is compiled separately into an LTO object that carries a contract stating what it requires and what it promises; PipeLink checks the contracts against each other, relocates every component by solving one joint placement of the SM's resources, and generates the transfers between them, and the vendor LTO backend folds the result into constants. Across 51 configurations on Hopper and Blackwell GPUs, linked kernels run within 0.2% of the same sources compiled as one unit. Linking makes new compositions practical: a CuTe attention operator inside a ThunderKittens decode megakernel speeds up decoding by 2.16$\times$, a four-owner MoE training megakernel matches the original monolith, one unchanged attention core yields sparse and context-parallel variants that outperform FlexAttention, vLLM and Transformer Engine, and a GEMM + SwiGLU composition whose consumer serves three GPU architectures runs 1.2$\times$ faster on Blackwell than cuBLAS with a compiled activation.
We study deterministic gossiping in unknown strongly connected directed radio networks with $n$ vertices, $m$ arcs, and distinct identifiers in $[N]$, where $N=n^{\mathcal O(1)}$. Messages are unrestricted and collision detection is unavailable. Our neighborhood-clearing procedure reuses an elected receiver to secure its entire closed in-neighborhood in $\mathcal O(n\log^2n\log\log n)$ rounds. Charging each distinct closed neighborhood once yields $\mathcal O(m\log^2n\log\log n)$ rounds under an explicit degree-concentration condition, for every polynomial identifier range. A complementary adaptive-pruning procedure combines quadratic contraction with completion sensitive to the number of unsecured sources. If $a\ge1$ sources remain at completion entry, permutation selectors reduce the balanced completion budget by a factor of $\log^{2/3}(a+1)$ relative to interleaved path selectors. A geometric execution attains $\widetilde{\mathcal O}(k^2+nτ+\min\{u^2,(nu)^{2/3}\})$ for a path cover of scale $k$, weight $τ$, and $u$ exceptions. For directed graphs with common in- and out-degree $Δ$, the portfolio achieves $\widetilde{\mathcal O}(\min\{n\sqrtΔ,n^2/Δ\})$. Every graph also admits $\widetilde{\mathcal O}(\min\{n^{4/3},N+(mn)^{3/5}\})$ rounds. Only $n,N$ are inputs; degrees, covers, and structural constants are used solely in the analysis. The unrestricted polynomial-identifier density extension remains open.
Sparse Mixture-of-Experts (MoE) models offer scalable capacity for Web applications, but uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory. We present LRMoE, a memory-aware MoE training system that jointly recalibrates backward demand and balances the retained workload. During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas. A capacity-aware routing algorithm distributes retained pairs among eligible instances of their original experts. Physical compaction then forms smaller communication payloads and variable-size grouped GEMMs, reducing backward traffic, computation, and execution-buffer demand. Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.
Wentao Huang, Qiming Huang, Yunhong Ji, Mian Lu, Yuqiang Chen, Bingsheng He, Kian-Lee Tan
Sparse general matrix-matrix multiplication (SpGEMM) is a key kernel in graph analytics, relational queries, and machine learning. Yet, its performance remains a significant challenge, due to the inherent sparsity and, more critically, the prevalent skew patterns often found in real-world workloads. This paper investigates this problem and demonstrates that an unbalanced workload distribution is the main culprit, harming performance through two co-existing root causes: first, multi-dimensional skew, where imbalance across computation cost, output density, and data locality cannot be captured by any single balancing metric; and second, cache contention, where dense substructures induce severe on-chip interference. To that end, we propose adaptive morsel scheduling (AMS), a dynamic scheduling framework that decouples skew dimensions via fixed-result-size morsels and integrates contention-aware load balancing (CALB) and data-conscious simultaneous multithreading activator (SMTact) to alleviate cache-level contention, alongside a hybrid accumulation mechanism (HYB) that adaptively selects the best accumulator per morsel. Evaluations across matrix squaring, relational queries, graph analytics, and other sparse matrix kernels show that AMS significantly outperforms state-of-the-art methods on skewed workloads while maintaining comparable performance on non-skewed ones. These gains, along with improved bandwidth utilization, remain valid across a wide range of workloads and hardware systems.
High-performance General Matrix Multiplication (GEMM) implementations rely on breaking down GEMM into carefully selected tiles to achieve peak performance. We present Waypoint, a new analytical model to automatically identify the optimal tile size for a given hardware and input matrix shape. Waypoint estimates the achievable performance for a given tile choice by capturing all the architectural effects including interactions with memory hierarchy, instruction-level parallelism, and multi-threading. Using the estimates as a proxy, we search through the space of possible tile sizes to identify the one that maximizes performance. Additionally, Waypoint helps hardware designers estimate realistic performance on experimental hardware optimizations. We build a new GEMM library OPAM for CPUs that uses Waypoint to automatically break down GEMM and achieve performance that matches state-of-the-art GEMM libraries and helps accelerate LLM inference on CPUs achieving $1.36 - 1.98 \times$ speedups.
Saarth Deshpande, Neeraja J. Yadwadkar, Venkat Arun
Microservice operators must reason about how applications would behave under changes that have not yet occurred: different workloads, resource allocations, or SLO targets. Evaluating every alternative experimentally is expensive, but predicting such counterfactuals is difficult because microservices contain many interacting components. End-to-end models must learn behavior over a combinatorial space of conditions, but composing per-component predictors causes prediction errors to accumulate. We address this challenge by changing what each component model must predict. Rather than predicting exact values, we learn simple, interpretable bounds that contain every observed behavior while remaining tight enough to support useful end-to-end conclusions. We build AGAR, a counterfactual performance analysis tool that composes these models to answer end-to-end SLO-feasibility questions. This allows us to formulate autoscaling as a counterfactual query: what is the smallest resource allocation under which an SLO is met? Our evaluation demonstrates that across three microservice applications, AGAR's bounds are actionable: its derived autoscaler reduces CPU allocation by up to 36.8% over state-of-the-art horizontal autoscalers while satisfying application SLOs. Further, AGAR is able to predict operational cost given workload traces, provide guidance on which service's startup time to optimize and localize performance bugs within each application.
Yangshen Deng, Jiaqi Ruan, Yeqi Huang, Nick Lau, Guangming Sheng, Luo Mai
Coding agents are beginning to develop the production-grade AI infrastructure that trains and serves large models, yet existing benchmarks cover only single kernels or single-GPU tasks, and grade with tests and prompts that misjudge both correct and incorrect patches. We introduce InfraDev, a benchmark of 36 tasks built from real-world pull requests to Megatron-LM, vLLM, and verl, covering infrastructure development for training, inference, and reinforcement learning. Each task asks an agent to reimplement a real change that spans many files and must run correctly across many GPUs. To grade such changes, we jointly synthesize hidden probes, which enforce the requirements the original tests leave unchecked, and a prompt that states every interface the probes call without revealing the solution. We also design a shared execution environment that reuses dependencies and artifacts across tasks, cutting storage by 52-57% and build time by 13-14% on Megatron-LM and vLLM. Across six frontier models, the best resolves only 38.9% of the tasks; the probes reject 32% of the submissions that pass every official test, and with the original pull request descriptions as prompts, no model resolves a single task. Failure analysis shows that structural and state errors dominate and that agents' own tests rarely check memory, time, or traffic. Turning these findings into an agent skill raises the resolved rate of a frontier model on the Megatron-LM tasks by 10.5%.
Edge-cloud synergies provide a promising paradigm for privacy-preserving deployment of foundation models in heterogeneous web environments, enabling lightweight edge models to support domain-specific multimodal we b services while cloud-hosted models facilitate collaborative knowledge sharing. However, in practical edge-cloud web environments, collaborative multimodal learning is challenged by modality heterogeneity (different modality combinations across domains) and model-structure heterogeneity (different modality-specific encoders/fusion modules). To address these issues, we propose ML-ECS, a collaborative multimodal learning framework that enables joint training between a server-based model and heterogeneous edge models. This framework consists of four components: (1) cross-modal contrastive learning (CCL) to align modality representations in a shared latent space, (2) adaptive multimodal tuning (AMT) to preserve domain-specific knowledge from local datasets, (3) modality-aware aggregation (MAA) to robustly aggregate while mitigating noise caused by missing modalities, and (4) SLM-enhanced CCL (SE-CCL) to facilitate bidirectional knowledge transfer between cloud and edge. Experimental results on various multimodal tasks show that ML-ECS consistently outperform state-of-the-art baselines under varying modality availability, achieving improvements of 1.55% to 16.75% in Rouge-LSum and improving both client- and server-side performance. In addition, by communicating only low-rank LoRA parameters and fused representations, ML-ECS achieves high communication efficiency, requiring only 0.71% of the total parameter volume.