LLM Serving

LLM: Large Language Model

Ask about this topic

Overview

Large language model serving studies how to run inference efficiently for many concurrent requests, balancing response speed, throughput, cost, and service-level objectives. Common approaches include continuous batching and prefill–decode scheduling, request routing, and managing or compressing KV caches; newer systems also use disaggregated hardware, speculative decoding, and workload-aware power control. Evaluations use request traces and production-like workloads to measure time to first token, token latency, throughput, quality, and resource use across models and hardware. A central challenge is that workloads and output lengths vary, so policies that improve throughput or cost can worsen tail latency, fairness, or response quality.

Latest papers

CardsList

216 papers31 papers in the last four weeks, up 94% on the four weeks before. 0.2% of all new papers.

Apr 10, 2026

LLM-Rosetta: A Hub-and-Spoke Intermediate Representation for Cross-Provider LLM API Translation

The rapid proliferation of Large Language Model (LLM) providers--each exposing proprietary API formats--has created a fragmented ecosystem where applications become tightly coupled to individual vendors. Switching or bridging providers requires O(N2)O(N^2) bilateral adapters, impeding portability and multi-provider architectures. We observe that despite substantial syntactic divergence, the major LLM APIs share a common semantic core: the practical challenge is the combinatorial surface of syntactic variations, not deep semantic incompatibility. Based on this finding, we present LLM-Rosetta, an open-source translation framework built on a hub-and-spoke Intermediate Representation (IR) that captures the shared semantic core--messages, content parts, tool calls, reasoning traces, and generation controls--in a 9-type content model and 10-type stream event schema. A modular Ops-composition converter architecture enables each API standard to be added independently. LLM-Rosetta supports bidirectional conversion (provider-to-IR-to-provider) for both request and response payloads, including chunk-level streaming with stateful context management. We implement converters for four API standards (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google GenAI), covering the vast majority of commercial providers. Empirical evaluation demonstrates lossless round-trip fidelity, correct streaming behavior, and sub-100 microsecond conversion overhead--competitive with LiteLLM's single-pass approach while providing bidirectionality and provider neutrality. LLM-Rosetta passes the Open Responses compliance suite and is deployed in production at Argonne National Laboratory. Code is available at https://github.com/Oaklight/llm-rosetta.

Peng Ding

University of Chicago

LLM Serving
AskPDF
Apr 9, 2026

Dual-Pool Token-Budget Routing for Cost-Efficient and Reliable LLM Serving

Production vLLM fleets typically provision each instance for the worst-case context length, leading to substantial KV-cache over-allocation and under-utilized concurrency. In practice, 80-95% of requests are short, yet are served under configurations optimized for long contexts, wasting 4-8×\times throughput capacity and triggering reliability issues such as OOM crashes, preemption, and request rejections. We identify a common root cause for these inefficiencies: configuration-traffic mismatch. We propose dual-pool token-budget routing, a lightweight dispatch mechanism that partitions a homogeneous fleet into two specialized pools: a high-throughput short-context pool and a high-capacity long-context pool. Each request is routed based on its estimated total token budget, computed using a per-category bytes-to-token ratio that is learned online via exponential moving average from usage.prompt_tokens feedback, eliminating the need for a tokenizer. We also develop a simple analytical model that predicts fleet-level cost savings from workload characteristics and measured throughput differences, enabling practitioners to estimate benefits prior to deployment. Evaluations on real-world traces from the Azure LLM Inference Dataset and LMSYS-Chat-1M, serving Llama-3-70B on A100 GPUs, show that our approach reduces GPU-hours by 31-42%, corresponding to $2.86M annual savings at fleet scale, while lowering preemption rates by 5.4×\times and improving P99 TTFT by 6%. A case study with Qwen3-235B-A22B on AMD MI300X at 10,000 req/s projects $15.4M in annual savings. The method incurs only O(1) dispatch overhead, adapts automatically to heterogeneous workloads, and composes seamlessly with existing optimizations such as PagedAttention, continuous batching, and prefill-decode disaggregation.

Xunzhuo Liu, Bowei He, Xue Liu +3

vLLM Semantic Router Project · MBZUAI · McGill University +2

LLM ServingLLM Inference EfficiencyLLM Inference SchedulingLLM Routing
AskPDF
Apr 9, 2026

Robust Length Prediction: A Perspective from Heavy-Tailed Prompt-Conditioned Distributions

Output-length prediction is important for efficient LLM serving, as it directly affects batching, memory reservation, and scheduling. For prompt-only length prediction, most existing methods use a one-shot sampled length as the label, implicitly treating each prompt as if it had one true target length. We show that this is unreliable: even under a fixed model and decoding setup, the same prompt induces a prompt-conditioned output length distribution, not a deterministic scalar, and this distribution is consistent with heavy-tailed behavior. Motivated by this, we cast length prediction as robust estimation from heavy-tailed prompt-conditioned length distributions. We propose prompt-conditioned length distribution (ProD) methods, which construct training targets from multiple independent generations of the same prompt. Two variants are developed to reuse the served LLM's hidden states: \mbox{ProD-M}, which uses a median-based target for robust point prediction, and ProD-D, which uses a distributional target that preserves prompt-conditioned uncertainty. We provide theoretical justifications by analyzing the estimation error under a surrogate model. Experiments across diverse scenarios show consistent gains in prediction quality.

Jing Wang, Yu-Yang Qian, Ke Xue +3

State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing 210023, China · School of Artificial Intelligence, Nanjing University, Nanjing 210023, China

LLM ServingRobust Estimation
AskPDF
Apr 8, 2026

Blink: CPU-Free LLM Inference by Delegating the Serving Stack to GPU and SmartNIC

Large Language Model (LLM) inference is rapidly becoming a core datacenter service, yet current serving stacks keep the host CPU on the critical path for orchestration and token-level control. This makes LLM performance sensitive to CPU interference, undermining application colocation and forcing operators to reserve CPU headroom, leaving substantial capacity unutilized. We introduce Blink, an end-to-end serving architecture that removes the host CPU from the steady-state inference path by redistributing responsibilities across a SmartNIC and a GPU. Blink offloads request handling to the SmartNIC, which delivers inputs directly into GPU memory via RDMA, and replaces host-driven scheduling with a persistent GPU kernel that performs batching, scheduling, and KV-cache management without CPU involvement. Evaluated against TensorRT-LLM, vLLM, and SGLang, Blink outperforms all baselines even in isolation, reducing pre-saturation P99 TTFT by up to 8.47×\times and P99 TPOT by up to 3.40×\times, improving decode throughput by up to 2.1×\times, and reducing energy per token by up to 48.6%\%. Under CPU interference, Blink maintains stable performance, while existing systems degrade by up to two orders of magnitude.

Mohammad Siavashi, Mariano Scazzariello, Gerald Q. Maguire +2

KTH Royal Institute of Technology, Stockholm, Sweden · RISE, Stockholm, Sweden

LLM ServingLLM InferenceLLM Inference EfficiencyLLM Inference Scheduling
AskPDF
Apr 8, 2026

Scalable Joint Resource Allocation for SLO-Constrained LLM Inference in Heterogeneous GPU Clouds

Serving large language model (LLM) inference in cloud environments requires jointly optimizing model selection, GPU provisioning, parallelism configuration, and workload routing under latency, accuracy, memory, and budget constraints. While mixed-integer linear programming (MILP) can model this problem, its computational cost limits frequent re-optimization under demand variability. Existing heuristics often optimize individual components separately and may become infeasible when system-wide constraints are enforced. This paper presents a scalable framework for SLO-constrained LLM inference. We formulate the problem as an MILP with a two-phase delay model capturing both prefill and autoregressive decoding under tensor and pipeline parallelism. To solve it efficiently, we develop two constraint-aware heuristics: a Greedy Heuristic (GH) and an Adaptive Greedy Heuristic (AGH). AGH extends GH through multi-start construction, local search, and GPU consolidation. Both methods maintain feasibility through parallelism-aware filtering, cost-based ranking, and adaptive parallelism scaling. Experiments based on the Azure LLM Inference Trace show that GH generates feasible solutions within one second, while AGH achieves near-optimal performance within three seconds and scales to large instances where exact solvers fail to converge. Under out-of-sample stress with up to 1.5x delay and accuracy inflation, AGH degrades gracefully through provisioned headroom, yielding substantially lower cost and SLO violations than cost-minimal MILP solutions. Across synthetic and real Azure workloads, AGH maintains SLO compliance at significantly lower cost than exact MILP solutions. These results demonstrate that high-quality allocations provide substantial robustness to demand variability while enabling rapid adaptation to workload changes.

Jiaming Cheng, Duong Tung Nguyen

The Ira A. Fulton Schools of Engineering, Arizona State University, Tempe, AZ, USA.

LLM ServingLLM Inference SchedulingLLM Inference ServingCost-Aware Inference
AskPDF
Apr 8, 2026

Foundry: Template-Based CUDA Graph Context Materialization for Fast LLM Serving Cold Start

Modern LLM service providers increasingly rely on autoscaling and parallelism reconfiguration to respond to rapidly changing workloads, but cold-start latency remains a major bottleneck. While recent systems have reduced model weight loading to seconds, CUDA graph capture still takes tens of seconds to minutes and often dominates startup. Unfortunately, CUDA graphs cannot be naively serialized: beyond graph topology, they are tightly coupled to execution context, including device addresses embedded in kernel arguments and kernel code lazily loaded during warmup. Existing approaches either rely on brittle kernel-specific patching or heavyweight process-level checkpoint/restore that are inflexible to dynamic parallelism switching. We present Foundry, a template-based CUDA graph context materialization system that persists both graph topology and execution context during an offline processing stage, and reconstructs executable graphs online with negligible overhead. Foundry enforces deterministic memory layouts, automatically extracts and reloads kernel binaries required by captured graphs, and reduces online reconstruction costs through topology-based templating. For distributed serving, Foundry further enables a single-GPU offline capture to generate templates for multi-GPU deployments by patching only rank-dependent communication state. Across dense and MoE models up to 235B parameters, Foundry reduces cold-start latency by up to 99%, cutting the initialization time of Qwen3-235B-A22B from 10 minutes to 3.9 seconds while preserving the throughput gains of CUDA graphs.

Xueshen Liu, Yongji Wu, Yuncheng Yao +3

University of Michigan · UC Berkeley · Duke

LLM ServingLLM Inference Acceleration
AskPDF
Apr 7, 2026

ForkKV: Scaling Multi-LoRA Agent Serving via Copy-on-Write Disaggregated KV Cache

The serving paradigm of large language models (LLMs) is rapidly shifting towards complex multi-agent workflows where specialized agents collaborate over massive shared contexts. While Low-Rank Adaptation (LoRA) enables the efficient co-hosting of these specialized agents on a single base model, it introduces a critical memory footprint bottleneck during serving. Specifically, unique LoRA activations cause Key-Value (KV) cache divergence across agents, rendering traditional prefix caching ineffective for shared contexts. This forces redundant KV cache maintenance, rapidly saturating GPU capacity and degrading throughput. To address this challenge, we introduce ForkKV, a serving system for multi-LoRA agent workflows centered around a novel memory management paradigm in OS: fork with copy-on-write (CoW). By exploiting the structural properties of LoRA, ForkKV physically decouples the KV cache into a massive shared component (analogous to the parent process's memory pages) and lightweight agent-specific components (the child process's pages). To support this mechanism, we propose a DualRadixTree architecture that allows newly forked agents to inherit the massive shared cache and apply CoW semantics for their lightweight unique cache. Furthermore, to guarantee efficient execution, we design ResidualAttention, a specialized kernel that reconstructs the disaggregated KV cache directly within on-chip SRAM. Comprehensive evaluations across diverse language models and practical datasets of different tasks demonstrate that ForkKV achieves up to 3.0x the throughput of state-of-the-art multi-LoRA serving systems with a negligible impact on generation quality.

Shao Wang, Rui Ren, Lin Gui

Shanghai Jiao Tong University, Shanghai, China

LLM ServingLLM Inference AccelerationDisaggregated LLM ServingKV-Cache Reuse
AskPDF
Mar 31, 2026

ParetoBandit: Budget-Paced Adaptive Routing for Non-Stationary LLM Serving

Multi-model LLM serving operates in a non-stationary, noisy environment: providers revise pricing, model quality can shift or regress without notice, and new models arrive regularly. More than a dozen recent methods have proposed learned routers to navigate the resulting quality--cost tradeoff across portfolios spanning a ∼\sim530×\times cost range. Despite this activity, two gaps in the current solution space limit routing effectiveness under these conditions: no existing router enforces a dollar-denominated cost ceiling in closed loop over an open-ended request stream, and none provides principled online adaptation to post-deployment shifts in pricing or model quality. We present ParetoBandit, an open-source adaptive router built on cost-aware contextual bandits that addresses both gaps. Its core contributions are: (1) an online primal--dual budget pacer that enforces a per-request cost ceiling without a known horizon, and (2) geometric forgetting on sufficient statistics that gives the bandit bounded memory for tracking quality and cost shifts. A hot-swap model registry further supports runtime model changes with budget-controlled exploration. On 1,824 benchmark prompts with a three-model portfolio, the router maintains budget compliance within 0.4%, adapts to price and quality shifts with up to +0.071 quality lift, and integrates a cold-started model within ∼\sim142 steps.

Annette Taberner-Miller

Independent Researcher.

Adaptive Model RoutingLLM RoutingLLM ServingNon-Stationary BanditsContextual Bandits
AskPDF
Mar 27, 2026

Rocks, Pebbles and Sand: Modality-aware Scheduling for Multimodal Large Language Model Inference

Multimodal Large Language Models (MLLMs) power platforms like ChatGPT, Gemini, and Copilot, enabling rich interactions across text, images, and videos. Unlike text-only workloads, multimodal requests are inherently heterogeneous, differing by orders of magnitude in latency and memory footprint. Existing LLM serving systems, designed for homogeneous text workloads, struggle under this heterogeneity: large image and video requests monopolize GPU resources, causing severe head-of-line blocking and degrading the responsiveness of lightweight, latency-critical text queries. Recent multimodal serving systems mitigate this interference through disaggregation, isolating encoding and inference stages across GPUs. While effective at large scale, such designs introduce severe GPU underutilization in more practical, moderate-scale deployments. In response, we build RPS-Serve a modality-aware scheduling framework that leverages the extreme heterogeneity in multimodal workloads and introduces a simple abstraction: videos behave like rocks, images like pebbles, and text like sand. Then, it lets sand flow quickly through pebbles and rocks, maximizing GPU utilization while ensuring fast responsiveness for lightweight requests and controlled execution for heavier ones. To achieve this, RPS-Serve estimates the GPU admission cost of each request, classifies requests into rocks, pebbles, and sand using lightweight clustering, and applies priority-based scheduling with aging to balance performance across modalities. Evaluation across state-of-the-art MLLMs shows that RPS-Serve increases GPU utilization by 65% and reduces TTFT by up to 9.3x, on average across models, compared to disaggregation-based multi-modal systems.

Konstantinos Papaioannou, Thaleia Dimitra Doudali

IMDEA Software Institute, Universidad Politécnica de Madrid

LLM ServingLLM Inference SchedulingEfficient Multimodal InferenceMultimodal Large Language Models
AskPDF
Mar 26, 2026

Runtime Burden Allocation for Structured LLM Routing in Agentic Expert Systems: A Full-Factorial Cross-Backend Methodology

Structured LLM routing is often treated as a prompt-engineering problem. We argue that it is, more fundamentally, a systems-level burden-allocation problem. As large language models (LLMs) become core control components in agentic AI systems, reliable structured routing must balance correctness, latency, and implementation cost under real deployment constraints. We show that this balance is shaped not only by prompts or schemas, but also by how structural work is allocated across the generation stack: whether output structure is emitted directly by the model, compressed during transport, or reconstructed locally after generation. We evaluate this formulation through a comprehensive full-factorial benchmark covering 48 deployment configurations and 15,552 requests across OpenAI, Gemini, and Llama backends. Our central finding is consequential: there is no universal best routing mode. Instead, backend-specific interaction effects dominate performance. Modes that remain highly reliable on Gemini and OpenAI can suffer substantial correctness degradation on Llama, while efficiency gains from compressed realization are strongly backend-dependent. Rather than presenting another isolated model comparison, this work contributes a deployable framework for reasoning about structured routing under heterogeneous backend conditions. We provide a cross-backend evaluation methodology and practical deployment guidance for navigating the correctness-cost-latency frontier in production-grade agentic expert systems.

Zhou Hanlin, Chan Huah Yong

School of Computer Sciences, Universiti Sains Malaysia (USM), Penang, Malaysia · Xiamen Software Vocational & Technical College, Xiamen, China

LLM RoutingLLM EvaluationLLM Inference EfficiencyLLM Serving
AskPDF
Mar 26, 2026

GhostServe: A Lightweight Checkpointing System in the Shadow for Fault-Tolerant LLM Serving

The rise of million-token, agent-based applications has placed unprecedented demands on large language model (LLM) inference services. The long-running nature of these tasks increases their susceptibility to hardware and software faults, leading to costly job failures, wasted resources, and degraded user experience. The stateful key-value (KV) cache, which grows with the sequence length, presents a central challenge as it is a critical and vulnerable component in distributed serving systems. In this work, we propose GhostServe, a novel checkpointing solution to facilitate fault-tolerant LLM serving. Specifically, GhostServe protects the streaming KV cache in the shadow by applying erasure coding to generate and store the parity shards in host memory. In the event of device failures, GhostServe enables fast reconstruction of the lost KV cache, allowing the inference process to resume seamlessly without costly full recomputation or state replication. Evaluations demonstrate that GhostServe reduces checkpointing latency by up to 2.7x and recovery latency by 2.1x for a single batch, and 1.2x median response latency compared to existing methods, in the presence of system failures, paving the way for high-availability and cost-effective LLM serving at scale.

Shakya Jayakody, Youpeng Zhao, Chinmay Dhanraj Nehate +1

University of Central Florida, Orlando, FL, USA.

LLM ServingLLM InferenceKV-Cache Management
AskPDF
Mar 25, 2026

Secure On-Premise Deployment of Open-Weights Large Language Models in Radiology: An Isolation-First Architecture with Prospective Pilot Evaluation

Purpose: To design, implement, evaluate, and report on the regulatory requirements of a self-hosted LLM infrastructure for radiology adhering to the principle of least privilege, emphasizing technical feasibility, network isolation, and clinical utility. Materials and Methods: The isolation-first, containerized LLM inference stack relies on strict network segmentation, host-enforced egress filtering, and active isolation monitoring preventing unauthorized external connectivity. An accompanying deployment package provides automated isolation and hardening tests. The system served the open-weights DeepSeek-R1 model via vLLM. In a one-week pilot phase, 22 residents and radiologists were free to use 10 predefined prompt-templates whenever they considered them useful in daily work. Afterward, they rated clinical utility and system stability on an 0-10 Likert scale and reported observed critical errors in model output. Results: The applied institutional governance pathway achieved approval from clinic management, compliance, data protection and information security officers for processing unanonymized PHI. The system was rated stable and user friendly during the pilot. Source text-anchored tasks, such as report corrections or simplifications, and radiology guideline recommendations received the highest utility ratings, whereas open-ended conclusion generation based on findings resulted in the highest frequency of critical errors, such as clinically relevant hallucinations or omissions. Conclusion: The proposed isolation-first on-premise architecture enabled overcoming regulatory borders, showed promising clinical utility in text-anchored tasks and is the current base to serve open-weights LLMs as an official service of a German University Hospital with over 10,000 employees. The deployment package were made publicly available (https://github.com/ukbonn/ukb-gpt).

Sebastian Nowak, Jann-Frederick Laß, Narine Mesropyan +8

Department of Diagnostic and Interventional Radiology, University Hospital Bonn · Department of Research-IT, University Hospital Bonn

LLM ServingLLM SecurityLLM Inference
AskPDF
Mar 24, 2026

StepCache: Step-Level Reuse with Lightweight Verification and Selective Patching for LLM Serving

We address LLM serving workloads where repeated requests share a common solution structure but differ in localized constraints, such as output schema, variable names, or numeric constants. Prior caching approaches typically reuse either full responses (semantic caching) or model-internal KV/prefix states, which are respectively brittle under partial changes or tightly coupled to specific backends. We present StepCache, a backend-agnostic step-level reuse layer that segments outputs into ordered steps, retrieves the best-matching cached request, verifies steps using lightweight task-aware checks, and regenerates only failing regions via selective patching. StepCache additionally supports strict structured-output enforcement for JSON, including single-step extraction, required-key constraints, and one-shot repair, as well as conservative skip-reuse fallbacks for semantic changes. For linear equations, StepCache promotes verification into correction via a bounded repair loop with a deterministic fallback that guarantees correctness when the backend model fails. In a CPU-only perturbation-heavy micro-benchmark on math and JSON variants, averaged over three seeds, StepCache reduces mean latency from 2.13 s to 0.67 s, median latency from 2.42 s to 0.01 s, and p95 latency from 3.38 s to 3.30 s. It also reduces total token usage from 36.1k to 27.3k and improves end-to-end correctness from 72.5% to 100% under task-specific checks and a stitched-output integrity check. Across requests, 79.7% take the reuse-only fast path, 5.4% require patching, and 14.9% trigger skip-reuse.

Azam Nouri

Department of Science, Technology & Mathematics, Lincoln University, USA

LLM ServingLLM Inference EfficiencyLLM Answer Verification
AskPDF
Mar 23, 2026

Chimera: Latency- and Performance-Aware Multi-agent Serving for Heterogeneous LLMs

Multi-agent applications often execute complex tasks as multi-stage workflows, where each stage is an LLM call whose output becomes part of context for subsequent steps. Existing LLM serving systems largely assume homogeneous clusters with identical model replicas. This design overlooks the potential of heterogeneous deployments, where models of different sizes and capabilities enable finer trade-offs between latency and performance. However, heterogeneity introduces new challenges in scheduling across models with diverse throughput and performance. We present Chimera, a predictive scheduling system for multi-agent workflow serving on heterogeneous LLM clusters that jointly improves end-to-end latency and task performance. Chimera applies semantic routing to estimate per-model confidence scores for each request, predicts the total remaining output length of the workflow, and estimates per-model congestion using in-flight predicted token volumes for load balancing. We evaluate Chimera on representative agentic workflows for code generation and math reasoning using multiple heterogeneous LLM configurations. Across comparable settings, Chimera traces the best latency-performance frontier, reducing end-to-end latency by 1.2--2.4×\times and improving task performance by 8.0-9.5 percentage points on average over competitive baselines including vLLM.

Kangqi Ni, Wenyue Hua, Xiaoxiang Shi +3

University of North Carolina, Chapel Hill · Microsoft · Carnegie Mellon University +2

LLM ServingLLM Inference SchedulingLLM Agent ServingLLM Agent Workflow OptimizationLLM Routing
AskPDF
Mar 20, 2026

SEAR: Schema-Based Evaluation and Routing for LLM Gateways

Evaluating production LLM responses and routing requests across providers in LLM gateways requires fine-grained quality signals and operationally grounded decisions. To address this gap, we present SEAR, a schema-based evaluation and routing system for multi-model, multi-provider LLM gateways. SEAR defines an extensible relational schema covering both LLM evaluation signals (context, intent, response characteristics, issue attribution, and quality scores) and gateway operational metrics (latency, cost, throughput), with cross-table consistency links across around one hundred typed, SQL-queryable columns. To populate the evaluation signals reliably, SEAR proposes self-contained signal instructions, in-schema reasoning, and multi-stage generation that produces database-ready structured outputs. Because signals are derived through LLM reasoning rather than shallow classifiers, SEAR captures complex request semantics, enables human-interpretable routing explanations, and unifies evaluation and routing in a single query layer. Across thousands of production sessions, SEAR achieves strong signal accuracy on human-labeled data and supports practical routing decisions, including large cost reductions with comparable quality.

Zecheng Zhang, Han Zheng, Yue Xu

Strukto.AI, USA · Infron.AI, USA

LLM RoutingLLM EvaluationLLM ServingStructured Output Generation
AskPDF
Mar 17, 2026

Efficient LLM Serving for Agentic Workflows: A Data Systems Perspective

Agentic workflows are composed of sequences of interdependent Large Language Model (LLM) calls, and they have become a dominant workload in modern AI systems. These workflows exhibit extensive redundancy from overlapping prompts and intermediate results due to speculative and parallel exploration. Existing LLM serving systems, such as vLLM, focus on optimizing individual inference calls and overlook cross-call dependencies, leading to significant inefficiencies. This paper rethinks LLM and agent serving from a data systems perspective and introduces Helium, a workflow-aware serving framework that models agentic workloads as query plans and treats LLM invocations as first-class operators. Helium integrates proactive caching and cache-aware scheduling to maximize reuse across prompts, KV states, and workflows. Through these techniques, Helium bridges classic query optimization principles with LLM serving, achieving up to 1.56x speedup over state-of-the-art agent serving systems on various workloads. Our results demonstrate that end-to-end optimization across workflows is essential for scalable and efficient LLM-based agents.

Noppanat Wadlom, Junyi Shen, Yao Lu

National University of Singapore Singapore

LLM ServingLLM Agent Workflow OptimizationAgentic WorkflowsLLM Inference Acceleration
AskPDF
Mar 13, 2026

Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity

Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from O(L∗sctx)O(L * s_ctx) bytes (GB-scale KV caches under stage-level disaggregation) to O(Nv∗d)O(N_v * d) bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster ($38k) improves Tokens/$ by 37% over a homogeneous baseline ($64k) without degrading latency.

Donglin Yu

Efficient Multimodal InferenceLLM ServingDisaggregated LLM ServingLLM Inference SchedulingMultimodal Large Language Models
AskPDF
Mar 13, 2026

98×\times Faster LLM Routing Without a Dedicated GPU: Flash Attention, Prompt Compression, and Near-Streaming for the vLLM Semantic Router

System-level routers that intercept LLM requests for safety classification, domain routing, and PII detection must be both fast and operationally lightweight: they should add minimal latency to every request, yet not require a dedicated GPU -- an expensive resource better used for LLM inference itself. When the router co-locates on the same GPU as vLLM serving instances, standard attention's O(n2)O(n^2) memory makes long-context classification (8K--32K tokens) impossible: at 8K tokens, three concurrent classifiers need ∼{\sim}4.5,GB for attention masks alone, far exceeding the memory left by vLLM. We present three staged optimizations for the vLLM Semantic Router, benchmarked on AMD Instinct MI300X, that solve both the latency and the memory problem. Stage~1: a custom CK Flash Attention operator for ONNX Runtime on ROCm reduces attention memory from O(n2)O(n^2) to O(n)O(n) and end-to-end (E2E) latency from 4{,}918,ms to 127,ms (\textbf{38.7×\times}), enabling 8K--32K tokens where SDPA OOMs. Stage~2: classical NLP prompt compression (TextRank, position weighting, TF-IDF, and novelty scoring) reduces all inputs to ∼{\sim}512 tokens without neural inference, capping both latency and GPU memory at a constant regardless of original prompt length (E2E 127→\to62,ms, \textbf{2.0×\times}). Stage~3: near-streaming body processing with adaptive chunking and zero-copy JSON eliminates serialization overhead (E2E 62→\to50,ms, \textbf{1.2×\times}). Cumulatively: \textbf{98×\times} improvement (4{,}918,ms to 50,ms), 16K-token routing in 108,ms, and a total router GPU footprint under 800,MB -- small enough to share a GPU with LLM serving and removing the need for a dedicated accelerator. Stage1 targets AMD ROCm (NVIDIA GPUs already have FlashAttention via cuDNN); Stages2 and~3 are hardware-agnostic.

Xunzhuo Liu, Bowei He, Xue Liu +3

vLLM Semantic Router Project · MBZUAI · McGill University +2

LLM RoutingLLM ServingLLM Inference AccelerationAttention Mechanisms
AskPDF
Mar 12, 2026

Characterizing Performance-Energy Trade-offs of Large Language Models in Multi-Request Workflows

Large language models (LLMs) are increasingly used in applications forming multi-request workflows like document summarization, search-based copilots, and multi-agent programming. While these workflows unlock richer functionality, they also amplify latency and energy demand during inference. Existing measurement and benchmarking efforts either focus on assessing LLM inference systems or consider single-request evaluations, overlooking workflow dependencies and cross-request interactions unique to multi-request workflows. Moreover, the energy usage of such interdependent LLM calls remains underexplored. To address these gaps, this paper presents the first systematic characterization of performance-energy trade-offs in multi-request LLM inference. We develop four representative workloads capturing sequential, interactive, agentic, and composite patterns common in modern deployments. Using an NVIDIA A100 testbed with state-of-the-art serving systems (vLLM and Parrot), we analyze how key energy knobs affect latency, throughput, and component-level energy use. Our findings reveal batch size as the most impactful lever, though benefits are workload dependent. While optimal batching benefits workloads with large shared prompts, it is ineffective for sequential summarization and only partially effective for multi-agent coding. GPU power capping provides modest but predictable savings, while output length induces linear energy scaling with limited efficiency gains. We further show that engine-level optimizations in vLLM maintain higher GPU utilization and efficiency, especially for decode-heavy workloads, while Parrot's workflow-aware scheduling achieves lower energy consumption under strict power constraints. These findings offer actionable guidelines for developers and system operators designing performance- and energy-aware LLM serving systems in emerging multi-request workflows.

Md. Monzurul Amin Ifath, Israat Haque

Dalhousie University, Halifax, NS, Canada

Parallel LLM InferenceLLM InferenceLLM ServingEnergy-Efficient ML
AskPDF
Mar 11, 2026

Improving LLM Performance Through Black-Box Online Tuning: A Case for Adding System Specs to Factsheets for Trusted AI

In this paper, we present a novel black-box online controller that uses only end-to-end measurements over short segments, without internal instrumentation, and hill climbing to maximize goodput, defined as the throughput of requests that satisfy the service-level objective. We provide empirical evidence that this design is well-founded. Using this advance in LLM serving as a concrete example, we then discuss the importance of integrating system performance and sustainability metrics into Factsheets for organizations adopting AI systems.

Yonas Atinafu, Henry Lin, Robin Cohen

David R. Cheriton School of Computer Science, University of Waterloo, Waterloo, Canada

LLM Inference ServingLLM ServingBlack-Box Optimization
AskPDF
Mar 10, 2026

FlexServe: A Fast and Secure LLM Serving System for Mobile Devices with Flexible Resource Isolation

Device-side Large Language Models (LLMs) have witnessed explosive growth, offering higher privacy and availability compared to cloud-side LLMs. During LLM inference, both model weights and user data are valuable, and attackers may even compromise the OS kernel to steal them. ARM TrustZone is the de facto hardware-based isolation technology on mobile devices, used to protect sensitive applications from a compromised OS. However, protecting LLM inference with TrustZone incurs significant overhead due to its inflexible isolation of memory and the NPU. To address these challenges, this paper introduces FlexServe, a fast and secure LLM serving system for mobile devices. It first introduces a Flexible Resource Isolation mechanism to construct Flexible Secure Memory (Flex-Mem) and Flexible Secure NPU (Flex-NPU). Both memory pages and the NPU can be efficiently switched between unprotected and protected modes. Based on these mechanisms, FlexServe designs a fast and secure LLM inference framework within TrustZone's secure world. The LLM-Aware Memory Management and Secure Inference Pipeline are introduced to accelerate inference. A Multi-Model Scheduler is proposed to optimize multi-model workflows. We implement a prototype of FlexServe and compare it with two TrustZone-based strawman designs. The results show that FlexServe achieves an average 10.05×10.05\times speedup in Time to First Token (TTFT) compared to the strawman, and an average 2.44×2.44\times TTFT speedup compared to an optimized strawman with pipeline and secure NPU enabled. For multi-model agent workflows, the end-to-end speedup is up to 24.30×24.30\times and 4.05×4.05\times compared to the strawman and optimized strawman, respectively.

Yinpeng Wu, Yitong Chen, Lixiang Wang +3

Institute of Parallel and Distributed Systems, Shanghai Jiao Tong University

LLM ServingOn-Device Language Model InferenceTrusted Execution EnvironmentsLLM SecurityLLM Inference Acceleration
AskPDF
Mar 6, 2026

MoEless: Efficient MoE LLM Serving with Serverless Experts

Large Language Models (LLMs) increasingly adopt Mixture-of-Experts (MoE) architectures to scale efficiently under stringent resource constraints. However, MoE's sparse activation causes severe expert load imbalance, where a few experts become stragglers while others remain underutilized, leading to inflated inference latency and cost. Existing solutions assume static, serverful model deployments, limiting expert elasticity and often incurring costly expert swapping or degraded output quality. We present MoEless, an efficient serverless MoE serving framework that mitigates expert load imbalance via elastic expert execution. MoEless leverages lightweight, layer-aware predictors to estimate incoming expert load distributions and proactively identify stragglers. We design optimized scaling and placement strategies to improve function locality, GPU utilization, and cross-expert load balance. MoEless is prototyped on top of Megatron-LM and deployed on an eight-GPU testbed. Experiments with open-source MoE models and real-world workloads show that MoEless reduces inference latency by 43% and inference cost by 84% compared to state-of-the-art solutions.

Hanfei Yu, Bei Ouyang, Shwai He +2

University of Maryland College Park

LLM Inference EfficiencyExpert Load BalancingLLM ServingMixture-of-Experts Inference
AskPDF
Feb 27, 2026

Data Driven Optimization of GPU efficiency for Distributed LLM-Adapter Serving

Large Language Model (LLM) adapters enable low-cost model specialization, but introduce complex caching and scheduling challenges in distributed serving systems where hundreds of adapters must be hosted concurrently. While prior work has largely focused on latency and throughput optimization, minimizing GPU resource requirements through near-peak utilization remains largely underexplored. This paper presents a data-driven pipeline that, for a given workload, computes an adapter placement that serves the workload with the minimum number of GPUs while avoiding request starvation and GPU memory errors. To that end, the approach identifies the maximum feasible throughput attainable on each GPU by leveraging accurate performance predictions learned from real serving behavior. The proposed pipeline integrates three components: (i) a Digital Twin (DT) tailored to LLM-adapter serving, (ii) a distilled machine learning (ML) model trained on DT-generated data, and (iii) a greedy placement algorithm that exploits ML-based performance estimates to maximize GPU efficiency. The DT emulates real system dynamics with high fidelity, achieving below 5% throughput estimation error while executing up to 90x faster than full LLM benchmarking across both predictable and unpredictable workloads. The learned ML models further accelerate performance estimation with marginal accuracy degradation, enabling scalable optimization. Experimental results demonstrate that the pipeline substantially improves GPU efficiency, reducing the number of GPUs required to sustain target workloads by 60% on average across the evaluated scenarios. Beyond GPU efficiency, the pipeline can be adapted to alternative objectives, such as latency minimization, highlighting its versatility for future large-scale LLM serving infrastructures.

Ferran Agullo, Joan Oliveras, Chen Wang +5

Barcelona Supercomputing Center (BSC), Barcelona, Spain · Universitat Politècnica de Catalunya - BarcelonaTech (UPC), Barcelona, Spain · IBM Research, New York, USA

LLM Inference EfficiencyLLM ServingDigital Twins
AskPDF
Feb 26, 2026

LLMServingSim 2.0: A Unified Simulator for Heterogeneous and Disaggregated LLM Serving Infrastructure

Large language model (LLM) serving infrastructures are undergoing a shift toward heterogeneity and disaggregation. Modern deployments increasingly integrate diverse accelerators and near-memory processing technologies, introducing significant hardware heterogeneity, while system software increasingly separates computation, memory, and model components across distributed resources to improve scalability and efficiency. As a result, LLM serving performance is no longer determined by hardware or software choices in isolation, but by their runtime interaction through scheduling, data movement, and interconnect behavior. However, understanding these interactions remains challenging, as existing simulators lack the ability to jointly model heterogeneous hardware and disaggregated serving techniques within a unified, runtime-driven framework. This paper presents LLMServingSim 2.0, a unified system-level simulator designed to make runtime-driven hardware-software interactions in heterogeneous and disaggregated LLM serving infrastructures explicit and analyzable. LLMServingSim 2.0 embeds serving decisions and hardware behavior into a single runtime loop, enabling interaction-aware modeling of batching, routing, offloading, memory, and power. The simulator supports extensible integration of emerging accelerators and memory systems through profile-based modeling, while capturing dynamic serving behavior and system-level effects. We validate LLMServingSim 2.0 against real deployments, showing that it reproduces key performance, memory, and power metrics with an average error of 0.95%, while maintaining simulation times of around 10 minutes even for complex configurations. These results demonstrate that LLMServingSim 2.0 provides a practical bridge between hardware innovation and serving-system design, enabling systematic exploration and co-design for next-generation LLM serving infrastructures.

Jaehong Cho, Hyunmin Choi, Guseul Heo +1

School of Computing KAIST Daejeon, South Korea

LLM ServingDisaggregated LLM ServingHardware-Software Co-DesignHeterogeneous Computing
AskPDF
Feb 23, 2026

vLLM Semantic Router: Signal Driven Decision Routing for Mixture-of-Modality Models

As large language models (LLMs) diversify across modalities, capabilities, and cost profiles, the problem of intelligent request routing: selecting the right model for each query at inference time, has become a critical systems challenge. We present vLLM Semantic Router, a signal-driven decision routing framework for Mixture-of-Modality (MoM) model deployments. The architecture follows two complementary Shannon-inspired views. In the information-theoretic regime, signal extraction reduces the entropy of "which model?" by distilling routing-relevant information from raw queries. In the Boolean-algebraic regime, the decision engine composes functionally complete routing policies from signal conditions. The central innovation is composable signal orchestration: thirteen heterogeneous signal types, spanning sub-millisecond heuristics and neural classifiers for semantics, safety, and modality, are composed through configurable Boolean decision rules into deployment-specific routing policies, so that fundamentally different scenarios (multi-cloud enterprise, privacy-regulated, cost-optimized) are expressed as different configurations over the same architecture. Matched decisions drive semantic model routing via thirteen selection algorithms, while per-decision plugin chains enforce safety constraints including a three-stage HaluGate hallucination detection pipeline and a lightweight episodic memory system with ReflectionGate for personalized multi-turn context. A typed neural-symbolic DSL specifies these routing policies and compiles them to multiple deployment targets, enabling configuration-first adaptation without code changes. Together, these components show that composable signal orchestration enables a single framework to serve diverse deployment scenarios with differentiated cost, privacy, and safety policies.

Xunzhuo Liu, Huamin Chen, Samzong Lu +30

Steve

Multimodal Model RoutingLLM RoutingAdaptive Model RoutingLLM Serving
AskPDF
Feb 23, 2026

ISO-Bench: Can Coding Agents Optimize Real-World Inference Workloads?

We introduce ISO-Bench, a benchmark for coding agents to test their capabilities on real-world inference optimization tasks. These tasks were taken from vLLM and SGLang, two of the most popular LLM serving frameworks. Each task provides an agent with a codebase and bottleneck description, whereby the agent must produce an optimization patch evaluated against expert human solutions. We curated 54 tasks from merged pull requests with measurable performance improvements. While existing benchmarks heavily use runtime-based metrics, such approaches can be gamed to pass tests without capturing the actual intent of the code changes. Therefore, we combine both hard (execution-based) and soft (LLM-based) metrics to show that both are necessary for complete evaluation. While evaluating both closed and open-source coding agents, we find no single agent dominates across codebases. Surprisingly, agents often identify correct bottlenecks but fail to execute working solutions. We also show that agents with identical underlying models differ substantially, suggesting scaffolding is as important as the model.

Ayush Nangia, Shikhar Mishra, Aman Gokrani +1

Coding Agent BenchmarksAI Coding AgentsAI Agent BenchmarksCode Generation EvaluationCode OptimizationLLM Inference Efficiency+1
AskPDF
Feb 22, 2026

AMV-L: Lifecycle-Managed Agent Memory for Tail-Latency Control in Long-Running LLM Systems

Long-running LLM agents require persistent memory to preserve state across interactions, yet most deployed systems manage memory with age-based retention (e.g., TTL). While TTL bounds item lifetime, it does not bound the computational footprint of memory on the request path: as retained items accumulate, retrieval candidate sets and vector similarity scans can grow unpredictably, yielding heavy-tailed latency and unstable throughput. We present AMV-L (Adaptive Memory Value Lifecycle), a memory-management framework that treats agent memory as a managed systems resource. AMV-L assigns each memory item a continuously updated utility score and uses value-driven promotion, demotion, and eviction to maintain lifecycle tiers; retrieval is restricted to a bounded, tier-aware candidate set that decouples the request-path working set from total retained memory. We implement AMV-L in a full-stack LLM serving system and evaluate it under identical long-running workloads against two baselines: TTL and an LRU working-set policy, with fixed prompt-injection caps. Relative to TTL, AMV-L improves throughput by 3.1x and reduces latency by 4.2x (median), 4.7x (p95), and 4.4x (p99), while reducing the fraction of requests exceeding 2s from 13.8% to 0.007%. Compared to LRU, AMV-L trades a small regression in median/p95 latency (+26% / +3%) for improved extreme-tail behavior (-15% p99; -98% >2s) and lower token overhead (approximately 6% fewer tokens/request), while matching retrieval quality (value means within approximately 0-2%). The gains arise primarily from bounding retrieval-set size and vector-search work, not from shortening prompts. Our results show that predictable performance for long-running LLM agents requires explicit control of memory working-set size and value-driven lifecycle management, rather than retention time alone.

Emmanuel Bamidele

Department of Computer Science, Georgia Institute of Technology

Memory Management for LLM AgentsAgent Memory ManagementLLM ServingMemory-Augmented Language Model AgentsLLM Inference Efficiency
AskPDF
Feb 18, 2026

FlowPrefill: Decoupling Preemption from Prefill Scheduling Granularity to Mitigate Head-of-Line Blocking in LLM Serving

The growing demand for large language models (LLMs) requires serving systems to handle many concurrent requests with diverse service level objectives (SLOs). This exacerbates head-of-line (HoL) blocking during the compute-intensive prefill phase, where long-running requests monopolize resources and delay higher-priority ones, leading to widespread time-to-first-token (TTFT) SLO violations. While chunked prefill enables interruptibility, it introduces an inherent trade-off between responsiveness and throughput: reducing chunk size improves response latency but degrades computational efficiency, whereas increasing chunk size maximizes throughput but exacerbates blocking. This necessitates an adaptive preemption mechanism. However, dynamically balancing execution granularity against scheduling overheads remains a key challenge. In this paper, we propose FlowPrefill, a TTFT-goodput-optimized serving system that resolves this conflict by decoupling preemption granularity from scheduling frequency. To achieve adaptive prefill scheduling, FlowPrefill introduces two key innovations: 1) Operator-Level Preemption, which leverages operator boundaries to enable fine-grained execution interruption without the efficiency loss associated with fixed small chunking; and 2) Event-Driven Scheduling, which triggers scheduling decisions only upon request arrival or completion events, thereby supporting efficient preemption responsiveness while minimizing control-plane overhead. Evaluation on real-world production traces shows that FlowPrefill improves maximum goodput by up to 5.6×\times compared to state-of-the-art systems while satisfying heterogeneous SLOs.

Chia-chi Hsieh, Zan Zong, Xinyang Chen +3

Tsinghua University, Beijing, China · University of Science and Technology Beijing, Beijing, China

LLM ServingLLM Inference SchedulingReal-Time Scheduling
AskPDF
Feb 2, 2026

Learning to Route and Schedule LLMs from User Retrials via Contextual Queueing Bandits

Explosive demands for LLMs often cause user queries to accumulate in server queues, requiring efficient routing (query-LLM matching) and scheduling (query prioritization) mechanisms. Several online algorithms are being deployed, but they overlook the following two key challenges inherent to conversational LLM services: (1) unsatisfied users may retry queries, increasing the server backlog, and (2) requests for explicit" feedback, such as ratings, degrade user experiences. In this paper, we develop a joint routing and scheduling algorithm that leverages implicit" feedback inferred from user retrial behaviors. The key idea is to propose and study the framework of contextual queueing bandits with multinomial logit feedback (CQB-MNL). CQB-MNL models query retrials, as well as context-based learning for user preferences over LLMs. Our algorithm, anytime CQB (ACQB), achieves efficient learning while maintaining queue stability by combining Thompson sampling with forced exploration at a decaying rate. We show that ACQB simultaneously achieves a cumulative regret of O~(t)\widetilde{\mathcal{O}}(\sqrt{t}) for routing and a queue length regret of O~(t−1/4)\widetilde{\mathcal{O}}(t^{-1/4}) for any large tt. For experiments, we refine query embeddings via contrastive learning while adopting a disjoint parameter model to learn LLM-specific parameters. Experiments on synthetic data, offline routing datasets (SPROUT, EmbedLLM, and RouterBench), and real user conversation logs (WildChat-1M) confirm that our methods improve routing, scheduling, and queue stability against strong online and offline-trained baselines.

Seoungbin Bae, Junyoung Son, Dabeen Lee

Department of Industrial & Systems Engineering, KAIST · Graduate School of Data Science, KAIST · Department of Mathematical Sciences, Seoul National University

LLM ServingLLM RoutingContextual BanditsLLM Request Scheduling
AskPDF
Jan 29, 2026

Test-Time Compute Games

Test-time compute has emerged as a promising strategy to enhance the reasoning abilities of large language models (LLMs). However, this strategy has in turn increased how much users pay cloud-based providers offering LLM-as-a-service, since providers charge users for the amount of test-time compute they use to generate an output. In our work, we show that the market of LLM-as-a-service is socially inefficient: providers have a financial incentive to increase the amount of test-time compute, even if this increase contributes little to the quality of the outputs. To address this inefficiency, we introduce a reverse second-price auction mechanism where providers bid their offered price and (expected) quality for the opportunity to serve a user, and users pay proportionally to the marginal value generated by the winning provider relative to the second-highest bidder. To illustrate and complement our theoretical results, we conduct experiments with multiple instruct models from the Llama\texttt{Llama} and Qwen\texttt{Qwen} families, as well as reasoning models distilled from DeepSeek-R1\texttt{DeepSeek-R1}, on math and science benchmark datasets.

Ander Artola Velasco, Dimitrios Rontogiannis, Stratis Tsirtsis +1

Max Planck Institute for Software Systems, Kaiserslautern, Germany · Hasso Plattner Institute, Potsdam, Germany

LLM ServingCost-Aware InferenceLLM Inference
AskPDF