ASPEN also achieves high resource utilization and memory reuse by letting each resource asynchronously traverse depthwise in the DNN graph to its full computing potential.

Serving large generative models such as LLMs and multi- modal transformers requires balancing user-facing SLOs (e.g., time-to-first-token, time-between-tokens) with provider goals of efficiency and cost reduction. Existing solutions rely on static provisioning or model-level autoscaling, both of which treat the model as a monolith. This coarse-grained resource management leads to degraded performance or significant resource underutilization due to poor adaptability to dynamic inference traffic that is common online. The root cause of this inefficiency lies in the internal structure of generative models: they are executed as graphs of interconnected operators. Through detailed characterization and systematic analysis, we find that operators are heterogeneous in their compute and memory footprints and exhibit diverse sensitivity to workload and resource factors such as batch size, sequence length, and traffic rate. This heterogeneity suggests that the operator, rather than the entire model, is the right granularity for scaling decisions. We propose an operator-level autoscaling framework, which allocates resources at finer (operator)-granularity, optimizing the scaling, batching, and placement based on individual operator profiles. Evaluated on production-scale traces, our approach preserves SLOs with up to 40% fewer GPUs and 35% less energy, or under fixed resources achieves 1.6x higher throughput with 5% less energy. These results show that the operator, rather than the model, is fundamentally a more effective unit for scaling large generative workloads.

The growing demand for computational resources in machine learning has made efficient resource allocation a critical challenge, especially in heterogeneous hardware clusters where devices vary in capability, age, and energy efficiency. Upgrading to the latest hardware is often infeasible, making sustainable use of existing, mixed-generation resources essential. In this paper, we propose a learning-based architecture for managing machine learning workloads in heterogeneous clusters. The system operates online, allocating resources to incoming training or inference requests while minimizing energy consumption and meeting performance requirements. It uses two neural networks: the first provides initial estimates of how well a new model will utilize different hardware types and how it will affect co-located models. An optimizer then allocates resources based on these estimates. After deployment, the system monitors real performance and uses this data to refine its predictions via a second neural network. This updated model improves estimates not only for the current hardware but also for hardware not initially allocated and for co-location scenarios not yet observed. The result is an adaptive, iterative approach that learns over time to make more effective resource allocation decisions in heterogeneous deep learning clusters.

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.

ASPEN also achieves high resource utilization and memory reuse by letting each resource asynchronously traverse depthwise in the DNN graph to its full computing potential.

Modern GPUs evolve rapidly, yet production compilers still rely on hand-crafted register allocation heuristics that require substantial re-tuning for each hardware generation. We introduce VeriLocc, a framework that combines large language models (LLMs) with formal compiler techniques to enable generalizable and verifiable register allocation across GPU architectures. VeriLocc fine-tunes an LLM to translate intermediate representations (MIRs) into target-specific register assignments, aided by static analysis for cross-architecture normalization and generalization and a verifier-guided regeneration loop to ensure correctness. Evaluated on matrix multiplication (GEMM) and multi-head attention (MHA), VeriLocc achieves 85-99% single-shot accuracy and near-100% pass@100. Case study shows that VeriLocc discovers more performant assignments than expert-tuned libraries, outperforming rocBLAS by over 10% in runtime.

We propose ELIS, a serving system for Large Language Models (LLMs) featuring an Iterative Shortest Remaining Time First (ISRTF) scheduler designed to efficiently manage inference tasks with the shortest remaining tokens. Current LLM serving systems often employ a first-come-first-served scheduling strategy, which can lead to the "head-of-line blocking" problem. To overcome this limitation, it is necessary to predict LLM inference times and apply a shortest job first scheduling strategy. However, due to the auto-regressive nature of LLMs, predicting the inference latency is challenging. ELIS addresses this challenge by training a response length predictor for LLMs using the BGE model, an encoder-based state-of-the-art model. Additionally, we have devised the ISRTF scheduling strategy, an optimization of shortest remaining time first tailored to existing LLM iteration batching. To evaluate our work in an industrial setting, we simulate streams of requests based on our study of real-world user LLM serving trace records. Furthermore, we implemented ELIS as a cloud-native scheduler system on Kubernetes to evaluate its performance in production environments. Our experimental results demonstrate that ISRTF reduces the average job completion time by up to 19.6%.

The proliferation of the Internet of Things (IoT) such as smartphones, wearables, drones, and smart speakers, as well as the gigantic amount of data they capture, have revolutionized the way we work, live, and interact with the world. Equipped with sensing, computing, networking, and communication capabilities, these devices are able to collect, analyze and transmit a wide range of data including images, videos, audio, texts, wireless signals, physiological signals from individuals and the physical world. In recent years, advancements in Artificial Intelligence (AI), particularly in deep learning (DL)/deep neural network (DNN), foundation models, and Generative AI, have propelled the integration of AI with IoT, making the concept of Artificial Intelligence of Things (AIoT) a reality. The synergy between IoT and modern AI enhances decision making, improves human-machine interactions, and facilitates more efficient operations, making AIoT one of the most exciting and promising areas that have the potential to fundamentally transform how people perceive and interact with the world. As illustrated in Figure 1, at its core, AIoT is grounded on three key components: sensing, computing, and networking & communication.

With the recent popularity of neural networks comes the need for efficient serving of inference workloads. A neural network inference workload can be represented as a computational graph with nodes as operators transforming multidimensional tensors. The tensors can be transposed and/or tiled in a combinatorially large number of ways, some configurations leading to accelerated inference. We propose TGraph, a neural graph architecture that allows screening for fast configurations of the target computational graph, thus representing an artificial intelligence (AI) tensor compiler in contrast to the traditional heuristics-based compilers. The proposed solution improves mean Kendall's $\tau$ across layout collections of TpuGraphs from 29.8% of the reliable baseline to 67.4% of TGraph. We estimate the potential CO$_2$ emission reduction associated with our work to be equivalent to over 50% of the total household emissions in the areas hosting AI-oriented data centers.

Large Language Models (LLMs) have demonstrated remarkable capabilities in important tasks such as natural language understanding, language generation, and complex reasoning and have the potential to make a substantial impact on our society. Such capabilities, however, come with the considerable resources they demand, highlighting the strong need to develop effective techniques for addressing their efficiency challenges.In this survey, we provide a systematic and comprehensive review of efficient LLMs research. We organize the literature in a taxonomy consisting of three main categories, covering distinct yet interconnected efficient LLMs topics from model-centric, data-centric, and framework-centric perspective, respectively. We have also created a GitHub repository where we compile the papers featured in this survey at https://github.com/AIoT-MLSys-Lab/Efficient-LLMs-Survey, and will actively maintain this repository and incorporate new research as it emerges. We hope our survey can serve as a valuable resource to help researchers and practitioners gain a systematic understanding of the research developments in efficient LLMs and inspire them to contribute to this important and exciting field.