Prism: Unleashing GPU Sharing for Cost-Efficient Multi-LLM Serving

Serving large language models (LLMs) is expensive, especially for providers hosting many models, making cost reduction essential. The unique workload patterns of serving multiple LLMs (i.e., multi-LLM serving) create new opportunities and challenges for this task. The long-tail popularity of models and their long idle periods present opportunities to improve utilization through GPU sharing. However, existing GPU sharing systems lack the ability to adjust their resource allocation and sharing policies at runtime, making them ineffective at meeting latency service-level objectives (SLOs) under rapidly fluctuating workloads. This paper presents Prism, a multi-LLM serving system that unleashes the full potential of GPU sharing to achieve both cost efficiency and SLO attainment. At its core, Prism tackles a key limitation of existing systems$\unicode{x2014}$the lack of $\textit{cross-model memory coordination}$, which is essential for flexibly sharing GPU memory across models under dynamic workloads. Prism achieves this with two key designs. First, it supports on-demand memory allocation by dynamically mapping physical to virtual memory pages, allowing flexible memory redistribution among models that space- and time-share a GPU. Second, it improves memory efficiency through a two-level scheduling policy that dynamically adjusts sharing strategies based on models' runtime demands. Evaluations on real-world traces show that Prism achieves more than $2\times$ cost savings and $3.3\times$ SLO attainment compared to state-of-the-art systems.

ConServe: Harvesting GPUs for Low-Latency and High-Throughput Large Language Model Serving

Many applications are leveraging large language models (LLMs) for complex tasks, and they generally demand low inference latency and high serving throughput for interactive online jobs such as chatbots. However, the tight latency requirement and high load variance of applications pose challenges to serving systems in achieving high GPU utilization. Due to the high costs of scheduling and preemption, today's systems generally use separate clusters to serve online and offline inference tasks, and dedicate GPUs for online inferences to avoid interference. This approach leads to underutilized GPUs because one must reserve enough GPU resources for the peak expected load, even if the average load is low. This paper proposes to harvest stranded GPU resources for offline LLM inference tasks such as document summarization and LLM benchmarking. Unlike online inferences, these tasks usually run in a batch-processing manner with loose latency requirements, making them a good fit for stranded resources that are only available shortly. To enable safe and efficient GPU harvesting without interfering with online tasks, we built ConServe, an LLM serving system that contains (1) an execution engine that preempts running offline tasks upon the arrival of online tasks, (2) an incremental checkpointing mechanism that minimizes the amount of recomputation required by preemptions, and (3) a scheduler that adaptively batches offline tasks for higher GPU utilization. Our evaluation demonstrates that ConServe achieves strong performance isolation when co-serving online and offline tasks but at a much higher GPU utilization. When colocating practical online and offline workloads on popular models such as Llama-2-7B, ConServe achieves 2.35$\times$ higher throughput than state-of-the-art online serving systems and reduces serving latency by 84$\times$ compared to existing co-serving systems.

Learning from Pattern Completion: Self-supervised Controllable Generation

The human brain exhibits a strong ability to spontaneously associate different visual attributes of the same or similar visual scene, such as associating sketches and graffiti with real-world visual objects, usually without supervising information. In contrast, in the field of artificial intelligence, controllable generation methods like ControlNet heavily rely on annotated training datasets such as depth maps, semantic segmentation maps, and poses, which limits the method's scalability. Inspired by the neural mechanisms that may contribute to the brain's associative power, specifically the cortical modularization and hippocampal pattern completion, here we propose a self-supervised controllable generation (SCG) framework. Firstly, we introduce an equivariant constraint to promote inter-module independence and intra-module correlation in a modular autoencoder network, thereby achieving functional specialization. Subsequently, based on these specialized modules, we employ a self-supervised pattern completion approach for controllable generation training. Experimental results demonstrate that the proposed modular autoencoder effectively achieves functional specialization, including the modular processing of color, brightness, and edge detection, and exhibits brain-like features including orientation selectivity, color antagonism, and center-surround receptive fields. Through self-supervised training, associative generation capabilities spontaneously emerge in SCG, demonstrating excellent generalization ability to various tasks such as associative generation on painting, sketches, and ancient graffiti. Compared to the previous representative method ControlNet, our proposed approach not only demonstrates superior robustness in more challenging high-noise scenarios but also possesses more promising scalability potential due to its self-supervised manner.

Individual brain parcellation: Review of methods, validations and applications

Individual brains vary greatly in morphology, connectivity and organization. The applicability of group-level parcellations is limited by the rapid development of precision medicine today because they do not take into account the variation of parcels at the individual level. Accurate mapping of brain functional regions at the individual level is pivotal for a comprehensive understanding of the variations in brain function and behaviors, early and precise identification of brain abnormalities, as well as personalized treatments for neuropsychiatric disorders. With the development of neuroimaging and machine learning techniques, studies on individual brain parcellation are booming. In this paper, we offer an overview of recent advances in the methodologies of individual brain parcellation, including optimization- and learning-based methods. Comprehensive evaluation metrics to validate individual brain mapping have been introduced. We also review the studies of how individual brain mapping promotes neuroscience research and clinical medicine. Finally, we summarize the major challenges and important future directions of individualized brain parcellation. Collectively, we intend to offer a thorough overview of individual brain parcellation methods, validations, and applications, along with highlighting the current challenges that call for an urgent demand for integrated platforms that integrate datasets, methods, and validations.

Nonparametric Automatic Differentiation Variational Inference with Spline Approximation

Automatic Differentiation Variational Inference (ADVI) is efficient in learning probabilistic models. Classic ADVI relies on the parametric approach to approximate the posterior. In this paper, we develop a spline-based nonparametric approximation approach that enables flexible posterior approximation for distributions with complicated structures, such as skewness, multimodality, and bounded support. Compared with widely-used nonparametric variational inference methods, the proposed method is easy to implement and adaptive to various data structures. By adopting the spline approximation, we derive a lower bound of the importance weighted autoencoder and establish the asymptotic consistency. Experiments demonstrate the efficiency of the proposed method in approximating complex posterior distributions and improving the performance of generative models with incomplete data.

VQPy: An Object-Oriented Approach to Modern Video Analytics

Video analytics is widely used in contemporary systems and services. At the forefront of video analytics are video queries that users develop to find objects of particular interest. Building upon the insight that video objects (e.g., human, animals, cars, etc.), the center of video analytics, are similar in spirit to objects modeled by traditional object-oriented languages, we propose to develop an object-oriented approach to video analytics. This approach, named VQPy, consists of a frontend$\unicode{x2015}$a Python variant with constructs that make it easy for users to express video objects and their interactions$\unicode{x2015}$as well as an extensible backend that can automatically construct and optimize pipelines based on video objects. We have implemented and open-sourced VQPy, which has been productized in Cisco as part of its DeepVision framework.

AnyOKP: One-Shot and Instance-Aware Object Keypoint Extraction with Pretrained ViT

Towards flexible object-centric visual perception, we propose a one-shot instance-aware object keypoint (OKP) extraction approach, AnyOKP, which leverages the powerful representation ability of pretrained vision transformer (ViT), and can obtain keypoints on multiple object instances of arbitrary category after learning from a support image. An off-the-shelf petrained ViT is directly deployed for generalizable and transferable feature extraction, which is followed by training-free feature enhancement. The best-prototype pairs (BPPs) are searched for in support and query images based on appearance similarity, to yield instance-unaware candidate keypoints.Then, the entire graph with all candidate keypoints as vertices are divided to sub-graphs according to the feature distributions on the graph edges. Finally, each sub-graph represents an object instance. AnyOKP is evaluated on real object images collected with the cameras of a robot arm, a mobile robot, and a surgical robot, which not only demonstrates the cross-category flexibility and instance awareness, but also show remarkable robustness to domain shift and viewpoint change.

Ethosight: A Reasoning-Guided Iterative Learning System for Nuanced Perception based on Joint-Embedding & Contextual Label Affinity

Traditional computer vision models often require extensive manual effort for data acquisition, annotation and validation, particularly when detecting subtle behavioral nuances or events. The difficulty in distinguishing routine behaviors from potential risks in real-world applications, such as differentiating routine shopping from potential shoplifting, further complicates the process. Moreover, these models may demonstrate high false positive rates and imprecise event detection when exposed to real-world scenarios that differ significantly from the conditions of the training data. To overcome these hurdles, we present Ethosight, a novel zero-shot computer vision system. Ethosight initiates with a clean slate based on user requirements and semantic knowledge of interest. Using localized label affinity calculations and a reasoning-guided iterative learning loop, Ethosight infers scene details and iteratively refines the label set. Reasoning mechanisms can be derived from large language models like GPT4, symbolic reasoners like OpenNARS\cite{wang2013}\cite{wang2006}, or hybrid systems. Our evaluations demonstrate Ethosight's efficacy across 40 complex use cases, spanning domains such as health, safety, and security. Detailed results and case studies within the main body of this paper and an appendix underscore a promising trajectory towards enhancing the adaptability and resilience of computer vision models in detecting and extracting subtle and nuanced behaviors.

Out-of-distribution forgetting: vulnerability of continual learning to intra-class distribution shift

Continual learning (CL) is an important technique to allow artificial neural networks to work in open environments. CL enables a system to learn new tasks without severe interference to its performance on old tasks, i.e., overcome the problems of catastrophic forgetting. In joint learning, it is well known that the out-of-distribution (OOD) problem caused by intentional attacks or environmental perturbations will severely impair the ability of networks to generalize. In this work, we reported a special form of catastrophic forgetting raised by the OOD problem in continual learning settings, and we named it out-of-distribution forgetting (OODF). In continual image classification tasks, we found that for a given category, introducing an intra-class distribution shift significantly impaired the recognition accuracy of CL methods for that category during subsequent learning. Interestingly, this phenomenon is special for CL as the same level of distribution shift had only negligible effects in the joint learning scenario. We verified that CL methods without dedicating subnetworks for individual tasks are all vulnerable to OODF. Moreover, OODF does not depend on any specific way of shifting the distribution, suggesting it is a risk for CL in a wide range of circumstances. Taken together, our work identified an under-attended risk during CL, highlighting the importance of developing approaches that can overcome OODF.

Emergence of Symbols in Neural Networks for Semantic Understanding and Communication

Being able to create meaningful symbols and proficiently use them for higher cognitive functions such as communication, reasoning, planning, etc., is essential and unique for human intelligence. Current deep neural networks are still far behind human's ability to create symbols for such higher cognitive functions. Here we propose a solution, named SEA-net, to endow neural networks with ability of symbol creation, semantic understanding and communication. SEA-net generates symbols that dynamically configure the network to perform specific tasks. These symbols capture compositional semantic information that enables the system to acquire new functions purely by symbolic manipulation or communication. In addition, we found that these self-generated symbols exhibit an intrinsic structure resembling that of natural language, suggesting a common framework underlying the generation and understanding of symbols in both human brains and artificial neural networks. We hope that it will be instrumental in producing more capable systems in the future that can synergize the strengths of connectionist and symbolic approaches for AI.