Papers

Fault-tolerant agreement protocols fail if replicas share a common flaw that simultaneously affects more replicas than the tolerable threshold. Therefore replicas should ideally fail independently, which can be achieved through diversification. However, in practice, often the same protocol implementation is shared by all replicas which is not surprising given that the provision of multiple diverse implementations is difficult and highly laborious. This poses a major risk, as a shared protocol implementation is a prime candidate for common bugs due to its complexity. With PatchyBFT, we demonstrate how, given a reference implementation, Large Language Models (LLMs) can be utilised for the automated and scalable generation of code that compiles, passes tests, and crucially differs semantically/binary-wise, that can re- place code in the reference implementation, thereby significantly reducing diversification costs. We demonstrate the feasibility of diversification of replication protocol implementations using LLMs by diversifying three implementations: PBFT, HotStuff, and Raft, showing how up to 65% of the codebase can be diversified.

Critical infrastructure assets such as bridges, tunnels, dams, and power line networks require timely and scalable inspection. While conventional manual inspection remains costly and hazardous, unmanned aerial vehicle (UAV)-based inspection has emerged as an efficient alternative for monitoring difficult-to-access structures. Existing UAV inspection pipelines have evolved from cloud-centric offline processing toward edge-based perception using lightweight object detectors such as YOLO for real- time defect localization. This article explores the transition toward fully edge-native semantic inspection powered by lightweight vision language models (VLMs), where UAVs move beyond object detection toward contextual structural understanding. It categorizes existing UAV inspection architectures, identifies their key system challenges and architectural requirements, and experimentally assesses the feasibility of semantic edge intelligence on NVIDIA Jetson UAV-class hardware using the COCO-Bridge dataset. The evaluation integrates a fine-tuned YOLO-26M for object localization and a lightweight SmolVLM-256 for semantic reasoning. Finally, it outlines future directions toward agentic, autonomous, trustworthy, and collaborative semantic UAV inspection across the edge-cloud continuum.

Large language models are increasingly trained on GPUs distributed across multiple regions, but geo-distributed training is challenging in practice. Real clusters often contain GPUs with different speeds and memory capacities, and they communicate over slow wide-area networks. Our analysis shows that this creates serious problems: existing synchronous methods preserve stable updates, but fast GPUs wait up to 20.9% of their runtime for slower ones, and all workers spend, on average, 65.8% of their runtime on synchronization. Recent asynchronous methods reduce waiting time but worsen the model accuracy due to stale updates. To address the problems, we present GeoMesh, a synchronous geo-distributed training framework for heterogeneous GPUs. GeoMesh balances per-worker workloads by assigning each GPU a suitable batch size and number of inner steps, so faster GPUs do more useful work instead of waiting. It also reduces communication volume by nearly 32x by exchanging compressed sign-based pseudo-gradients with lightweight magnitude and token count. Across heterogeneous GPUs and Azure-derived WAN, GeoMesh reduces time-to-target perplexity by up to 70.2% over representative baselines and lowers straggler- and WAN-induced GPU idle by up to 8.0x and 5.6x, respectively, while preserving comparable zero-shot accuracy.

Distributed model training at scale is frequently interrupted by transient network failures, conventionally forcing cluster managers to abort all processes and roll back to the latest checkpoint. While periodic checkpointing provides durability, frequent snapshotting introduces severe storage backpressure: our measurements on a 1.216B-parameter decoder reveal that per-update asynchronous checkpointing incurs up to a +656.7% latency overhead, consumes 32.5 GiB of host memory, and generates 3.39 TB/hour of storage traffic. To eliminate this overhead, we present AccelPact, a parallel runtime enabling zero-I/O in-memory fault recovery for sharded distributed training. When communication fails at a committed optimizer step, device memory remains quiescent and uncorrupted, yet continuation fails because frameworks like PyTorch FSDP cache internal communication handles across module wrappers and parameter hierarchies. AccelPact resolves this dependency invalidation via a non-invasive reference-rebinding mechanism coordinated by an out-of-band Gloo consensus protocol. On 16 NVIDIA RTX 5880 GPUs training full-parameter Mistral-7B, AccelPact eliminates checkpoint replay, yielding a 1.197x whole-run goodput improvement over cold restart and 1.194x over NVRx checkpoint restoration at checkpoint age 5, rising to 1.698x at age 18. Across ten successive fault injections, all 16 ranks maintain bit-identical parameter state with zero numerical drift. Across 4-to-16 GPU cluster topologies, reference rebinding executes in constant time (0.493-0.518 ms). Operating directly on native communicator instances, AccelPact requires zero application-code modifications and avoids compiler graph breaks under torch.compile, providing an efficient foundation for resilient deep learning.

LLM serving is typically offered as a shared, multi-tenant service, where high-demand workloads from one client can cause latency SLO violations for others. Existing solutions for performance isolation equalize client throughput in the long run, for example through queueing and batching fairness. However, these approaches do not provide latency isolation guarantees; as a result, well-behaved clients can still experience significant degradation to their token-level latencies. In this paper, we present FairInference, which provides the novel {\delta}-token fairness guarantee: for a well-behaved client, if a token is generated in d time units in isolation, it will be generated within d + {\delta} time units in multi-tenant execution, providing strong latency isolation guarantees for LLM serving. To achieve this, FairInference addresses a key challenge of LLM serving: bounding delays from sharing GPU resources without support for fine-grained scheduling or resource allocation. In FairInference, the scheduler enforces per-token deadlines, while bounding the delays from GPU compute sharing and accounting for the additional delays introduced by the shared KV caching in GPU memory. We show that FairInference effectively bounds token-level latency spikes for well-behaved clients and improves overall throughput compared to state-of-the-art LLM serving systems.

Frontier open-weight language models increasingly use Mixture-of-Experts (MoE) architectures to expand model capacity while activating only a small subset of experts per token. Local inference must nevertheless keep the complete expert pool available, which remains far beyond consumer-grade RAM and VRAM capacity even after quantization. SSDs provide practical capacity at this scale, but turning that capacity into executable model memory requires efficient expert delivery, coordinated management of SSD, RAM, and VRAM, and CPU–GPU hybrid execution under bounded bandwidth. We present SSD-LLaMA, an SSD-native local MoE inference system that addresses these challenges with an SSD I/O pipeline optimized for expert delivery, a native three-tier storage hierarchy that delivers and retains experts dynamically, and balanced CPU–GPU hybrid execution. SSD-LLaMA executes every selected expert without pruning or substitution. Across three frontier MoE model families, SSD-LLaMA improves prefill token rate by 1.52$\times$–4.19$\times$ and decode token rate by 2.10$\times$–15.58$\times$ over the evaluated baselines. We also achieve higher than 1 token/s for running trillion-parameter model with a single RTX 5090 and no more than 32GB RAM.

Open-source video generative models ship almost exclusively as PyTorch/CUDA reference implementations. This leaves Cloud TPU pods without a production-ready inference path, despite offering large, cost-effective accelerator memory pools ideal for long-sequence spatiotemporal attention. We present vidax, an open-source JAX/Flax inference engine and zero-copy PyTorch-to-JAX weight translator for modern video generation architectures. vidax covers a diverse set of spatiotemporal models — including Diffusion Transformers, omnimodal Mixture-of-Transformers, 3D VAEs, text encoders, and native samplers — with zero PyTorch dependency in the execution path. The framework unifies 1D tensor parallelism with DeepSpeed-Ulysses sequence parallelism on a single JAX sharding mesh, integrates TPU flash-attention kernels, and implements per-layer weight offloading to support reference resolutions that exceed single-device memory. We benchmark compile times, latency, and peak memory utilization on TPU v4-8 hardware, and document real-world numerical bugs surfaced during checkpoint translation. vidax is released open-source as a baseline for JAX and TPU video generation research.

Private large language model (LLM) fine-tuning is increasingly important for organizations that need to adapt models using sensitive data, but it often exceeds the memory capacity of commodity datacenter accelerators. Apple Silicon offers a different design point through large unified memory and lower complete-system cost, while recent Apple software support enables distributed execution over RDMA-over-Thunderbolt (TB). This paper studies whether Apple Silicon can serve as a practical platform for private LLM fine-tuning. We characterize RDMA-over-TB communication on Mac Studio nodes, showing that the measured bandwidth is far below nominal TB specifications. Next, we extend Apple's implementation with multi-trunk communication, persistent worker threads, and CPU-side gradient overlap to better exploit multiple direct TB links for LLM fine-tuning workloads. Finally, on a four-node Mac Studio cluster that fine-tunes a Qwen3-9B, our optimizations improve weak-scaling throughput by up to 1.6X over the single-trunk, non-overlapped baseline and reach 936 tokens/s for sequence length 17408. We further compare Apple Silicon with an NVIDIA H100 platform to quantify the trade-off between memory capacity, throughput, and acquisition cost, showing that Apple Silicon can provide a cost-effective solution for private LLM fine-tuning.

Poor indoor air quality can cause up to five times more direct health problems to occupants than outdoor air. In particular, it may cause headaches, fatigue, eye/throat irritation, and long-time exposure is linked to respiratory and heart as well as some forms of cancer. Despite the importance of indoor health and well-being, most current monitoring devices and systems (usually for offices and workspaces) are passive. The Environmental Quality Monitor (EnQyMo) platform is a generic Internet of Things (IoT) middleware designed to process several sensor data related to air quality in indoor spaces and correlate this data with health exposure risks of users/workplace employees. Using Bluetooth Low Energy (BLE) beacons and a mobile IoT middleware it is able to identify the (smartphone) users exposed to these polluted air or high CO2 (carbon dioxide) levels, and generate location-specific alarms only to the users at the places with the unhealthy air conditions. At the core of EnQyMo is an agency of Large Language Models (LLMs) capable of interpreting regulatory standards and scientific literature to automatically identify critical health exposure levels.

Long-context LLM inference is bottlenecked by attention, whose repeated KV-cache reads make decoding memory-bound. Self-speculative decoding alleviates this by drafting tokens with sparse attention and verifying them with full attention, but existing batched methods remain synchronized: all requests in a batch share a single draft-verify schedule, even though the optimal draft length varies widely across requests and changes dynamically within each request. We propose ASPIRE, a non-synchronized batched self-speculative decoding framework built on three components. First, a unified mixed forward allows drafting and verifying requests to coexist in the same batched forward pass, removing the need for global draft-verify phases. Second, a lightweight online speculation scheduler uses per-request acceptance-rate estimates and a batch-aware cost model to let each request independently choose when to verify. Third, an intra-draft refresh layer performs full attention at a single designated layer during drafting, updating the sparse context at every draft step to reduce staleness during drafting. Across three models and five reasoning and long-context benchmarks, ASPIRE achieves $1.70$-$4.58\times$ speedup in decoding throughput over autoregressive baselines and improves average speedup by approximately $27\%$ over the strongest prior self-speculative baselines.

Dynamic Image Graph Construction (DIGC) is the primary performance bottleneck in FPGA acceleration of Vision Graph Neural Networks (ViGs), reconstructing graph connectivity at every layer through irregular, memory-intensive computation. Existing FPGA accelerators optimize DIGC but still execute it unconditionally, making repeated graph reconstruction a persistent source of latency and energy consumption. We propose the SpecReuse algorithm, which computes compact spectral descriptors of intermediate features and reuses previously constructed graphs when descriptor drift remains below a calibrated threshold. We further present the SpecReuse accelerator, an FPGA architecture that realizes graph reuse through lightweight hardware for spectral descriptor extraction and reuse control while remaining compatible with existing graph-construction accelerators. Experimental results demonstrate up to a $2.69\times$ speedup in end-to-end inference and approximately 58–65% lower energy per inference with negligible FPGA resource overhead and minimal loss in classification accuracy.

When agentic sessions run to a million tokens with many sessions resident at once, the KV cache and the index that ranks it live in host memory, and the scan that ranks all n keys for a top-k step becomes the traffic that bounds decoding. We present Fathom, a key scan in which each query decides how many bits of each key channel to read. The 4-bit K cache is stored channel-major as bit planes, so a prefix of t planes is exactly the channel's t-bit quantizer, and the query spends its bit budget by reverse water-filling over the variance-weighted importance of its channels. At one million tokens on Qwen3-8B a decode step is 1.67x faster in GPU time than with the 136-bit scans of Double Sparsity, Loki and SparQ r=32, and in the same GPU time as SparQ's 68-bit read (r=16) Fathom reads 18% fewer bytes with lower attention error on six of seven model and context settings. On RULER-style tasks every per-token scan matches exact top-k decoding, and on real coding-agent sessions Fathom reaches the step agreement of the most accurate 136-bit scan at 92 bits. The store is the 4-bit K copy a quantized serving stack already holds, and the method is not faster when the index is resident in GPU memory.

Durable-execution frameworks commonly journal execution steps of a long-running function and reconstruct state by deterministic replay; the journal is therefore the workflow's authoritative state, and in current deployments it typically resides with whichever provider hosts the run. This paper examines what changes if the journal is instead an authenticated, append-only log replicated among the parties to the workflow. We propose an architecture in which runs are chains of single-writer epoch cores; executor succession is governed by quorum-finalized closure certificates; durability, timeout, and checkpoint decisions are carried by explicit certificates rather than implicit trust; and large payloads travel a separate content-addressed distribution plane. A threat model separates what signatures establish (authorship and order) from what they do not (truth of recorded effects). We state the design as a set of typed artifacts and invariants, identify which claims are established by existing systems and which are proposals requiring validation, and define the prototype experiments, including specific adversarial cases, that would validate or falsify the proposal's central mechanisms.

Multi-agent systems split a task across a tree of agents and justify the split with folklore: smaller contexts, cleaner separation, parallelism. We ask what the split does to how much of what the leaves discover reaches the root. Model a decomposition as a tree in which an agent handed $b$ items keeps any one with probability $r(b)$. If $r(b)=1/b$, every tree delivers exactly one finding, for every task size and every shape; we verify this to $2.4 \times 10^{-15}$ on 20,000 random irregular trees. If $r(b)=Cb^{-\delta}$, a depth-$k$ tree over $N$ findings yields $C^k N^{1-\delta}$: task size and architecture separate, and architecture contributes only $C \le 1$ per level, so flat is optimal for yield and no arrangement of agents escapes the exponent $\delta$. On 600 production deep-research traces $\delta = 0.34$ [0.30, 0.38], by three identifications that do not share a failure mode. At a hop where item boundaries come from the tool rather than a text heuristic, and where $b=1$ occurs 550 times, $C = 0.571$ [0.527, 0.615] is observed rather than extrapolated, over 16,082 hops. A tier also costs alignment: on 1,012 annotated multi-agent traces one brief in sixteen goes off-target, giving $\mu = 0.939$ and a per-tier penalty $C\mu = 0.536$. Depth is bought on two other axes. The root context is the only state that persists and the only one that cannot cheaply forget, and depth cuts its exposure from $N$ items to $N^{1/k}$. Depth is also cheaper: production flat agents bill as $N^{1.39}$, not the $N^2$ an append-only context predicts, and at equal spend two tiers overtake flat at 403 findings. Across every parameter we measured the model says 0.7% to 11.3% of production sessions are worth delegating, against 7.8% that do. A hazard model on 743,819 production tool calls finds that delegation does not respond to a filling context and is instead an opening move.

Despite the strategies of modern precision agriculture to leverage the integration of legacy agricultural systems, the challenges of IoT data fragmentation, farmers' sovereignty preservation, and limited interoperability still persist. This paper presents our work, conducted within the Horizon Europe DS2 project (DataSpace, DataShare 2.0), that applies an interoperability-oriented framework supporting participants of different agricultural data spaces to share data products and services under secure, sovereign, and transparent methods. The suggested methodology follows a layered reference architecture, where each layer consists of independent operational modules that facilitate the inter-sector data exchange between DigiAgro and AgroScience Data Spaces. The result of this work is an automated ecosystem for sharing diverse farm IoT measurements, satellite images and metrics, weather forecasts, and analytics services across distinct data spaces, aiming to generate accurate recommendations on crop practices, such as irrigation schedules, that farmers and agronomists will rely on to increase crop production while maintaining sustainability.

Reliable and total-order broadcasts in the Byzantine-fault model are well studied, but adding causal order has received comparatively little attention, largely due to the complexity that stems from actions of Byzantine processes. Existing solutions almost exclusively build causal ordering on top of total-order broadcast. The combination of causal order with reliable broadcast remains rare, and the few solutions that exist adopt the classical definition of causality based on events occurring at individual processes (the happened-before relation). This definition is not sufficient to enforce causal ordering among broadcast messages: Byzantine processes can lie about, omit, and forge dependency information and thereby violate the causal order among self-reported events. Such manipulations remain indistinguishable from correct behavior to any single observer. We demonstrate the issue and its consequences via a front-running attack. To close this gap, we extend the notion of reliable broadcast to externalize local potential knowledge. We use this to formalize the first complete definition of causal message ordering in reliable broadcast under Byzantine faults. Unlike the classical formalization, this notion is grounded in the joint observations of a sufficiently large group of correct processes rather than a process's own view. Building on this definition, we characterize the properties of a Byzantine reliable broadcast channel that guarantees causal ordering. We then present an efficient protocol that satisfies these properties: it is resilient to the optimal number of $f < n/3$ Byzantine faults and for one instance that broadcasts payload message~$m$, it has bit complexity $O(n^2(|m| + \lambda + n))$, where $\lambda$ denotes the maximal size of a unique (cryptographic) label for~$m$. Finally, we prove the protocol achieves Byzantine reliable broadcast with causal ordering.

Federated learning lets multiple participants train a shared model without pooling raw data, by exchanging locally trained model updates instead. Federated averaging assumes that averaging local models is a reasonable way to solve one shared problem when participants' data are broadly similar. Work on non-IID federated learning has shown that this assumption can withstand differences in label and feature distributions. We ask whether it survives a different strain specific to graph neural networks, where client graphs differ not in label or feature distribution but in structure itself, requiring the same shared weights to operate over fundamentally different topologies. We call the resulting harm structural negative transfer. In a federation of real citation networks and synthetic structural proxies, a structurally atypical client lost more than half its achievable accuracy simply by joining. In an initial six-client federation, two label-free structural statistics computable before training were strongly associated with this harm. Expanding to twenty clients showed that degree divergence remained associated with harm, although more weakly, and survived removal of domain contrast. Spectral divergence did not replicate, which we trace to a confound caused by the composition of the reference pool used for leave-one-out statistics. A causal intervention isolating topology found no significant effect. A degree-normalization mechanism held across twenty-four seeds but did not explain the harm when corrected. The best of five candidate fixes beat a tuned baseline only until a matched, structurally blind control was applied, after which the gain disappeared. What survives is a modest, partially replicated, degree-specific signal that is not yet a validated predictor at scale.

The contest in large-scale AI computing is no longer whether one processor can be made stronger, but whether a million of them can still behave as a single computer. This paper argues that two extensions are required. First, classic BSP, nested recursively, becomes a plan that holds at every scale: each superstep consists of parallel work, then exchange and aggregation, then a barrier, and at every layer the participating units are peers. Second, the von Neumann single-machine architecture, extended past its master–slave habit, becomes the Nested Parallel von Neumann Architecture: a nest of peer-equal layers, from chip package to autonomous zone, joined end to end by one memory-semantic interconnect, the Unified Bus. The software nest and the hardware nest correspond layer by layer, and the $\tau$ Scaling law folds time at every layer. We examine which workloads the nested structure serves, AI training above all but much of classic HPC as well, and describe the hardware decisions that make the nesting physical. Many processors, still one computer.

Random access into compressed data is normally bought with density. We measure the exchange rate. Across four formats and nine axes on a common corpus, the cost of cutting a 254 MB archive into independently addressable 16 KiB units is 1.632% of the archive for an absolute-offset format against 6.57% for seekable zstd, and the gap widens as the unit shrinks: at 4 KiB, 5.33% against 10.06%. Because the cost is small, several properties follow that are usually unavailable: splitting an archive is free and occasionally profitable (-0.28% on tiled input), append needs no format change, seek latency does not depend on position, and one archive is read by both a CPU and a GPU decoder. We give three structural results with proofs and bit-perfect verification - that the repeat-distance chain of an LZ77 parse forms a substitution monoid and is therefore prefix-scannable without touching the bitstream, that self-overlapping matches are periodic rather than chained, and that dependency depth admits an encoder-enforced bound - and we report each measured limit together with the mechanism that sets it. Seventeen rejected directions are listed with their numbers, including one that improved density by 26% and was declined. Every claim carries a level: reproducible by command, measured with a stated reason, or estimated. The measurement tool is released separately (DOI 10.5281/zenodo.22713364) with 435 provenanced records.

Multi-tenant GPU clusters frequently remain underutilized even when tenants experience long queueing delays, because quota control, queue ordering, preemption, and GPU sharing are driven by different local signals. We present DeepShare, a scheduler that uses a continuous tenant-assurance signal to coordinate these decisions at runtime. DeepShare combines elastic quota borrowing, tenant-specific runtime prediction, cost-aware best-effort preemption, and interference-aware MPS colocation, while using the same assurance signal to decide when borrowed capacity should be reclaimed and when sharing should become more conservative. In trace-driven experiments on 23,859 Venus jobs and 3,200 internal jobs, DeepShare achieves an average GPU utilization of 70.58%, a 29.5% improvement over the strongest non-intrusive sharing baseline, while reducing average queueing delay by 46%. On a 16-GPU Kubernetes testbed, it reduces the average job completion time by 34% and maintains 93% QoS compliance for guaranteed tenants. These results show that treating tenant assurance as a runtime control loop achieves a more advantageous utilization-QoS trade-off than optimizing quotas, scheduling, and resource sharing independently.

Message Passing Interface (MPI) has been the dominant programming model for High Performance Computing (HPC) for three decades, and as HPC workloads increasingly migrate to cloud infrastructure for scalability and cost efficiency, MPI applications must contend with an execution environment fundamentally unlike traditional supercomputers: ephemeral resources, dynamic pricing and preemptable instances. In such a volatile setting, the ability to relocate running MPI processes between nodes without restarting the job is a necessity for cost-effective, resilient execution. Existing approaches either require restarting the entire job from a global checkpoint, or transparently intercepting the full MPI stack at prohibitive complexity. To address these challenges, we propose \name, a cooperative migration system for MPI that enables selective process group migration on-the-fly. When a cloud instance is scheduled for preemption, only the affected ranks are relocated while the remaining processes briefly quiesce and resume in place, avoiding the cost of a full-job checkpoint. \name tackles this through a cooperative protocol between the MPI process management runtime and rank processes. We expose an XMPI\_quiesce interface built atop the MPI Sessions API that allows applications to mark safe migration points, and we extend the Hydra process manager to orchestrate the full migration lifecycle: rank quiescence, CRIU checkpoint/restore, proxy relaunch on the target node, and seamless rank reconnection. We evaluate and show that the cooperative quiesce phase accounts for less than 1.4% of total migration downtime, and that this downtime is governed by the migrating node's rank count alone, independent of job size, and the instrumentation introduces no measurable overhead during normal execution.

Hidden Semi-Markov Models (HSMMs) are fundamental probabilistic models widely adopted across diverse domains, from computational biology to finance and signal processing. The Viterbi algorithm decodes the most likely state sequence given an HSMM and can be applied iteratively for ab initio model learning. However, existing Viterbi implementations remain sequential, and GPU-accelerated solutions are entirely absent, making HSMM decoding impractical for large-scale workloads. We present a tensor-based formulation of the Viterbi algorithm for HSMMs, restructuring the inner loops into tensor operations that naturally map onto SIMD units and massively parallel architectures. Building on this formulation, we provide optimized implementations spanning single- and multi-core CPUs, and, for the first time, GPU. Experimental evaluation demonstrates speedups of up to 14x on a single core, over 200x with multi-core, and over 570x on GPU over the state-of-the-art sequential baseline, establishing a new performance baseline for large-scale HSMM decoding.

LLM agents execute long-horizon workflows where each model response determines the progress of subsequent tool interactions and environment transitions. Unlike chatbot serving, where TTFT and TPOT SLO constraints are critical, agentic workloads are completion-oriented and increasingly governed by job completion time (JCT). This shift challenges existing LLM serving designs optimized around token SLOs. Through a systematic exploration of scheduling and parallelism, we uncover a previously overlooked principle for agent serving: JCT is governed by the balance between prefill and decode efficiency. A prefill-prioritized scheduling policy achieves the best TTFT and the highest decode throughput, yet fails to attain the lowest JCT. This principle further reshapes the parallelism landscape: we show that pipeline parallelism (PP), long overlooked because it offers little decode-latency advantage, can reduce JCT by providing a more favorable balance between prefill and decode efficiency. Based on these insights, we build PipeSwift, an optimized open-source pipeline-parallel runtime integrated with a tailored micro-batch partitioning strategy co-designed with schedule considering the above trade-off, and pipeline-integrated multi-token prediction. Evaluated on real coding and web-search agent trajectories with two 360B+ MoE models on 64 H800 GPUs, PipeSwift reduces overall JCT by up to 1.45$\times$ over SGLang wide-EP, 2.33$\times$ over vLLM PP2, and 1.54$\times$ over today's state-of-the-art open-source PD-disaggregated deployment. PipeSwift is open-sourced at https://github.com/wsjdsg/PipeSwift.

Federated learning (FL) in healthcare faces pronounced heterogeneity and temporal concept drift across clinical centers, where evolving patient populations and care practices shift data distributions. Existing approaches rely on persistent global communication, incurring substantial bandwidth overhead while risking negative transfer from poorly aligned peers. We propose Adaptive Bayesian Partner Selection (ABPS), a peer-to-peer framework that governs who collaborates, when, and at what cost. Each center maintains a Beta-Bernoulli posterior over prospective peers' Shapley marginal utility, ranks candidates with an Upper Confidence Bound (UCB) criterion, and forms collaborations through a lightweight propose-reject mechanism, with the option to abstain from communication when no mutually beneficial partner exists. The framework admits a stochastic decision interpretation, yielding finite-sample concentration guarantees and O(kappa log T) regret in partner selection, along with conditions under which intentional isolation is optimal under negative transfer. Lightweight extensions (head personalization, bfloat16 quantized communication, and a tunable active-set size) further improve efficiency, and a goal-aware metadata filter enables institution-specific collaboration strategies. On binary in-hospital mortality prediction over the first 24 hours of an ICU stay, with 230 non-IID clinical centers drawn from MIMIC-IV, the full ABPS-X variant matches the strongest federated baseline (FedDyn, AUROC 0.758) at 0.09x the communication cost of FedAvg, with reduced variability. A diversity-driven configuration activates intentional isolation for a substantial fraction of centers. These results show that adaptive, utility-aware collaboration reduces communication without sacrificing accuracy when centers are numerous and small, offering a scalable paradigm for healthcare FL.

In agentic distributed systems, an agent may be authorized to mutate external infrastructure while lacking evidence that the mutation is ready to execute. Cognitive Admission Control (CAC) makes this evidence requirement explicit. A policy maps a typed action and its modeled risk to assurance obligations specifying predicates, evidence classes, scope, freshness, and witness-set constraints. A deterministic evaluator distinguishes satisfied, violated, and unresolved obligations; unresolved conditions produce targeted evidence-acquisition requests. Successful admission produces a certificate binding the action, its witness manifest, and dispatch-time guards. We formalize the admission calculus and the assumptions connecting it to mediated execution. The guarantees are policy-relative: physical safety additionally requires sound evidence, an adequate environment model, and preservation of relevant conditions through the effect. A TypeScript prototype is evaluated in 2,730 controlled local trials with independent effect observation and matched fault schedules. Across 390 CAC trials, 120 effects complete without modeled harm and no harmful effects occur. A live-policy baseline achieves the same completion count but admits the constructed correlated-witness failure. Mechanism ablations isolate guard, evidence-class, structural-cut, and remediation behavior. A further 9,000 measurements exercise the complete local dispatch path with persistent replay protection. These results establish tested implementation behaviors and local costs, not production failure rates or comparisons of language-model capability.

Filtered approximate nearest neighbor search, i.e. returning the top-k vectors nearest to a query vector among those satisfying one or more attribute predicates, has become a fundamental operation in modern vector search systems. Graph-based solutions employ beam search to solve a batch of queries in parallel for high throughput and employ high fixed beamwidth of 100 or greater for ensuring high recall. We observe that, given a batch of queries, more than half of the queries across multiple data sets can be solved precisely with a beamwidth of just 50 or less. Therefore, existing systems based on fixed high beamwidth sacrifice throughput to achieve high recall by forcing every query to search as thoroughly as the hardest query in the batch even though majority of queries can be resolved by a shallow search. In this paper we present a filtered ANNS engine for a single GPU named BOA that uses online beamwidth adaptation to customize the search effort across queries within a batch under multi-attribute range filters. We address the recall throughput tradeoff with a multi-phase search: all queries are first evaluated under a narrow beam, and only those with uncertain results are progressively refined with wider beamwidths. This renders recall largely insensitive to the starting beamwidth, whereas prior methods must use a fixed high beamwidth for high recall. BOA+ overlaps execution of phases to further enhance throughput. Our experiments show that, for 10,000 queries, online adaptation achieves 94.05% to 99.96% recall with average beamwidth ranging from 22 to 77, while a non-adaptive approach requires a fixed beamwidth of 500 to achieve similar or lower recall. Consequently, adaptivity increases throughput by 7x to 12.5x

Robust aggregation methods for federated learning quietly rest on a fragile assumption: that whoever shows up in a given round is a fair sample of the full population. In practice, they rarely are. When only a handful of clients participate per round, even a modest fraction of adversaries can dominate that sample and silently invalidate the finite-sample guarantees that coordinate-wise median, Krum, Bulyan, and trimmed mean all depend on. We introduce SWB-DM to address this directly. SWB treats each slice of a client update as a one-dimensional distribution, computes a trimmed Wasserstein barycenter across clients, and recovers coordinate identity via a medoid-based gauge-fixing step – a heuristic we developed and do not claim it belongs to standard optimal-transport theory. DeMoA-style delayed momentum then caches updates across the full client population each round, decoupling robustness from whoever happened to be sampled. Trim ratio calibration is not cosmetic: under-trimming causes collapse at corruption levels a properly calibrated model survives. Across 448 CIFAR-10 configurations, plus CIFAR-100, FEMNIST, and a 500-client scalability run, we find several mechanistically distinct failure modes. Even-sample coordinate-wise median degrades to a deterministic wrong answer. Krum silently violates its own n greater than 2f+2 precondition and diverges without warning. Bulyan's n greater than or equal to 4f+3 threshold produces a sharp pass/fail boundary. On attacks, IPM defeats order-statistic defenses – including SWB – more reliably than ALIE, confirmed through delta-space measurements against a convergence bound. SWB-DM's cache carries a real warm-up cost, but extending all baselines to the same round budget shows its CIFAR-10 gains are disproportionately large. On CIFAR-100, FLTrust benefits more – for reasons entirely unrelated to caching.

Modern AI and HPC systems integrate accelerators, high-speed networks, and management controllers at rack scale. Developing software for this infrastructure typically requires access to scarce, costly hardware, while software abstractions can obscure how workloads depend on resources across servers and accelerators. This paper presents Cnuas, an open-source, experimental rack-scale emulation platform whose baseline architecture follows the Open Compute Project (OCP) Open Rack v3 specifications. Through functional emulation, it supports experimentation, learning and software development within academic and industrial research and development, rather than matching the throughput or latency of physical hardware. Its web-based user interface visualizes racks, devices and their interconnections to help developers build a system-level mental model of the infrastructure supporting their workloads. At its core, CnuasNIC and CnuasSwitch implement a guest-visible remote direct memory access (RDMA) adapter and a host-resident hybrid software switch supporting both RoCEv2 and native InfiniBand. The platform also provides a dedicated AI/ML accelerator (GPU) peer fabric and OpenBMC-based rack management with executable power supply and battery backup firmware over RS-485. These components support the study of device, driver, and firmware interfaces on commodity hosts. The accelerator software stack remains an early research prototype, and facility modeling with OpenUSD is an exploratory extension. The paper presents the architecture, interfaces, and bounded prototype results as a basis for community collaboration across the core platform and its extensions.

This paper presents a systematic examination and experimental comparison of the prominent Federated Learning (FL) frameworks FedML, Flower, Substra, and OpenFL. The frameworks are evaluated experimentally by implementing Federated Learning over a varying number of clients, emphasizing a thorough analysis of scalability and key performance metrics. The study assesses the impact of increasing client counts on total training time, loss and accuracy values, and CPU and RAM usage. Results indicate distinct performance characteristics among the frameworks, with Flower displaying an unusually high loss, FedML achieving a notably low accuracy range of 66% to 79%, and Substra demonstrating good resource efficiency, albeit with an exponential growth in total training time. Notably, OpenFL emerges as the most scalable platform, demonstrating consistent accuracy, loss, and training time across different client counts. OpenFL's stable CPU and RAM underscore its reliability in real-world scenarios. This comprehensive analysis provides valuable insights into the relative performance of FL frameworks, offering good understanding of their capabilities and providing guidance for their effective deployment across diverse user bases.

Amid the rapid advancement of physical-world intelligence, cloud-edge collaborative large language models (LLMs) have emerged as a promising roadmap for practical LLM deployment. However, existing cloud-edge paradigms struggle to balance global consensus with local personalization, which fails to satisfy the need for a unified knowledge foundation on the cloud and domain-specific adaptation at the edge. To address this, we introduce, for the first time, a personalized bilevel optimization framework that formalizes cloud-edge LLM collaboration as a dual structure: the upper level optimizes edge-side personalization, while the lower level governs cloud-side knowledge transfer, reaching cloud-edge evolving in coordination. We then propose CIDERS, an efficient solver that decomposes the model into a learnable backbone and a messenger. While the cloud performs knowledge transfer to the learnable backbone, the key lies in embedding global trajectories into each local personalization step via consensus-variate correction to reconcile personalization with consensus. We provide a comprehensive theoretical analysis, including a geometric characterization of the local trajectory and a full convergence guarantee, revealing an explicit trade-off structure between personalization and global convergence. Extensive experiments demonstrate that CIDERS consistently outperforms competitive baselines on the compressed edge path, with 3.1x and 1.7x gains on mathematical reasoning and code generation, respectively, and a 10% relative gain on instruction metrics. Mechanism experiments attribute these gains to early consensus-corrected coordination and task-aware distillation. Overall, CIDERS offers a viable path toward consensus-guided continuous personalization in cloud-edge LLM systems.