Serving Large Language Models (LLMs) is critical for AI-powered applications, yet it demands substantial computational resources, particularly in memory bandwidth and computational throughput. Low-precision computation has emerged as a key technique to improve efficiency while reducing resource consumption. Existing approaches for generating low-precision kernels are limited to weight bit widths that are powers of two and suffer from suboptimal performance because of high-level GPU programming abstractions. These abstractions restrict critical optimizations, such as fine-grained register management and optimized memory access patterns, that are essential for efficient low-precision computations. In this paper, we introduce Tilus, a domain-specific language designed for General-Purpose GPU (GPGPU) computing that supports low-precision data types with arbitrary bit widths from 1 to 8 while maintaining GPU programmability. Tilus features a thread-block-level programming model, a hierarchical memory space, a novel algebraic layout system, and extensive support for diverse low-precision data types. Tilus programs are compiled into highly efficient GPU programs through automatic vectorization and instruction selection. Extensive experiments demonstrate that Tilus efficiently supports a full spectrum of low-precision data types, and outperforms state-of-the-art low-precision kernels. Compared to existing compilers such as Triton and Ladder, as well as hand-optimized kernels such as QuantLLM and Marlin, Tilus achieves performance improvements of: 1.75x, 2.61x, 1.29x and 1.03x, respectively. We open-source Tilus at https://github.com/NVIDIA/tilus.

Composed Image Retrieval (CIR) is an emerging yet challenging task that allows users to search for target images using a multimodal query, comprising a reference image and a modification text specifying the user’s desired changes to the reference image. Given its significant academic and practical value, CIR has become a rapidly growing area of interest in the computer vision and machine learning communities, particularly with the advances in deep learning. To the best of our knowledge, there is currently no comprehensive review of CIR to provide a timely overview of this field. Therefore, we synthesize insights from over 150 publications in top conferences and journals, including ACM TOIS, SIGIR, and CVPR. In particular, we systematically categorize existing supervised CIR and zero-shot CIR models using a fine-grained taxonomy. For a comprehensive review, we also briefly discuss approaches for tasks closely related to CIR, such as attribute-based CIR and dialog-based CIR. Additionally, we summarize benchmark datasets for evaluation and analyze existing supervised and zero-shot CIR methods by comparing experimental results across multiple datasets. Furthermore, we present promising future directions in this field, offering practical insights for researchers interested in further exploration.

Dynamic shape computations have become critical in modern machine learning workloads, especially in emerging large language models. The success of these models has driven the demand for their universal deployment across a diverse set of backend environments. In this paper, we present Relax, a compiler abstraction for optimizing end-to-end dynamic machine learning workloads. Relax introduces a cross-level abstraction that encapsulates computational graphs, loop-level tensor programs, and external library calls in a single representation. Relax also introduces first-class symbolic shape annotations to track dynamic shape computations globally across the program, enabling dynamic shape-aware cross-level optimizations. We build an end-to-end compilation framework using the proposed approach to optimize dynamic shape models. Experimental results on LLMs show that Relax delivers performance competitive with state-of-the-art systems across various GPUs and enables deployment of emerging models to a broader set of emerging environments, including mobile phones, embedded devices, and web browsers.

Composed Image Retrieval (CIR) is a challenging task that aims to retrieve the target image with a multimodal query, i.e., a reference image, and its complementary modification text. As previous supervised or zero-shot learning paradigms all fail to strike a good trade-off between the model’s generalization ability and retrieval performance, recent researchers have introduced the task of few-shot CIR (FS-CIR) and proposed a textual inversion-based network based on pretrained CLIP model to realize it. Despite its promising performance, the approach encounters two key limitations: simply relying on the few annotated samples for CIR model training and indiscriminately selecting training triplets for CIR model fine-tuning. To address these two limitations, we propose a novel two-stage pseudo triplet guided few-shot CIR scheme, dubbed PTG-FSCIR. In the first stage, we propose an attentive masking and captioning-based pseudo triplet generation method, to construct pseudo triplets from pure image data and use them to fulfill the CIR-task specific pretraining. In the second stage, we propose a challenging triplet-based CIR fine-tuning method, where we design a pseudo modification text-based sample challenging score estimation strategy and a robust top range-based random sampling strategy for sampling robust challenging triplets to promote the model fine-tuning. Notably, our scheme is plug-and-play and compatible with any existing supervised CIR models. We test our scheme across two backbones on three public datasets (i.e., FashionIQ, CIRR, and Birds-to-Words), achieving maximum improvements of 13.3%, 22.2%, and 17.4% respectively, demonstrating our scheme’s efficacy.

Recent advancements in generative models have enabled high-fidelity text-to-image generation. However, open-source image-editing models still lag behind their proprietary counterparts, primarily due to limited high-quality data and insufficient benchmarks.To overcome these limitations, we introduce ImgEdit, a large-scale, high-quality image-editing dataset comprising one million carefully curated edit pairs, which contain both novel and complex single-turn edits, as well as challenging multi-turn tasks.To ensure the data quality, we employ a multi-stage pipeline that integrates a cutting-edge vision-language model, a detection model, a segmentation model, alongside task-specific in-painting procedures and strict post-processing. ImgEdit surpasses existing datasets in both task novelty and data quality.Using ImgEdit, we train ImgEdit-E1, an editing model using Vision Language Model to process the reference image and editing prompt, which outperforms existing open-source models on multiple tasks, highlighting the value of ImgEdit and model design.For comprehensive evaluation, we introduce ImgEdit-Bench, a benchmark designed to evaluate image editing performance in terms of instruction adherence, editing quality, and detail preservation.It includes a basic testsuite, a challenging single-turn suite, and a dedicated multi-turn suite. We evaluate both open-source and proprietary models, as well as ImgEdit-E1, providing deep analysis and actionable insights into the current behavior of image-editing models.

Due to their strong generalizable multimodal processing and reasoning capabilities, Multimodal Large Language Models (MLLMs) have demonstrated significant potential as universal image retrievers, effectively addressing diverse real-world image retrieval tasks.Nevertheless, pioneering studies, while promising, overlook the potential of fine-grained context modeling and disentangled fine-tuning objectives in enhancing MLLMs' retrieval performance, particularly for complex tasks such as long-text-to-image retrieval, visual dialog retrieval, and composed image retrieval (CIR).Therefore, in this work, we propose an automated fine-grained multimodal quintuple dataset construction pipeline and a novel two-stage fine-grained multimodal fine-tuning strategy.The dataset generation pipeline produces a comprehensive CIR dataset with fine-grained image captions and modification text, facilitating fine-grained context modeling.Beyond the previously entangled fine-tuning paradigm, our approach separates the fine-tuning process into two distinct stages: (1) fine-grained context reasoning-oriented fine-tuning and (2) fine-grained retrieval-oriented fine-tuning.These stages aim to sequentially enhance the model's context understanding and query-target alignment capabilities, thereby improving retrieval performance.Extensive experiments across five datasets encompassing diverse and complex image retrieval tasks demonstrate the remarkable superiority of our method over existing approaches in *Xuemeng Song (sxmustc

As a key economic crop, jujube’s external morphology directly affects quality grading and market value. However, traditional inspection methods relying on manual sampling or 2D image analysis suffer from inefficiency and limited feature characterization, particularly in quantifying complex geometric traits such as irregular wrinkles and localized depressions on jujube surfaces. Existing 3D reconstruction techniques have been applied in agricultural product inspection but face challenges in widespread adoption due to high costs and low resolution. This study proposes a single-view high-resolution RGB 3D reconstruction method for jujube based on generative artificial intelligence. Specifically, we designed a two-stage single-view 3D reconstruction framework and modified the cross-attention layers in the U-shaped network architecture of a stable video diffusion mode to meet the requirements of high-resolution and clear texture reconstruction for jujube. Additionally, we improved training and inference efficiency through parallel computing and achieved automation integration with Unmanned Agents. The proposed method successfully reconstructed 3D models from single-view 1,024 1,024 RGB images of jujube. Our model achieves a PSNR of 23.52 and an SSIM of 0.86 on the public dataset.This approach provides a low-cost, high-precision 3D digital solution for non-destructive phenotyping of jujube, offering practical value for advancing intelligent sorting and quality evaluation in agricultural production.

Kernel orchestration is the task of mapping the computation defined in different operators of a deep neural network (DNN) to the execution of GPU kernels on modern hardware platforms. Prior approaches optimize kernel orchestration by greedily applying operator fusion, which fuses the computation of multiple operators into a single kernel, and miss a variety of optimization opportunities in kernel orchestration.

Deploying deep learning models on various devices has become an important topic. The wave of hardware specialization brings a diverse set of acceleration primitives for multi-dimensional ten- sor computations. These new acceleration primitives, along with the emerging machine learning models, bring tremendous engineering challenges. In this paper, we present TensorIR, a compiler abstraction for optimizing programs with these tensor computation primitives. TensorIR generalizes the loop nest representation used in existing machine learning compilers to bring tensor computation as the first-class citizen. Finally, we build an end-to-end framework on top of our abstraction to automatically optimize deep learning models for given tensor computation primitives. Experimental results show that TensorIR compilation automatically uses the tensor computation primitives for given hardware backends and delivers performance that is competitive to state-of-art hand-optimized systems across platforms.

Automatic optimization for tensor programs becomes increasingly important as we deploy deep learning in various environments, and efficient optimization relies on a rich search space and effective search. Most existing efforts adopt a search space which lacks the ability to efficiently enable domain experts to grow the search space. This paper introduces MetaSchedule, a domain-specific probabilistic programming language abstraction to construct a rich search space of tensor programs. Our abstraction allows domain experts to analyze the program, and easily propose stochastic choices in a modular way to compose program transformation accordingly. We also build an end-to-end learning-driven framework to find an optimized program for a given search space. Experimental results show that MetaSchedule can cover the search space used in the state-of-the-art tensor program optimization frameworks in a modular way. Additionally, it empowers domain experts to conveniently grow the search space and modularly enhance the system, which brings 48% speedup on end-to-end deep learning workloads.