Mixture of Experts (MoE) LLMs face significant obstacles due to their massive parameter scale, which imposes memory, storage, and deployment challenges. Although recent expert merging methods promise greater efficiency by consolidating multiple experts, they are fundamentally hindered by parameter conflicts arising from expert specialization. In this paper, we present Sub-MoE, a novel MoE compression framework via Subspace Expert Merging. Our key insight is to perform joint Singular Value Decomposition (SVD) on concatenated expert weights, reducing conflicting parameters by extracting shared $U$-matrices while enabling effective merging of the expert-specific $V$ components. Specifically, Sub-MoE consists of two innovative phases: (1) Adaptive Expert Clustering, which groups functionally coherent experts via K-means clustering based on cosine similarity of expert outputs; and (2) Subspace Expert Merging, which first enforces Experts Union Decomposition to derive the shared $U$-matrix across experts in the same group, then pursues frequency-based merging for individual $V$-matrices, and finalizes expert reconstruction using the merged $V$-matrix. In this way, we align and fuse experts in a shared subspace, and can be extended with intra-expert compression for further inference optimization. Extensive experiments on Mixtral, DeepSeek, and Qwen-1.5|3 MoE LLMs demonstrate that our Sub-MoE significantly outperforms existing expert pruning and merging methods. Notably, our Sub-MoE maintains 96\%|86\% of original performance with 25\%|50\% expert reduction on Mixtral-8x7B in zero-shot benchmarks. Code will be released at https://github.com/lliai/MoERazor.
Large Language Model (LLM) libraries have emerged as the foundational infrastructure powering today's AI revolution, serving as the backbone for LLM deployment, inference optimization, fine-tuning, and production serving across diverse applications. Despite their critical role in the LLM ecosystem, these libraries face frequent quality issues and bugs that threaten the reliability of AI systems built upon them. To address this knowledge gap, we present the first comprehensive empirical investigation into bug characteristics and testing practices in modern LLM libraries. We examine 313 bug-fixing commits extracted across two widely-adopted LLM libraries: HuggingFace Transformers and vLLM.Through rigorous manual analysis, we establish comprehensive taxonomies categorizing bug symptoms into 5 types and root causes into 14 distinct categories.Our primary discovery shows that API misuse has emerged as the predominant root cause (32.17%-48.19%), representing a notable transition from algorithm-focused defects in conventional deep learning frameworks toward interface-oriented problems. Additionally, we examine 7,748 test functions to identify 7 distinct test oracle categories employed in current testing approaches, with predefined expected outputs (such as specific tensors and text strings) being the most common strategy. Our assessment of existing testing effectiveness demonstrates that the majority of bugs escape detection due to inadequate test cases (41.73%), lack of test drivers (32.37%), and weak test oracles (25.90%). Drawing from these findings, we offer some recommendations for enhancing LLM library quality assurance.
A significant portion of the energy consumed by Large Language Models (LLMs) arises from their inference processes; hence developing energy-efficient methods for inference is crucial. While several techniques exist for inference optimization, output compression remains relatively unexplored, with only a few preliminary efforts addressing this aspect. In this work, we first benchmark 12 decoder-only LLMs across 5 datasets, revealing that these models often produce responses that are substantially longer than necessary. We then conduct a comprehensive quality assessment of LLM responses, formally defining six information categories present in LLM responses. We show that LLMs often tend to include redundant or additional information besides the minimal answer. To address this issue of long responses by LLMs, we explore several simple and intuitive prompt-engineering strategies. Empirical evaluation shows that appropriate prompts targeting length reduction and controlling information content can achieve significant energy optimization between 25-60\% by reducing the response length while preserving the quality of LLM responses.
Large language model (LLM) agents have shown promising performance in generating code for solving complex data science problems. Recent studies primarily focus on enhancing in-context learning through improved search, sampling, and planning techniques, while overlooking the importance of the order in which problems are tackled during inference. In this work, we develop a novel inference-time optimization framework, referred to as DSMentor, which leverages curriculum learning -- a strategy that introduces simpler task first and progressively moves to more complex ones as the learner improves -- to enhance LLM agent performance in challenging data science tasks. Our mentor-guided framework organizes data science tasks in order of increasing difficulty and incorporates a growing long-term memory to retain prior experiences, guiding the agent's learning progression and enabling more effective utilization of accumulated knowledge. We evaluate DSMentor through extensive experiments on DSEval and QRData benchmarks. Experiments show that DSMentor using Claude-3.5-Sonnet improves the pass rate by up to 5.2% on DSEval and QRData compared to baseline agents. Furthermore, DSMentor demonstrates stronger causal reasoning ability, improving the pass rate by 8.8% on the causality problems compared to GPT-4 using Program-of-Thoughts prompts. Our work underscores the importance of developing effective strategies for accumulating and utilizing knowledge during inference, mirroring the human learning process and opening new avenues for improving LLM performance through curriculum-based inference optimization.
Sparse Mixture of Experts (MoE) large language models (LLMs) are gradually becoming the mainstream approach for ultra-large-scale models. Existing optimization efforts for MoE models have focused primarily on coarse-grained MoE architectures. With the emergence of DeepSeek Models, fine-grained MoE models are gaining popularity, yet research on them remains limited. Therefore, we want to discuss the efficiency dynamic under different service loads. Additionally, fine-grained models allow deployers to reduce the number of routed experts, both activated counts and total counts, raising the question of how this reduction affects the trade-off between MoE efficiency and performance. Our findings indicate that while deploying MoE models presents greater challenges, it also offers significant optimization opportunities. Reducing the number of activated experts can lead to substantial efficiency improvements in certain scenarios, with only minor performance degradation. Reducing the total number of experts provides limited efficiency gains but results in severe performance degradation. Our method can increase throughput by at least 10\% without any performance degradation. Overall, we conclude that MoE inference optimization remains an area with substantial potential for exploration and improvement.
This work presents SimpleAR, a vanilla autoregressive visual generation framework without complex architecure modifications. Through careful exploration of training and inference optimization, we demonstrate that: 1) with only 0.5B parameters, our model can generate 1024x1024 resolution images with high fidelity, and achieve competitive results on challenging text-to-image benchmarks, e.g., 0.59 on GenEval and 79.66 on DPG; 2) both supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) training could lead to significant improvements on generation aesthectics and prompt alignment; and 3) when optimized with inference acceleraton techniques like vLLM, the time for SimpleAR to generate an 1024x1024 image could be reduced to around 14 seconds. By sharing these findings and open-sourcing the code, we hope to reveal the potential of autoregressive visual generation and encourage more participation in this research field. Code is available at https://github.com/wdrink/SimpleAR.
Large Language Models (LLMs) are indispensable in today's applications, but their inference procedure -- generating responses by processing text in segments and using a memory-heavy Key-Value (KV) cache -- demands significant computational resources, particularly under memory constraints. This paper formulates LLM inference optimization as a multi-stage online scheduling problem where sequential prompt arrivals and KV cache growth render conventional scheduling ineffective. We develop a fluid dynamics approximation to provide a tractable benchmark that guides algorithm design. Building on this, we propose the Waiting for Accumulated Inference Threshold (WAIT) algorithm, which uses multiple thresholds to schedule incoming prompts optimally when output lengths are known, and extend it to Nested WAIT for cases with unknown output lengths. Theoretical analysis shows that both algorithms achieve near-optimal performance against the fluid benchmark in heavy traffic conditions, balancing throughput, latency, and Time to First Token (TTFT). Experiments with the Llama-7B model on an A100 GPU using both synthetic and real-world datasets demonstrate improved throughput and reduced latency relative to established baselines like vLLM and Sarathi. This work bridges operations research and machine learning, offering a rigorous framework for the efficient deployment of LLMs under memory constraints.
Motion expression video segmentation is designed to segment objects in accordance with the input motion expressions. In contrast to the conventional Referring Video Object Segmentation (RVOS), it places emphasis on motion as well as multi-object expressions, making it more arduous. Recently, Large Multimodal Models (LMMs) have begun to shine in RVOS due to their powerful vision-language perception capabilities. In this work, we propose a simple and effective inference optimization method to fully unleash the potential of LMMs in referring video segmentation. Firstly, we use Sa2VA as our baseline, which is a unified LMM for dense grounded understanding of both images and videos. Secondly, we uniformly sample the video frames during the inference process to enhance the model's understanding of the entire video. Finally, we integrate the results of multiple expert models to mitigate the erroneous predictions of a single model. Our solution achieved 61.98% J&F on the MeViS test set and ranked 1st place in the 4th PVUW Challenge MeViS Track at CVPR 2025.
The increasing computational demands of transformer models in time series classification necessitate effective optimization strategies for energy-efficient deployment. Our study presents a systematic investigation of optimization techniques, focusing on structured pruning and quantization methods for transformer architectures. Through extensive experimentation on three distinct datasets (RefrigerationDevices, ElectricDevices, and PLAID), we quantitatively evaluate model performance and energy efficiency across different transformer configurations. Our experimental results demonstrate that static quantization reduces energy consumption by 29.14% while maintaining classification performance, and L1 pruning achieves a 63% improvement in inference speed with minimal accuracy degradation. Our findings provide valuable insights into the effectiveness of optimization strategies for transformer-based time series classification, establishing a foundation for efficient model deployment in resource-constrained environments.
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.
The rapid development of deep neural networks (DNNs) is inherently accompanied by the problem of high computational costs. To tackle this challenge, dynamic voltage frequency scaling (DVFS) is emerging as a promising technology for balancing the latency and energy consumption of DNN inference by adjusting the computing frequency of processors. However, most existing models of DNN inference time are based on the CPU-DVFS technique, and directly applying the CPU-DVFS model to DNN inference on GPUs will lead to significant errors in optimizing latency and energy consumption. In this paper, we propose a DVFS-aware latency model to precisely characterize DNN inference time on GPUs. We first formulate the DNN inference time based on extensive experiment results for different devices and analyze the impact of fitting parameters. Then by dividing DNNs into multiple blocks and obtaining the actual inference time, the proposed model is further verified. Finally, we compare our proposed model with the CPU-DVFS model in two specific cases. Evaluation results demonstrate that local inference optimization with our proposed model achieves a reduction of no less than 66% and 69% in inference time and energy consumption respectively. In addition, cooperative inference with our proposed model can improve the partition policy and reduce the energy consumption compared to the CPU-DVFS model.
We investigate a family of gradient flows of positive and probability measures, focusing on the Hellinger-Kantorovich (HK) geometry, which unifies transport mechanism of Otto-Wasserstein, and the birth-death mechanism of Hellinger (or Fisher-Rao). A central contribution is a complete characterization of global exponential decay behaviors of entropy functionals (e.g. KL, $\chi^2$) under Otto-Wasserstein and Hellinger-type gradient flows. In particular, for the more challenging analysis of HK gradient flows on positive measures -- where the typical log-Sobolev arguments fail -- we develop a specialized shape-mass decomposition that enables new analysis results. Our approach also leverages the (Polyak-)\L{}ojasiewicz-type functional inequalities and a careful extension of classical dissipation estimates. These findings provide a unified and complete theoretical framework for gradient flows and underpin applications in computational algorithms for statistical inference, optimization, and machine learning.
The emergence of large-scale Mixture of Experts (MoE) models represents a significant advancement in artificial intelligence, offering enhanced model capacity and computational efficiency through conditional computation. However, deploying and running inference on these models presents significant challenges in computational resources, latency, and energy efficiency. This comprehensive survey analyzes optimization techniques for MoE models across the entire system stack. We first establish a taxonomical framework that categorizes optimization approaches into model-level, system-level, and hardware-level optimizations. At the model level, we examine architectural innovations including efficient expert design, attention mechanisms, various compression techniques such as pruning, quantization, and knowledge distillation, as well as algorithm improvement including dynamic routing strategies and expert merging methods. At the system level, we investigate distributed computing approaches, load balancing mechanisms, and efficient scheduling algorithms that enable scalable deployment. Furthermore, we delve into hardware-specific optimizations and co-design strategies that maximize throughput and energy efficiency. This survey provides both a structured overview of existing solutions and identifies key challenges and promising research directions in MoE inference optimization. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE inference optimization research, we have established a repository accessible at https://github.com/MoE-Inf/awesome-moe-inference/.
Machine learning models deployed on edge devices have enabled numerous exciting new applications, such as humanoid robots, AR glasses, and autonomous vehicles. However, the computing resources available on these edge devices are not catching up with the ever-growing number of parameters in these models. As the models become bigger and more complicated, the novel yet sophisticated structure challenges the inference runtime optimization. We present FluidML, a generic runtime memory management and optimization framework that can flexibly transform the model execution blueprint to achieve faster and more memory-efficient inference. Evaluations across different platforms show that FluidML can consistently reduce the end-to-end inference latency by up to 25.38% for popular language models and reduce peak memory usage by up to 41.47%, compared to state-of-the-art approaches. FluidML is of ~30K line of codes, built for general-purpose usage, and will be released as an open-source inference runtime optimization framework to the community.
Recent research has challenged the necessity of complex deep learning architectures for time series forecasting, demonstrating that simple linear models can often outperform sophisticated approaches. Building upon this insight, we introduce a novel architecture the Temporal Linear Net (TLN), that extends the capabilities of linear models while maintaining interpretability and computational efficiency. TLN is designed to effectively capture both temporal and feature-wise dependencies in multivariate time series data. Our approach is a variant of TSMixer that maintains strict linearity throughout its architecture. TSMixer removes activation functions, introduces specialized kernel initializations, and incorporates dilated convolutions to handle various time scales, while preserving the linear nature of the model. Unlike transformer-based models that may lose temporal information due to their permutation-invariant nature, TLN explicitly preserves and leverages the temporal structure of the input data. A key innovation of TLN is its ability to compute an equivalent linear model, offering a level of interpretability not found in more complex architectures such as TSMixer. This feature allows for seamless conversion between the full TLN model and its linear equivalent, facilitating both training flexibility and inference optimization.
The evolving capabilities of large language models are accompanied by growing sizes and deployment costs, necessitating effective inference optimisation techniques. We propose a novel pruning method utilising centrality measures from graph theory, reducing both the computational requirements and the memory footprint of these models. Specifically, we devise a method for creating a weighted directed acyclical graph representation of multilayer perceptrons to which we apply a modified version of the weighted PageRank centrality measure to compute node importance scores. In combination with uniform pruning this leads to structured sparsity. We call this pruning method MLPRank. Furthermore we introduce an extension to decoder-only transformer models and call it LLMRank. For both variants we demonstrate a strong performance. With MLPRank on average leading to 6.09 % higher accuracy retention than three popular baselines and 13.42 % with LLMRank compared to two popular baselines. Code is available at https://github.com/amazon-science/llm-rank-pruning.
Balancing mutually diverging performance metrics, such as, processing latency, outcome accuracy, and end device energy consumption is a challenging undertaking for deep learning model inference in ad-hoc edge environments. In this paper, we propose EdgeRL framework that seeks to strike such balance by using an Advantage Actor-Critic (A2C) Reinforcement Learning (RL) approach that can choose optimal run-time DNN inference parameters and aligns the performance metrics based on the application requirements. Using real world deep learning model and a hardware testbed, we evaluate the benefits of EdgeRL framework in terms of end device energy savings, inference accuracy improvement, and end-to-end inference latency reduction.
This paper works on Binary Neural Networks (BNNs), a promising avenue for efficient deep learning, offering significant reductions in computational overhead and memory footprint to full precision networks. However, the challenge of energy-intensive training and the drop in performance have been persistent issues. Tackling the challenge, prior works focus primarily on task-related inference optimization. Unlike prior works, this study offers an innovative methodology integrating BNNs with cyclic precision training, introducing the CycleBNN. This approach is designed to enhance training efficiency while minimizing the loss in performance. By dynamically adjusting precision in cycles, we achieve a convenient trade-off between training efficiency and model performance. This emphasizes the potential of our method in energy-constrained training scenarios, where data is collected onboard and paves the way for sustainable and efficient deep learning architectures. To gather insights on CycleBNN's efficiency, we conduct experiments on ImageNet, CIFAR-10, and PASCAL-VOC, obtaining competitive performances while using 96.09\% less operations during training on ImageNet, 88.88\% on CIFAR-10 and 96.09\% on PASCAL-VOC. Finally, CycleBNN offers a path towards faster, more accessible training of efficient networks, accelerating the development of practical applications. The PyTorch code is available at \url{https://github.com/fedeloper/CycleBNN/}
Sparse Mixture of Expert (SMoE) models have emerged as a scalable alternative to dense models in language modeling. These models use conditionally activated feedforward subnetworks in transformer blocks, allowing for a separation between total model parameters and per-example computation. However, large token-routed SMoE models face a significant challenge: during inference, the entire model must be used for a sequence or a batch, resulting in high latencies in a distributed setting that offsets the advantages of per-token sparse activation. Our research explores task-specific model pruning to inform decisions about designing SMoE architectures, mainly modulating the choice of expert counts in pretraining. We investigate whether such pruned models offer advantages over smaller SMoE models trained from scratch, when evaluating and comparing them individually on tasks. To that end, we introduce an adaptive task-aware pruning technique UNCURL to reduce the number of experts per MoE layer in an offline manner post-training. Our findings reveal a threshold pruning factor for the reduction that depends on the number of experts used in pretraining, above which, the reduction starts to degrade model performance. These insights contribute to our understanding of model design choices when pretraining with SMoE architectures, particularly useful when considering task-specific inference optimization for later stages.
This report examines the fine-tuning of Large Language Models (LLMs), integrating theoretical insights with practical applications. It outlines the historical evolution of LLMs from traditional Natural Language Processing (NLP) models to their pivotal role in AI. A comparison of fine-tuning methodologies, including supervised, unsupervised, and instruction-based approaches, highlights their applicability to different tasks. The report introduces a structured seven-stage pipeline for fine-tuning LLMs, spanning data preparation, model initialization, hyperparameter tuning, and model deployment. Emphasis is placed on managing imbalanced datasets and optimization techniques. Parameter-efficient methods like Low-Rank Adaptation (LoRA) and Half Fine-Tuning are explored for balancing computational efficiency with performance. Advanced techniques such as memory fine-tuning, Mixture of Experts (MoE), and Mixture of Agents (MoA) are discussed for leveraging specialized networks and multi-agent collaboration. The report also examines novel approaches like Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO), which align LLMs with human preferences, alongside pruning and routing optimizations to improve efficiency. Further sections cover validation frameworks, post-deployment monitoring, and inference optimization, with attention to deploying LLMs on distributed and cloud-based platforms. Emerging areas such as multimodal LLMs, fine-tuning for audio and speech, and challenges related to scalability, privacy, and accountability are also addressed. This report offers actionable insights for researchers and practitioners navigating LLM fine-tuning in an evolving landscape.