Search arXiv⌕ Search

arXiv subjects

Dimitrios S. Nikolopoulos

Publications and source records attributed to Dimitrios S. Nikolopoulos.

At least 19 recordsLinked to original sources

Intent-Aware Caching for Efficient LLM Serving

Modern LLM serving systems use prefix caching to accelerate multi-turn conversational workloads. They cache the key-value states of previous conversation prefixes and reuse them when subsequent requests share the same prefix, to avoid redundant prefill computation and reduce serving latency. However, achieving a high cache hit rate remains challenging because cache reuse depends on whether a conversation continues and how soon the next turn arrives. Our analysis of real workloads shows that user intent is a useful signal for this behavior, as different intents exhibit different continuation probabilities and inter-turn delays. Based on this key insight, we propose Intent-Aware Caching (IAC), an online eviction policy that classifies requests by intent, learns lightweight reuse statistics from recent serving logs and leverages these statistics to guide an oracle-inspired eviction strategy. Extensive evaluation on real workloads (WildChat and ShareChat) shows that IAC reaches a 34.6% cache hit rate. This substantially outperforms both the previous learning-based approach (18.8%) and traditional methods such as least recently used (LRU, 12.2%), while reducing average serving latency by up to 46.1%.

cs.AI↗

Rethinking KV Cache Eviction for Agentic LLM Serving

Key-value (KV) caches impose substantial memory pressure on large language model serving systems, requiring cache eviction when capacity is insufficient. Existing mechanisms typically evict KV from the ends of cached sequences to preserve prefix sharing across requests. However, agent workloads repeatedly reuse long histories across successive turns. Recomputing the KV cache for tokens near the end of a sequence is substantially more expensive than for earlier tokens because their attention computation involves longer histories. To address this problem, we propose preserving shared prefixes and evicting the KV cache from the beginning of the remaining history while retaining the cached suffix. We build Penelope around this strategy by co-designing the eviction policy, KV recomputation, and unified GPU-CPU cache management. Returning requests reuse the KV cache retained in GPU or CPU memory and recompute only missing segments while preserving reuse of the cached suffix. We implement Penelope in SGLang and evaluate it on three Qwen3 model sizes using real agent traces. Compared with SGLang, Penelope reduces mean time to first token by up to 60% and job completion time by up to 37% with a GPU-only cache, and by up to 76% and 60% with a two-tier GPU-CPU cache.

cs.DC↗

Factorization-Guided Training and Inference for Diffusion-Drafted Speculative Decoding

Speculative decoding accelerates large language model inference through parallel verification of tokens proposed by a drafter. Diffusion drafters further parallelize proposal generation, but greater proposal capacity or higher mean accepted tokens (MAT) need not yield higher throughput. Because verification accepts only a consecutive prefix, early rejections can render later correct predictions ineffective, limiting MAT gains, while even higher MAT may not improve throughput if per-round latency increases. We introduce DScout, which uses two complementary factorizations to guide diffusion drafter training and inference for higher throughput. For training, DScout factorizes position-wise loss into inclusion, weighting, and prediction loss. This guides the design and combination of mixed-horizon training, intra-block unmasking, and acceptance-weighted loss to align supervision with prefix acceptance and improve MAT. For inference, DScout factorizes throughput as the ratio of MAT to per-round latency, guiding target-conditioning depth, drafter depth, and inference block size selection to identify a high-throughput configuration for each workload. Across six benchmarks, DScout achieves $3.84\times$ and $4.14\times$ mean speedups over autoregressive decoding on Qwen3-4B and Qwen3-8B, respectively, improving mean speedup over the state-of-the-art baseline by 9.1% and 7.1%. Across models and concurrency settings, our inference configuration policy provides up to 9.75% additional mean speedup over DScout's fixed configuration.

cs.LG↗

A Framework for Accelerating Transformer Inference on RISC-V for Edge AI

This work presents a framework for accelerating transformer-based language models (LMs) on resource-constrained IoT devices. The framework targets compact LMs: BERT-Tiny (B-Ty), MobileBERT (M-Bt), MiniLM (M-Lm), Electra (E-Lt) and DeBERTa (D-Bt) -- selected for their architectural diversity and use in edge inference scenarios. The proposed flow derives lightweight instruction set extensions tailored to the non-obvious computational patterns of these models. In addition, a custom instruction is introduced to accelerate the address generation stage of batch matrix multiplication, achieving a performance improvement of 15.39--21.74% with modest ASIC overheads of 6.79% in area and 2.33% in power. To further enhance performance without incurring additional processor core hardware cost, an optional compiler-directed loop unrolling strategy is employed, trading increased code size for overall reduced execution time. Evaluation on the Synopsys trv32p3f RISC-V core demonstrates inference speedups of up to 2.19x, and FPGA implementation on the AMD Zynq UltraScale+ ZCU102 shows a 32.94% area overhead at 75 MHz, whereas the ASIC implementation using the TSMC 28 nm library incurs a 21.08% area overhead while operating at 250 MHz.

cs.AR↗

ZTA-Q: an Open-source RISC-V Platform for Accurate Quantized CNN Inference

Low-precision inference is widely adopted in edge AI to reduce computational cost and memory footprint. However, existing open-source accelerator platforms provide limited end-to-end support for CNNs following the standard TensorFlow Lite integer inference scheme. This paper presents ZTA-Q, an open-source RISC-V-based platform that enables accurate deployment of TensorFlow Lite INT8 models. In addition to extending operator support, ZTA-Q provides a configurable post-processing datapath for studying how circuit-level approximations, including reduced multiplier precision, shared shift scaling, and simplified rounding, affect model accuracy. The proposed system is implemented on a Digilent Arty A7-100T FPGA and operates at 83.3 MHz. Evaluations on representative CNN models show that with LUT, register, and DSP overheads of 26.3%, 12.6%, and 150%, respectively, ZTA-Q limits the degradation in both top-1 and top-5 accuracy to within 0.25 percentage points.

cs.AR↗

XMPIaaS: Towards Cloud Native MPI via Cooperative Process Migration

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 \texttt{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.

cs.DC↗

Serving Masked Diffusion LLMs: Characterization and Design Principles from Real Hardware

Masked diffusion language models (dLLMs) can in principle generate text faster than autoregressive (AR) models, since they denoise many tokens at once. Recent systems have begun building serving infrastructure for dLLMs, but none first measure how these models behave under real, concurrent serving load. Serving systems built without this grounding risk carrying over assumptions from AR serving that may not hold for dLLMs. We characterize dLLM serving to close this gap, using LLaDA-8B-Instruct with a D2F (Discrete Diffusion Forcing) LoRA adapter on a single NVIDIA H200 GPU, evaluated on GSM8K and HumanEval. We report three findings. First, request difficulty, the number of denoising steps a request needs, is discrete rather than continuous: requests fall into 11 fixed step-count levels (178 + 29k), and no signal we test predicts the level before generation starts (best R2 = 0.150). Second, benchmarks with short generation budgets below 320 tokens understate serving variance, since requests are cut off before the latency spread appears. Third, only 24% of single-request wall-clock time is GPU computation; the rest is CPU-side dispatch overhead. Batching mainly helps by amortizing this overhead: sharing one forward pass per denoising step improves throughput by 16.0x at batch size 16 over a per-request-dispatch baseline. We also argue structurally that output quality should not degrade with batch size, stating three assumptions this rests on; we measure 74 to 76% GSM8K accuracy at single-request scale. Finally, we derive a batch-timeout rule for fixed-fill synchronized batching under Poisson arrivals. Together, these results show that serving diffusion language models needs parallelism at the level of each denoising step, which differs from AR serving in how admission and eviction interact with an already shared forward pass.

cs.AI↗

CAI-DLLM: Convergence Aware Inference for Diffusion Language Models

Diffusion language models can generate many tokens in parallel, but they still require repeated denoising steps during inference. This makes generation costly, especially when the model continues to recompute tokens that are already stable. To address these limitations, we propose CAI-DLLM, a training-free inference method that uses first-step confidence to guide denoising and reduce inference time. Specifically, CAI-DLLM commits easy tokens earlier, allocates more denoising steps to harder tokens, and adjusts decoding schedules across output blocks. As it relies only on first-step confidence signals, it does not require retraining, extra predictors, or weight updates. We evaluate CAI-DLLM on LLaDA-8B-Instruct and Dream-7B-Instruct across math, code, reasoning, commonsense, and long-context tasks. CAI-DLLM achieves up to 18.2x wall clock inference speedup on LLaDA GSM8K while improving accuracy from 76.27% to 77.41%, and up to 13.1x speedup on Dream HumanEval while achieving higher pass@1 than no-cache inference, 48.17% compared with 46.95%. On harder reasoning tasks, speedups reach 44.8x, with a largest accuracy drop of 4.4 points, while energy consumption is reduced by up to 95.3%.

cs.AI↗

FACT: Compositional Kernel Synthesis with a Three-Stage Agentic Workflow

Deep learning compilers and vendor libraries deliver strong baseline performance but their performance is bounded by finite, engineer-curated catalogs. When these omit needed optimizations, practitioners substitute hand-written CUDA or CUTLASS, demanding expertise in GPU microarchitecture and C++ template metaprogramming. Recent LLM-based agents target kernel generation in raw CUDA, forcing rediscovery of optimizations already encoded in mature libraries. We present FACT (Framework for Agentic CUTLASS Transpilation), a three-stage agent-driven workflow optimizing PyTorch modules through multi-pattern composition while grounding synthesis in CUTLASS C++. Pattern discovery inspects the traced graph, matches subgraphs to optimization rules, retrieves vetted examples, and outputs prioritized patterns. Pattern realization implements each pattern as a CUTLASS kernel, verifies, and auto-tunes. Pattern composition assembles extensions into an optimized module for benchmarking. We evaluate the workflow on KernelBench across NVIDIA A100 and H100 GPUs. On Level 1 GEMM problems (square, batched, large-K matrix multiply), auto-tuned CUTLASS kernels achieve 1.06x-1.18x speedups on A100 and 0.84x-1.80x performance variations on H100 over cuBLAS. On Level 3 transformer blocks against PyTorch eager baseline, FACT achieves 2.03x speedup on MiniGPT (vs. Inductor: 1.89x, TensorRT: 1.85x) and 1.41x on Llama 3 8B (vs. Inductor: 1.17x, TensorRT: 1.18x). Our framework couples agentic graph-level pattern discovery with architecture-specific auto-tuning and a dynamic pattern registry, offering a practical path from traced PyTorch modules to deployable kernels.

cs.DC↗

ConfigSpec: Profiling-Based Configuration Selection for Distributed Edge--Cloud Speculative LLM Serving

Speculative decoding enables collaborative Large Language Model (LLM) inference across cloud and edge by separating lightweight token drafting from heavyweight verification. While prior systems show performance and cost benefits, practical deployment requires navigating a large configuration space spanning draft model variants, quantisation levels, speculative lengths, and heterogeneous edge devices. This paper presents ConfigSpec, a configurationselection framework for distributed speculative LLM serving. ConfigSpec profiles edge devices and draft-target alignment, and models drafting throughput, acceptance rate, and power to evaluate goodput, verification cost efficiency, and energy efficiency across the joint configuration space. Our analysis across three edge platforms and two LLM families reveals structurally conflicting optima. Firstly, goodput is maximised by the smallest, fastest draft model at device-dependent speculative lengths (K*=2-10). Secondly, both cost and energy efficiency converge to K=2 due to a dominant bonus-token effect-with cost favouring the largest drafter for its high acceptance rate and energy favouring the smallest for its low power draw. These conflicts confirm that no single fixed configuration can simultaneously optimise all objectives, underscoring the need for profiling-based configuration selection in disaggregated edge-cloud LLM inference.

cs.DC↗

WISP: Waste- and Interference-Suppressed Distributed Speculative LLM Serving at the Edge via Dynamic Drafting and SLO-Aware Batching

As Large Language Models (LLMs) become increasingly accessible to end users, an ever-growing number of inference requests are initiated from edge devices and computed on centralized GPU clusters. However, the resulting exponential growth in computation workload is placing significant strain on data centers, while edge devices remain largely underutilized, leading to imbalanced workloads and resource inefficiency across the network. Integrating edge devices into the LLM inference process via speculative decoding helps balance the workload between the edge and the cloud, while maintaining lossless prediction accuracy. In this paper, we identify and formalize two critical bottlenecks that limit the efficiency and scalability of distributed speculative LLM serving: Wasted Drafting Time and Verification Interference. To address these challenges, we propose WISP, an efficient and SLO-aware distributed LLM inference system that consists of an intelligent speculation controller, a verification time estimator, and a verification batch scheduler. These components collaboratively enhance drafting efficiency and optimize verification request scheduling on the server. Extensive numerical results show that WISP improves system capacity by up to 2.1x and 4.1x, and increases system goodput by up to 1.94x and 3.7x, compared to centralized serving and SLED, respectively.

cs.DC↗

Modality Inflation: Energy Characterization and Optimization Opportunities for MLLM Inference

Multimodal large language models (MLLMs) are built on text-only LLMs by incorporating additional modalities, enabling multimodal understanding and a broader range of applications. However, these additions introduce a previously unexplored energy trade-off across modalities that remains poorly understood, as most prior work focuses on text-only models. In this paper, we examine modality inflation, a key source of inefficiency in which multimodal inputs increase inference workloads through extra encoding stages and expanded token sequences. We provide the first detailed, stage-level analysis of energy consumption in MLLM inference by breaking the pipeline into vision encoding, prefill, and decoding stages. Using four representative MLLMs evaluated on NVIDIA A100 GPU, we quantify the additional energy required for multimodal inference compared to text-only baselines, observing overheads ranging from 17% to 94% across models for identical inputs. Our results show that energy bottlenecks differ widely across model architectures, stemming either from compute-heavy vision encoders or from the downstream impact of large visual token sequences during prefill. By examining GPU power traces, we further uncover substantial GPU underutilization during multimodal execution and show that input complexity leads to markedly different energy scaling behaviors across models. Finally, we demonstrate that stage-wise dynamic voltage and frequency scaling (DVFS) is an effective optimization, allowing energy savings with only modest performance impact. Together, these findings offer practical insights and concrete guidance for designing more energy-efficient multimodal LLM serving systems.

cs.DC↗

APEX: Asynchronous Parallel CPU-GPU Execution for Online LLM Inference on Constrained GPUs

Deploying large language models (LLMs) for online inference is often constrained by limited GPU memory, particularly due to the growing KV cache during auto-regressive decoding. Hybrid GPU-CPU execution has emerged as a promising solution by offloading KV cache management and parts of attention computation to the CPU. However, a key bottleneck remains: existing schedulers fail to effectively overlap CPU-offloaded tasks with GPU execution during the latency-critical, bandwidth-bound decode phase. This particularly penalizes real-time, decode-heavy applications (e.g., chat, Chain-of-Thought reasoning) which are currently underserved by existing systems, especially under memory pressure typical of edge or low-cost deployments. We present APEX, a novel, profiling-informed scheduling strategy that maximizes CPU-GPU parallelism during hybrid LLM inference. Unlike systems relying on static rules or purely heuristic approaches, APEX dynamically dispatches compute across heterogeneous resources by predicting execution times of CPU and GPU subtasks to maximize overlap while avoiding scheduling overheads. We evaluate APEX on diverse workloads and GPU architectures (NVIDIA T4, A10), using LLaMa-2-7B and LLaMa-3.1-8B models. Compared to GPU-only schedulers like vLLM, APEX improves throughput by 84% - 96% on T4 and 11% - 89% on A10 GPUs, while preserving latency. Against the best existing hybrid schedulers, it delivers up to 72% (T4) and 37% (A10) higher throughput in long-output settings. APEX significantly advances hybrid LLM inference efficiency on such memory-constrained hardware and provides a blueprint for scheduling in heterogeneous AI systems, filling a critical gap for efficient real-time LLM applications.

cs.DC↗

Taming the Memory Footprint Crisis: System Design for Production Diffusion LLM Serving

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to Autoregressive Models (ARMs), utilizing parallel decoding to overcome sequential bottlenecks. However, existing research focuses primarily on kernel-level optimizations, lacking a holistic serving framework that addresses the unique memory dynamics of diffusion processes in production. We identify a critical "memory footprint crisis" specific to dLLMs, driven by monolithic logit tensors and the severe resource oscillation between compute-bound "Refresh" phases and bandwidth-bound "Reuse" phases. To bridge this gap, we present dLLM-Serve, an efficient dLLM serving system that co-optimizes memory footprint, computational scheduling, and generation quality. dLLM-Serve introduces Logit-Aware Activation Budgeting to decompose transient tensor peaks, a Phase-Multiplexed Scheduler to interleave heterogeneous request phases, and Head-Centric Sparse Attention to decouple logical sparsity from physical storage. We evaluate dLLM-Serve on diverse workloads (LiveBench, Burst, OSC) and GPUs (RTX 4090, L40S). Relative to the state-of-the-art baseline, dLLM-Serve improves throughput by 1.61$\times$-1.81$\times$ on the consumer-grade RTX 4090 and 1.60$\times$-1.74$\times$ on the server-grade NVIDIA L40S, while reducing tail latency by nearly 4$\times$ under heavy contention. dLLM-Serve establishes the first blueprint for scalable dLLM inference, converting theoretical algorithmic sparsity into tangible wall-clock acceleration across heterogeneous hardware. The code is available at https://github.com/chosen-ox/dLLM-Serve.

cs.DC↗

DiffPro: Joint Timestep and Layer-Wise Precision Optimization for Efficient Diffusion Inference

Diffusion models produce high quality images but inference is costly due to many denoising steps and heavy matrix operations. We present DiffPro, a post-training, hardware-faithful framework that works with the exact integer kernels used in deployment and jointly tunes timesteps and per-layer precision in Diffusion Transformers (DiTs) to reduce latency and memory without any training. DiffPro combines three parts: a manifold-aware sensitivity metric to allocate weight bits, dynamic activation quantization to stabilize activations across timesteps, and a budgeted timestep selector guided by teacher-student drift. In experiments DiffPro achieves up to 6.25x model compression, fifty percent fewer timesteps, and 2.8x faster inference with Delta FID <= 10 on standard benchmarks, demonstrating practical efficiency gains. DiffPro unifies step reduction and precision planning into a single budgeted deployable plan for real-time energy-aware diffusion inference.

cs.LG↗

MARVEL: An End-to-End Framework for Generating Model-Class Aware Custom RISC-V Extensions for Lightweight AI

Deploying deep neural networks (DNNs) on resource-constrained IoT devices remains a challenging problem, often requiring hardware modifications tailored to individual AI models. Existing accelerator-generation tools, such as AMD's FINN, do not adequately address extreme resource limitations faced by IoT endpoints operating in bare-metal environments without an operating system (OS). To overcome these constraints, we propose MARVEL-an automated, end-to-end framework that generates custom RISC-V ISA extensions tailored to specific DNN model classes, with a primary focus on convolutional neural networks (CNNs). The proposed method profiles high-level DNN representations in Python and generates an ISA-extended RISC-V core with associated compiler tools for efficient deployment. The flow leverages (1) Apache TVM for translating high-level Python-based DNN models into optimized C code, (2) Synopsys ASIP Designer for identifying compute-intensive kernels, modeling, and generating a custom RISC-V and (3) Xilinx Vivado for FPGA implementation. Beyond a model class specific RISC-V, our approach produces an optimized bare-metal C implementation, eliminating the need for an OS or extensive software dependencies. Unlike conventional deployment pipelines relying on TensorFlow/PyTorch runtimes, our solution enables seamless execution in highly resource-constrained environments. We evaluated the flow on popular DNN models such as LeNet-5*, MobileNetV1, ResNet50, VGG16, MobileNetV2 and DenseNet121 using the Synopsys trv32p3 RISC-V core as a baseline. Results show a 2x speedup in inference and upto 2x reduction in energy per inference at a 28.23% area overhead when implemented on an AMD Zynq UltraScale+ ZCU104 FPGA platform.

cs.AR↗

QPART: Adaptive Model Quantization and Dynamic Workload Balancing for Accuracy-aware Edge Inference

As machine learning inferences increasingly move to edge devices, adapting to diverse computational capabilities, hardware, and memory constraints becomes more critical. Instead of relying on a pre-trained model fixed for all future inference queries across diverse edge devices, we argue that planning an inference pattern with a request-specific model tailored to the device's computational capacity, accuracy requirements, and time constraints is more cost-efficient and robust to diverse scenarios. To this end, we propose an accuracy-aware and workload-balanced inference system that integrates joint model quantization and inference partitioning. In this approach, the server dynamically responds to inference queries by sending a quantized model and adaptively sharing the inference workload with the device. Meanwhile, the device's computational power, channel capacity, and accuracy requirements are considered when deciding. Furthermore, we introduce a new optimization framework for the inference system, incorporating joint model quantization and partitioning. Our approach optimizes layer-wise quantization bit width and partition points to minimize time consumption and cost while accounting for varying accuracy requirements of tasks through an accuracy degradation metric in our optimization model. To our knowledge, this work represents the first exploration of optimizing quantization layer-wise bit-width in the inference serving system, by introducing theoretical measurement of accuracy degradation. Simulation results demonstrate a substantial reduction in overall time and power consumption, with computation payloads decreasing by over 80% and accuracy degradation kept below 1%.

cs.DC↗

MARCO: Multi-Agent Code Optimization with Real-Time Knowledge Integration for High-Performance Computing

Large language models (LLMs) have transformed software development through code generation capabilities, yet their effectiveness for high-performance computing (HPC) remains limited. HPC code requires specialized optimizations for parallelism, memory efficiency, and architecture-specific considerations that general-purpose LLMs often overlook. We present MARCO (Multi-Agent Reactive Code Optimizer), a novel framework that enhances LLM-generated code for HPC through a specialized multi-agent architecture. MARCO employs separate agents for code generation and performance evaluation, connected by a feedback loop that progressively refines optimizations. A key innovation is MARCO's web-search component that retrieves real-time optimization techniques from recent conference proceedings and research publications, bridging the knowledge gap in pre-trained LLMs. Our extensive evaluation on the LeetCode 75 problem set demonstrates that MARCO achieves a 14.6\% average runtime reduction compared to Claude 3.5 Sonnet alone, while the integration of the web-search component yields a 30.9\% performance improvement over the base MARCO system. These results highlight the potential of multi-agent systems to address the specialized requirements of high-performance code generation, offering a cost-effective alternative to domain-specific model fine-tuning.

cs.DC↗