A practical large language model (LLM) service may involve a long system prompt, which specifies the instructions, examples, and knowledge documents of the task and is reused across requests. However, the long system prompt causes throughput/latency bottlenecks as the cost of generating the next token grows w.r.t the sequence length. This paper aims to improve the efficiency of LLM services that involve long system prompts. Our key observation is that handling these system prompts requires heavily redundant memory accesses in existing causal attention computation algorithms. Specifically, for batched requests, the cached hidden states (i.e., key-value pairs) of system prompts are transferred from off-chip DRAM to on-chip SRAM multiple times, each corresponding to an individual request. To eliminate such a redundancy, we propose RelayAttention, an attention algorithm that allows reading these hidden states from DRAM exactly once for a batch of input tokens. RelayAttention is a free lunch: it maintains the generation quality while requiring no model retraining, as it is based on a mathematical reformulation of causal attention. We have observed significant performance improvements to a production-level system, vLLM, through integration with RelayAttention. The improvements are even more profound with longer system prompts. © 2024 Association for Computational Linguistics

Inverse rendering aims at recovering both geometry and materials of objects. It provides a more compatible re-construction for conventional rendering engines, compared with the neural radiance fields (NeRFs). On the other hand, existing NeRF-based inverse rendering methods can-not handle glossy objects with local light interactions well, as they typically oversimplify the illumination as a 2D environmental map, which assumes infinite lights only. Observing the superiority of NeRFs in recovering radiance fields, we propose a novel 5D Neural Plenoptic Function (NeP) based on NeRFs and ray tracing, such that more accurate lighting-object interactions can be formulated via the ren-dering equation. We also design a material-aware cone sampling strategy to efficiently integrate lights inside the BRDF lobes with the help of pre-filtered radiance fields. Our method has two stages: the geometry of the target ob-ject and the pre-filtered environmental radiance fields are reconstructed in the first stage, and materials of the target object are estimated in the second stage with the proposed NeP and material-aware cone sampling strategy. Exten-sive experiments on the proposed real-world and synthetic datasets demonstrate that our method can reconstruct high-fidelity geometry/materials of challenging glossy objects with complex lighting interactions from nearby objects. Project webpage: https://whyy.si.te/paper/nep

Convolutions (Convs) and multi-head self-attentions (MHSAs) are typically considered alternatives to each other for building vision backbones. Although some works try to integrate both, they apply the two operators simultaneously at the finest pixel granularity. With Convs responsible for per-pixel feature extraction already, the question is whether we still need to include the heavy MHSAs at such a fine-grained level. In fact, this is the root cause of the scalability issue w.r.t. the input resolution for vision transformers. To address this important problem, we propose in this work to use MSHAs and Convs in parallel \textbf{at different granularity levels} instead. Specifically, in each layer, we use two different ways to represent an image: a fine-grained regular grid and a coarse-grained set of semantic slots. We apply different operations to these two representations: Convs to the grid for local features, and MHSAs to the slots for global features. A pair of fully differentiable soft clustering and dispatching modules is introduced to bridge the grid and set representations, thus enabling local-global fusion. Through extensive experiments on various vision tasks, we empirically verify the potential of the proposed integration scheme, named \textit{GLMix}: by offloading the burden of fine-grained features to light-weight Convs, it is sufficient to use MHSAs in a few (e.g., 64) semantic slots to match the performance of recent state-of-the-art backbones, while being more efficient. Our visualization results also demonstrate that the soft clustering module produces a meaningful semantic grouping effect with only IN1k classification supervision, which may induce better interpretability and inspire new weakly-supervised semantic segmentation approaches. Code will be available at \url{https://github.com/rayleizhu/GLMix}.

Sepsis is a life-threatening condition characterized by an exaggerated immune response to pathogens, leading to organ damage and high mortality rates in the intensive care unit. Although deep learning has achieved impressive performance on prediction and classification tasks in medicine, it requires large amounts of data and lacks explainability, which hinder its application to sepsis diagnosis. We introduce a deep learning framework, called scCaT, which blends the capsulating architecture with Transformer to develop a sepsis diagnostic model using single-cell RNA sequencing data and transfers it to bulk RNA data. The capsulating architecture effectively groups genes into capsules based on biological functions, which provides explainability in encoding gene expressions. The Transformer serves as a decoder to classify sepsis patients and controls. Our model achieves high accuracy with an AUROC of 0.93 on the single-cell test set and an average AUROC of 0.98 on seven bulk RNA cohorts. Additionally, the capsules can recognize different cell types and distinguish sepsis from control samples based on their biological pathways. This study presents a novel approach for learning gene modules and transferring the model to other data types, offering potential benefits in diagnosing rare diseases with limited subjects.

In this paper, we present a Neural Preset technique to address the limitations of existing color style transfer methods, including visual artifacts, vast memory requirement, and slow style switching speed. Our method is based on two core designs. First, we propose Deterministic Neural Color Mapping (DNCM) to consistently operate on each pixel via an image-adaptive color mapping matrix, avoiding artifacts and supporting high-resolution inputs with a small memory footprint. Second, we develop a two-stage pipeline by dividing the task into color normalization and stylization, which allows efficient style switching by extracting color styles as presets and reusing them on normalized input images. Due to the unavailability of pairwise datasets, we describe how to train Neural Preset via a self-supervised strategy. Various advantages of Neural Preset over existing methods are demonstrated through comprehensive evaluations. Besides, we show that our trained model can naturally support multiple applications without fine-tuning, including low-light image enhancement, underwater image correction, image dehazing, and image harmonization. The project page is: https://ZHKKKe.github.io/NeuralPreset.

As the core building block of vision transformers, attention is a powerful tool to capture long-range dependency. However, such power comes at a cost: it incurs a huge computation burden and heavy memory footprint as pairwise token interaction across all spatial locations is computed. A series of works attempt to alleviate this problem by introducing handcrafted and content-agnostic sparsity into attention, such as restricting the attention operation to be inside local windows, axial stripes, or dilated windows. In contrast to these approaches, we propose a novel dynamic sparse attention via bi-level routing to enable a more flexible allocation of computations with content awareness. Specifically, for a query, irrelevant key-value pairs are first filtered out at a coarse region level, and then fine-grained token-to-token attention is applied in the union of remaining candidate regions (i.e., routed regions). We provide a simple yet effective implementation of the proposed bilevel routing attention, which utilizes the sparsity to save both computation and memory while involving only GPU-friendly dense matrix multiplications. Built with the proposed bi-level routing attention, a new general vision transformer, named BiFormer, is then presented. As BiFormer attends to a small subset of relevant tokens in a query adaptive manner without distraction from other irrelevant ones, it enjoys both good performance and high computational efficiency, especially in dense prediction tasks. Empirical results across several computer vision tasks such as image classification, object detection, and semantic segmentation verify the effectiveness of our design. Code is available at https://github.com/rayleizhu/BiFormer.

Although CNNs have achieved remarkable progress on the shadow detection task, they tend to make mistakes in dark non-shadow regions and relatively bright shadow regions. They are also susceptible to brightness change. These two phenomenons reveal that deep shadow detectors heavily depend on the intensity cue, which we refer to as intensity bias. In this paper, we propose a novel feature decomposition and reweighting scheme to mitigate this intensity bias, in which multi-level integrated features are decomposed into intensity-variant and intensity-invariant components through self-supervision. By reweighting these two types of features, our method can reallocate the attention to the corresponding latent semantics and achieves balanced exploitation of them. Extensive experiments on three popular datasets show that the proposed method outperforms state-of-the-art shadow detectors.