Large multimodal models (LMMs) demonstrate impressive capabilities in understanding images, videos, and audio beyond text. However, efficiently serving LMMs in production environments poses significant challenges due to their complex architectures and heterogeneous characteristics across their multi-stage inference pipelines. We present the first comprehensive systems analysis of two prominent LMM architectures, decoder-only and cross-attention, across six representative open-source models, revealing key systems design implications. We also present an in-depth analysis of production LMM inference traces, uncovering unique workload characteristics, including variable, heavy-tailed request distributions and bursty traffic patterns. Based on these insights, we propose ModServe, a modular LMM serving system that decouples stages for independent optimization and adaptive scaling. ModServe dynamically reconfigures stages and handles bursty traffic with modality-aware scheduling and autoscaling to meet tail latency SLOs while minimizing costs. ModServe achieves 3.3-5.5x higher throughput (leading to 25-41.3% cost saving) while meeting SLOs on a 128-GPU cluster with production traces.
In the context of object detection, sliding-window classifiers and single-shot Convolutional Neural Network (CNN) meta-architectures typically yield multiple overlapping candidate windows with similar high scores around the true location of a particular object. Non-Maximum Suppression (NMS) is the process of selecting a single representative candidate within this cluster of detections, so as to obtain a unique detection per object appearing on a given picture. In this paper, we present a highly scalable NMS algorithm for embedded GPU architectures that is designed from scratch to handle workloads featuring thousands of simultaneous detections on a given picture. Our kernels are directly applicable to other sequential NMS algorithms such as FeatureNMS, Soft-NMS or AdaptiveNMS that share the inner workings of the classic greedy NMS method. The obtained performance results show that our parallel NMS algorithm is capable of clustering 1024 simultaneous detected objects per frame in roughly 1 ms on both NVIDIA Tegra X1 and NVIDIA Tegra X2 on-die GPUs, while taking 2 ms on NVIDIA Tegra K1. Furthermore, our proposed parallel greedy NMS algorithm yields a 14x-40x speed up when compared to state-of-the-art NMS methods that require learning a CNN from annotated data.
The substantial computational and memory demands of Large Language Models (LLMs) hinder their deployment. Block Floating Point (BFP) has proven effective in accelerating linear operations, a cornerstone of LLM workloads. However, as sequence lengths grow, nonlinear operations, such as Attention, increasingly become performance bottlenecks due to their quadratic computational complexity. These nonlinear operations are predominantly executed using inefficient floating-point formats, which renders the system challenging to optimize software efficiency and hardware overhead. In this paper, we delve into the limitations and potential of applying BFP to nonlinear operations. Given our findings, we introduce a hardware-software co-design framework (DB-Attn), including: (i) DBFP, an advanced BFP version, overcomes nonlinear operation challenges with a pivot-focus strategy for diverse data and an adaptive grouping strategy for flexible exponent sharing. (ii) DH-LUT, a novel lookup table algorithm dedicated to accelerating nonlinear operations with DBFP format. (iii) An RTL-level DBFP-based engine is implemented to support DB-Attn, applicable to FPGA and ASIC. Results show that DB-Attn provides significant performance improvements with negligible accuracy loss, achieving 74% GPU speedup on Softmax of LLaMA and 10x low overhead performance improvement over SOTA designs.
Large language models exhibit a remarkable capacity in language generation and comprehension. These advances enable AI systems to produce more human-like and emotionally engaging text. However, these models rely on a large number of parameters, requiring significant computational resources for training and inference. In some scenarios, accessing these resources can be challenging (e.g., budget or hardware limitations). Techniques like reducing precision bits can make models more memory-efficient, reducing the computational resources needed, at the cost of reduced accuracy. This paper addresses the trade-off between different quantization values, GPU RAM utilization, and text quality in affective text generation (e.g., "I really enjoy running in the snow-covered forest"). To evaluate, we use an emotion classifier and ten seed prompts to generate affective text. We test three setups of precision bits (8, 16, and 32) across five open-weight language models from two different families. Our findings demonstrate that bit reductions lead to memory savings, achieving a reduction of 76%. However, this optimization comes with a trade-off, leading to a decrease of up to 10 pp in F1 score for larger models and an increase of 10 pp for smaller models, along with roughly double the inference time. In terms of text quality, larger models at lower quantization levels generally outperform smaller, higher-precision models -- while requiring similar memory.
Zeroth-order optimization (ZO) has demonstrated remarkable promise in efficient fine-tuning tasks for Large Language Models (LLMs). In particular, recent advances incorporate the low-rankness of gradients, introducing low-rank ZO estimators to further reduce GPU memory consumption. However, most existing works focus solely on the low-rankness of each individual gradient, overlooking a broader property shared by all gradients throughout the training, i.e., all gradients approximately reside within a similar subspace. In this paper, we consider two factors together and propose a novel low-rank ZO estimator, TeZO, which captures the low-rankness across both the model and temporal dimension. Specifically, we represent ZO perturbations along the temporal dimension as a 3D tensor and employ Canonical Polyadic Decomposition (CPD) to extract each low-rank 2D matrix, significantly reducing the training cost. TeZO can also be easily extended to the Adam variant while consuming less memory than MeZO-SGD, and requiring about only 35% memory of MeZO-Adam. Both comprehensive theoretical analysis and extensive experimental research have validated its efficiency, achieving SOTA-comparable results with lower overhead of time and memory.
Efficient real-world deployments of large language models (LLMs) rely on Key-Value (KV) caching for processing and generating long outputs, reducing the need for repetitive computation. For large contexts, Key-Value caches can take up tens of gigabytes of device memory, as they store vector representations for each token and layer. Recent work has shown that the cached vectors can be compressed through quantization, pruning or merging, but these techniques often compromise quality towards higher compression rates. In this work, we aim to improve Key & Value compression by exploiting two observations: 1) the inherent dependencies between keys and values across different layers, and 2) high-compression mechanisms for internal network states. We propose AQUA-KV, an adaptive quantization for Key-Value caches that relies on compact adapters to exploit existing dependencies between Keys and Values, and aims to "optimally" compress the information that cannot be predicted. AQUA-KV significantly improves compression rates, while maintaining high accuracy on state-of-the-art LLM families. On Llama 3.2 LLMs, we achieve near-lossless inference at 2-2.5 bits per value with under $1\%$ relative error in perplexity and LongBench scores. AQUA-KV is one-shot, simple, and efficient: it can be calibrated on a single GPU within 1-6 hours, even for 70B models.
This study aims to enlarge our current knowledge on application of brain-inspired network science principles for training artificial neural networks (ANNs) with sparse connectivity. Dynamic sparse training (DST) can reduce the computational demands in ANNs, but faces difficulties to keep peak performance at high sparsity levels. The Cannistraci-Hebb training (CHT) is a brain-inspired method for growing connectivity in DST. CHT leverages a gradient-free, topology-driven link regrowth, which has shown ultra-sparse (1% connectivity or lower) advantage across various tasks compared to fully connected networks. Yet, CHT suffers two main drawbacks: (i) its time complexity is O(Nd^3) - N node network size, d node degree - hence it can apply only to ultra-sparse networks. (ii) it selects top link prediction scores, which is inappropriate for the early training epochs, when the network presents unreliable connections. We propose a GPU-friendly approximation of the CH link predictor, which reduces the computational complexity to O(N^3), enabling a fast implementation of CHT in large-scale models. We introduce the Cannistraci-Hebb training soft rule (CHTs), which adopts a strategy for sampling connections in both link removal and regrowth, balancing the exploration and exploitation of network topology. To improve performance, we integrate CHTs with a sigmoid gradual density decay (CHTss). Empirical results show that, using 1% of connections, CHTs outperforms fully connected networks in MLP on visual classification tasks, compressing some networks to < 30% nodes. Using 5% of the connections, CHTss outperforms fully connected networks in two Transformer-based machine translation tasks. Using 30% of the connections, CHTss achieves superior performance compared to other dynamic sparse training methods in language modeling, and it surpasses the fully connected counterpart in zero-shot evaluations.
The rapid growth of Large Language Models has driven demand for effective model compression techniques to reduce memory and computation costs. Low-rank pruning has gained attention for its tensor coherence and GPU compatibility across all densities. However, low-rank pruning has struggled to match the performance of semi-structured pruning, often doubling perplexity (PPL) at similar densities. In this paper, we propose Pivoting Factorization (PIFA), a novel lossless meta low-rank representation that unsupervisedly learns a compact form of any low-rank representation, effectively eliminating redundant information. PIFA identifies pivot rows (linearly independent rows) and expresses non-pivot rows as linear combinations, achieving an additional 24.2\% memory savings and 24.6\% faster inference over low-rank layers at r/d = 0.5, thereby significantly enhancing performance at the same density. To mitigate the performance degradation caused by low-rank pruning, we introduce a novel, retraining-free low-rank reconstruction method that minimizes error accumulation (M). MPIFA, combining M and PIFA into an end-to-end framework, significantly outperforms existing low-rank pruning methods and, for the first time, achieves performance comparable to semi-structured pruning, while surpassing it in GPU efficiency and compatibility.
Recent advancements in diffusion models have been leveraged to address inverse problems without additional training, and Diffusion Posterior Sampling (DPS) (Chung et al., 2022a) is among the most popular approaches. Previous analyses suggest that DPS accomplishes posterior sampling by approximating the conditional score. While in this paper, we demonstrate that the conditional score approximation employed by DPS is not as effective as previously assumed, but rather aligns more closely with the principle of maximizing a posterior (MAP). This assertion is substantiated through an examination of DPS on 512x512 ImageNet images, revealing that: 1) DPS's conditional score estimation significantly diverges from the score of a well-trained conditional diffusion model and is even inferior to the unconditional score; 2) The mean of DPS's conditional score estimation deviates significantly from zero, rendering it an invalid score estimation; 3) DPS generates high-quality samples with significantly lower diversity. In light of the above findings, we posit that DPS more closely resembles MAP than a conditional score estimator, and accordingly propose the following enhancements to DPS: 1) we explicitly maximize the posterior through multi-step gradient ascent and projection; 2) we utilize a light-weighted conditional score estimator trained with only 100 images and 8 GPU hours. Extensive experimental results indicate that these proposed improvements significantly enhance DPS's performance. The source code for these improvements is provided in https://github.com/tongdaxu/Rethinking-Diffusion-Posterior-Sampling-From-Conditional-Score-Estimator-to-Maximizing-a-Posterior.
Quality-Diversity optimization comprises a family of evolutionary algorithms aimed at generating a collection of diverse and high-performing solutions. MAP-Elites (ME), a notable example, is used effectively in fields like evolutionary robotics. However, the reliance of ME on random mutations from Genetic Algorithms limits its ability to evolve high-dimensional solutions. Methods proposed to overcome this include using gradient-based operators like policy gradients or natural evolution strategies. While successful at scaling ME for neuroevolution, these methods often suffer from slow training speeds, or difficulties in scaling with massive parallelization due to high computational demands or reliance on centralized actor-critic training. In this work, we introduce a fast, sample-efficient ME based algorithm capable of scaling up with massive parallelization, significantly reducing runtimes without compromising performance. Our method, ASCII-ME, unlike existing policy gradient quality-diversity methods, does not rely on centralized actor-critic training. It performs behavioral variations based on time step performance metrics and maps these variations to solutions using policy gradients. Our experiments show that ASCII-ME can generate a diverse collection of high-performing deep neural network policies in less than 250 seconds on a single GPU. Additionally, it operates on average, five times faster than state-of-the-art algorithms while still maintaining competitive sample efficiency.
In this methods article, we provide a flexible but easy-to-use implementation of Direct Coupling Analysis (DCA) based on Boltzmann machine learning, together with a tutorial on how to use it. The package \texttt{adabmDCA 2.0} is available in different programming languages (C++, Julia, Python) usable on different architectures (single-core and multi-core CPU, GPU) using a common front-end interface. In addition to several learning protocols for dense and sparse generative DCA models, it allows to directly address common downstream tasks like residue-residue contact prediction, mutational-effect prediction, scoring of sequence libraries and generation of artificial sequences for sequence design. It is readily applicable to protein and RNA sequence data.
Spot spraying represents an efficient and sustainable method for reducing the amount of pesticides, particularly herbicides, used in agricultural fields. To achieve this, it is of utmost importance to reliably differentiate between crops and weeds, and even between individual weed species in situ and under real-time conditions. To assess suitability for real-time application, different object detection models that are currently state-of-the-art are compared. All available models of YOLOv8, YOLOv9, YOLOv10, and RT-DETR are trained and evaluated with images from a real field situation. The images are separated into two distinct datasets: In the initial data set, each species of plants is trained individually; in the subsequent dataset, a distinction is made between monocotyledonous weeds, dicotyledonous weeds, and three chosen crops. The results demonstrate that while all models perform equally well in the metrics evaluated, the YOLOv9 models, particularly the YOLOv9s and YOLOv9e, stand out in terms of their strong recall scores (66.58 % and 72.36 %), as well as mAP50 (73.52 % and 79.86 %), and mAP50-95 (43.82 % and 47.00 %) in dataset 2. However, the RT-DETR models, especially RT-DETR-l, excel in precision with reaching 82.44 \% on dataset 1 and 81.46 % in dataset 2, making them particularly suitable for scenarios where minimizing false positives is critical. In particular, the smallest variants of the YOLO models (YOLOv8n, YOLOv9t, and YOLOv10n) achieve substantially faster inference times down to 7.58 ms for dataset 2 on the NVIDIA GeForce RTX 4090 GPU for analyzing one frame, while maintaining competitive accuracy, highlighting their potential for deployment in resource-constrained embedded computing devices as typically used in productive setups.
Tree-based Genetic Programming (TGP) is a key evolutionary algorithm widely used in symbolic regression, feature engineering, and scientific modeling. Its high computational demands make GPU acceleration essential for scalable and high-performance evolutionary computation. However, GPU acceleration of TGP faces three key challenges: inefficient tree encoding, highly heterogeneous genetic operations, and limited parallelism in fitness evaluation. To address these challenges, we introduce EvoGP, a comprehensive GPU-accelerated TGP framework. First, we design a tensorized encoding scheme to represent tree with different structures as tensors with the same shape, optimizing memory access and enabling efficient parallel execution. Second, we propose a unified parallel framework for genetic operations by leveraging shared computational primitives and implementing dedicated CUDA kernels for scalable performance. Third, we present a fully parallel fitness evaluation strategy for symbolic regression, exploiting both population-level and data-level parallelism to maximize GPU utilization. Moreover, we implement a comprehensive library to provide rich algorithm operators and benchmark problems. EvoGP is extensively tested on various tasks, including symbolic regression, classification, and robotics control, demonstrating its versatility and effectiveness across diverse application scenarios. Experimental results show that EvoGP achieves up to a 140.89x speedup over the state-of-the-art GPU-based TGP implementation, while maintaining or exceeding the accuracy of baseline methods. EvoGP is open-source and accessible at: https://github.com/EMI-Group/evogp.
We present CrowdSplat, a novel approach that leverages 3D Gaussian Splatting for real-time, high-quality crowd rendering. Our method utilizes 3D Gaussian functions to represent animated human characters in diverse poses and outfits, which are extracted from monocular videos. We integrate Level of Detail (LoD) rendering to optimize computational efficiency and quality. The CrowdSplat framework consists of two stages: (1) avatar reconstruction and (2) crowd synthesis. The framework is also optimized for GPU memory usage to enhance scalability. Quantitative and qualitative evaluations show that CrowdSplat achieves good levels of rendering quality, memory efficiency, and computational performance. Through these experiments, we demonstrate that CrowdSplat is a viable solution for dynamic, realistic crowd simulation in real-time applications.
Recent advancements in fine-tuning Vision-Language Foundation Models (VLMs) have garnered significant attention for their effectiveness in downstream few-shot learning tasks.While these recent approaches exhibits some performance improvements, they often suffer from excessive training parameters and high computational costs. To address these challenges, we propose a novel Block matrix-based low-rank adaptation framework, called Block-LoRA, for fine-tuning VLMs on downstream few-shot tasks. Inspired by recent work on Low-Rank Adaptation (LoRA), Block-LoRA partitions the original low-rank decomposition matrix of LoRA into a series of sub-matrices while sharing all down-projection sub-matrices. This structure not only reduces the number of training parameters, but also transforms certain complex matrix multiplication operations into simpler matrix addition, significantly lowering the computational cost of fine-tuning. Notably, Block-LoRA enables fine-tuning CLIP on the ImageNet few-shot benchmark using a single 24GB GPU. We also show that Block-LoRA has the more tighter bound of generalization error than vanilla LoRA. Without bells and whistles, extensive experiments demonstrate that Block-LoRA achieves competitive performance compared to state-of-the-art CLIP-based few-shot methods, while maintaining a low training parameters count and reduced computational overhead.
Evolutionary Reinforcement Learning (EvoRL) has emerged as a promising approach to overcoming the limitations of traditional reinforcement learning (RL) by integrating the Evolutionary Computation (EC) paradigm with RL. However, the population-based nature of EC significantly increases computational costs, thereby restricting the exploration of algorithmic design choices and scalability in large-scale settings. To address this challenge, we introduce $\texttt{$\textbf{EvoRL}$}$, the first end-to-end EvoRL framework optimized for GPU acceleration. The framework executes the entire training pipeline on accelerators, including environment simulations and EC processes, leveraging hierarchical parallelism through vectorization and compilation techniques to achieve superior speed and scalability. This design enables the efficient training of large populations on a single machine. In addition to its performance-oriented design, $\texttt{$\textbf{EvoRL}$}$ offers a comprehensive platform for EvoRL research, encompassing implementations of traditional RL algorithms (e.g., A2C, PPO, DDPG, TD3, SAC), Evolutionary Algorithms (e.g., CMA-ES, OpenES, ARS), and hybrid EvoRL paradigms such as Evolutionary-guided RL (e.g., ERL, CEM-RL) and Population-Based AutoRL (e.g., PBT). The framework's modular architecture and user-friendly interface allow researchers to seamlessly integrate new components, customize algorithms, and conduct fair benchmarking and ablation studies. The project is open-source and available at: https://github.com/EMI-Group/evorl.
The rapid growth of generative AI and its integration into everyday workflows have significantly increased the demand for large language model (LLM) inference services. While proprietary models remain popular, recent advancements in open-source LLMs have positioned them as strong contenders. However, deploying these models is often constrained by the high costs and limited availability of GPU resources. In response, this paper presents the design of a decentralized offline serving system for LLM inference. Utilizing idle GPU resources, our proposed system, DeServe, decentralizes access to LLMs at a lower cost. DeServe specifically addresses key challenges in optimizing serving throughput in high-latency network environments. Experiments demonstrate that DeServe achieves a 6.7x-12.6x improvement in throughput over existing serving system baselines in such conditions.
Large Language Models (LLMs) have become the new foundation for many applications, reshaping human society like a storm. Disaggregated inference, which separates prefill and decode stages, is a promising approach to improving hardware utilization and service quality. However, due to inefficient inter-node communication, existing systems restrict disaggregated inference to a single node, limiting resource allocation flexibility and reducing service capacity. This paper introduces KVDirect, which optimizes KV cache transfer to enable a distributed disaggregated LLM inference. KVDirect achieves this through the following contributions. First, we propose a novel tensor-centric communication mechanism that reduces the synchronization overhead in traditional distributed GPU systems. Second, we design a custom communication library to support dynamic GPU resource scheduling and efficient KV cache transfer. Third, we introduce a pull-based KV cache transfer strategy that reduces GPU resource idling and improves latency. Finally, we implement KVDirect as an open-source LLM inference framework. Our evaluation demonstrates that KVDirect reduces per-request latency by 55% compared to the baseline across diverse workloads under the same resource constraints.
Mixture-of-Experts (MoE) models, though highly effective for various machine learning tasks, face significant deployment challenges on memory-constrained devices. While GPUs offer fast inference, their limited memory compared to CPUs means not all experts can be stored on the GPU simultaneously, necessitating frequent, costly data transfers from CPU memory, often negating GPU speed advantages. To address this, we present DAOP, an on-device MoE inference engine to optimize parallel GPU-CPU execution. DAOP dynamically allocates experts between CPU and GPU based on per-sequence activation patterns, and selectively pre-calculates predicted experts on CPUs to minimize transfer latency. This approach enables efficient resource utilization across various expert cache ratios while maintaining model accuracy through a novel graceful degradation mechanism. Comprehensive evaluations across various datasets show that DAOP outperforms traditional expert caching and prefetching methods by up to 8.20x and offloading techniques by 1.35x while maintaining accuracy.
It has long been a problem to arrange and execute irregular workloads on massively parallel devices. We propose a general framework for statically batching irregular workloads into a single kernel with a runtime task mapping mechanism on GPUs. We further apply this framework to Mixture-of-Experts (MoE) model inference and implement an optimized and efficient CUDA kernel. Our MoE kernel achieves up to 91% of the peak Tensor Core throughput on NVIDIA H800 GPU and 95% on NVIDIA H20 GPU.