With the development of large language models (LLMs), it has become increasingly important to optimize hardware usage and improve throughput. In this paper, we study the inference optimization of the serving system that deploys LLMs. To optimize system throughput and maximize hardware utilization, we formulate the inference optimization problem as a mixed-integer programming (MIP) model and propose a hybrid offline-online method as solution. The offline method improves large-scale inference systems by introducing a Minimizing Makespan Bin Packing Problem. We further provide a theoretical lower bound computation method. Then, we propose an online sorting and preemptive scheduling method to better utilize hardware. In the online iteration scheduling process, a Lagrangian method is applied to evaluate the cost efficiency of inserting prefill stages versus decode stages at each iteration and dynamically determine when to preempt decoding tasks and insert prefill tasks. Experiments using real-world data from the LLaMA-65B model and the GSM8K dataset demonstrate that system utilization improves from 80.2% to 89.1%, and the total inference time decreases from 201.00 to 190.58 seconds. A 100-cases study shows that our method consistently outperforms the baseline method and improves the utilization rate by 8.0% on average. Finally, we discuss potential future extensions, including stochastic modeling, reinforcement learning-based schedulers, and dynamic decision-making strategies for system throughput and hardware utilization.
Transformers and large language models (LLMs) have revolutionized machine learning, with attention mechanisms at the core of their success. As the landscape of attention variants expands, so too do the challenges of optimizing their performance, particularly across different hardware platforms. Current optimization strategies are often narrowly focused, requiring extensive manual intervention to accommodate changes in model configurations or hardware environments. In this paper, we introduce AttentionEngine, a comprehensive framework designed to streamline the optimization of attention mechanisms across heterogeneous hardware backends. By decomposing attention computation into modular operations with customizable components, AttentionEngine enables flexible adaptation to diverse algorithmic requirements. The framework further automates kernel optimization through a combination of programmable templates and a robust cross-platform scheduling strategy. Empirical results reveal performance gains of up to 10x on configurations beyond the reach of existing methods. AttentionEngine offers a scalable, efficient foundation for developing and deploying attention mechanisms with minimal manual tuning. Our code has been open-sourced and is available at https://github.com/microsoft/AttentionEngine.
Large language models (LLMs) are widely used for natural language understanding and text generation. An LLM model relies on a time-consuming step called LLM decoding to generate output tokens. Several prior works focus on improving the performance of LLM decoding using parallelism techniques, such as batching and speculative decoding. State-of-the-art LLM decoding has both compute-bound and memory-bound kernels. Some prior works statically identify and map these different kernels to a heterogeneous architecture consisting of both processing-in-memory (PIM) units and computation-centric accelerators. We observe that characteristics of LLM decoding kernels (e.g., whether or not a kernel is memory-bound) can change dynamically due to parameter changes to meet user and/or system demands, making (1) static kernel mapping to PIM units and computation-centric accelerators suboptimal, and (2) one-size-fits-all approach of designing PIM units inefficient due to a large degree of heterogeneity even in memory-bound kernels. In this paper, we aim to accelerate LLM decoding while considering the dynamically changing characteristics of the kernels involved. We propose PAPI (PArallel Decoding with PIM), a PIM-enabled heterogeneous architecture that exploits dynamic scheduling of compute-bound or memory-bound kernels to suitable hardware units. PAPI has two key mechanisms: (1) online kernel characterization to dynamically schedule kernels to the most suitable hardware units at runtime and (2) a PIM-enabled heterogeneous computing system that harmoniously orchestrates both computation-centric processing units and hybrid PIM units with different computing capabilities. Our experimental results on three broadly-used LLMs show that PAPI achieves 1.8$\times$ and 11.1$\times$ speedups over a state-of-the-art heterogeneous LLM accelerator and a state-of-the-art PIM-only LLM accelerator, respectively.
Decarbonization plans promote the transition to heat pumps (HPs), creating new opportunities for their energy flexibility in demand response programs, solar photovoltaic integration and optimization of distribution networks. This paper reviews scheduling-based and real-time optimization methods for controlling HPs with a focus on energy flexibility in distribution networks. Scheduling-based methods fall into two categories: rule-based controllers (RBCs), which rely on predefined control rules without explicitly seeking optimal solutions, and optimization models, which are designed to determine the optimal scheduling of operations. Real-time optimization is achieved through model predictive control (MPC), which relies on a predictive model to optimize decisions over a time horizon, and reinforcement learning (RL), which takes a model-free approach by learning optimal strategies through direct interaction with the environment. The paper also examines studies on the impact of HPs on distribution networks, particularly those leveraging energy flexibility strategies. Key takeaways suggest the need to validate control strategies for extreme cold-weather regions that require backup heaters, as well as develop approaches designed for demand charge schemes that integrate HPs with other controllable loads. From a grid impact assessment perspective, studies have focused primarily on RBCs for providing energy flexibility through HP operation, without addressing more advanced methods such as real-time optimization using MPC or RL-based algorithms. Incorporating these advanced control strategies could help identify key limitations, including the impact of varying user participation levels and the cost-benefit trade-offs associated with their implementation.
Pioneering companies such as Waymo have deployed robo-taxi services in several U.S. cities. These robo-taxis are electric vehicles, and their operations require the joint optimization of ride matching, vehicle repositioning, and charging scheduling in a stochastic environment. We model the operations of the ride-hailing system with robo-taxis as a discrete-time, average reward Markov Decision Process with infinite horizon. As the fleet size grows, the dispatching is challenging as the set of system state and the fleet dispatching action set grow exponentially with the number of vehicles. To address this, we introduce a scalable deep reinforcement learning algorithm, called Atomic Proximal Policy Optimization (Atomic-PPO), that reduces the action space using atomic action decomposition. We evaluate our algorithm using real-world NYC for-hire vehicle data and we measure the performance using the long-run average reward achieved by the dispatching policy relative to a fluid-based reward upper bound. Our experiments demonstrate the superior performance of our Atomic-PPO compared to benchmarks. Furthermore, we conduct extensive numerical experiments to analyze the efficient allocation of charging facilities and assess the impact of vehicle range and charger speed on fleet performance.
In light of the rapid expansion of civil aviation, addressing the delays and congestion phenomena in the vicinity of metroplex caused by the imbalance between air traffic flow and capacity is crucial. This paper first proposes a bi-level optimization model for the collaborative flight sequencing of arrival and departure flights in the metroplex with multiple airports, considering both the runway systems and TMA (Terminal Control Area) entry/exit fixes. Besides, the model is adaptive to various traffic scenarios. The genetic algorithm is employed to solve the proposed model. The Shanghai TMA, located in China, is used as a case study, and it includes two airports, Shanghai Hongqiao International Airport and Shanghai Pudong International Airport. The results demonstrate that the model can reduce arrival delay by 51.52%, departure delay by 18.05%, and the runway occupation time of departure flights by 23.83%. Furthermore, the model utilized in this study significantly enhances flight scheduling efficiency, providing a more efficient solution than the traditional FCFS (First Come, First Served) approach. Additionally, the algorithm employed offers further improvements over the NSGA II algorithm.
This paper presents a hierarchical framework for demand response optimization in air separation units (ASUs) that combines reinforcement learning (RL) with linear model predictive control (LMPC). We investigate two control architectures: a direct RL approach and a control-informed methodology where an RL agent provides setpoints to a lower-level LMPC. The proposed RL-LMPC framework demonstrates improved sample efficiency during training and better constraint satisfaction compared to direct RL control. Using an industrial ASU case study, we show that our approach successfully manages operational constraints while optimizing electricity costs under time-varying pricing. Results indicate that the RL-LMPC architecture achieves comparable economic performance to direct RL while providing better robustness and requiring fewer training samples to converge. The framework offers a practical solution for implementing flexible operation strategies in process industries, bridging the gap between data-driven methods and traditional control approaches.
The Aircraft Landing Problem (ALP) is one of the challenging problems in aircraft transportation and management. The challenge is to schedule the arriving aircraft in a sequence so that the cost and delays are optimized. There are various solution approaches to solving this problem, most of which are based on operations research algorithms and meta-heuristics. Although traditional methods perform better on one or the other factors, there remains a problem of solving real-time rescheduling and computational scalability altogether. This paper presents a novel deep reinforcement learning (DRL) framework that combines graph neural networks with actor-critic architectures to address the ALP. This paper introduces three key contributions: A graph-based state representation that efficiently captures temporal and spatial relationships between aircraft, a specialized actor-critic architecture designed to handle multiple competing objectives in landing scheduling, and a runway balance strategy that ensures efficient resource utilization while maintaining safety constraints. The results show that the trained algorithm can be tested on different problem sets and the results are competitive to operation research algorithms. The experimental results on standard benchmark data sets demonstrate a 99.95% reduction in computational time compared to Mixed Integer Programming (MIP) and 38% higher runway throughput over First Come First Serve (FCFS) approaches. Therefore, the proposed solution is competitive to traditional approaches and achieves substantial advancements. Notably, it does not require retraining, making it particularly suitable for industrial deployment. The frameworks capability to generate solutions within 1 second enables real-time rescheduling, addressing critical requirements of air traffic management.
As large language models (LLMs) become integral to code-related tasks, a central question emerges: do LLMs truly understand program execution semantics? We introduce EquiBench, a new benchmark for evaluating LLMs through equivalence checking, i.e., determining whether two programs produce identical outputs for all possible inputs. Unlike prior code generation benchmarks, this task directly tests a model's understanding of code execution semantics. EquiBench consists of 2400 program pairs across four languages and six categories. These pairs are generated through program analysis, compiler scheduling, and superoptimization, ensuring high-confidence labels, nontrivial difficulty, and full automation. The transformations span syntactic edits, structural modifications, and algorithmic changes, covering a broad spectrum of semantic variation. We evaluate 19 state-of-the-art LLMs and find that in the most challenging categories, the best accuracies are 63.8% and 76.2%, only modestly above the 50% random baseline. Further analysis reveals that models often rely on syntactic similarity rather than exhibiting robust reasoning over execution semantics, highlighting fundamental limitations.
Large language model (LLM) applications are evolving beyond simple chatbots into dynamic, general-purpose agentic programs, which scale LLM calls and output tokens to help AI agents reason, explore, and solve complex tasks. However, existing LLM serving systems ignore dependencies between programs and calls, missing significant opportunities for optimization. Our analysis reveals that programs submitted to LLM serving engines experience long cumulative wait times, primarily due to head-of-line blocking at both the individual LLM request and the program. To address this, we introduce Autellix, an LLM serving system that treats programs as first-class citizens to minimize their end-to-end latencies. Autellix intercepts LLM calls submitted by programs, enriching schedulers with program-level context. We propose two scheduling algorithms-for single-threaded and distributed programs-that preempt and prioritize LLM calls based on their programs' previously completed calls. Our evaluation demonstrates that across diverse LLMs and agentic workloads, Autellix improves throughput of programs by 4-15x at the same latency compared to state-of-the-art systems, such as vLLM.
Heuristics have achieved great success in solv- ing combinatorial optimization problems (COPs). However, heuristics designed by humans re- quire too much domain knowledge and testing time. Given the fact that Large Language Mod- els (LLMs) possess strong capabilities to under- stand and generate content, and a knowledge base that covers various domains, which offer a novel way to automatically optimize heuristics. There- fore, we propose Planning of Heuristics (PoH), an optimization method that integrates the self- reflection of LLMs with the Monte Carlo Tree Search (MCTS), a well-known planning algo- rithm. PoH iteratively refines generated heuristics by evaluating their performance and providing im- provement suggestions. Our method enables to it- eratively evaluate the generated heuristics (states) and improve them based on the improvement sug- gestions (actions) and evaluation results (rewards), by effectively simulating future states to search for paths with higher rewards. In this paper, we apply PoH to solve the Traveling Salesman Prob- lem (TSP) and the Flow Shop Scheduling Prob- lem (FSSP). The experimental results show that PoH outperforms other hand-crafted heuristics and Automatic Heuristic Design (AHD) by other LLMs-based methods, and achieves the signifi- cant improvements and the state-of-the-art per- formance of our proposed method in automating heuristic optimization with LLMs to solve COPs.
Automating planning with LLMs presents transformative opportunities for traditional industries, yet remains underexplored. In commercial construction, the complexity of automated scheduling often requires manual intervention to ensure precision. We propose CONSTRUCTA, a novel framework leveraging LLMs to optimize construction schedules in complex projects like semiconductor fabrication. CONSTRUCTA addresses key challenges by: (1) integrating construction-specific knowledge through static RAG; (2) employing context-sampling techniques inspired by architectural expertise to provide relevant input; and (3) deploying Construction DPO to align schedules with expert preferences using RLHF. Experiments on proprietary data demonstrate performance improvements of +42.3% in missing value prediction, +79.1% in dependency analysis, and +28.9% in automated planning compared to baseline methods, showcasing its potential to revolutionize construction workflows and inspire domain-specific LLM advancements.
In the astronomical observation field, determining the allocation of observation resources of the telescope array and planning follow-up observations for targets of opportunity (ToOs) are indispensable components of astronomical scientific discovery. This problem is computationally challenging, given the online observation setting and the abundance of time-varying factors that can affect whether an observation can be conducted. This paper presents ROARS, a reinforcement learning approach for online astronomical resource-constrained scheduling. To capture the structure of the astronomical observation scheduling, we depict every schedule using a directed acyclic graph (DAG), illustrating the dependency of timing between different observation tasks within the schedule. Deep reinforcement learning is used to learn a policy that can improve the feasible solution by iteratively local rewriting until convergence. It can solve the challenge of obtaining a complete solution directly from scratch in astronomical observation scenarios, due to the high computational complexity resulting from numerous spatial and temporal constraints. A simulation environment is developed based on real-world scenarios for experiments, to evaluate the effectiveness of our proposed scheduling approach. The experimental results show that ROARS surpasses 5 popular heuristics, adapts to various observation scenarios and learns effective strategies with hindsight.
Stand-alone perception systems in autonomous driving suffer from limited sensing ranges and occlusions at extended distances, potentially resulting in catastrophic outcomes. To address this issue, collaborative perception is envisioned to improve perceptual accuracy by using vehicle-to-everything (V2X) communication to enable collaboration among connected and autonomous vehicles and roadside units. However, due to limited communication resources, it is impractical for all units to transmit sensing data such as point clouds or high-definition video. As a result, it is essential to optimize the scheduling of communication links to ensure efficient spectrum utilization for the exchange of perceptual data. In this work, we propose a deep reinforcement learning-based V2X user scheduling algorithm for collaborative perception. Given the challenges in acquiring perceptual labels, we reformulate the conventional label-dependent objective into a label-free goal, based on characteristics of 3D object detection. Incorporating both channel state information (CSI) and semantic information, we develop a double deep Q-Network (DDQN)-based user scheduling framework for collaborative perception, named SchedCP. Simulation results verify the effectiveness and robustness of SchedCP compared with traditional V2X scheduling methods. Finally, we present a case study to illustrate how our proposed algorithm adaptively modifies the scheduling decisions by taking both instantaneous CSI and perceptual semantics into account.
Effective parking management is essential for ensuring accessibility, safety, and convenience in master-planned communities, particularly in active adult neighborhoods experiencing rapid growth. Accurately assessing parking utilization is a crucial first step in planning for future demand, but data collection methods can be costly and labor-intensive. This paper presents a low-cost yet highly accurate methodology for measuring parking utilization using road tubes connected to portable traffic counters from JAMAR Technologies, Inc. By integrating results from JAMAR's analysis tool with custom Python scripting, the methodology enables precise parking lot counts through parameter optimization and automated error correction. The system's efficiency allows for scalable deployment without significant manual observation, reducing both costs and disruptions to daily operations. Using Tellico Village as a case study, this research demonstrates that community planners can obtain actionable parking insights on a limited budget, empowering them to make informed decisions about capacity expansion, traffic flow improvements, and facility scheduling. The findings underscore the feasibility of leveraging cost-effective technology to optimize infrastructure planning and ensure long-term resident satisfaction as communities grow.
The rapid rise of large language models (LLMs) in text streaming services has introduced significant cost and Quality of Experience (QoE) challenges in serving millions of daily requests, especially in meeting Time-To-First-Token (TTFT) and Time-Between-Token (TBT) requirements for real-time interactions. Our real-world measurements show that both server-based and on-device deployments struggle to meet diverse QoE demands: server deployments face high costs and last-hop issues (e.g., Internet latency and dynamics), while on-device LLM inference is constrained by resources. We introduce DiSCo, a device-server cooperative scheduler designed to optimize users' QoE by adaptively routing requests and migrating response generation between endpoints while maintaining cost constraints. DiSCo employs cost-aware scheduling, leveraging the predictable speed of on-device LLM inference with the flexible capacity of server-based inference to dispatch requests on the fly, while introducing a token-level migration mechanism to ensure consistent token delivery during migration. Evaluations on real-world workloads -- including commercial services like OpenAI GPT and DeepSeek, and open-source deployments such as LLaMA3 -- show that DiSCo can improve users' QoE by reducing tail TTFT (11-52\%) and mean TTFT (6-78\%) across different model-device configurations, while dramatically reducing serving costs by up to 84\% through its migration mechanism while maintaining comparable QoE levels.
Combinatorial optimization problems, such as scheduling and route planning, are crucial in various industries but are computationally intractable due to their NP-hard nature. Neural Combinatorial Optimization methods leverage machine learning to address these challenges but often depend on sequential decision-making, which is prone to error accumulation as small mistakes propagate throughout the process. Inspired by self-evaluation techniques in Large Language Models, we propose a novel framework that generates and evaluates subsets of assignments, moving beyond traditional stepwise approaches. Applied to the Job-Shop Scheduling Problem, our method integrates a heterogeneous graph neural network with a Transformer to build a policy model and a self-evaluation function. Experimental validation on challenging, well-known benchmarks demonstrates the effectiveness of our approach, surpassing state-of-the-art methods.
The integration of renewable energy sources in microgrids introduces significant operational challenges due to their intermittent nature and the mismatch between generation and demand patterns. Effective demand response (DR) strategies are crucial for maintaining system stability and economic efficiency, particularly in microgrids with high renewable penetration. This paper presents a comprehensive mixed-integer linear programming (MILP) framework for optimizing DR operations in a microgrid with solar generation and battery storage systems. The framework incorporates load classification, dynamic price thresholding, and multi-period coordination for optimal DR event scheduling. Analysis across seven distinct operational scenarios demonstrates consistent peak load reduction of 10\% while achieving energy cost savings ranging from 13.1\% to 38.0\%. The highest performance was observed in scenarios with high solar generation, where the framework achieved 38.0\% energy cost reduction through optimal coordination of renewable resources and DR actions. The results validate the framework's effectiveness in managing diverse operational challenges while maintaining system stability and economic efficiency.
In our daily lives and industrial settings, we often encounter dynamic problems that require reasoning over time and metric constraints. These include tasks such as scheduling, routing, and production sequencing. Dynamic logics have traditionally addressed these needs but often lack the flexibility and integration required for comprehensive problem modeling. This research aims to extend Answer Set Programming (ASP), a powerful declarative problem-solving approach, to handle dynamic domains effectively. By integrating concepts from dynamic, temporal, and metric logics into ASP, we seek to develop robust systems capable of modeling complex dynamic problems and performing efficient reasoning tasks, thereby enhancing ASPs applicability in industrial contexts.
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.