Speculative decoding has emerged as a powerful method to improve latency and throughput in hosting large language models. However, most existing implementations focus on generating a single sequence. Real-world generative AI applications often require multiple responses and how to perform speculative decoding in a batched setting while preserving its latency benefits poses non-trivial challenges. This paper describes a system of batched speculative decoding that sets a new state of the art in multi-sequence generation latency and that demonstrates superior GPU utilization as well as quality of generations within a time budget. For example, for a 7.8B-size model on a single A100 GPU and with a batch size of 8, each sequence is generated at an average speed of 5.8ms per token, the overall throughput being 1.1K tokens per second. These results represent state-of-the-art latency and a 2.15X speed-up over optimized regular decoding. Within a time budget that regular decoding does not finish, our system is able to generate sequences with HumanEval Pass@First of 43% and Pass@All of 61%, far exceeding what's feasible with single-sequence speculative decoding. Our peak GPU utilization during decoding reaches as high as 15.8%, more than 3X the highest of that of regular decoding and around 10X of single-sequence speculative decoding.
Open-source simulation tools play a crucial role for neuromorphic application engineers and hardware architects to investigate performance bottlenecks and explore design optimizations before committing to silicon. Reconfigurable Architecture for Neuromorphic Computing (RANC) is one such tool that offers ability to execute pre-trained Spiking Neural Network (SNN) models within a unified ecosystem through both software-based simulation and FPGA-based emulation. RANC has been utilized by the community with its flexible and highly parameterized design to study implementation bottlenecks, tune architectural parameters or modify neuron behavior based on application insights and study the trade space on hardware performance and network accuracy. In designing architectures for use in neuromorphic computing, there are an incredibly large number of configuration parameters such as number and precision of weights per neuron, neuron and axon counts per core, network topology, and neuron behavior. To accelerate such studies and provide users with a streamlined productive design space exploration, in this paper we introduce the GPU-based implementation of RANC. We summarize our parallelization approach and quantify the speedup gains achieved with GPU-based tick-accurate simulations across various use cases. We demonstrate up to 780 times speedup compared to serial version of the RANC simulator based on a 512 neuromorphic core MNIST inference application. We believe that the RANC ecosystem now provides a much more feasible avenue in the research of exploring different optimizations for accelerating SNNs and performing richer studies by enabling rapid convergence to optimized neuromorphic architectures.
Large Language Models (LLMs), despite their remarkable performance across a wide range of tasks, necessitate substantial GPU memory and consume significant computational resources. Beyond the memory taken up by model weights, the memory used by the KV cache rises linearly with sequence length, becoming a primary bottleneck for inference. In this paper, we introduce an innovative method for optimizing the KV cache, which considerably minimizes its memory footprint. Upon thorough investigation, we discover that in most Transformer models, (i) there is a striking similarity between adjacent tokens' query vectors, and (ii) the attention calculation of the current query can rely exclusively on the attention information of a small fraction of preceding queries. Based on these observations, we present CORM, a KV cache eviction policy that dynamically retains essential key-value pairs for inference without the need for model fine-tuning. Our validation shows that CORM reduces the inference memory usage of KV cache by up to 70\% with negligible performance degradation across six tasks in LongBench. Furthermore, we demonstrate that CORM is compatible with GQA for further compression rate.
Large language models (LLMs) are increasingly integrated into many online services, yet they remain cost-prohibitive to deploy due to the requirement of expensive GPU instances. Prior work has addressed the high cost of LLM serving by improving the inference engine, but less attention has been given to selecting the most cost-efficient GPU type(s) for a specific LLM service. There is a large and growing landscape of GPU types and, within these options, higher cost does not always lead to increased performance. Instead, through a comprehensive investigation, we find that three key LLM service characteristics (request size, request rate, SLO) strongly influence GPU cost efficiency, and differing GPU types are most cost efficient for differing LLM service settings. As a result, the most cost-efficient allocation for a given service is typically a mix of heterogeneous GPU types. Based on this analysis, we introduce M\'elange, a GPU allocation framework that navigates these diverse LLM service characteristics and heterogeneous GPU option space to automatically and efficiently derive the minimal-cost GPU allocation for a given LLM service. We formulate the GPU allocation task as a cost-aware bin packing problem where GPUs are bins and items are slices of the service workload. Our formulation's constraints account for a service's unique characteristics, allowing M\'elange to be flexible to support diverse service settings and heterogeneity-aware to adapt the GPU allocation to a specific service. Compared to using only a single GPU type, M\'elange reduces deployment costs by up to 77% in conversational settings, 33% in document-based settings, and 51% in a mixed setting.
Large Language Models (LLMs) have made remarkable progress in processing extensive contexts, with the Key-Value (KV) cache playing a vital role in enhancing their performance. However, the growth of the KV cache in response to increasing input length poses challenges to memory and time efficiency. To address this problem, this paper introduces SnapKV, an innovative and fine-tuning-free approach that efficiently minimizes KV cache size while still delivering comparable performance in real-world applications. We discover that each attention head in the model consistently focuses on specific prompt attention features during generation. Meanwhile, this robust pattern can be obtained from an 'observation' window located at the end of the prompts. Drawing on this insight, SnapKV automatically compresses KV caches by selecting clustered important KV positions for each attention head. Our approach significantly reduces the growing computational overhead and memory footprint when processing long input sequences. Specifically, SnapKV achieves a consistent decoding speed with a 3.6x increase in generation speed and an 8.2x enhancement in memory efficiency compared to the baseline when processing inputs of 16K tokens. At the same time, it maintains comparable performance to the baseline models across 16 long sequence datasets. Moreover, SnapKV can process up to 380K context tokens on a single A100-80GB GPU using HuggingFace implementation with minor changes, exhibiting only a negligible accuracy drop in the Needle-in-a-Haystack test. Further comprehensive studies suggest SnapKV's potential for practical applications.
Fine-tuning Large Language Models (LLMs) is a common practice to adapt pre-trained models for specific applications. While methods like LoRA have effectively addressed GPU memory constraints during fine-tuning, their performance often falls short, especially in multi-task scenarios. In contrast, Mixture-of-Expert (MoE) models, such as Mixtral 8x7B, demonstrate remarkable performance in multi-task learning scenarios while maintaining a reduced parameter count. However, the resource requirements of these MoEs remain challenging, particularly for consumer-grade GPUs with less than 24GB memory. To tackle these challenges, we propose MixLoRA, an approach to construct a resource-efficient sparse MoE model based on LoRA. MixLoRA inserts multiple LoRA-based experts within the feed-forward network block of a frozen pre-trained dense model and employs a commonly used top-k router. Unlike other LoRA-based MoE methods, MixLoRA enhances model performance by utilizing independent attention-layer LoRA adapters. Additionally, an auxiliary load balance loss is employed to address the imbalance problem of the router. Our evaluations show that MixLoRA improves about 9% accuracy compared to state-of-the-art PEFT methods in multi-task learning scenarios. We also propose a new high-throughput framework to alleviate the computation and memory bottlenecks during the training and inference of MOE models. This framework reduces GPU memory consumption by 40% and token computation latency by 30% during both training and inference.
Transforming large pre-trained low-resolution diffusion models to cater to higher-resolution demands, i.e., diffusion extrapolation, significantly improves diffusion adaptability. We propose tuning-free CutDiffusion, aimed at simplifying and accelerating the diffusion extrapolation process, making it more affordable and improving performance. CutDiffusion abides by the existing patch-wise extrapolation but cuts a standard patch diffusion process into an initial phase focused on comprehensive structure denoising and a subsequent phase dedicated to specific detail refinement. Comprehensive experiments highlight the numerous almighty advantages of CutDiffusion: (1) simple method construction that enables a concise higher-resolution diffusion process without third-party engagement; (2) fast inference speed achieved through a single-step higher-resolution diffusion process, and fewer inference patches required; (3) cheap GPU cost resulting from patch-wise inference and fewer patches during the comprehensive structure denoising; (4) strong generation performance, stemming from the emphasis on specific detail refinement.
Evaluating retrieval-augmented generation (RAG) presents challenges, particularly for retrieval models within these systems. Traditional end-to-end evaluation methods are computationally expensive. Furthermore, evaluation of the retrieval model's performance based on query-document relevance labels shows a small correlation with the RAG system's downstream performance. We propose a novel evaluation approach, eRAG, where each document in the retrieval list is individually utilized by the large language model within the RAG system. The output generated for each document is then evaluated based on the downstream task ground truth labels. In this manner, the downstream performance for each document serves as its relevance label. We employ various downstream task metrics to obtain document-level annotations and aggregate them using set-based or ranking metrics. Extensive experiments on a wide range of datasets demonstrate that eRAG achieves a higher correlation with downstream RAG performance compared to baseline methods, with improvements in Kendall's $\tau$ correlation ranging from 0.168 to 0.494. Additionally, eRAG offers significant computational advantages, improving runtime and consuming up to 50 times less GPU memory than end-to-end evaluation.
Global placement is a fundamental step in VLSI physical design. The wide use of 2D processing element (PE) arrays in machine learning accelerators poses new challenges of scalability and Quality of Results (QoR) for state-of-the-art academic global placers. In this work, we develop DG-RePlAce, a new and fast GPU-accelerated global placement framework built on top of the OpenROAD infrastructure, which exploits the inherent dataflow and datapath structures of machine learning accelerators. Experimental results with a variety of machine learning accelerators using a commercial 12nm enablement show that, compared with RePlAce (DREAMPlace), our approach achieves an average reduction in routed wirelength by 10% (7%) and total negative slack (TNS) by 31% (34%), with faster global placement and on-par total runtimes relative to DREAMPlace. Empirical studies on the TILOS MacroPlacement Benchmarks further demonstrate that post-route improvements over RePlAce and DREAMPlace may reach beyond the motivating application to machine learning accelerators.
Federated Learning (FL) is an emerging machine learning paradigm that enables the collaborative training of a shared global model across distributed clients while keeping the data decentralized. Recent works on designing systems for efficient FL have shown that utilizing serverless computing technologies, particularly Function-as-a-Service (FaaS) for FL, can enhance resource efficiency, reduce training costs, and alleviate the complex infrastructure management burden on data holders. However, current serverless FL systems still suffer from the presence of stragglers, i.e., slow clients that impede the collaborative training process. While strategies aimed at mitigating stragglers in these systems have been proposed, they overlook the diverse hardware resource configurations among FL clients. To this end, we present Apodotiko, a novel asynchronous training strategy designed for serverless FL. Our strategy incorporates a scoring mechanism that evaluates each client's hardware capacity and dataset size to intelligently prioritize and select clients for each training round, thereby minimizing the effects of stragglers on system performance. We comprehensively evaluate Apodotiko across diverse datasets, considering a mix of CPU and GPU clients, and compare its performance against five other FL training strategies. Results from our experiments demonstrate that Apodotiko outperforms other FL training strategies, achieving an average speedup of 2.75x and a maximum speedup of 7.03x. Furthermore, our strategy significantly reduces cold starts by a factor of four on average, demonstrating suitability in serverless environments.
Underwater images taken from autonomous underwater vehicles (AUV's) often suffer from low light, high turbidity, poor contrast, motion-blur and excessive light scattering and hence require image enhancement techniques for object recognition. Machine learning methods are being increasingly used for object recognition under such adverse conditions. These enhanced object recognition methods of images taken from AUV's has potential applications in underwater pipeline and optical fibre surveillance, ocean bed resource extraction, ocean floor mapping, underwater species exploration, etc. While the classical machine learning methods are very efficient in terms of accuracy, they require large datasets and high computational time for image classification. In the current work, we use quantum-classical hybrid machine learning methods for real-time under-water object recognition on-board an AUV for the first time. We use real-time motion-blurred and low-light images taken from an on-board camera of AUV built in-house and apply existing hybrid machine learning methods for object recognition. Our hybrid methods consist of quantum encoding and flattening of classical images using quantum circuits and sending them to classical neural networks for image classification. The results of hybrid methods carried out using Pennylane based quantum simulators both on GPU and using pre-trained models on an on-board NVIDIA GPU chipset are compared with results from corresponding classical machine learning methods. We observe that the hybrid quantum machine learning methods show an efficiency greater than 65\% and reduction in run-time by one-thirds and require 50\% smaller dataset sizes for training the models compared to classical machine learning methods. We hope that our work opens up further possibilities in quantum enhanced real-time computer vision in autonomous vehicles.
Recent works on audio-driven talking head synthesis using Neural Radiance Fields (NeRF) have achieved impressive results. However, due to inadequate pose and expression control caused by NeRF implicit representation, these methods still have some limitations, such as unsynchronized or unnatural lip movements, and visual jitter and artifacts. In this paper, we propose GaussianTalker, a novel method for audio-driven talking head synthesis based on 3D Gaussian Splatting. With the explicit representation property of 3D Gaussians, intuitive control of the facial motion is achieved by binding Gaussians to 3D facial models. GaussianTalker consists of two modules, Speaker-specific Motion Translator and Dynamic Gaussian Renderer. Speaker-specific Motion Translator achieves accurate lip movements specific to the target speaker through universalized audio feature extraction and customized lip motion generation. Dynamic Gaussian Renderer introduces Speaker-specific BlendShapes to enhance facial detail representation via a latent pose, delivering stable and realistic rendered videos. Extensive experimental results suggest that GaussianTalker outperforms existing state-of-the-art methods in talking head synthesis, delivering precise lip synchronization and exceptional visual quality. Our method achieves rendering speeds of 130 FPS on NVIDIA RTX4090 GPU, significantly exceeding the threshold for real-time rendering performance, and can potentially be deployed on other hardware platforms.
Diffusion models have significantly advanced the state of the art in image, audio, and video generation tasks. However, their applications in practical scenarios are hindered by slow inference speed. Drawing inspiration from the approximation strategies utilized in consistency models, we propose the Sub-path Linear Approximation Model (SLAM), which accelerates diffusion models while maintaining high-quality image generation. SLAM treats the PF-ODE trajectory as a series of PF-ODE sub-paths divided by sampled points, and harnesses sub-path linear (SL) ODEs to form a progressive and continuous error estimation along each individual PF-ODE sub-path. The optimization on such SL-ODEs allows SLAM to construct denoising mappings with smaller cumulative approximated errors. An efficient distillation method is also developed to facilitate the incorporation of more advanced diffusion models, such as latent diffusion models. Our extensive experimental results demonstrate that SLAM achieves an efficient training regimen, requiring only 6 A100 GPU days to produce a high-quality generative model capable of 2 to 4-step generation with high performance. Comprehensive evaluations on LAION, MS COCO 2014, and MS COCO 2017 datasets also illustrate that SLAM surpasses existing acceleration methods in few-step generation tasks, achieving state-of-the-art performance both on FID and the quality of the generated images.
Spatiotemporal networks' observational capabilities are crucial for accurate data gathering and informed decisions across multiple sectors. This study focuses on the Spatiotemporal Ranged Observer-Observable Bipartite Network (STROOBnet), linking observational nodes (e.g., surveillance cameras) to events within defined geographical regions, enabling efficient monitoring. Using data from Real-Time Crime Camera (RTCC) systems and Calls for Service (CFS) in New Orleans, where RTCC combats rising crime amidst reduced police presence, we address the network's initial observational imbalances. Aiming for uniform observational efficacy, we propose the Proximal Recurrence approach. It outperformed traditional clustering methods like k-means and DBSCAN by offering holistic event frequency and spatial consideration, enhancing observational coverage.
A series of graph filtering (GF)-based collaborative filtering (CF) showcases state-of-the-art performance on the recommendation accuracy by using a low-pass filter (LPF) without a training process. However, conventional GF-based CF approaches mostly perform matrix decomposition on the item-item similarity graph to realize the ideal LPF, which results in a non-trivial computational cost and thus makes them less practical in scenarios where rapid recommendations are essential. In this paper, we propose Turbo-CF, a GF-based CF method that is both training-free and matrix decomposition-free. Turbo-CF employs a polynomial graph filter to circumvent the issue of expensive matrix decompositions, enabling us to make full use of modern computer hardware components (i.e., GPU). Specifically, Turbo-CF first constructs an item-item similarity graph whose edge weights are effectively regulated. Then, our own polynomial LPFs are designed to retain only low-frequency signals without explicit matrix decompositions. We demonstrate that Turbo-CF is extremely fast yet accurate, achieving a runtime of less than 1 second on real-world benchmark datasets while achieving recommendation accuracies comparable to best competitors.
Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.
Data assimilation is a core component of numerical weather prediction systems. The large quantity of data processed during assimilation requires the computation to be distributed across increasingly many compute nodes, yet existing approaches suffer from synchronisation overhead in this setting. In this paper, we exploit the formulation of data assimilation as a Bayesian inference problem and apply a message-passing algorithm to solve the spatial inference problem. Since message passing is inherently based on local computations, this approach lends itself to parallel and distributed computation. In combination with a GPU-accelerated implementation, we can scale the algorithm to very large grid sizes while retaining good accuracy and compute and memory requirements.
Characterizing and predicting the training performance of modern machine learning (ML) workloads on compute systems with compute and communication spread between CPUs, GPUs, and network devices is not only the key to optimization and planning but also a complex goal to achieve. The primary challenges include the complexity of synchronization and load balancing between CPUs and GPUs, the variance in input data distribution, and the use of different communication devices and topologies (e.g., NVLink, PCIe, network cards) that connect multiple compute devices, coupled with the desire for flexible training configurations. Built on top of our prior work for single-GPU platforms, we address these challenges and enable multi-GPU performance modeling by incorporating (1) data-distribution-aware performance models for embedding table lookup, and (2) data movement prediction of communication collectives, into our upgraded performance modeling pipeline equipped with inter-and intra-rank synchronization for ML workloads trained on multi-GPU platforms. Beyond accurately predicting the per-iteration training time of DLRM models with random configurations with a geomean error of 5.21% on two multi-GPU platforms, our prediction pipeline generalizes well to other types of ML workloads, such as Transformer-based NLP models with a geomean error of 3.00%. Moreover, even without actually running ML workloads like DLRMs on the hardware, it is capable of generating insights such as quickly selecting the fastest embedding table sharding configuration (with a success rate of 85%).
This paper addresses the problem of detecting time series outliers, focusing on systems with repetitive behavior, such as industrial robots operating on production lines.Notable challenges arise from the fact that a task performed multiple times may exhibit different duration in each repetition and that the time series reported by the sensors are irregularly sampled because of data gaps. The anomaly detection approach presented in this paper consists of three stages.The first stage identifies the repetitive cycles in the lengthy time series and segments them into individual time series corresponding to one task cycle, while accounting for possible temporal distortions.The second stage computes a prototype for the cycles using a GPU-based barycenter algorithm, specifically tailored for very large time series.The third stage uses the prototype to detect abnormal cycles by computing an anomaly score for each cycle.The overall approach, named WarpEd Time Series ANomaly Detection (WETSAND), makes use of the Dynamic Time Warping algorithm and its variants because they are suited to the distorted nature of the time series.The experiments show that \wetsand scales to large signals, computes human-friendly prototypes, works with very little data, and outperforms some general purpose anomaly detection approaches such as autoencoders.
With large language models (LLMs) widely deployed in long content generation recently, there has emerged an increasing demand for efficient long-sequence inference support. However, key-value (KV) cache, which is stored to avoid re-computation, has emerged as a critical bottleneck by growing linearly in size with the sequence length. Due to the auto-regressive nature of LLMs, the entire KV cache will be loaded for every generated token, resulting in low utilization of computational cores and high latency. While various compression methods for KV cache have been proposed to alleviate this issue, they suffer from degradation in generation quality. We introduce TriForce, a hierarchical speculative decoding system that is scalable for long sequence generation. This approach leverages the original model weights and dynamic sparse KV cache via retrieval as a draft model, which serves as an intermediate layer in the hierarchy and is further speculated by a smaller model to reduce its drafting latency. TriForce not only facilitates impressive speedups for Llama2-7B-128K, achieving up to 2.31$\times$ on an A100 GPU but also showcases scalability in handling even longer contexts. For the offloading setting on two RTX 4090 GPUs, TriForce achieves 0.108s/token$\unicode{x2014}$only half as slow as the auto-regressive baseline on an A100, which attains 7.78$\times$ on our optimized offloading system. Additionally, TriForce performs 4.86$\times$ than DeepSpeed-Zero-Inference on a single RTX 4090 GPU. TriForce's robustness is highlighted by its consistently outstanding performance across various temperatures. The code is available at https://github.com/Infini-AI-Lab/TriForce.