Search

As the scale of models and training data continues to grow, there is an expanding reliance on more GPUs to train large-scale models, which inevitably increases the likelihood of encountering dynamic stragglers that some devices lag behind in performance occasionally. However, hybrid parallel training, one of the de facto paradigms to train large models, is typically sensitive to the stragglers. This paper presents Malleus, a straggler-resilient hybrid parallel training framework for large-scale models. Malleus captures the dynamic straggler issues at the nuanced, per-GPU granularity during training. Once a shift in the GPU ability is detected, Malleus adaptively adjusts the parallelization of GPU devices, pipeline stages, model layers, and training data through a novel planning algorithm, accommodating the dynamic stragglers in real time. In addition, Malleus seamlessly and efficiently migrates the model states to fulfill the adjusted parallelization plan on the fly, without sacrificing the stability of the training tasks. Empirical results on large language models with up to 110B parameters show that Malleus consistently outperforms existing parallel training frameworks under various straggler situations, delivering on average 2.63-5.28 times of efficiency improvement.

Large Language Models (LLMs) have exhibited significant potential in performing diverse tasks, including the ability to call functions or use external tools to enhance their performance. While current research on function calling by LLMs primarily focuses on single-turn interactions, this paper addresses the overlooked necessity for LLMs to engage in multi-turn function calling--critical for handling compositional, real-world queries that require planning with functions but not only use functions. To facilitate this, we introduce an approach, BUTTON, which generates synthetic compositional instruction tuning data via bottom-up instruction construction and top-down trajectory generation. In the bottom-up phase, we generate simple atomic tasks based on real-world scenarios and build compositional tasks using heuristic strategies based on atomic tasks. Corresponding functions are then developed for these compositional tasks. The top-down phase features a multi-agent environment where interactions among simulated humans, assistants, and tools are utilized to gather multi-turn function calling trajectories. This approach ensures task compositionality and allows for effective function and trajectory generation by examining atomic tasks within compositional tasks. We produce a dataset BUTTONInstruct comprising 8k data points and demonstrate its effectiveness through extensive experiments across various LLMs.

Human feedback is crucial in the interactions between humans and Large Language Models (LLMs). However, existing research primarily focuses on benchmarking LLMs in single-turn dialogues. Even in benchmarks designed for multi-turn dialogues, the user inputs are often independent, neglecting the nuanced and complex nature of human feedback within real-world usage scenarios. To fill this research gap, we introduce FB-Bench, a fine-grained, multi-task benchmark designed to evaluate LLMs' responsiveness to human feedback in real-world usage scenarios. Drawing from the two main interaction scenarios, FB-Bench comprises 734 meticulously curated samples, encompassing eight task types, five deficiency types of response, and nine feedback types. We extensively evaluate a broad array of popular LLMs, revealing significant variations in their performance across different interaction scenarios. Further analysis indicates that task, human feedback, and deficiencies of previous responses can also significantly impact LLMs' responsiveness. Our findings underscore both the strengths and limitations of current models, providing valuable insights and directions for future research. Both the toolkits and the dataset of FB-Bench are available at https://github.com/PKU-Baichuan-MLSystemLab/FB-Bench.

Large language models (LLMs) like GPT-4, PaLM, and LLaMA have shown significant improvements in various reasoning tasks. However, smaller models such as Llama-3-8B and DeepSeekMath-Base still struggle with complex mathematical reasoning because they fail to effectively identify and correct reasoning errors. Recent reflection-based methods aim to address these issues by enabling self-reflection and self-correction, but they still face challenges in independently detecting errors in their reasoning steps. To overcome these limitations, we propose SuperCorrect, a novel two-stage framework that uses a large teacher model to supervise and correct both the reasoning and reflection processes of a smaller student model. In the first stage, we extract hierarchical high-level and detailed thought templates from the teacher model to guide the student model in eliciting more fine-grained reasoning thoughts. In the second stage, we introduce cross-model collaborative direct preference optimization (DPO) to enhance the self-correction abilities of the student model by following the teacher's correction traces during training. This cross-model DPO approach teaches the student model to effectively locate and resolve erroneous thoughts with error-driven insights from the teacher model, breaking the bottleneck of its thoughts and acquiring new skills and knowledge to tackle challenging problems. Extensive experiments consistently demonstrate our superiority over previous methods. Notably, our SuperCorrect-7B model significantly surpasses powerful DeepSeekMath-7B by 7.8%/5.3% and Qwen2.5-Math-7B by 15.1%/6.3% on MATH/GSM8K benchmarks, achieving new SOTA performance among all 7B models. Code: https://github.com/YangLing0818/SuperCorrect-llm, Comment: Project: https://github.com/YangLing0818/SuperCorrect-llm

Advanced diffusion models like RPG, Stable Diffusion 3 and FLUX have made notable strides in compositional text-to-image generation. However, these methods typically exhibit distinct strengths for compositional generation, with some excelling in handling attribute binding and others in spatial relationships. This disparity highlights the need for an approach that can leverage the complementary strengths of various models to comprehensively improve the composition capability. To this end, we introduce IterComp, a novel framework that aggregates composition-aware model preferences from multiple models and employs an iterative feedback learning approach to enhance compositional generation. Specifically, we curate a gallery of six powerful open-source diffusion models and evaluate their three key compositional metrics: attribute binding, spatial relationships, and non-spatial relationships. Based on these metrics, we develop a composition-aware model preference dataset comprising numerous image-rank pairs to train composition-aware reward models. Then, we propose an iterative feedback learning method to enhance compositionality in a closed-loop manner, enabling the progressive self-refinement of both the base diffusion model and reward models over multiple iterations. Theoretical proof demonstrates the effectiveness and extensive experiments show our significant superiority over previous SOTA methods (e.g., Omost and FLUX), particularly in multi-category object composition and complex semantic alignment. IterComp opens new research avenues in reward feedback learning for diffusion models and compositional generation. Code: https://github.com/YangLing0818/IterComp, Comment: Project: https://github.com/YangLing0818/IterComp

Visual-language models have advanced the development of universal models, yet their application in medical imaging remains constrained by specific functional requirements and the limited data. Current general-purpose models are typically designed with task-specific branches and heads, which restricts the shared feature space and the flexibility of model. To address these challenges, we have developed a decomposed-composed universal medical imaging paradigm (UniMed) that supports tasks at all levels. To this end, we first propose a decomposed decoder that can predict two types of outputs -- pixel and semantic, based on a defined input queue. Additionally, we introduce a composed decoder that unifies the input and output spaces and standardizes task annotations across different levels into a discrete token format. The coupled design of these two components enables the model to flexibly combine tasks and mutual benefits. Moreover, our joint representation learning strategy skilfully leverages large amounts of unlabeled data and unsupervised loss, achieving efficient one-stage pretraining for more robust performance. Experimental results show that UniMed achieves state-of-the-art performance on eight datasets across all three tasks and exhibits strong zero-shot and 100-shot transferability. We will release the code and trained models upon the paper's acceptance.

Large language models (LLMs) have demonstrated exceptional performance across a wide range of tasks and domains, with data preparation playing a critical role in achieving these results. Pre-training data typically combines information from multiple domains. To maximize performance when integrating data from various domains, determining the optimal data proportion is essential. However, state-of-the-art (SOTA) LLMs rarely disclose details about their pre-training data, making it difficult for researchers to identify ideal data proportions. In this paper, we introduce a new topic, \textit{data proportion detection}, which enables the automatic estimation of pre-training data proportions by analyzing the generated outputs of LLMs. We provide rigorous theoretical proofs, practical algorithms, and preliminary experimental results for data proportion detection. Based on these findings, we offer valuable insights into the challenges and future directions for effective data proportion detection and data management.

Recently, using large language models (LLMs) for data augmentation has led to considerable improvements in unsupervised sentence embedding models. However, existing methods encounter two primary challenges: limited data diversity and high data noise. Current approaches often neglect fine-grained knowledge, such as entities and quantities, leading to insufficient diversity. Additionally, unsupervised data frequently lacks discriminative information, and the generated synthetic samples may introduce noise. In this paper, we propose a pipeline-based data augmentation method via LLMs and introduce the Gaussian-decayed gradient-assisted Contrastive Sentence Embedding (GCSE) model to enhance unsupervised sentence embeddings. To tackle the issue of low data diversity, our pipeline utilizes knowledge graphs (KGs) to extract entities and quantities, enabling LLMs to generate more diverse, knowledge-enriched samples. To address high data noise, the GCSE model uses a Gaussian-decayed function to limit the impact of false hard negative samples, enhancing the model's discriminative capability. Experimental results show that our approach achieves state-of-the-art performance in semantic textual similarity (STS) tasks, using fewer data samples and smaller LLMs, demonstrating its efficiency and robustness across various models.

Recent foundation models are capable of handling multiple machine learning (ML) tasks and multiple data modalities with the unified base model structure and several specialized model components. However, the development of such multi-task (MT) multi-modal (MM) models poses significant model management challenges to existing training systems. Due to the sophisticated model architecture and the heterogeneous workloads of different ML tasks and data modalities, training these models usually requires massive GPU resources and suffers from sub-optimal system efficiency. In this paper, we investigate how to achieve high-performance training of large-scale MT MM models through data heterogeneity-aware model management optimization. The key idea is to decompose the model execution into stages and address the joint optimization problem sequentially, including both heterogeneity-aware workload parallelization and dependency-driven execution scheduling. Based on this, we build a prototype system and evaluate it on various large MT MM models. Experiments demonstrate the superior performance and efficiency of our system, with speedup ratio up to 71% compared to state-of-the-art training systems.

In recent years, Large Language Models (LLMs) have demonstrated significant improvements across a variety of tasks, one of which is the long-context capability. The key to improving long-context performance lies in effective data organization and management strategies that integrate data from multiple domains and optimize the context window during training. Through extensive experimental analysis, we identified three key challenges in designing effective data management strategies that enable the model to achieve long-context capability without sacrificing performance in other tasks: (1) a shortage of long documents across multiple domains, (2) effective construction of context windows, and (3) efficient organization of large-scale datasets. To address these challenges, we introduce DataSculpt, a novel data management framework designed for long-context training. We first formulate the organization of training data as a multi-objective combinatorial optimization problem, focusing on attributes including relevance, homogeneity, integrity, and efficiency. Specifically, our approach utilizes a coarse-to-fine methodology to optimize training data organization both efficiently and effectively. We begin by clustering the data based on semantic similarity (coarse), followed by a multi-objective greedy search within each cluster to score and concatenate documents into various context windows (fine). Our comprehensive evaluations demonstrate that DataSculpt significantly enhances long-context training performance, resulting in improvements of 18.09% in retrieval augmentation, 21.23% in summarization, 21.27% in reading comprehension, and a 3.81% increase in code completion, while also maintaining overall model proficiency with a 4.88% improvement.

Training large language model (LLM) is a computationally intensive task, which is typically conducted in data centers with homogeneous high-performance GPUs. This paper explores an alternative approach by deploying the training computation across heterogeneous GPUs to enable better flexibility and efficiency for heterogeneous resource utilization. To achieve this goal, we propose a novel system, FlashFlex, that can flexibly support an asymmetric partition of the parallel training computations across the scope of data-, pipeline-, and tensor model parallelism. We further formalize the allocation of asymmetric partitioned training computations over a set of heterogeneous GPUs as a constrained optimization problem and propose an efficient solution based on a hierarchical graph partitioning algorithm. Our approach can adaptively allocate asymmetric training computations across GPUs, fully leveraging the available computational power. We conduct extensive empirical studies to evaluate the performance of FlashFlex, where we find that when training LLMs at different scales (from 7B to 30B), FlashFlex can achieve comparable training MFU when running over a set of heterogeneous GPUs compared with the state of the art training systems running over a set of homogeneous high-performance GPUs with the same amount of total peak FLOPS. The achieved smallest gaps in MFU are 11.61% and 0.30%, depending on whether the homogeneous setting is equipped with and without RDMA. Our implementation is available at https://github.com/Relaxed-System-Lab/FlashFlex.

Large Language Models (LLMs) have become instrumental across various applications, with the customization of these models to specific scenarios becoming increasingly critical. System message, a fundamental component of LLMs, is consist of carefully crafted instructions that guide the behavior of model to meet intended goals. Despite the recognized potential of system messages to optimize AI-driven solutions, there is a notable absence of a comprehensive benchmark for evaluating how well different LLMs follow these system messages. To fill this gap, we introduce SysBench, a benchmark that systematically analyzes system message following ability in terms of three challenging aspects: constraint complexity, instruction misalignment and multi-turn stability. In order to enable effective evaluation, SysBench constructs multi-turn user conversations covering various interaction relationships, based on six common types of constraints from system messages in real-world scenarios. Our dataset contains 500 system messages from various domains, each paired with 5 turns of user conversations, which have been manually formulated and checked to guarantee high quality. SysBench provides extensive evaluation across various LLMs, measuring their ability to follow specified constraints given in system messages. The results highlight both the strengths and weaknesses of existing models, offering key insights and directions for future research. The open source library SysBench is available at https://github.com/PKU-Baichuan-MLSystemLab/SysBench.

With the development of Multimodal Large Language Models (MLLMs), the evaluation of multimodal models in the context of mathematical problems has become a valuable research field. Multimodal visual-textual mathematical reasoning serves as a critical indicator for evaluating the comprehension and complex multi-step quantitative reasoning abilities of MLLMs. However, previous multimodal math benchmarks have not sufficiently integrated visual and textual information. To address this gap, we proposed MathScape, a new benchmark that emphasizes the understanding and application of combined visual and textual information. MathScape is designed to evaluate photo-based math problem scenarios, assessing the theoretical understanding and application ability of MLLMs through a categorical hierarchical approach. We conduct a multi-dimensional evaluation on 11 advanced MLLMs, revealing that our benchmark is challenging even for the most sophisticated models. By analyzing the evaluation results, we identify the limitations of MLLMs, offering valuable insights for enhancing model performance.

The adeptness of Large Language Models (LLMs) in comprehending and following natural language instructions is critical for their deployment in sophisticated real-world applications. Existing evaluations mainly focus on fragmented constraints or narrow scenarios, but they overlook the comprehensiveness and authenticity of constraints from the user's perspective. To bridge this gap, we propose CFBench, a large-scale Comprehensive Constraints Following Benchmark for LLMs, featuring 1,000 curated samples that cover more than 200 real-life scenarios and over 50 NLP tasks. CFBench meticulously compiles constraints from real-world instructions and constructs an innovative systematic framework for constraint types, which includes 10 primary categories and over 25 subcategories, and ensures each constraint is seamlessly integrated within the instructions. To make certain that the evaluation of LLM outputs aligns with user perceptions, we propose an advanced methodology that integrates multi-dimensional assessment criteria with requirement prioritization, covering various perspectives of constraints, instructions, and requirement fulfillment. Evaluating current leading LLMs on CFBench reveals substantial room for improvement in constraints following, and we further investigate influencing factors and enhancement strategies. The data and code are publicly available at https://github.com/PKU-Baichuan-MLSystemLab/CFBench, Comment: 15 pages, 10 figures

Recently, with the rise of web images, managing and understanding large-scale image datasets has become increasingly important. Vision Large Language Models (VLLMs) have recently emerged due to their robust vision-understanding capabilities. However, training these models requires vast amounts of data, posing challenges to efficiency, effectiveness, data quality, and privacy. In this paper, we introduce SynthVLM, a novel data synthesis pipeline for VLLMs. Unlike existing methods that generate captions from images, SynthVLM employs advanced diffusion models and high-quality captions to automatically generate and select high-resolution images from captions, creating precisely aligned image-text pairs. Leveraging these pairs, we achieve state-of-the-art (SoTA) performance on various vision question answering tasks, maintaining high alignment quality and preserving advanced language abilities. Moreover, SynthVLM surpasses traditional GPT-4 Vision-based caption generation methods in performance while significantly reducing computational overhead. Crucially, our method's reliance on purely generated data ensures the preservation of privacy, achieving SoTA performance with just 100k data points (only 18% of the official dataset size).

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing the reuse of memory across transformer layers. Empirical results demonstrate that MEMO achieves an average of 2.42x and 2.26x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.

In recent years, the rise of Large Language Models (LLMs) has spurred a growing demand for plug-and-play AI systems. Among the various AI techniques, prompt engineering stands out as particularly significant. However, users often face challenges in writing prompts due to the steep learning curve and significant time investment, and existing automatic prompt engineering (APE) models can be difficult to use. To address this issue, we propose PAS, an LLM-based plug-and-play APE system. PAS utilizes LLMs trained on high-quality, automatically generated prompt complementary datasets, resulting in exceptional performance. In comprehensive benchmarks, PAS achieves state-of-the-art (SoTA) results compared to previous APE models, with an average improvement of 6.09 points. Moreover, PAS is highly efficient, achieving SoTA performance with only 9000 data points. Additionally, PAS can autonomously generate prompt augmentation data without requiring additional human labor. Its flexibility also allows it to be compatible with all existing LLMs and applicable to a wide range of tasks. PAS excels in human evaluations, underscoring its suitability as a plug-in for users. This combination of high performance, efficiency, and flexibility makes PAS a valuable system for enhancing the usability and effectiveness of LLMs through improved prompt engineering.

Recently, with the rise of web videos, managing and understanding large-scale video datasets has become increasingly important. Video Large Language Models (VideoLLMs) have emerged in recent years due to their strong video understanding capabilities. However, training and inference processes for VideoLLMs demand vast amounts of data, presenting significant challenges to data management, particularly regarding efficiency, robustness, and effectiveness. In this work, we present KeyVideoLLM, a text-video frame similarity-based keyframe selection method designed to manage VideoLLM data efficiently, robustly, and effectively. Specifically, KeyVideoLLM achieves a remarkable data compression rate of up to 60.9 times, substantially lowering disk space requirements, which proves its high efficiency. Additionally, it maintains a 100% selection success rate across all video formats and scales, enhances processing speed by up to 200 times compared to existing keyframe selection methods, and does not require hyperparameter tuning. Beyond its outstanding efficiency and robustness, KeyVideoLLM further improves model performance in video question-answering tasks during both training and inference stages. Notably, it consistently achieved the state-of-the-art (SoTA) experimental results on diverse datasets.

Flow matching (FM) is a general framework for defining probability paths via Ordinary Differential Equations (ODEs) to transform between noise and data samples. Recent approaches attempt to straighten these flow trajectories to generate high-quality samples with fewer function evaluations, typically through iterative rectification methods or optimal transport solutions. In this paper, we introduce Consistency Flow Matching (Consistency-FM), a novel FM method that explicitly enforces self-consistency in the velocity field. Consistency-FM directly defines straight flows starting from different times to the same endpoint, imposing constraints on their velocity values. Additionally, we propose a multi-segment training approach for Consistency-FM to enhance expressiveness, achieving a better trade-off between sampling quality and speed. Preliminary experiments demonstrate that our Consistency-FM significantly improves training efficiency by converging 4.4x faster than consistency models and 1.7x faster than rectified flow models while achieving better generation quality. Our code is available at: https://github.com/YangLing0818/consistency_flow_matching, Comment: Code: https://github.com/YangLing0818/consistency_flow_matching

In recent years, with the rapid advancements in large language models (LLMs), achieving excellent empathetic response capability has become a crucial prerequisite. Consequently, managing and understanding large-scale video datasets has gained increasing importance. However, empathetic data are typically trained without any quality selection, leading to inefficient data usage and wasted computational resources. Additionally, using raw data can result in low performance in empathetic dialogues. In this work, we present Efficient-Empathy, a sensibility and rationality score-based data selection algorithm that automatically selects sensibility and rationality data while discarding low-quality data. With only the sensibility data (59% of the full dataset), our trained sensibility model efficiently achieves state-of-the-art (SoTA) performance. Furthermore, with multiple data selection hyperparameters, the sensibility model demonstrates SoTA performance, showcasing the robustness of our method. By integrating sensibility and rationality data with a MoE structure, we achieve even higher performance, demonstrating the effectiveness of our Efficient-Empathy algorithm.

As the field of Large Language Models (LLMs) continues to evolve, the context length in inference is steadily growing. Key-Value Cache (KVCache), a crucial component in LLM inference, has now become the primary memory bottleneck due to limited GPU memory. Current methods selectively determine suitable keys and values for self-attention computation in LLMs to address the issue. However, they either fall short in maintaining model quality or result in high serving latency. Drawing inspiration from advanced embedding retrieval techniques used in the database community, we consider the storage and searching of KVCache as a typical embedding retrieval problem. We propose PQCache, which employs Product Quantization (PQ) to manage KVCache, maintaining model quality while ensuring low serving latency. During the prefilling phase, we apply PQ to tokens' keys for each LLM layer and head. During the autoregressive decoding phase, for each newly generated token, we first identify important tokens through Maximum Inner-Product Search (MIPS) using PQ codes and centroids, then fetch the corresponding key-value pairs for self-attention computation. Through meticulous design of overlapping and caching, we minimize any additional computation and communication overhead during both phases. Extensive experiments show that PQCache achieves both effectiveness and efficiency. It maintains model quality even when only 1/5 of the tokens are involved in attention, while attaining acceptable system latency.

Graph neural networks (GNNs) have attracted widespread attention for their impressive capability of graph representation learning. However, the increasing prevalence of large-scale graphs presents a significant challenge for GNN training due to their computational demands, limiting the applicability of GNNs in various scenarios. In response to this challenge, graph condensation (GC) is proposed as a promising acceleration solution, focusing on generating an informative compact graph that enables efficient training of GNNs while retaining performance. Despite the potential to accelerate GNN training, existing GC methods overlook the quality of large training graphs during both the training and inference stages. They indiscriminately emulate the training graph distributions, making the condensed graphs susceptible to noises within the training graph and significantly impeding the application of GC in intricate real-world scenarios. To address this issue, we propose robust graph condensation (RobGC), a plug-and-play approach for GC to extend the robustness and applicability of condensed graphs in noisy graph structure environments. Specifically, RobGC leverages the condensed graph as a feedback signal to guide the denoising process on the original training graph. A label propagation-based alternating optimization strategy is in place for the condensation and denoising processes, contributing to the mutual purification of the condensed graph and training graph. Additionally, as a GC method designed for inductive graph inference, RobGC facilitates test-time graph denoising by leveraging the noise-free condensed graph to calibrate the structure of the test graph. Extensive experiments show that RobGC is compatible with various GC methods, significantly boosting their robustness under different types and levels of graph structural noises.

Diffusion models have demonstrated great success in text-to-video (T2V) generation. However, existing methods may face challenges when handling complex (long) video generation scenarios that involve multiple objects or dynamic changes in object numbers. To address these limitations, we propose VideoTetris, a novel framework that enables compositional T2V generation. Specifically, we propose spatio-temporal compositional diffusion to precisely follow complex textual semantics by manipulating and composing the attention maps of denoising networks spatially and temporally. Moreover, we propose an enhanced video data preprocessing to enhance the training data regarding motion dynamics and prompt understanding, equipped with a new reference frame attention mechanism to improve the consistency of auto-regressive video generation. Extensive experiments demonstrate that our VideoTetris achieves impressive qualitative and quantitative results in compositional T2V generation. Code is available at: https://github.com/YangLing0818/VideoTetris, Comment: NeurIPS 2024. Code: https://github.com/YangLing0818/VideoTetris

We introduce Buffer of Thoughts (BoT), a novel and versatile thought-augmented reasoning approach for enhancing accuracy, efficiency and robustness of large language models (LLMs). Specifically, we propose meta-buffer to store a series of informative high-level thoughts, namely thought-template, distilled from the problem-solving processes across various tasks. Then for each problem, we retrieve a relevant thought-template and adaptively instantiate it with specific reasoning structures to conduct efficient reasoning. To guarantee the scalability and stability, we further propose buffer-manager to dynamically update the meta-buffer, thus enhancing the capacity of meta-buffer as more tasks are solved. We conduct extensive experiments on 10 challenging reasoning-intensive tasks, and achieve significant performance improvements over previous SOTA methods: 11% on Game of 24, 20% on Geometric Shapes and 51% on Checkmate-in-One. Further analysis demonstrate the superior generalization ability and model robustness of our BoT, while requiring only 12% of the cost of multi-query prompting methods (e.g., tree/graph of thoughts) on average. Notably, we find that our Llama3-8B+BoT has the potential to surpass Llama3-70B model. Our project is available at: https://github.com/YangLing0818/buffer-of-thought-llm, Comment: NeurIPS 2024 Spotlight. Project: https://github.com/YangLing0818/buffer-of-thought-llm

Multimodal large language models (MLLMs) enhance the capabilities of standard large language models by integrating and processing data from multiple modalities, including text, vision, audio, video, and 3D environments. Data plays a pivotal role in the development and refinement of these models. In this survey, we comprehensively review the literature on MLLMs from a data-centric perspective. Specifically, we explore methods for preparing multimodal data during the pretraining and adaptation phases of MLLMs. Additionally, we analyze the evaluation methods for the datasets and review the benchmarks for evaluating MLLMs. Our survey also outlines potential future research directions. This work aims to provide researchers with a detailed understanding of the data-driven aspects of MLLMs, fostering further exploration and innovation in this field.

Graphs play an increasingly important role in various big data applications. However, existing graph data structures cannot simultaneously address the performance bottlenecks caused by the dynamic updates, large scale, and high query complexity of current graphs. This paper proposes a novel data structure for large-scale dynamic graphs called CuckooGraph. It does not require any prior knowledge of the upcoming graphs, and can adaptively resize to the most memory-efficient form while requiring few memory accesses for very fast graph data processing. The key techniques of CuckooGraph include TRANSFORMATION and DENYLIST. TRANSFORMATION fully utilizes the limited memory by designing related data structures that allow flexible space transformations to smoothly expand/tighten the required space depending on the number of incoming items. DENYLIST efficiently handles item insertion failures and further improves processing speed. Our experimental results show that compared with the most competitive solution Spruce, CuckooGraph achieves about $33\times$ higher insertion throughput while requiring only about $68\%$ of the memory space.

In current deep learning tasks, Adam style optimizers such as Adam, Adagrad, RMSProp, Adafactor, and Lion have been widely used as alternatives to SGD style optimizers. These optimizers typically update model parameters using the sign of gradients, resulting in more stable convergence curves. The learning rate and the batch size are the most critical hyperparameters for optimizers, which require careful tuning to enable effective convergence. Previous research has shown that the optimal learning rate increases linearly or follows similar rules with batch size for SGD style optimizers. However, this conclusion is not applicable to Adam style optimizers. In this paper, we elucidate the connection between optimal learning rates and batch sizes for Adam style optimizers through both theoretical analysis and extensive experiments. First, we raise the scaling law between batch sizes and optimal learning rates in the sign of gradient case, in which we prove that the optimal learning rate first rises and then falls as the batch size increases. Moreover, the peak value of the surge will gradually move toward the larger batch size as training progresses. Second, we conducted experiments on various CV and NLP tasks and verified the correctness of the scaling law.

Graph Neural Networks (GNNs) have demonstrated effectiveness in various graph-based tasks. However, their inefficiency in training and inference presents challenges for scaling up to real-world and large-scale graph applications. To address the critical challenges, a range of algorithms have been proposed to accelerate training and inference of GNNs, attracting increasing attention from the research community. In this paper, we present a systematic review of acceleration algorithms in GNNs, which can be categorized into three main topics based on their purpose: training acceleration, inference acceleration, and execution acceleration. Specifically, we summarize and categorize the existing approaches for each main topic, and provide detailed characterizations of the approaches within each category. Additionally, we review several libraries related to acceleration algorithms in GNNs and discuss our Scalable Graph Learning (SGL) library. Finally, we propose promising directions for future research. A complete summary is presented in our GitHub repository: https://github.com/PKU-DAIR/SGL/blob/main/Awsome-GNN-Acceleration.md., Comment: 9 pages,3 figures

Large language models (LLMs) suffer from low efficiency as the mismatch between the requirement of auto-regressive decoding and the design of most contemporary GPUs. Specifically, billions to trillions of parameters must be loaded to the GPU cache through its limited memory bandwidth for computation, but only a small batch of tokens is actually computed. Consequently, the GPU spends most of its time on memory transfer instead of computation. Recently, parallel decoding, a type of speculative decoding algorithms, is becoming more popular and has demonstrated impressive efficiency improvement in generation. It introduces extra decoding heads to large models, enabling them to predict multiple subsequent tokens simultaneously and verify these candidate continuations in a single decoding step. However, this approach deviates from the training objective of next token prediction used during pre-training, resulting in a low hit rate for candidate tokens. In this paper, we propose a new speculative decoding algorithm, Clover, which integrates sequential knowledge into the parallel decoding process. This enhancement improves the hit rate of speculators and thus boosts the overall efficiency. Clover transmits the sequential knowledge from pre-speculated tokens via the Regressive Connection, then employs an Attention Decoder to integrate these speculated tokens. Additionally, Clover incorporates an Augmenting Block that modifies the hidden states to better align with the purpose of speculative generation rather than next token prediction. The experiment results demonstrate that Clover outperforms the baseline by up to 91% on Baichuan-Small and 146% on Baichuan-Large, respectively, and exceeds the performance of the previously top-performing method, Medusa, by up to 37% on Baichuan-Small and 57% on Baichuan-Large, respectively.

Optimizing the Key-Value (KV) cache of the Large Language Model (LLM) has been considered critical to saving the cost of inference. Most of the existing KV-cache compression algorithms attempted to sparsify the sequence of tokens by taking advantage of the different importance of tokens. However, most of these methods treat all layers equally, allocating the same KV budget to each layer. This approach is suboptimal, as some layers may be less sensitive to input tokens yet still receive the same budget as others. In this work, we found that by identifying the importance of attention layers, we could optimize the KV-cache jointly from two dimensions, i.e., sequence-wise and layer-wise. Based on our observations regarding layer-wise importance in inference, we propose SqueezeAttention to precisely optimize the allocation of KV-cache budget among layers on-the-fly and then incorporate three representative sequence-wise algorithms to compress the KV-cache for each layer with its very own budget. Specifically, we first measure each layer's importance by calculating the cosine similarity of the input prompt differences before and after the self-attention layers. Based on this similarity, we then categorize the layers into two groups and adjust their KV budgets accordingly. By optimizing the KV-cache from both sequence's and layer's dimensions, SqueezeAttention achieves around 30% to 70% of the memory reductions and up to 2.2 times of throughput improvements in a wide range of LLMs and benchmarks. The code is available at https://github.com/hetailang/SqueezeAttention.

With the proliferation of spatio-textual data, Top-k KNN spatial keyword queries (TkQs), which return a list of objects based on a ranking function that evaluates both spatial and textual relevance, have found many real-life applications. Existing geo-textual indexes for TkQs use traditional retrieval models like BM25 to compute text relevance and usually exploit a simple linear function to compute spatial relevance, but its effectiveness is limited. To improve effectiveness, several deep learning models have recently been proposed, but they suffer severe efficiency issues. To the best of our knowledge, there are no efficient indexes specifically designed to accelerate the top-k search process for these deep learning models. To tackle these issues, we propose a novel technique, which Learns to Index the Spatio-Textual data for answering embedding based spatial keyword queries (called LIST). LIST is featured with two novel components. Firstly, we propose a lightweight and effective relevance model that is capable of learning both textual and spatial relevance. Secondly, we introduce a novel machine learning based Approximate Nearest Neighbor Search (ANNS) index, which utilizes a new learning-to-cluster technique to group relevant queries and objects together while separating irrelevant queries and objects. Two key challenges in building an effective and efficient index are the absence of high-quality labels and unbalanced clustering results. We develop a novel pseudo-label generation technique to address the two challenges. Experimental results show that LIST significantly outperforms state-of-the-art methods on effectiveness, with improvements up to 19.21% and 12.79% in terms of NDCG@1 and Recall@10, and is three orders of magnitude faster than the most effective baseline.

Advancements in model algorithms, the growth of foundational models, and access to high-quality datasets have propelled the evolution of Artificial Intelligence Generated Content (AIGC). Despite its notable successes, AIGC still faces hurdles such as updating knowledge, handling long-tail data, mitigating data leakage, and managing high training and inference costs. Retrieval-Augmented Generation (RAG) has recently emerged as a paradigm to address such challenges. In particular, RAG introduces the information retrieval process, which enhances the generation process by retrieving relevant objects from available data stores, leading to higher accuracy and better robustness. In this paper, we comprehensively review existing efforts that integrate RAG technique into AIGC scenarios. We first classify RAG foundations according to how the retriever augments the generator, distilling the fundamental abstractions of the augmentation methodologies for various retrievers and generators. This unified perspective encompasses all RAG scenarios, illuminating advancements and pivotal technologies that help with potential future progress. We also summarize additional enhancements methods for RAG, facilitating effective engineering and implementation of RAG systems. Then from another view, we survey on practical applications of RAG across different modalities and tasks, offering valuable references for researchers and practitioners. Furthermore, we introduce the benchmarks for RAG, discuss the limitations of current RAG systems, and suggest potential directions for future research. Github: https://github.com/PKU-DAIR/RAG-Survey., Comment: Citing 353 papers, 22 pages, 1 table, 12 figures. Project: https://github.com/PKU-DAIR/RAG-Survey

Diffusion models have demonstrated exceptional efficacy in various generative applications. While existing models focus on minimizing a weighted sum of denoising score matching losses for data distribution modeling, their training primarily emphasizes instance-level optimization, overlooking valuable structural information within each mini-batch, indicative of pair-wise relationships among samples. To address this limitation, we introduce Structure-guided Adversarial training of Diffusion Models (SADM). In this pioneering approach, we compel the model to learn manifold structures between samples in each training batch. To ensure the model captures authentic manifold structures in the data distribution, we advocate adversarial training of the diffusion generator against a novel structure discriminator in a minimax game, distinguishing real manifold structures from the generated ones. SADM substantially improves existing diffusion transformers (DiT) and outperforms existing methods in image generation and cross-domain fine-tuning tasks across 12 datasets, establishing a new state-of-the-art FID of 1.58 and 2.11 on ImageNet for class-conditional image generation at resolutions of 256x256 and 512x512, respectively., Comment: Accepted by CVPR 2024

Conditional diffusion models have exhibited superior performance in high-fidelity text-guided visual generation and editing. Nevertheless, prevailing text-guided visual diffusion models primarily focus on incorporating text-visual relationships exclusively into the reverse process, often disregarding their relevance in the forward process. This inconsistency between forward and reverse processes may limit the precise conveyance of textual semantics in visual synthesis results. To address this issue, we propose a novel and general contextualized diffusion model (ContextDiff) by incorporating the cross-modal context encompassing interactions and alignments between text condition and visual sample into forward and reverse processes. We propagate this context to all timesteps in the two processes to adapt their trajectories, thereby facilitating cross-modal conditional modeling. We generalize our contextualized diffusion to both DDPMs and DDIMs with theoretical derivations, and demonstrate the effectiveness of our model in evaluations with two challenging tasks: text-to-image generation, and text-to-video editing. In each task, our ContextDiff achieves new state-of-the-art performance, significantly enhancing the semantic alignment between text condition and generated samples, as evidenced by quantitative and qualitative evaluations. Our code is available at https://github.com/YangLing0818/ContextDiff, Comment: ICLR 2024. Project: https://github.com/YangLing0818/ContextDiff

Diffusion models have achieved remarkable advancements in text-to-image generation. However, existing models still have many difficulties when faced with multiple-object compositional generation. In this paper, we propose RealCompo, a new training-free and transferred-friendly text-to-image generation framework, which aims to leverage the respective advantages of text-to-image models and spatial-aware image diffusion models (e.g., layout, keypoints and segmentation maps) to enhance both realism and compositionality of the generated images. An intuitive and novel balancer is proposed to dynamically balance the strengths of the two models in denoising process, allowing plug-and-play use of any model without extra training. Extensive experiments show that our RealCompo consistently outperforms state-of-the-art text-to-image models and spatial-aware image diffusion models in multiple-object compositional generation while keeping satisfactory realism and compositionality of the generated images. Notably, our RealCompo can be seamlessly extended with a wide range of spatial-aware image diffusion models and stylized diffusion models. Our code is available at: https://github.com/YangLing0818/RealCompo, Comment: NeurIPS 2024. Project: https://github.com/YangLing0818/RealCompo

Diffusion models have exhibit exceptional performance in text-to-image generation and editing. However, existing methods often face challenges when handling complex text prompts that involve multiple objects with multiple attributes and relationships. In this paper, we propose a brand new training-free text-to-image generation/editing framework, namely Recaption, Plan and Generate (RPG), harnessing the powerful chain-of-thought reasoning ability of multimodal LLMs to enhance the compositionality of text-to-image diffusion models. Our approach employs the MLLM as a global planner to decompose the process of generating complex images into multiple simpler generation tasks within subregions. We propose complementary regional diffusion to enable region-wise compositional generation. Furthermore, we integrate text-guided image generation and editing within the proposed RPG in a closed-loop fashion, thereby enhancing generalization ability. Extensive experiments demonstrate our RPG outperforms state-of-the-art text-to-image diffusion models, including DALL-E 3 and SDXL, particularly in multi-category object composition and text-image semantic alignment. Notably, our RPG framework exhibits wide compatibility with various MLLM architectures (e.g., MiniGPT-4) and diffusion backbones (e.g., ControlNet). Our code is available at: https://github.com/YangLing0818/RPG-DiffusionMaster, Comment: ICML 2024. Project: https://github.com/YangLing0818/RPG-DiffusionMaster

Diffusion models are a new class of generative models, and have dramatically promoted image generation with unprecedented quality and diversity. Existing diffusion models mainly try to reconstruct input image from a corrupted one with a pixel-wise or feature-wise constraint along spatial axes. However, such point-based reconstruction may fail to make each predicted pixel/feature fully preserve its neighborhood context, impairing diffusion-based image synthesis. As a powerful source of automatic supervisory signal, context has been well studied for learning representations. Inspired by this, we for the first time propose ConPreDiff to improve diffusion-based image synthesis with context prediction. We explicitly reinforce each point to predict its neighborhood context (i.e., multi-stride features/tokens/pixels) with a context decoder at the end of diffusion denoising blocks in training stage, and remove the decoder for inference. In this way, each point can better reconstruct itself by preserving its semantic connections with neighborhood context. This new paradigm of ConPreDiff can generalize to arbitrary discrete and continuous diffusion backbones without introducing extra parameters in sampling procedure. Extensive experiments are conducted on unconditional image generation, text-to-image generation and image inpainting tasks. Our ConPreDiff consistently outperforms previous methods and achieves a new SOTA text-to-image generation results on MS-COCO, with a zero-shot FID score of 6.21., Comment: Accepted by NeurIPS 2023

Given the sheer volume of contemporary e-commerce applications, recommender systems (RSs) have gained significant attention in both academia and industry. However, traditional cloud-based RSs face inevitable challenges, such as resource-intensive computation, reliance on network access, and privacy breaches. In response, a new paradigm called on-device recommender systems (ODRSs) has emerged recently in various industries like Taobao, Google, and Kuaishou. ODRSs unleash the computational capacity of user devices with lightweight recommendation models tailored for resource-constrained environments, enabling real-time inference with users' local data. This tutorial aims to systematically introduce methodologies of ODRSs, including (1) an overview of existing research on ODRSs; (2) a comprehensive taxonomy of ODRSs, where the core technical content to be covered span across three major ODRS research directions, including on-device deployment and inference, on-device training, and privacy/security of ODRSs; (3) limitations and future directions of ODRSs. This tutorial expects to lay the foundation and spark new insights for follow-up research and applications concerning this new recommendation paradigm., Comment: Technical tutorial; to appear at The Web Conference 2024

Recently, the growing memory demands of embedding tables in Deep Learning Recommendation Models (DLRMs) pose great challenges for model training and deployment. Existing embedding compression solutions cannot simultaneously meet three key design requirements: memory efficiency, low latency, and adaptability to dynamic data distribution. This paper presents CAFE, a Compact, Adaptive, and Fast Embedding compression framework that addresses the above requirements. The design philosophy of CAFE is to dynamically allocate more memory resources to important features (called hot features), and allocate less memory to unimportant ones. In CAFE, we propose a fast and lightweight sketch data structure, named HotSketch, to capture feature importance and report hot features in real time. For each reported hot feature, we assign it a unique embedding. For the non-hot features, we allow multiple features to share one embedding by using hash embedding technique. Guided by our design philosophy, we further propose a multi-level hash embedding framework to optimize the embedding tables of non-hot features. We theoretically analyze the accuracy of HotSketch, and analyze the model convergence against deviation. Extensive experiments show that CAFE significantly outperforms existing embedding compression methods, yielding 3.92% and 3.68% superior testing AUC on Criteo Kaggle dataset and CriteoTB dataset at a compression ratio of 10000x. The source codes of CAFE are available at GitHub.

Contrastive learning (CL) has recently gained prominence in the domain of recommender systems due to its great ability to enhance recommendation accuracy and improve model robustness. Despite its advantages, this paper identifies a vulnerability of CL-based recommender systems that they are more susceptible to poisoning attacks aiming to promote individual items. Our analysis indicates that this vulnerability is attributed to the uniform spread of representations caused by the InfoNCE loss. Furthermore, theoretical and empirical evidence shows that optimizing this loss favors smooth spectral values of representations. This finding suggests that attackers could facilitate this optimization process of CL by encouraging a more uniform distribution of spectral values, thereby enhancing the degree of representation dispersion. With these insights, we attempt to reveal a potential poisoning attack against CL-based recommender systems, which encompasses a dual-objective framework: one that induces a smoother spectral value distribution to amplify the InfoNCE loss's inherent dispersion effect, named dispersion promotion; and the other that directly elevates the visibility of target items, named rank promotion. We validate the threats of our attack model through extensive experimentation on four datasets. By shedding light on these vulnerabilities, our goal is to advance the development of more robust CL-based recommender systems. The code is available at \url{https://github.com/CoderWZW/ARLib}., Comment: 12 pages, 7 figures

Learnable embedding vector is one of the most important applications in machine learning, and is widely used in various database-related domains. However, the high dimensionality of sparse data in recommendation tasks and the huge volume of corpus in retrieval-related tasks lead to a large memory consumption of the embedding table, which poses a great challenge to the training and deployment of models. Recent research has proposed various methods to compress the embeddings at the cost of a slight decrease in model quality or the introduction of other overheads. Nevertheless, the relative performance of these methods remains unclear. Existing experimental comparisons only cover a subset of these methods and focus on limited metrics. In this paper, we perform a comprehensive comparative analysis and experimental evaluation of embedding compression. We introduce a new taxonomy that categorizes these techniques based on their characteristics and methodologies, and further develop a modular benchmarking framework that integrates 14 representative methods. Under a uniform test environment, our benchmark fairly evaluates each approach, presents their strengths and weaknesses under different memory budgets, and recommends the best method based on the use case. In addition to providing useful guidelines, our study also uncovers the limitations of current methods and suggests potential directions for future research.

Graph neural networks (GNNs) have exhibited exceptional efficacy in a diverse array of applications. However, the sheer size of large-scale graphs presents a significant challenge to real-time inference with GNNs. Although existing Scalable GNNs leverage linear propagation to preprocess the features and accelerate the training and inference procedure, these methods still suffer from scalability issues when making inferences on unseen nodes, as the feature preprocessing requires the graph to be known and fixed. To further accelerate Scalable GNNs inference in this inductive setting, we propose an online propagation framework and two novel node-adaptive propagation methods that can customize the optimal propagation depth for each node based on its topological information and thereby avoid redundant feature propagation. The trade-off between accuracy and latency can be flexibly managed through simple hyper-parameters to accommodate various latency constraints. Moreover, to compensate for the inference accuracy loss caused by the potential early termination of propagation, we further propose Inception Distillation to exploit the multi-scale receptive field information within graphs. The rigorous and comprehensive experimental study on public datasets with varying scales and characteristics demonstrates that the proposed inference acceleration framework outperforms existing state-of-the-art graph inference acceleration methods in terms of accuracy and efficiency. Particularly, the superiority of our approach is notable on datasets with larger scales, yielding a 75x inference speedup on the largest Ogbn-products dataset., Comment: 2024 IEEE 40th International Conference on Data Engineering (ICDE). arXiv admin note: substantial text overlap with arXiv:2211.00495