LLM Inference Serving

Modern LLM serving is no longer homogeneous or monolithic. Production systems now combine disaggregated execution, complex parallelism, runtime optimizations, and stateful workloads such as reasoning, agents, and RL rollouts. Simulation is attractive for exploring this growing design space, yet existing simulators lack the architectural completeness and decision-grade fidelity it demands. Their monolithic-replica abstractions are ill-suited to disaggregated serving, while average-case analytical proxies can distort SLA predictions and even reverse optimization conclusions. We present Frontier, a discrete-event simulator for modern LLM inference serving. Frontier features a disaggregated abstraction. It captures the structure and dynamics of modern serving systems by modeling co-location, Prefill-Decode Disaggregation (PDD), and Attention-FFN Disaggregation (AFD) with role-specific cluster workers, incorporating key runtime optimizations (e.g., CUDA Graphs, speculative decoding) within the scheduler-batch-engine loop, and supporting stateful requests for emerging workloads. It further provides accurate and generalizable predictions of computation, communication, and memory costs across diverse serving scenarios with complex workload compositions. On 16-H800 GPU testbed, Frontier achieves an average throughput error below 4%. Compared with state-of-the-art simulators, it reduces end-to-end latency error from 44.9% to 6.4% under co-location and from 51.7% to 2.6% under disaggregation. It scales to over 1K GPUs on commodity CPUs and enables new use cases such as SLA-dependent Pareto frontier exploration, heterogeneous disaggregated allocation, agentic reasoning scheduling validation, and RL post-training reconfiguration. We release Frontier at https://github.com/NetX-lab/Frontier.

Linear attention has recently gained significant attention for long-context inference due to its constant decoding cost with respect to context length. However, existing serving systems typically serve linear attention by recurrently computing and updating a large linear attention state in every decoding step. Since the state is much larger than the per-token key and value, recurrent decoding incurs substantial memory access and becomes inefficient for serving linear attention. In this paper, we propose KVBuffer, an IO-aware serving mechanism for linear attention. By buffering recent keys and values, KVBuffer enables serving systems to compute linear attention outputs in more flexible and memory-efficient ways. For decoding, KVBuffer enables chunkwise computation, which reduces average memory access and decoding latency by deferring state updates and applying them in batch. For speculative decoding, KVBuffer verifies draft tokens in parallel and avoids storing temporary states. For short contexts, KVBuffer computes attention outputs directly from buffered keys and values, without creating or updating the linear attention state. We implement KVBuffer in SGLang for Qwen3-Next. Our evaluations show that KVBuffer can reduce linear attention decoding latency by up to 45.17% and increase the maximum number of serving requests by 5x for speculative decoding when verifying four draft tokens.

Conventional transformer inference engines are request-driven, paying an O(n) prefill cost on every query. In streaming workloads, where data arrives continuously and queries probe an ever-growing context, this cost is prohibitive. We introduce a data-driven computational model centred on stateful sessions: a persistent KV cache advanced incrementally as new data arrives, so prefill is moved off the critical path and query latency becomes O(|q|), independent of accumulated context size. Building on this, Flash Queries reclaim idle GPU cycles between data arrivals to pre-evaluate registered questions and return cached answers before the user asks, a pattern that is structurally impossible in stateless engines because they discard intermediate state between requests. A multi-tenant continuous-batching scheduler with cell-budget admission and prefix-aware grouped prefill lets dozens of stateful sessions coexist on a single GPU while preserving full quadratic self-attention. On streaming market-data benchmarks the reference implementation achieves up to 5.9x speedup over conventional inference engines (vLLM, SGLang, TensorRT-LLM, llama.cpp), holding query latency constant as accumulated context grows.

Today's inference-time workloads increasingly depend on timely access to a model's internal states. We present DMI-Lib, a high-speed deep model inspector that treats internal observability as a first-class systems primitive, decoupling it from the inference hot path via an asynchronous observability substrate built from Ring^2, a GPU-CPU memory abstraction for capturing and staging tensors, and a policy-controlled host backend that exports them. DMI-Lib enables the placement of observation points across a rich space of internal signals and diverse inference backends while preserving serving optimizations and adhering to tight GPU memory budgets. Our experiments demonstrate that DMI-Lib incurs only 0.4%--6.8% overhead in offline batch inference and an average of 6% in moderate online serving, reducing latency overhead by 2x-15x compared to existing baselines with similar observability features. DMI-Lib is open-sourced at https://github.com/ProjectDMX/DMI.

Static-graph LLM decoders provide predictable launches, fixed tensor shapes, and low submission overhead, but online decoding exposes highly irregular KV-cache behavior: request lengths differ, EOS events arrive asynchronously, and logical histories fragment over time. Dynamic runtimes recover flexibility through paged KV management and step-level scheduling, while static-graph executors often over-reserve memory and suffer burst-time latency outliers. This paper studies whether much of this variability can be absorbed below a fixed decode interface. We present KV-RM, a runtime design that regularizes KV-cache movement beneath a static-graph LLM decoder. KV-RM decouples logical KV histories from physical storage, tracks active KV state through a block pager, and materializes each decode step through a single committed descriptor. A merge-staged transport path coalesces non-contiguous KV mappings into a small number of large transfer groups before a fixed-shape attention kernel consumes them. Optional bounded far-history summaries can be enabled under the same interface, but the core design does not depend on them. On a 2-GPU NVIDIA A100 node, KV-RM improves mixed-length decoding throughput and tail latency relative to a static-graph baseline, reduces reserved KV memory across workload families, and removes severe burst-time latency spikes under production-trace replay. These results suggest that KV-cache movement, rather than kernel shape, can be an effective boundary for recovering runtime flexibility in static-graph LLM serving.

Enterprise AI backends increasingly admit heterogeneous execution requests across model deployment, inference, evaluation, data movement, and agentic workflows. In many systems, those requests arrive in service-specific shapes, which makes it difficult to attach shared admission-time behavior such as logging, governance hints, resource accounting, authorization-aware policy hooks, and later runtime review without rebuilding the same contract in each subsystem. This paper introduces the execution envelope, a normalized internal admission object that records who is asking for what kind of execution, what resources were requested, what policy-relevant scope accompanied the request, and what the backend ultimately granted. The proposal is intentionally narrow. It does not replace service-specific request models, perform scheduling, or introduce a new authority token. Instead, it defines a descriptive admission seam that can be threaded through real backend paths before backend-specific resolution begins. I formalize the distinction between requested and granted resources, specify the field families, invariants, and lifecycle of the envelope, work through POST /serving/deploy_model as an initial proving ground, and position the design relative to usage control, analyzable authorization, admission control, and cluster scheduling. The central claim is that a shared execution-admission contract is a useful missing primitive for modern AI backends because it creates one place to attach governance and observability without pretending to solve placement, policy, and runtime execution in a single step.

Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the branch externality and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce TAPER, a per-step admission controller that treats extra branches as opportunistic work, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, TAPER improves goodput by $1.77\times$ over IRP-Off and by $1.48\times$ over IRP-Eager, while maintaining over $95\%$ SLO attainment.

Prefix caching can reduce LLM inference latency by reusing KV caches across requests with shared prompts, but cluster-scale reuse is challenging because caches are partitioned across nodes. We propose a decentralized, prefix-cache-aware routing scheme for peer-to-peer LLM serving. Each node maintains a local radix tree of its own cached prefixes and asynchronously refreshed estimates of peer caches using periodic anti-entropy. Requests are routed to the node with the longest estimated prefix match, without centralized coordination or KV-cache transfer. Stale metadata only causes cache misses, not incorrect outputs, making weak consistency sufficient for correctness. Evaluation on simulated MMLU workloads show that decentralized routing improves latency under low communication delay and skewed prefix distributions, while high network latency and affinity-induced hotspots limit its benefits.

Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $δ$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.

The rapid adoption of large language models (LLMs) has created significant challenges for efficient inference at scale. Unlike traditional workloads, LLM inference is constrained by both computation and the memory overhead of key-value (KV) caching, which accelerates decoding but quickly exhausts GPU memory. In this paper, we introduce the first queueing-theoretic framework that explicitly incorporates both computation and GPU memory constraints into the analysis of LLM inference. Based on this framework, we derive rigorous stability and instability conditions that determine whether an LLM inference service can sustain incoming demand without unbounded queue growth. This result offers a powerful tool for system deployment, potentially addressing the core challenge of GPU provisioning. By combining an estimated request arrival rate with our derived stable service rate, operators can calculate the necessary cluster size to avoid both costly over-purchasing and performance-violating under-provisioning. We further validate our theoretical predictions through extensive experiments in real GPU production environments. Our results show that the predicted stability conditions are highly accurate, with deviations typically within 10%.

Public inference benchmarks compare AI systems at the model and provider level, but the unit at which deployment decisions are actually made is the endpoint: the (provider, model, stock-keeping-unit) tuple at which a specific quantization, decoding strategy, region, and serving stack is exposed. We introduce TokenArena, a continuous benchmark that measures inference at endpoint granularity along five core axes (output speed, time to first token, workload-blended price, effective context, and quality on the live endpoint) and synthesizes them, together with a modeled energy estimate, into three headline composites: joules per correct answer, dollars per correct answer, and endpoint fidelity (output-distribution similarity to a first-party reference). The framework's novelty is empirical and methodological. Across 78 endpoints serving 12 model families, the same model on different endpoints differs in mean accuracy by up to 12.5 points on math and code, in fingerprint similarity to first party by up to 12 points, in tail latency by an order of magnitude, and in modeled joules per correct answer by a factor of 6.2. We further show that workload-aware blended pricing reorders the leaderboard substantially: 7 of 10 top-ranked endpoints under the chat preset (3:1 input:output) fall out of the top 10 under the retrieval-augmented preset (20:1), and the reasoning preset (1:5) elevates frontier closed models that the chat preset penalizes on price. We release the framework, schema, probe and eval harness, and a v1.0 leaderboard snapshot under CC BY 4.0. TokenArena is a methodology, not a single ranking; we publish full provenance and limitations and welcome external replication.

Mixture-of-Experts (MoE) models offer high capacity with efficient inference cost by activating a small subset of expert models per input. However, deploying MoE models requires all experts to reside in memory, creating a gap between the resource used by activated experts and the provisioned resources. This underutilization is further pronounced in multi-tenant scenarios. In this paper, we propose FaaSMoE, a multi-tenant MoE serving architecture built on Function-as-a-Service (FaaS) platforms. FaaSMoE decouples the control and execution planes of MoE by deploying experts as stateless FaaS functions, enabling on-demand and scale-to-zero expert invocation across tenants. FaaSMoE further supports configurable expert granularity within functions, trading off per-expert elasticity for reduced invocation overhead. We implement a prototype with an open-source edge-oriented FaaS platform and evaluate it using Qwen1.5-moe-2.7B under multi-tenant workloads. Compared to a full-model baseline, FaaSMoE uses less than one third of the resources, demonstrating a practical and resource-efficient path towards scalable MoE serving in a multi-tenant environment.

Modern enterprise AI applications increasingly rely on compound AI systems - architectures that compose multiple models, retrievers, and tools to accomplish complex tasks. Deploying such systems in production demands inference infrastructure that can efficiently serve concurrent, heterogeneous model invocations while maintaining cost-effectiveness and low latency. This paper presents a production deployment study of a modular, platform-agnostic inference architecture developed at Salesforce to support compound AI use cases including Agentforce (autonomous AI agents) and ApexGuru (AI-powered code analysis). The system integrates serverless execution, dynamic autoscaling, and MLOps pipelines to deliver consistent low-latency inference across multi-component agent workflows. We report production results demonstrating over 50% reduction in tail latency (P95), up to 3.9x throughput improvement, and 30 to 40% cost savings compared to prior static deployments. We further present a novel analysis of compound-system-specific challenges including multi-model fan-out overhead, cascading cold-start propagation, and heterogeneous scaling dynamics that emerge uniquely when serving agentic workloads. Through detailed case studies and operational lessons, we illustrate how the architecture enables compound AI systems to scale model invocations in parallel, handle bursty multi-agent workloads, and support rapid model iteration - capabilities essential for operationalizing agentic AI at enterprise scale.

Power has become a central bottleneck for AI inference. This problem is becoming more urgent as agentic AI emerges as a major workload class, yet prior power-management techniques focus almost entirely on single-turn LLM serving. Our analysis shows that agentic serving behaves fundamentally differently: each request carries long-lived context that evolves across tool-interleaved turns, and lowering GPU frequency can push the system into a thrashing regime where memory pressure sharply worsens both performance and power efficiency. These observations show that power optimization for agentic serving requires rethinking. We present KAIROS, a context-aware power optimization system for agentic AI serving. KAIROS uses agent context as a first-class control signal to jointly manage GPU frequency, per-instance concurrency, and multi-instance request placement. This enables KAIROS to save power when memory headroom exists while avoiding thrashing and preserving performance targets. At a high level, KAIROS tracks requests at agent granularity, adapts local control to context growth and agent progress, and routes agents across instances to jointly improve power efficiency and memory stability. Evaluated across diverse software and data engineering agentic tasks, KAIROS achieves an average of 27% (up to 39.8%) power reduction while meeting the performance targets.

Agentic workflows carry out complex tasks by orchestrating multiple large language models (LLMs) and tools. Serving them at a target throughput with low latency is hard because they are written in arbitrary agentic frameworks and their execution times are unpredictable: execution branches, fans out, or recurs in data-dependent ways. Since their LLMs often outnumber the available GPUs, they also oversubscribe GPUs. We describe Scepsy, a serving system that schedules arbitrary multi-LLM agentic workflows onto a GPU cluster. Scepsy exploits the insight that, while the end-to-end latency of an agentic workflow is unpredictable, each LLM's fraction of execution time is comparatively stable across requests. Scepsy profiles each LLM under different parallelism degrees and combines the profiles with these fractions into an Aggregate LLM Pipeline, a lightweight throughput and latency predictor for allocations. To minimize latency at a target throughput, Scepsy uses the Aggregate LLM Pipeline to search over fractional GPU shares, tensor parallelism degrees, and replica counts. A hierarchical heuristic then places the chosen allocation onto the cluster, minimizing fragmentation and respecting network topology. On realistic agentic workflows, Scepsy achieves up to 2.5x higher throughput before saturation and 1.0-3.3x lower latency than systems that optimize LLMs independently or rely on user-specified allocations.

An agent meets a model as served: through an endpoint with a price card, a shared cache and other tenants, or on whatever hardware a self-hosted model runs. Benchmarks rank the weights. We introduce a Tetris benchmark that measures what agents get from models as served: every move is scored against an oracle, and every agent receives the same pieces. In five pre-specified experiments with nine open-weight models on one serverless provider, plus an open decision model self-hosted on a CPU, we find that price and size do not predict decision quality; that resending history costs almost nothing when cached input is free, would cost eleven to twelve times more if it were not, and makes play worse; that deployments keep between 4 and more than 64 agent contexts warm, in line with their throughput rather than the model's KV-cache size; that an agent's own long requests slow its slowest short decisions more than tenfold, which a simple admission rule cuts by 59% without hurting play; and that the decision model plays mid-pack but takes 27 s per move on a CPU, about 650 times longer than reported on a GPU. Architecture predicts some of what an agent sees; the deployment sets the rest.

Large language models(LLMs) are now used to power complex multi-turn agentic workflows. Existing systems run agentic inference by loosely assembling isolated components: an LLM inference engine (e.g., vLLM) and a tool orchestrator (e.g., Kubernetes). Although agentic workflows involve multiple LLM and tool requests, these systems schedule and allocate resources separately on a per-request basis, without end-to-end knowledge of the workflow. This leads to sub-optimal management of KV cache and tool execution environments. To address the challenges, we propose ThunderAgent, a fast, simple, and program-aware agentic inference system. We first abstract agentic workflows as LLM Programs, enabling a unified view of heterogeneous resources, including KV caches, system states, and external tool assets such as disk memory and network ports. Built upon this abstraction, ThunderAgent introduces a program-aware scheduler and a tool resource manager designed to maximize KV cache hit rates, mitigate memory imbalances, and enable asynchronous environment preparation. Evaluations across coding, routing, and scientific discovery agents demonstrate that ThunderAgent achieves 1.5-3.6x throughput improvements in serving, 1.8-3.9x in RL rollout, and up to 4.2x disk memory savings compared to state-of-the-art inference systems. To facilitate reproducibility and support future development, we open-source the system implementations of the whole ThunderAgent at: https://github.com/Agentic-Kinetics/ThunderAgent.

LLM inference is inherently expensive, even a modest slowdown can translate into substantial operating costs and severe availability risks. Recently, a growing body of research known as latency attacks focuses on crafting inputs to trigger worst-case output lengths. However, we report a contrary finding that these algorithmic-level latency attacks are largely ineffective against modern LLM serving systems. We reveal that system-level optimization such as continuous batching provides a logical isolation to mitigate contagious latency impact on co-located users. Thus, in this paper, we shift our focus from the algorithm to the system layer, and introduce a new Fill and Squeeze attack strategy targeting the state transition of the scheduler. Fill'' first exhausts the global KV cache to induce Head-of-Line blocking, while Squeeze'' forces the system into repetitive preemption. By manipulating output lengths using different attack prompts, and leveraging side-channel probing of memory status, we demonstrate that the attack can succeed in a practical black-box setting with much less cost. Extensive evaluations on vLLM indicate up to $75-742\times$ TTFT degradation relative to benign baselines and $1.5-4\times$ average slowdown on Time Per Output Token compared to existing attacks with 30-40% lower attack cost. Code: https://github.com/Phil-Fan/FS-attack

Speculative decoding can substantially accelerate LLM inference, but realizing its benefits in practice is challenging due to evolving workloads. We present TIDE (Temporal Incremental Draft Engine), a serving-engine-native framework that integrates online draft adaptation directly into high-performance LLM inference systems. TIDE reuses target model's intermediate hidden states generated during inference as training signals for draft adaptation, thereby avoiding additional target model computation and serving-time overhead. It employs adaptive runtime control to activate speculation and draft model training only when beneficial. TIDE exploits heterogeneous clusters by mapping inference and training to appropriate GPU classes. Across diverse real-world workloads, TIDE achieves up to 1.66$\times$ throughput over no-speculation baselines while recovering performance on misaligned workloads where static draft models degrade throughput. TIDE also reduces training time by up to 3.02$\times$ and storage requirements by 24$\times$ compared to existing draft training approaches, and improves system throughput by up to 1.22$\times$ on heterogeneous GPU clusters.

Efficient scheduling of large language model (LLM) inference tasks is critical for achieving low latency and high throughput, a challenge that is becoming increasingly acute with the rise of reasoning-capable LLMs whose generation lengths are highly variable. Traditional strategies like First Come, First-Serve (FCFS) often suffer from Head-of-Line (HOL) blocking, where long-running tasks delay shorter ones queued behind them. In this paper, we introduce PARS, a prompt-aware LLM task scheduler that mitigates HOL blocking by approximating shortest-job-first (SJF) scheduling through pairwise ranking with a margin ranking loss. PARS effectively predicts response-length-based task ordering directly from prompts, thereby optimizing scheduling decisions with minimal overhead. In addition, it integrates seamlessly with vLLM, a state-of-the-art LLM serving system, for the research community. Extensive experiments across multiple LLM models and real-world inference use cases, including chat, math, and code generation, demonstrate that PARS significantly reduces latency by up to 15.7x compared to the vLLM default scheduler. Cross-model evaluations demonstrate that our design generalizes effectively, allowing effective scheduling across diverse LLMs without requiring model-specific retraining.

LLMs are increasingly used world-wide from daily tasks to agentic systems and data analytics, requiring significant GPU resources. While LLM inference systems are capable of serving millions of requests from multiple users, they often lack theoretical models to determine whether they achieve the performance upper bounds of underlying hardware resources. Beyond online workload serving, merely analyzing existing systems-or developing yet another one-is both GPU-intensive and labor-intensive. This paper provides a comprehensive survey of LLM inference systems, focusing on their cache management policies and availability. We then show that simulations can be an effective tool to save GPU hours in the development and analysis phase of inference systems, revealing useful insights for developing better inference techniques, unlike how existing studies used simulations to find the best parameters inside a given system. Finally, we provide theoretical tools to estimate the optimal performance and formulate new ideas. Based on the theoretical analysis, especially on the cache management in LLM inference, we propose a simple yet effective cache replacement policy that can be easily plugged into existing preemptive schedulers and systems. We show that such a simple policy inspired from database systems can substantially save GPU hours in actual inference systems on online workloads. We share our experience submitting a journal paper to a database venue in November 2025 for anyone considering a similar path.