Large language models (LLMs) deliver superior performance but require substantial computational resources and operate with relatively low efficiency, while smaller models can efficiently handle simpler tasks with fewer resources. LLM routing is a crucial paradigm that dynamically selects the most suitable large language models from a pool of candidates to process diverse inputs, ensuring optimal resource utilization while maintaining response quality. Existing routing frameworks typically model this as a locally optimal decision-making problem, selecting the presumed best-fit LLM for each query individually, which overlook global budget constraints, resulting in ineffective resource allocation. To tackle this problem, we introduce OmniRouter, a fundamentally controllable routing framework for multi-LLM serving. Instead of making per-query greedy choices, OmniRouter models the routing task as a constrained optimization problem, assigning models that minimize total cost while ensuring the required performance level. Specifically, a hybrid retrieval-augmented predictor is designed to predict the capabilities and costs of LLMs and a constrained optimizer is employed to control globally optimal query-model allocation. Experiments show that OmniRouter achieves up to 6.30% improvement in response accuracy while simultaneously reducing computational costs by at least 10.15% compared to competitive router baselines. The code and the dataset are available at https://github.com/agiresearch/OmniRouter.
Data and pipeline parallelism are ubiquitous for training of Large Language Models (LLM) on distributed nodes. Driven by the need for cost-effective training, recent work explores efficient communication arrangement for end to end training. Motivated by LLM's resistance to layer skipping and layer reordering, in this paper, we explore stage (several consecutive layers) skipping in pipeline training, and challenge the conventional practice of sequential pipeline execution. We derive convergence and throughput constraints (guidelines) for pipelining with skipping and swapping pipeline stages. Based on these constraints, we propose SkipPipe, the first partial pipeline framework to reduce the end-to-end training time for LLMs while preserving the convergence. The core of SkipPipe is a path scheduling algorithm that optimizes the paths for individual microbatches and reduces idle time (due to microbatch collisions) on the distributed nodes, complying with the given stage skipping ratio. We extensively evaluate SkipPipe on LLaMa models from 500M to 8B parameters on up to 20 nodes. Our results show that SkipPipe reduces training iteration time by up to $55\%$ compared to full pipeline. Our partial pipeline training also improves resistance to layer omission during inference, experiencing a drop in perplexity of only $7\%$ when running only half the model. Our code is available at https://github.com/gensyn-ai/skippipe.
Long-horizon combinatorial optimization problems (COPs), such as the Flexible Job-Shop Scheduling Problem (FJSP), often involve complex, interdependent decisions over extended time frames, posing significant challenges for existing solvers. While Rolling Horizon Optimization (RHO) addresses this by decomposing problems into overlapping shorter-horizon subproblems, such overlap often involves redundant computations. In this paper, we present L-RHO, the first learning-guided RHO framework for COPs. L-RHO employs a neural network to intelligently fix variables that in hindsight did not need to be re-optimized, resulting in smaller and thus easier-to-solve subproblems. For FJSP, this means identifying operations with unchanged machine assignments between consecutive subproblems. Applied to FJSP, L-RHO accelerates RHO by up to 54% while significantly improving solution quality, outperforming other heuristic and learning-based baselines. We also provide in-depth discussions and verify the desirable adaptability and generalization of L-RHO across numerous FJSP variates, distributions, online scenarios and benchmark instances. Moreover, we provide a theoretical analysis to elucidate the conditions under which learning is beneficial.
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.
This study explores the integration of Blackout Diffusion into the DIFUSCO framework for combinatorial optimization, specifically targeting the Traveling Salesman Problem (TSP). Inspired by the success of discrete-time diffusion models (D3PM) in maintaining structural integrity, we extend the paradigm to a continuous-time framework, leveraging the unique properties of Blackout Diffusion. Continuous-time modeling introduces smoother transitions and refined control, hypothesizing enhanced solution quality over traditional discrete methods. We propose three key improvements to enhance the diffusion process. First, we transition from a discrete-time-based model to a continuous-time framework, providing a more refined and flexible formulation. Second, we refine the observation time scheduling to ensure a smooth and linear transformation throughout the diffusion process, allowing for a more natural progression of states. Finally, building upon the second improvement, we further enhance the reverse process by introducing finer time slices in regions that are particularly challenging for the model, thereby improving accuracy and stability in the reconstruction phase. Although the experimental results did not exceed the baseline performance, they demonstrate the effectiveness of these methods in balancing simplicity and complexity, offering new insights into diffusion-based combinatorial optimization. This work represents the first application of Blackout Diffusion to combinatorial optimization, providing a foundation for further advancements in this domain. * The code is available for review at https://github.com/Giventicket/BlackoutDIFUSCO.
The exponential growth in model sizes has significantly increased the communication burden in Federated Learning (FL). Existing methods to alleviate this burden by transmitting compressed gradients often face high compression errors, which slow down the model's convergence. To simultaneously achieve high compression effectiveness and lower compression errors, we study the gradient compression problem from a novel perspective. Specifically, we propose a systematical algorithm termed Extended Single-Step Synthetic Features Compressing (E-3SFC), which consists of three sub-components, i.e., the Single-Step Synthetic Features Compressor (3SFC), a double-way compression algorithm, and a communication budget scheduler. First, we regard the process of gradient computation of a model as decompressing gradients from corresponding inputs, while the inverse process is considered as compressing the gradients. Based on this, we introduce a novel gradient compression method termed 3SFC, which utilizes the model itself as a decompressor, leveraging training priors such as model weights and objective functions. 3SFC compresses raw gradients into tiny synthetic features in a single-step simulation, incorporating error feedback to minimize overall compression errors. To further reduce communication overhead, 3SFC is extended to E-3SFC, allowing double-way compression and dynamic communication budget scheduling. Our theoretical analysis under both strongly convex and non-convex conditions demonstrates that 3SFC achieves linear and sub-linear convergence rates with aggregation noise. Extensive experiments across six datasets and six models reveal that 3SFC outperforms state-of-the-art methods by up to 13.4% while reducing communication costs by 111.6 times. These findings suggest that 3SFC can significantly enhance communication efficiency in FL without compromising model performance.
We show that learning-rate schedules for large model training behave surprisingly similar to a performance bound from non-smooth convex optimization theory. We provide a bound for the constant schedule with linear cooldown; in particular, the practical benefit of cooldown is reflected in the bound due to the absence of logarithmic terms. Further, we show that this surprisingly close match between optimization theory and practice can be exploited for learning-rate tuning: we achieve noticeable improvements for training 124M and 210M Llama-type models by (i) extending the schedule for continued training with optimal learning-rate, and (ii) transferring the optimal learning-rate across schedules.
This paper proposes Genetic Algorithm with Border Trades (GAB), a novel modification of the standard genetic algorithm that enhances exploration by incorporating new chromosome patterns in the breeding process. This approach significantly mitigates premature convergence and improves search diversity. Empirically, GAB achieves up to 8x higher fitness and 10x faster convergence on complex job scheduling problems compared to standard Genetic Algorithms, reaching average fitness scores of 888 versus 106 in under 20 seconds. On the classic Flip-Flop problem, GAB consistently finds optimal or near-optimal solutions in fewer generations, even as input sizes scale to thousands of bits. These results highlight GAB as a highly effective and computationally efficient alternative for solving large-scale combinatorial optimization problems.
Recent advances in neural models have shown considerable promise in solving Traveling Salesman Problems (TSPs) without relying on much hand-crafted engineering. However, while non-autoregressive (NAR) approaches benefit from faster inference through parallelism, they typically deliver solutions of inferior quality compared to autoregressive ones. To enhance the solution quality while maintaining fast inference, we propose DEITSP, a diffusion model with efficient iterations tailored for TSP that operates in a NAR manner. Firstly, we introduce a one-step diffusion model that integrates the controlled discrete noise addition process with self-consistency enhancement, enabling optimal solution prediction through simultaneous denoising of multiple solutions. Secondly, we design a dual-modality graph transformer to bolster the extraction and fusion of features from node and edge modalities, while further accelerating the inference with fewer layers. Thirdly, we develop an efficient iterative strategy that alternates between adding and removing noise to improve exploration compared to previous diffusion methods. Additionally, we devise a scheduling framework to progressively refine the solution space by adjusting noise levels, facilitating a smooth search for optimal solutions. Extensive experiments on real-world and large-scale TSP instances demonstrate that DEITSP performs favorably against existing neural approaches in terms of solution quality, inference latency, and generalization ability. Our code is available at $\href{https://github.com/DEITSP/DEITSP}{https://github.com/DEITSP/DEITSP}$.
We consider an online channel scheduling problem for a single transmitter-receiver pair equipped with $N$ arbitrarily varying wireless channels. The transmission rates of the channels might be non-stationary and could be controlled by an oblivious adversary. At every slot, incoming data arrives at an infinite-capacity data queue located at the transmitter. A scheduler, which is oblivious to the current channel rates, selects one of the $N$ channels for transmission. At the end of the slot, the scheduler only gets to know the transmission rate of the selected channel. The objective is to minimize the queue length regret, defined as the difference between the queue length at some time $T$ achieved by an online policy and the queue length obtained by always transmitting over the single best channel in hindsight. We propose a weakly adaptive Multi-Armed Bandit (MAB) algorithm for minimizing the queue length regret in this setup. Unlike previous works, we do not make any stability assumptions about the queue or the arrival process. Hence, our result holds even when the queueing process is unstable. Our main observation is that the queue length regret can be upper bounded by the regret of a MAB policy that competes against the best channel in hindsight uniformly over all sub-intervals of $[T]$. As a technical contribution of independent interest, we then propose a weakly adaptive adversarial MAB policy which achieves $\tilde{O}(\sqrt{N}T^{\frac{3}{4}})$ regret with high probability, implying the same bound for queue length regret.
Recent advances in Large Language Models (LLMs) have enabled the development of Video-LLMs, advancing multimodal learning by bridging video data with language tasks. However, current video understanding models struggle with processing long video sequences, supporting multi-turn dialogues, and adapting to real-world dynamic scenarios. To address these issues, we propose StreamChat, a training-free framework for streaming video reasoning and conversational interaction. $\StreamChat$ leverages a novel hierarchical memory system to efficiently process and compress video features over extended sequences, enabling real-time, multi-turn dialogue. Our framework incorporates a parallel system scheduling strategy that enhances processing speed and reduces latency, ensuring robust performance in real-world applications. Furthermore, we introduce StreamBench, a versatile benchmark that evaluates streaming video understanding across diverse media types and interactive scenarios, including multi-turn interactions and complex reasoning tasks. Extensive evaluations on StreamBench and other public benchmarks demonstrate that StreamChat significantly outperforms existing state-of-the-art models in terms of accuracy and response times, confirming its effectiveness for streaming video understanding. Code is available at StreamChat: https://github.com/hmxiong/StreamChat.
Large language models (LLMs) are remarked by their substantial computational requirements. To mitigate the cost, researchers develop specialized CUDA kernels, which often fuse several tensor operations to maximize the utilization of GPUs as much as possible. However, those specialized kernels may still leave performance on the table as CUDA assembly experts show that manual optimization of GPU SASS schedules can lead to better performance, and trial-and-error is largely employed to manually find the best GPU SASS schedules. In this work, we employ an automatic approach to optimize GPU SASS schedules, which thus can be integrated into existing compiler frameworks. The key to automatic optimization is training an RL agent to mimic how human experts perform manual scheduling. To this end, we formulate an assembly game, where RL agents can play to find the best GPU SASS schedules. The assembly game starts from a \textit{-O3} optimized SASS schedule, and the RL agents can iteratively apply actions to mutate the current schedules. Positive rewards are generated if the mutated schedules get higher throughput by executing on GPUs. Experiments show that CuAsmRL can further improve the performance of existing specialized CUDA kernels transparently by up to $26\%$, and on average $9\%$. Moreover, it is used as a tool to reveal potential optimization moves learned automatically.
Vehicle-to-Infrastructure (V2I) technology enables information exchange between vehicles and road infrastructure. Specifically, when a vehicle approaches a roadside unit (RSU), it can exchange information with the RSU to obtain accurate data that assists in driving. With the release of the 3rd Generation Partnership Project (3GPP) Release 16, which includes the 5G New Radio (NR) Vehicle-to-Everything (V2X) standards, vehicles typically adopt mode-2 communication using sensing-based semi-persistent scheduling (SPS) for resource allocation. In this approach, vehicles identify candidate resources within a selection window and exclude ineligible resources based on information from a sensing window. However, vehicles often drive at different speeds, resulting in varying amounts of data transmission with RSUs as they pass by, which leads to unfair access. Therefore, it is essential to design an access scheme that accounts for different vehicle speeds to achieve fair access across the network. This paper formulates an optimization problem for vehicular networks and proposes a multi-objective optimization scheme to address it by adjusting the selection window in the SPS mechanism of 5G NR V2I mode-2. Simulation results demonstrate the effectiveness of the proposed scheme
The increasing demand for flexible and efficient urban transportation solutions has spotlighted the limitations of traditional Demand Responsive Transport (DRT) systems, particularly in accommodating diverse passenger needs and dynamic urban environments. Autonomous Mobility-on-Demand (AMoD) systems have emerged as a promising alternative, leveraging connected and autonomous vehicles (CAVs) to provide responsive and adaptable services. However, existing methods primarily focus on either vehicle scheduling or path planning, which often simplify complex urban layouts and neglect the necessity for simultaneous coordination and mutual avoidance among CAVs. This oversimplification poses significant challenges to the deployment of AMoD systems in real-world scenarios. To address these gaps, we propose CoDriveVLM, a novel framework that integrates high-fidelity simultaneous dispatching and cooperative motion planning for future AMoD systems. Our method harnesses Vision-Language Models (VLMs) to enhance multi-modality information processing, and this enables comprehensive dispatching and collision risk evaluation. The VLM-enhanced CAV dispatching coordinator is introduced to effectively manage complex and unforeseen AMoD conditions, thus supporting efficient scheduling decision-making. Furthermore, we propose a scalable decentralized cooperative motion planning method via consensus alternating direction method of multipliers (ADMM) focusing on collision risk evaluation and decentralized trajectory optimization. Simulation results demonstrate the feasibility and robustness of CoDriveVLM in various traffic conditions, showcasing its potential to significantly improve the fidelity and effectiveness of AMoD systems in future urban transportation networks. The code is available at https://github.com/henryhcliu/CoDriveVLM.git.
For self-scheduling cascaded hydropower (S-CHP) facilities, medium-term planning is a critical step that coordinates water availability over the medium-term horizon, providing water usage guidance for their short-term operations in wholesale market participation. Typically, medium-term planning strategies (e.g., reservoir storage targets at the end of each short-term period) are determined by either optimization methods or rules of thumb. However, with the integration of variable renewable energy sources (VRESs), optimization-based methods suffer from deviations between the anticipated and actual reservoir storage, while rules of thumb could be financially conservative, thereby compromising short-term operating profitability in wholesale market participation. This paper presents a deep reinforcement learning (DRL)-based framework to derive medium-term planning policies for VRES-integrated S-CHPs (VS-CHPs), which can leverage contextual information underneath individual short-term periods and train planning policies by their induced short-term operating profits in wholesale market participation. The proposed DRL-based framework offers two practical merits. First, its planning strategies consider both seasonal requirements of reservoir storage and needs for short-term operating profits. Second, it adopts a multi-parametric programming-based strategy to accelerate the expensive training process associated with multi-step short-term operations. Finally, the DRL-based framework is evaluated on a real-world VS-CHP, demonstrating its advantages over current practice.
Residential load forecasting (RLF) is crucial for resource scheduling in power systems. Most existing methods utilize all given load records (dense data) to indiscriminately extract the dependencies between historical and future time series. However, there exist important regular patterns residing in the event-related associations among different appliances (sparse knowledge), which have yet been ignored. In this paper, we propose an Event-Response Knowledge Guided approach (ERKG) for RLF by incorporating the estimation of electricity usage events for different appliances, mining event-related sparse knowledge from the load series. With ERKG, the event-response estimation enables portraying the electricity consumption behaviors of residents, revealing regular variations in appliance operational states. To be specific, ERKG consists of knowledge extraction and guidance: i) a forecasting model is designed for the electricity usage events by estimating appliance operational states, aiming to extract the event-related sparse knowledge; ii) a novel knowledge-guided mechanism is established by fusing such state estimates of the appliance events into the RLF model, which can give particular focuses on the patterns of users' electricity consumption behaviors. Notably, ERKG can flexibly serve as a plug-in module to boost the capability of existing forecasting models by leveraging event response. In numerical experiments, extensive comparisons and ablation studies have verified the effectiveness of our ERKG, e.g., over 8% MAE can be reduced on the tested state-of-the-art forecasting models.
Transformers, driven by attention mechanisms, form the foundation of large language models (LLMs). As these models scale up, efficient GPU attention kernels become essential for high-throughput and low-latency inference. Diverse LLM applications demand flexible and high-performance attention solutions. We present FlashInfer: a customizable and efficient attention engine for LLM serving. FlashInfer tackles KV-cache storage heterogeneity using block-sparse format and composable formats to optimize memory access and reduce redundancy. It also offers a customizable attention template, enabling adaptation to various settings through Just-In-Time (JIT) compilation. Additionally, FlashInfer's load-balanced scheduling algorithm adjusts to dynamism of user requests while maintaining compatibility with CUDAGraph which requires static configuration. FlashInfer have been integrated into leading LLM serving frameworks like SGLang, vLLM and MLC-Engine. Comprehensive kernel-level and end-to-end evaluations demonstrate FlashInfer's ability to significantly boost kernel performance across diverse inference scenarios: compared to state-of-the-art LLM serving solutions, FlashInfer achieve 29-69% inter-token-latency reduction compared to compiler backends for LLM serving benchmark, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.
Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: \href{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.
Artificial Intelligence Generated Content (AIGC) has gained significant popularity for creating diverse content. Current AIGC models primarily focus on content quality within a centralized framework, resulting in a high service delay and negative user experiences. However, not only does the workload of an AIGC task depend on the AIGC model's complexity rather than the amount of data, but the large model and its multi-layer encoder structure also result in a huge demand for computational and memory resources. These unique characteristics pose new challenges in its modeling, deployment, and scheduling at edge networks. Thus, we model an offloading problem among edges for providing real AIGC services and propose LAD-TS, a novel Latent Action Diffusion-based Task Scheduling method that orchestrates multiple edge servers for expedited AIGC services. The LAD-TS generates a near-optimal offloading decision by leveraging the diffusion model's conditional generation capability and the reinforcement learning's environment interaction ability, thereby minimizing the service delays under multiple resource constraints. Meanwhile, a latent action diffusion strategy is designed to guide decision generation by utilizing historical action probability, enabling rapid achievement of near-optimal decisions. Furthermore, we develop DEdgeAI, a prototype edge system with a refined AIGC model deployment to implement and evaluate our LAD-TS method. DEdgeAI provides a real AIGC service for users, demonstrating up to 29.18% shorter service delays than the current five representative AIGC platforms. We release our open-source code at https://github.com/ChangfuXu/DEdgeAI/.
Speech brain-computer interfaces aim to decipher what a person is trying to say from neural activity alone, restoring communication to people with paralysis who have lost the ability to speak intelligibly. The Brain-to-Text Benchmark '24 and associated competition was created to foster the advancement of decoding algorithms that convert neural activity to text. Here, we summarize the lessons learned from the competition ending on June 1, 2024 (the top 4 entrants also presented their experiences in a recorded webinar). The largest improvements in accuracy were achieved using an ensembling approach, where the output of multiple independent decoders was merged using a fine-tuned large language model (an approach used by all 3 top entrants). Performance gains were also found by improving how the baseline recurrent neural network (RNN) model was trained, including by optimizing learning rate scheduling and by using a diphone training objective. Improving upon the model architecture itself proved more difficult, however, with attempts to use deep state space models or transformers not yet appearing to offer a benefit over the RNN baseline. The benchmark will remain open indefinitely to support further work towards increasing the accuracy of brain-to-text algorithms.