With the increasing adoption of AI applications such as large language models and computer vision AI, the computational demands on AI inference systems are continuously rising, making the enhancement of task processing capacity using existing hardware a primary objective in edge clouds. We propose EPARA, an end-to-end AI parallel inference framework in edge, aimed at enhancing the edge AI serving capability. Our key idea is to categorize tasks based on their sensitivity to latency/frequency and requirement for GPU resources, thereby achieving both request-level and service-level task-resource allocation. EPARA consists of three core components: 1) a task-categorized parallelism allocator that decides the parallel mode of each task, 2) a distributed request handler that performs the calculation for the specific request, and 3) a state-aware scheduler that periodically updates service placement in edge clouds. We implement a EPARA prototype and conduct a case study on the EPARA operation for LLMs and segmentation tasks. Evaluation through testbed experiments involving edge servers, embedded devices, and microcomputers shows that EPARA achieves up to 2.1$\times$ higher goodput in production workloads compared to prior frameworks, while adapting to various edge AI inference tasks.

Large language model (LLM) inference systems face a fundamental tension between minimizing Time-to-First-Token (TTFT) latency for new requests and maintaining a high, steady token generation rate (low Time-Per-Output-Token, or TPOT) for ongoing requests. Existing stall-free batching schedulers proposed by Sarathi, while effective at preventing decode stalls, introduce significant computational unfairness. They prioritize decode tasks excessively, simultaneously leading to underutilized decode slack and unnecessary prefill queuing delays, which collectively degrade the system's overall quality of service (QoS). This work identifies the root cause of this unfairness: the non-monotonic nature of Time-Between-Tokens (TBT) as a scheduling metric and the rigid decode-prioritizing policy that fails to adapt to dynamic workload bursts. We therefore propose FairBatching, a novel LLM inference scheduler that enforces fair resource allocation between prefill and decode tasks. It features an adaptive batch capacity determination mechanism, which dynamically adjusts the computational budget to improve the GPU utilization without triggering SLO violations. Its fair and dynamic batch formation algorithm breaks away from the decode-prioritizing paradigm, allowing computation resources to be reclaimed from bursting decode tasks to serve prefill surges, achieving global fairness. Furthermore, FairBatching provides a novel load estimation method, enabling more effective coordination with upper-level schedulers. Implemented and evaluated on realistic traces, FairBatching significantly reduces TTFT tail latency by up to 2.29x while robustly maintaining TPOT SLOs, achieving overall 20.0% improvement in single-node capacity and 54.3% improvement in cluster-level capacity.

LLMs have low GPU efficiency and high latency due to autoregressive decoding. Speculative decoding (SD) mitigates this using a small draft model to speculatively generate multiple tokens, which are then verified in parallel by a target model. However, when speculation accuracy is low, the overhead from rejected tokens can offset the benefits, limiting SD's effectiveness, especially at large batch sizes. To address this, we propose Speculative Verification (SV), an efficient augmentation to SD that dynamically predicts speculation accuracy and adapts the verification length to maximize throughput. SV introduces a companion model - a small auxiliary model similar in size to the draft model - to estimate the alignment between draft and target model distributions. By maximizing the information gain from quantifying this alignment, SV refines verification decisions, reducing wasted computation on rejected tokens and improving decoding efficiency. Moreover, SV requires no modifications to the draft or target models and is compatible with existing SD variants. We extensively evaluated SV on publicly available LLMs across three NLP tasks using nine combinations of draft, companion, and target models, including 13B-72B target models and three types of variations: base (no finetuning), instruction-tuned, and task fine-tuned. Across all experiments and batch sizes (4-80), SV consistently outperforms both SD and standard decoding with the target model. It improves SD performance by up to 2$\times$, with an average speedup of 1.4 $\times$ in large-batch settings (batch sizes 32-80). These results demonstrate SV's robustness, scalability, and practical utility for efficient LLM inference.

Serving large language models (LLMs) to millions of users requires efficient resource allocation and parallelism strategies. It is a labor intensive trial-and-error process to find such a strategy. We present BestServe, a novel framework for ranking serving strategies by estimating goodput under various operating scenarios. Supporting both collocated and disaggregated architectures, BestServe leverages an inference simulator built on an adapted roofline model and CPU-GPU dispatch dynamics. Our framework determines the optimal strategy in minutes on a single standard CPU, eliminating the need for costly benchmarking, while achieving predictions within a $20\%$ error margin. It appeals to be practical for rapid deployment planning because of its lightweight design and strong extensibility.

The rapid advancement of Large Language Models (LLMs) has driven the need for more efficient serving strategies. In this context, efficiency refers to the proportion of requests that meet their Service Level Objectives (SLOs), particularly for Time To First Token (TTFT) and Time Between Tokens (TBT). However, existing systems often prioritize one metric at the cost of the other. We present Ascendra, an LLM serving system designed to meet both TTFT and TBT SLOs simultaneously. The core insight behind Ascendra is that a request's urgency evolves as it approaches its deadline. To leverage this, Ascendra partitions GPU resources into two types of instances: low-priority and high-priority. Low-priority instances maximize throughput by processing requests out of arrival order, but at the risk of request starvation. To address this, Ascendra employs a performance model to predict requests at risk of missing their SLOs and proactively offloads them to high-priority instances. High-priority instances are optimized for low-latency execution and handle urgent requests nearing their deadlines. This partitioned architecture enables Ascendra to effectively balance high throughput and low latency. Extensive evaluation shows that Ascendra improves system throughput by up to 1.7x compared to vLLM and Sarathi-Serve while meeting both TTFT and TBT SLOs.

The rise of compound AI serving -- integrating multiple operators in a pipeline that may span edge and cloud tiers -- enables end-user applications such as autonomous driving, generative AI-powered meeting companions, and immersive gaming. Achieving high service goodput -- i.e., meeting service level objectives (SLOs) for pipeline latency, accuracy, and costs -- requires effective planning of operator placement, configuration, and resource allocation across infrastructure tiers. However, the diverse SLO requirements, varying edge capabilities, and high query volumes create an enormous planning search space, rendering current solutions fundamentally limited for real-time serving and cost-efficient deployments. This paper presents Circinus, an SLO-aware query planner for large-scale compound AI workloads. Circinus novelly decomposes multi-query planning and multi-dimensional SLO objectives while preserving global decision quality. By exploiting plan similarities within and across queries, it significantly reduces search steps. It further improves per-step efficiency with a precision-aware plan profiler that incrementally profiles and strategically applies early stopping based on imprecise estimates of plan performance. At scale, Circinus selects query-plan combinations to maximize global SLO goodput. Evaluations in real-world settings show that Circinus improves service goodput by 3.2-5.0$\times$, accelerates query planning by 4.2-5.8$\times$, achieving query response in seconds, while reducing deployment costs by 3.2-4.0$\times$ over state of the arts even in their intended single-tier deployments.

Recent years have seen the emergence of machine learning (ML) workloads deployed in warehouse-scale computing (WSC) settings, also known as ML fleets. As the computational demands placed on ML fleets have increased due to the rise of large models and growing demand for ML applications, it has become increasingly critical to measure and improve the efficiency of such systems. However, there is not yet an established methodology to characterize ML fleet performance and identify potential performance optimizations accordingly. This paper presents a large-scale analysis of an ML fleet based on Google's TPUs, introducing a framework to capture fleet-wide efficiency, systematically evaluate performance characteristics, and identify optimization strategies for the fleet. We begin by defining an ML fleet, outlining its components, and analyzing an example Google ML fleet in production comprising thousands of accelerators running diverse workloads. Our study reveals several critical insights: first, ML fleets extend beyond the hardware layer, with model, data, framework, compiler, and scheduling layers significantly impacting performance; second, the heterogeneous nature of ML fleets poses challenges in characterizing individual workload performance; and third, traditional utilization-based metrics prove insufficient for ML fleet characterization. To address these challenges, we present the "ML Productivity Goodput" (MPG) metric to measure ML fleet efficiency. We show how to leverage this metric to characterize the fleet across the ML system stack. We also present methods to identify and optimize performance bottlenecks using MPG, providing strategies for managing warehouse-scale ML systems in general. Lastly, we demonstrate quantitative evaluations from applying these methods to a real ML fleet for internal-facing Google TPU workloads, where we observed tangible improvements.

Reliability is a fundamental challenge in operating large-scale machine learning (ML) infrastructures, particularly as the scale of ML models and training clusters continues to grow. Despite decades of research on infrastructure failures, the impact of job failures across different scales remains unclear. This paper presents a view of managing two large, multi-tenant ML clusters, providing quantitative analysis, operational experience, and our own perspective in understanding and addressing reliability concerns at scale. Our analysis reveals that while large jobs are most vulnerable to failures, smaller jobs make up the majority of jobs in the clusters and should be incorporated into optimization objectives. We identify key workload properties, compare them across clusters, and demonstrate essential reliability requirements for pushing the boundaries of ML training at scale. We hereby introduce a taxonomy of failures and key reliability metrics, analyze 11 months of data from two state-of-the-art ML environments with over 150 million A100 GPU hours and 4 million jobs. Building on our data, we fit a failure model to project Mean Time to Failure for various GPU scales. We further propose a method to estimate a related metric, Effective Training Time Ratio, as a function of job parameters, and we use this model to gauge the efficacy of potential software mitigations at scale. Our work provides valuable insights and future research directions for improving the reliability of AI supercomputer clusters, emphasizing the need for flexible, workload-agnostic, and reliability-aware infrastructure, system software, and algorithms.

Large language models (LLMs) have achieved remarkable performance and are widely deployed in various applications, while the serving of LLM inference has raised concerns about user experience and serving throughput. Accordingly, service level objectives (SLOs) and goodput-the number of requests that meet SLOs per second-are introduced to evaluate the performance of LLM serving. However, existing metrics fail to capture the nature of user experience. We observe two ridiculous phenomena in existing metrics: 1) delaying token delivery can smooth the tail time between tokens (tail TBT) of a request and 2) dropping the request that fails to meet the SLOs midway can improve goodput. In this paper, we revisit SLO and goodput metrics in LLM serving and propose a unified metric framework smooth goodput including SLOs and goodput to reflect the nature of user experience in LLM serving. The framework can adapt to specific goals of different tasks by setting parameters. We re-evaluate the performance of different LLM serving systems under multiple workloads based on this unified framework and provide possible directions for future optimization of existing strategies. We hope that this framework can provide a unified standard for evaluating LLM serving and foster researches in the field of LLM serving optimization to move in a cohesive direction.

Reducing the inference latency of large language models (LLMs) is crucial, and speculative decoding (SD) stands out as one of the most effective techniques. Rather than letting the LLM generate all tokens directly, speculative decoding employs effective proxies to predict potential outputs, which are then verified by the LLM without compromising the generation quality. Yet, deploying SD in real online LLM serving systems (with continuous batching) does not always yield improvement -- under higher request rates or low speculation accuracy, it paradoxically increases latency. Furthermore, there is no best speculation length work for all workloads under different system loads. Based on the observations, we develop a dynamic framework SmartSpec. SmartSpec dynamically determines the best speculation length for each request (from 0, i.e., no speculation, to many tokens) -- hence the associated speculative execution costs -- based on a new metric called goodput, which characterizes the current observed load of the entire system and the speculation accuracy. We show that SmartSpec consistently reduces average request latency by up to 3.2x compared to non-speculative decoding baselines across different sizes of target models, draft models, request rates, and datasets. Moreover, SmartSpec can be applied to different styles of speculative decoding, including traditional, model-based approaches as well as model-free methods like prompt lookup and tree-style decoding.