Papers

We consider the dynamic scheduling problem in a multi-class M/G/N + M queue with convex delay costs and impatient customers that have exponential abandonment times. We apply the Whittle index approach to find a reasonable heuristic solution for this tricky problem. By assuming exponential abandonment times, we are able to make a very straightforward use of the results of the corresponding problem with patient customers presented in Queueing Systems, Vol. 110, Article no. 2, 2026. Our main theoretical achievements are proving that the closed version of the corresponding discrete-time problem is indexable and deriving an explicit expression for the Whittle index. The discrete-time results are utilized to develop the Whittle index policy for the original continuous-time scheduling problem.

Autoregressive image generation has emerged as a paradigm for multimodal AI systems due to its compatibility with transformer-based LLM serving infrastructures. However, generating thousands of visual tokens per request makes decoding increasingly bottlenecked by KV cache accesses during attention computation. Sparse attention is particularly attractive for this workload because many visual generation applications tolerate moderate quality degradation in exchange for improved performance and efficiency. While sparse attention has been extensively explored for text-based LLM inference, it remains unclear whether its sparsity assumptions generalize effectively to autoregressive image generation. We present the first systematic characterization of attention sparsity in autoregressive image generation across diverse workloads and representative open-source models. Our analysis reveals several distinguishing properties, including a pronounced prefill-decode asymmetry, strong attention concentration on prompt and local tokens, and a unique diagonal attention sparsity pattern arising from the spatial locality of visual tokens. Motivated by these observations, we propose a diagonal-aware sparse attention mechanism that selectively skips KV entries along the diagonal attention direction within a recent window. Implemented on top of a GPU-based serving system using FlexGen, FlashAttention-2, and custom kernels, our approach achieves up to 3.1x throughput and 1.19x latency improvements with less than 2% quality degradation compared to dense inference.

Repeated prompt prefixes are increasingly common in LLM serving workloads, appearing in system prompts, templated retrieval-augmented generation pipelines, agent frameworks, and multi-turn conversations. Modern inference runtimes such as vLLM and TensorRT-LLM provide mechanisms for reusing previously computed KV-cache state across requests, yet it remains unclear when prefix reuse materially improves serving performance on contemporary accelerators and when its benefits are limited by scheduling, cache granularity, concurrency, or memory pressure. This paper presents PrefixBench-H100, a reproducible benchmark and measurement framework for characterizing prefix reuse on a single NVIDIA H100. PrefixBench-H100 combines controlled synthetic traces with chat-style and retrieval-style workloads, and evaluates two widely used LLM serving runtimes under matched workload conditions. The benchmark varies shared-prefix length, suffix diversity, request arrival pattern, concurrency, output length, and cache configuration, while collecting time-to-first-token, inter-token latency, end-to-end latency, throughput, cache-hit statistics, GPU memory usage, and selected profiling traces. The goal of PrefixBench-H100 is not to introduce a new caching algorithm, but to expose the practical operating envelope of prefix reuse for H100-class LLM serving. The study identifies the regime where prefix reuse provides substantial first-token latency reductions and the regime where cache pressure erodes them, while showing that cache effectiveness itself is largely insensitive to concurrency and output length; the cross-runtime differences that remain arise above the cache, in the scheduling layer.

Hybrid terrestrial-satellite quantum networks combine terrestrial fiber infrastructure with free-space links to enable long-distance entanglement distribution. Entanglement can be routed over different combinations of terrestrial and satellite links, resulting in paths with different entanglement generation rates (EGRs) and fidelities. Existing routing algorithms, however, typically reduce routing to a single-objective problem by optimizing either EGR or fidelity, or by optimizing one while constraining the other, and thus do not explicitly capture the trade-off between the two. In this paper, we present a bi-objective routing framework for hybrid quantum networks that jointly optimizes end-to-end EGR and fidelity. We formulate routing as a Pareto optimization problem and show that it possesses a special mathematical structure: the bottleneck nature of end-to-end EGR and the multiplicative nature of end-to-end fidelity allow the problem to be transformed into a MAXMIN-MINSUM bi-criterion path problem. This transformation enables the exact polynomial-time computation of a minimal complete Pareto set using Martins' bicriterion routing algorithm. Our results demonstrate that hybrid networks expose substantially richer Pareto frontiers than terrestrial-only networks and that the proposed framework outperforms representative single-objective routing policies by allowing applications to select paths based on their EGR and fidelity requirements.

Large language model (LLM) inference deployed on Kubernetes faces an autoscaling challenge that conventional web services do not: new serving replicas take two to ten minutes to start because multi-gigabyte model weights must be loaded, which makes purely reactive scaling structurally late. We ask a sharp question: among the components of a predictive autoscaler, which ones actually matter under such long actuation delays? We answer it by decomposing predictive autoscaling into four factors—token-aware demand tracking, startup-delay lookahead, a bounded uncertainty margin, and plant-state observation—and measuring each factor's contribution in isolation on production-derived heavy-tailed workloads generated by ServeGen. Our main finding is that a simple exponentially weighted moving average (EWMA) predictor with delay-aware lookahead and an upper confidence bound (UCB) margin captures most of the benefit, reducing time-to-first-token (TTFT) service-level-objective (SLO) violations from 53% (reactive, queries-per-second based) to 0.5% across five random seeds; lookahead alone is the single largest factor, a 14$\times$ reduction. Kalman filter variants do not consistently improve the cost–SLO tradeoff. Controlled experiments isolate the reason token granularity is necessary: context length, through key–value cache pressure, degrades TTFT far more than request rate at matched throughput. Finally, a validation on a real Kubernetes cluster (Qwen2.5-7B, A100, vLLM) confirms the central mechanism: a delay-aware lookahead controller cuts TTFT violations from 63.5% to 3.7% relative to reactive KEDA scaling. We distinguish throughout which findings are specific to Kubernetes actuation and which are general to LLM serving.

Test-time scaling can improve large language model reasoning by generating and combining multiple candidate responses. In sampling-based methods, the inference budget is often described by the number of generated candidates, N. However, N tells us how many candidates are generated, not how they are executed. The same candidate budget can be produced in one batched generation call or split across several sequential calls with smaller batch sizes. We first study the effect of increasing N on reasoning accuracy using Phi-3-mini and Qwen2.5-1.5B on 500 GSM8K prompts. As expected, increasing N from 1 to 8 improves accuracy by 8.4 percentage points for Phi-3-mini and 18.4 points for Qwen2.5-1.5B. However, accuracy alone does not show the systems cost of using a larger candidate budget. We therefore fix N = 8 and compare four generation schedules: 1x8, 2x4, 4x2, and 8x1, where axb denotes a generation calls with b candidates per call. We measure latency, throughput, GPU-hours, and gross GPU-device energy while keeping the total candidate count fixed. On A100 GPUs, eight serial calls use 4.64-4.86x as much gross GPU-device energy and have 5.77-6.12x the P95 latency of one batched call with eight candidates. The same pattern appears across three independently scheduled A100 nodes per model and in short-output SciQ/V100 experiments. These results show that candidate count alone is not enough to describe the systems cost of multi-candidate test-time scaling. When candidates are independent and memory allows it, fewer generation calls with larger batch sizes are more efficient. Evaluations should therefore report not only candidate count and accuracy, but also generation schedule and GPU-level systems metrics.

Sparse tensor contractions are bottlenecked by sparse-sparse coiteration loops that resist standard loop vectorization. We present Splyce, an auto-vectorization framework in MLIR that overcomes this through a dual-path execution model. By decoupling coordinate intersection from pointer management via selective predication, Splyce inherently eliminates data-dependent branches as a side effect, allowing modern superscalar engines to maximize instruction-level parallelism and hide memory latency. Beyond simple branch elimination, our transformation exposes independent computation that can be executed concurrently, increasing functional-unit utilization that would otherwise be constrained by sequential dependencies. Evaluation across foundational sparse tensor kernels demonstrates performance ranging from 1.96X to 2.86X on synthetic inputs, with consistent speedups sustained across a vast majority of irregular real-world datasets from the SuiteSparse collection. Ultimately, Splyce demonstrates that by converting unpredictable control-flow into a predictable data stream, compiler-driven speculation can effectively reconcile the memory efficiency of compressed storage with the execution-unit throughput of modern superscalar architectures.

Analytical performance models — derivations of throughput or speedup from hardware parameters — make claims independently verifiable and expose binding constraints, yet rarely accompany architecture papers because building one by hand takes weeks of expert effort. We present Rosetta, a multi-agent LLM pipeline that automatically generates first-principles analytical models from research paper PDFs. Given a paper as sole input, Rosetta produces a mathematical specification, an executable Python model, and a plain-English interpretation — all autonomously, with zero human intervention. The formalization process itself is the primary value: it surfaces implicit assumptions and identifies missing parameters. Four design decisions address failure modes of naïve LLM-based generation: a scientific constitution that prohibits circular reasoning, verify-repair loops with independent critic agents, dual verification separating functional correctness from scientific validity, and a best-of-$N$ ensemble that exploits LLM stochasticity. We evaluate Rosetta across three complementary tracks: expert evaluation of 12 landmark papers (CS1), automated scoring of 97 unfiltered ISCA 2025 and HPCA 2026 papers (CS2), and author self-evaluation by six active research groups (CS3). Across CS1, specification quality scores 4–5/5 on 10 of 12 papers with zero significant hallucinations; across CS2, 56% of fit-screened papers reach Tier~A insight quality. The strongest finding comes from CS3: Rosetta's output led to revised claims and new experiments in active submissions, and five of six author-evaluators said they would use it again.

The generalized switch is a highly flexible queueing model, covering multiclass, multiserver, and multiresource queueing systems as well as a wide variety of stochastic networks. Although many scheduling policies have been developed for this model, they almost entirely address unknown job duration settings. How to optimally use known job durations in the generalized switch has remained open. Moreover, optimizing mean response time remains open in both settings. We introduce the first policy to guarantee heavy-traffic optimal mean response time in the generalized switch, our Smallest Equalizing Bucket (SEB) policy. The key challenge in designing an optimal scheduling policy is that we must simultaneously prioritize small jobs and also minimize resource waste, all while fitting within the generalized switch's service options. SEB overcomes this challenge by grouping jobs into duration-based buckets and enforcing an "equalizing" service structure that keeps each bucket balanced while still prioritizing the smallest jobs. We prove SEB's heavy-traffic optimality. Simulations further confirm the effectiveness of SEB-inspired heuristics.

Approximate multipliers can reduce hardware area and energy consumption in Deep Neural Network (DNN) inference; however, they introduce computational errors. Assessing the accuracy of numerous approximate multiplier designs across diverse DNN models and large-scale datasets remains challenging due to prohibitive evaluation times. This overhead primarily stems from the slow emulation of approximate multiplier behavior using look-up tables (LUTs) on CPU and GPU platforms. Moreover, the resulting accuracy degradation must be carefully quantified and, if necessary, mitigated (e.g., through retraining), further increasing the overall evaluation cost. To address these challenges, we propose FAME, an FPGA-based platform for evaluating approximate multipliers. The platform exploits the reconfigurable logic of Field-Programmable Gate Arrays (FPGAs) to implement approximate multipliers directly in hardware, eliminating the need for LUT-based emulation on CPU/GPU platforms and thereby enabling efficient DNN inference while significantly reducing evaluation time on large datasets. Furthermore, we introduce a pattern-guided DNN retraining technique to mitigate accuracy degradation induced by approximate multipliers. Specifically, retraining is guided by multiplier-specific patterns to effectively recover potential accuracy losses. We evaluate FAME using two DNN models, ResNet-18 and MobileNetV2, on the ImageNet dataset across 27 approximate multipliers. During inference, our approach achieves up to a 3.47x speedup in approximate multiplier evaluation compared to prior LUT-based emulation methods. Furthermore, the proposed retraining technique improves accuracy by up to 65.5% over existing retraining approaches for the evaluated multipliers. The code is publicly available at: https://github.com/gicLAB/FAME

Local agents need memory for model execution and working history. We present JustFit, an MLX runtime that coordinates their overlapping allocations: KVExec executes and checkpoints four-bit KV with bounded workspace, PhaseSwap loads phase-dependent components, and StateTrans preserves history across execution modes. With Qwen3.8-27B MXFP4 on a 24 GiB M4 Pro MacBook, JustFit reaches 327,680 retained positions across two requests, 10.67 times the evaluated baseline's 30,720-position single-request record. One request completes the full 262,144-position native window at median 5.986 tokens/s. Each shape completes 16,384 outputs per request in three fresh processes: B1 uses one cold build and two prefix extensions; B2 uses three ordered prefix extensions (Section 4). Image encoding can proceed while preserving a live 196,608-input text request. A controlled, repetitive 32K+6K workload reaches median 18.284 tokens/s at 15,626 MiB; a separate AIME 2026 evaluation scores 29/30. Coordinating execution and state lifetimes makes longer histories feasible on personal hardware.

Model serving is one of the largest cost drivers in production recommender systems. Maximizing its throughput requires navigating a deeply layered hierarchy: GPU kernels, the ML framework computation graph, the model server, and on-demand feature processing – each demanding specialized domain expertise. Such cross-layer expertise is inherently difficult to acquire, and does not scale with a workload that continuously grows and evolves, leaving significant cost efficiency gains unrealized. While recent AI agents have demonstrated human expert level efficiency in standalone GPU kernel optimization, automated tuning and optimization for the rest of the serving stack remain largely unexplored. We present FlashVector, an agentic system that optimizes performance across all layers of the model serving stack. The key contribution is an extensible framework to generalize the single kernel optimization agent paradigm to heterogeneous technical stacks, and to deliver performance improvements holistically. After deployment in Unity's Vector advertising platform, FlashVector achieved up to 2x throughput increase and up to 1.98x latency speedup on model server, and up to 1.6x throughput increase on feature store. These optimizations were discovered not only at the GPU kernel and computation graph levels, but also across the other components of the model serving stack, such as the model server (NVIDIA Triton's C++ codebase) and the on-demand feature transformation service (Python codebase), demonstrating the extensibility of the framework to more complex system architectures.

Software performance analysis and prediction requires integrating multiple signals, as code structure alone cannot capture runtime behavior shaped by execution frequency, resource contention, and I/O patterns. We present a critical-path-aware performance analysis methodology that automatically discovers recurring performance patterns by synthesizing static code features, dynamic execution traces, and kernel-level resource data. In a preliminary study across six real-world C/C++ applications (SQLite, OpenSSL, Zstandard, FFmpeg, cURL, and jq), we first empirically confirm that static complexity metrics explain only 10.4% of the variance ($\rho^2$) in critical path execution time, quantifying a gap that, while theoretically expected, had not been measured systematically across applications. Motivated by this finding, we analyze nearly 80,000 critical execution paths and address two research questions. First, we discover 13 distinct performance archetypes: recurring behavioral patterns that appear consistently across different applications, independent of their domain or implementation. Five of these patterns are near-universal and appear in at least five of the six applications studied. Notably, three of these archetypes are present in all six applications, and together, these common patterns account for 56.4% of all observed paths. Each archetype maps to specific resource profiles and optimization strategies that transfer across domains. Second, we leverage these archetypes within a multi-signal regression detection framework that triangulates path structure, resource consumption, and archetype deviations, achieving an F1-score of 0.867 and a 60.4% improvement over resource-only methods.

We extend the analysis of strong aggregation to general compatibility graphs, focusing on item counts rather than positions, and exploring generalized greedy matching disciplines. We prove that under a condition of automorphism-based transition consistency, the associated Markov chain is strongly aggregable for an arbitrary graph with a non-trivial automorphism group. Furthermore, we extend our analysis to non-greedy matching disciplines, distinguishing scenarios where compatible items can or cannot coexist within the same state. This result is illustrated with a simple compatibility graph with a rich automorphism structure: the odd rings. For all scenarios, we investigate the strong aggregation properties of the resulting Markov chains. This work enhances the theoretical understanding of lumpability in stochastic matching models and provides a foundation for analyzing complex graph structures.

Age of information (AoI) has become a standard design objective for timely monitoring as it measures the freshness of the latest update at a receiver. AoI alone, however, is insufficient in goal-oriented applications where decisions depend on consecutive observations. Age dispersion complements AoI by measuring the generation-time separation between consecutive updates. In this paper, we study joint freshness and dispersion control for a generate-at-will status update link modeled as a finite-state Markov wireless channel. The objective is to minimize the long-run probability that either AoI or age dispersion exceeds a prescribed threshold under an average transmission rate constraint. For channel-dependent randomized transmission policies, we derive exact matrix expressions for the stationary joint distribution of AoI and dispersion. For reversible channels, mean dispersion is lower bounded by the reciprocal of the delivery throughput, and the difference is an explicit non-negative variance term for which we establish the exact equality condition. We then formulate the adaptive joint-threshold control problem as a constrained Markov decision process and prove that it admits an exact finite-state representation whose size scales linearly with the number of channel states and the threshold levels. Simulation results show that the controller reduces joint threshold violations by 37.3% relative to adaptive AoI-only control under the same transmission budget, while increasing mean AoI by 8.3%.

Interocular differences in visual perception can induce a variety of effects when fused by the brain. For example, prior works have found that carefully crafted binocular differences in local detail can improve contrast. It has also been found that when the frequency content of two stimuli are slightly different, blur suppression leads to a fused percept that is typically dominated by the sharper image. In this paper, we develop a psychophysical framework to measure the perception of natural image stimuli with interocular frequency differences across the visual field. To this end, we study the effect of dichoptic foveation, which we define as the application of blur to one eye and a simultaneous sharpening filter to the other. Stimuli were viewed in a virtual reality (VR) head-mounted display (HMD) and placed at different retinal eccentricities. Study data were scaled to a perceptual just objectionable difference (JOD) scale, and a 4D model was fit to it; our results suggest that interocular frequency differences can be well described by a simple computational model. We applied the model in a realistic VR scenario with free exploration of 360{\deg} videos to improve a base foveated rendering system by enhancing high frequency information dichoptically.

As artificial intelligence's capabilities improve, it is increasingly viewed as a general scientific method. But how true are these claims? Does AI outperform all techniques, or only some, and how is this changing? To assess the claims, we assemble a corpus of 2,507 head-to-head comparisons between AI and other scientific analysis techniques across 27 scientific disciplines from papers published between 2000 and early 2025. We find a profound dichotomy. Relative to traditional statistics, AI often outperforms, but at a significantly higher computational cost. But there are also nearly a quarter of cases where AI is both more expensive and performs worse than traditional statistical techniques and this fraction has been stable for a decade. Relative to scientific computing, AI often underperforms, but at lower computational cost. This has begun to change: since 2020, AI's performance against scientific computing has notably strengthened and it now outperforms on more than half of comparisons. These patterns suggest that AI is therefore not a universal replacement for existing methods, but rather a valuable – and improving – part of a new AI-enabled scientific frontier.

Eight-bit integer (INT8) post-training quantization is the default recipe for edge deployment, under a widely held assumption: INT8 makes inference faster at a small, predictable accuracy cost, and a model quantized once can be carried to any target. We test that assumption with a controlled measurement study across seven hardware classes – ARM and x86 CPUs, a discrete GPU, an NVIDIA Jetson AGX Orin iGPU and its NVDLA cores, and two vendor NPUs (Qualcomm Hexagon HTP, DEEPX DX-M1) – holding the ONNX artifact and the quantization scales fixed so the integer kernel or ISA is the only free variable. Portability fails on three axes. (1) The sign of the INT8 speedup is set by the CPU's dot-product ISA (ARM dotprod/SDOT, x86 VNNI): cores that have it speed up by up to 2.1x, cores that lack it slow down by 1.7x, for the identical model and runtime. (2) INT8 outputs are not portable, and the rule is an invariance rather than a gradient: FP32 predictions are bit-identical for every pair (1000/1000), while INT8 predictions agree 1000/1000 exactly when two targets share an integer kernel and 958-965/1000 whenever they do not – independent of whether the boundary is CPU<->CPU or CPU<->accelerator, and invisible to top-1 accuracy, which is preserved. (3) Vendor NPUs own quantization: a bring-your-own QDQ graph fails silently on one NPU (external scales ignored, accuracy 0.75 -> 0.005 while it compiles, profiles and runs without error) and loudly on the other (the compiler refuses the graph), so only the vendor's native path yields a correct engine. We further show that edge-NPU latency regimes are set by output/device-to-host transfer size rather than compute, and locate the transition with a fixed-compute sweep. We release the scripts and 32 reports. "Quantize once, deploy anywhere" is unsafe for embedded and automotive deployment, where per-input determinism and redundancy matter.

As the complexity of High Performance Computing (HPC) ecosys- tems continually increases, achieving optimal performance becomes a challenge. Traditional performance autotuning techniques pro- vide promising means to navigate this complexity, these techniques remain computationally intensive and require many evaluations to find optimal configurations. This work proposes an autotuning framework that designs a machine learning-based ensemble LLVM Intermediate Representa- tion (IR) ranker, Neural Configuration Scorer (NCS). NCS ranks the performance of IRs sampled by a transfer-learning-based autotuner, improving the efficiency of the tuning process by reducing tuning overheads and circumventing subpar evaluations. By leveraging knowledge from related tasks, we are able to effectively exploit the transfer relationship to access high-performing configurations in fewer samples than traditional techniques that rely upon itera- tive refinement. Our framework can achieve similar performance improvements as state-of-the-art autotuning techniques with up to 61.67% fewer evaluations, averaging 27.85% fewer evaluations across various HPC benchmarks.

We present the enablement, correctness recovery, and performance engineering of DeepSeek-V4-Flash inference on AMD Instinct MI250 GPUs using the gfx90a/CDNA2 architecture. The system integrates native safetensors loading, tensor and expert parallelism, FP4 routed mixture-of-experts computation, FP8 dense projections, sparse attention, HIP graph execution, and OpenAI-compatible serving within SGLang. An initially fast execution path was found to be numerically incorrect because of a routed-expert W2 layout mismatch. We identify the output permutation, repair the weight layout at load time, and establish fixed-token and hash-based correctness checks before further optimization. On the corrected path, decode performance is improved through packed FP4 weights, INT8 activation quantization, CDNA2 dot-product instructions, peer-read all-reduce, and topology-aware kernel geometry. Prefill is accelerated using CDNA2 MFMA kernels, improved packed-weight reuse, reduced sparse-attention overhead, larger chunks, and retuned expert sorting. On four MI250 GCDs, TP4/EP1 native autoregressive decode reaches approximately 74.5 tok/s, while a 4,604-token prompt reaches 2.061-2.062 s TTFT, or approximately 2,234 input tok/s. The results show that efficient DeepSeek-V4-Flash inference on CDNA2 is limited not only by memory bandwidth, but also by FP4 execution-format mismatch, low-M utilization, and per-layer synchronization costs.

Managed language runtimes often provide concurrent garbage collectors so that latency-critical applications with large working sets can keep running while most collection work proceeds in the background. ZGC is a production-quality, generational, concurrent collector in OpenJDK with sub-millisecond pause times. While ZGC is designed to run concurrently, frequent and excessive collections with ZGC can still slow the mutators due to synchronization costs and interference with shared computing resources. Hence, the ZGC scheduler is conservative by default, and in most cases, will grow the heap toward the maximum allowed before scheduling a collection. While this approach minimizes collection effort, it can be wasteful, or even harmful, if the maximum heap size is not well tuned to the actual working set. We propose Opportunistic ZGC (OppZGC), a feedback-directed ZGC scheduling policy that constrains the heap dynamically and automatically, without per-application tuning. OppZGC identifies periods when CPU cores are underutilized and leverages them for concurrent collection with ZGC. We describe the design and implementation of OppZGC in OpenJDK's HotSpot Java VM and evaluate it with standard and latency-sensitive benchmarks from DaCapo Chopin and SPECjbb. OppZGC limits heap usage when there is CPU capacity sufficient for additional collections, and avoids scheduling extra collections when they would substantially degrade performance. Overall, it reduces maximum heap usage for our DaCapo benchmarks by between 61% and 90%, on average, depending on configuration, with minimal impact on throughput and request latency compared to default ZGC.

Adding answer options can lower multiple-choice scores without improving assessment validity. Turkish MMLU Pro examines this distinction using 12,000 Turkish-source questions across 58 sections. Each question retains its stem, five original options and source key, and receives five options copied from other questions in the same section. Sentence-embedding retrieval proposes candidates; a language model selects existing identifiers. Deterministic verification reconstructs all 60,000 additions. A 25-model calibration exposes scoring and generation-budget effects. Five evaluations produce source-key accuracies of 34.8%-81.4%. On 981 shared questions, one API-served model falls from 93.7% with five choices to 83.1% with ten; 102 of 115 lost correct responses select borrowed options. The decrease is 24.4 percentage points on heuristically flagged negative stems and 5.9 points elsewhere. A completed human-checked audit of 200 sampled questions, with undocumented reviewer tool use, yields 47 and 31 multiple-answer judgments across the two record sets, 25 of the latter unresolved. These records support concern about ambiguity, while their dependence and incomplete reviewer-method documentation limit validation. Because order and labels also change, the paired comparison measures augmentation as implemented. The contribution is a traceable construction and an analysis of its validity limits, not evidence that lower ten-choice scores measure knowledge better.

Large language model (LLM) inference in AI datacenters creates a coupled control problem between GPU serving and facility cooling. Raising ambient temperature setpoints can reduce cooling energy and carbon, but also shrinks thermal headroom, induces GPU throttling, and leads to Service-Level-Objective (SLO) violations. In this paper, we study joint cooling–computing control for LLM inference: minimizing per-job GPU-plus-cooling energy while satisfying thermal safety and latency SLO constraints. We present ETCInfer, an energy-efficient, thermal-aware scheduler that selects a pre-job Computer Room Air Conditioner (CRAC) setpoint and adapts per-GPU frequency and micro-batch size during execution. ETCInfer builds compact physics-informed control models by calibrating GPU heat generation, chassis heat dissipation, CRAC power, and prefill/decode latency relations from telemetry. These models estimate hidden thermal states and time-to-throttle, enabling the scheduler to evaluate energy, temperature, and latency before applying an action. We formulate this joint setpoint–frequency–micro-batch control problem as a partially observable Markov decision process and design ETCAdapter, a learning-based controller that minimizes per-job energy under thermal safety and SLO constraints. We implement ETCInfer as a coordination layer over typical inference and cluster management stacks. Evaluation across real-trace simulation and validation experiments shows that ETCInfer reduces total job energy by up to 33.1%, thermal throttle exposure by up to 92.9%, and keeps SLO violation rates below 0.7% even at ambient temperatures up to $48^{\circ}\mathrm{C}$.

External cache transfers can succeed while a hybrid language model resumes from an inconsistent state. We examine the full 45-layer GLM-5.3-Flash model, using the RedHatAI/ GLM-5.3-Flash-NVFP4 quantized checkpoint with vLLM and LMCache under four-way tensor parallelism. A complete-hit recovery mismatch restored state for the full prompt while the scheduler credited one fewer token. We aligned recovery through strict-prefix lookup and established a numerical comparison using shared computation corrections, matched checkpoint scheduling, and fixed per-rank kernel configurations. In a nine-length serial workload, agreement with the modified recomputation control improved from 34/36 to 36/36 generations, each containing 64 token IDs. A separate instrumented run passed recorded transfer-page, effective-tail, and delayed-save checks. Three additional synthetic templates passed 72 paired 256-token continuations across two fresh-container runs. A subsequent serial performance study preserved output equality across 120 requests; among the measured trials, CPU reload reduced time to first token by 46-64% and total request time by 1.9-7.0% relative to modified cold recomputation. The contribution is an experimentally validated integration repair applying an existing checkpoint-alignment principle. The evidence is confined to one model revision and controlled configuration; it does not establish general determinism, task-quality equivalence, concurrent-serving gains, or capacity beyond GPU memory.

Shared host-memory caching can avoid repeated prefill when a request moves between inference replicas. Its usefulness depends on both correct state transfer and lost prefix locality. We study two single-GPU 27B vLLM replicas sharing a 256 GiB LMCache pool. After adopting an existing packed-page patch, we isolate a raw-pointer fallback that omits the dependency on the current CUDA stream. Controlled byte tests fail under an imposed delay and pass when the dependency is restored; the existing mixed allocator provides a working deployment path. Full-pool allocation checks and service regression complete the validation. A four-block OFF-ON-ON-OFF comparison contains 768 measured requests within two block pairs. Median cross-replica time to first content token falls from 31.715 to 0.605 seconds at 128k input and from 92.047 to 0.790 seconds at 256k. Six-turn synthetic sessions alternating replicas improve by approximately 35% and 45% at initial contexts of 32k and 128k, while fixed placement shows little benefit. This engineering case study identifies practical validation steps and the locality conditions in which shared caching pays off.

We predict single-sequence model throughput from GGUF metadata using roofline-shaped predictors with quantization-specific scale factors fitted on reference models. The scored cohort comprises 318 phase-depth measurements from 53 host-file configurations on two Apple M4 Max systems and an NVIDIA RTX 5080. On host-specific held-out sets of four, five, and two configurations, an active-parameter decode model obtains 13.1%, 14.4%, and 36.1% mean absolute percentage error (MAPE), versus 49.4%, 55.3%, and 51.9% when charging total parameters. Leave-one-host-out coefficients fitted on the other two systems yield 11.6%, 16.8%, and 36.0% test MAPE. A low-bit model ladder changes ordering across runtime stacks. The P2 prefill baseline gives 18.7%, 22.2%, and 108.2% test MAPE. GGUF structure helps on all three systems, but fitted efficiencies are not universal.

Deep Learning (DL)-based channel estimation has shown high accuracy and low latency in terrestrial 5G NR, but Low Earth Orbit (LEO) Non-Terrestrial Networks (NTNs) introduce Doppler and synchronization impairments that may require NTN-specific architectures. We test whether High-Efficiency Learning-based channel Estimation using dual Neural Attention (HELENA), originally designed for terrestrial channels, remains effective after NTN retraining and suitable across high-performance and power-constrained inference platforms. Its unchanged architecture is trained on paired receiver-compensated (NTN-1) and residual-impaired (NTN-2) datasets and compared with eight terrestrial-origin models trained on the same NTN data and the NTN-specific MDELAN-SISO. HELENA achieves the lowest observed SNR-averaged NMSE among the DL estimators in both conditions, including 55.8-62.7% lower linear-scale NMSE than MDELAN-SISO. All DL models degrade in NTN-2, demonstrating the challenge posed by residual Doppler and its associated impairments. On an RTX PRO 4500, HELENA achieves 0.0595 ms 99th-percentile (P99) inference latency, 88.1% below the 0.5 ms budget, with lower energy than its closest attention-based competitors. On a 10 W Jetson Orin NX, it retains a favorable accuracy-energy trade-off, but no model meets the P99 budget. Thus, HELENA needs no NTN-specific redesign for the evaluated task, while embedded tail latency remains an open challenge.

Metastability—a phenomenon where systems remain trapped in quasi-stable states before abruptly transitioning under rare perturbations—is ubiquitous in physical systems. Although metastability is a widely observed phenomenon, its identification and analysis present significant challenges. To address these challenges, we propose a novel framework for analyzing metastability using Koopman theory. We use a finite set of system trajectories to learn a representation of the dynamics that defines a latent space in which the system evolves linearly, thereby enabling a systematic characterization of metastable behavior through the spectral properties of the linear mapping. Empirical evaluations demonstrate that our approach is capable of anticipating metastable behavior significantly earlier than its actual manifestation, even with $10\%$ of the simulation duration. Moreover, we establish that the dominant eigenvalue of the learned Koopman matrix in the latent space serves as a critical indicator for detecting metastability across both single-server and multi-server configurations.

Modern mobile inference runs on heterogeneous platforms combining mobile GPUs with multiple CPU core clusters. Existing optimizations typically exploit either inter-operator parallelism, by assigning entire operators to CPU cores or to the GPU, or intra-operator parallelism, by partitioning each operator for CPU-GPU co-execution. We consider these two forms of parallelism together, to improve inference latency of tasks that can be represented by a static DAG of operators with predefined input/output tensor shapes (e.g., CNNs or vision transformers). We define the problem of partition-aware DAG scheduling for mobile heterogeneous inference, illustrating that the best strategy depends on the structure of the inference DAG, thus motivating a joint formulation capturing operator partition choices, device assignment, and execution order. We propose an online iterative search framework, which decomposes large DAGs into stages, focuses search on critical operators, and uses latency predictors to estimate partitioned execution without exhaustive profiling. Across representative mobile inference workloads, our approach achieves latency close to an offline solution while keeping scheduling overhead to a fraction of the model initialization cost, allowing platform-specific scheduling at deployment time.

Predicting the runtime of noisy quantum circuit simulations is important for scheduling, resource allocation, and performance optimization. However, accurate prediction is challenging because backend-aware transpilation can substantially alter the original circuit structure, while the backend-derived noise model and simulator execution behavior can introduce additional runtime variation. We study the effectiveness of graph neural networks (GNNs) and conventional regression methods in predicting Qiskit Aer simulation runtime measured after transpilation. We construct a dataset from a benchmark pool of 1,402 unique circuits spanning 22 circuit families, two Qiskit fake-backend configurations, and four transpiler optimization levels. Specifically, we compare a source GNN using original circuit information, hybrid GNN combining source-level graph with post-transpilation features, and transpiled GNN using only transpiled circuit information, along with five regression models. In the overall-model setting, the transpiled GNN achieves the strongest performance among the graph-based representations at all four optimization levels, obtaining $R^2$ values of 0.974, 0.713, 0.745, and 0.605 for optimization levels 0 through 3, respectively. However, under backend-specific evaluation, the advantage of GNN decreases, with conventional regression models matching or outperforming the GNNs in several settings. These results indicate that post-transpilation information is useful, while the value of explicit graph modeling depends on the backend and optimization level.