Zirui Liu

Keyu Duan, Zirui Liu, Peihao Wang et al.

Large-scale graph training is a notoriously challenging problem for graph neural networks (GNNs). Due to the nature of evolving graph structures into the training process, vanilla GNNs usually fail to scale up, limited by the GPU memory space. Up to now, though numerous scalable GNN architectures have been proposed, we still lack a comprehensive survey and fair benchmark of this reservoir to find the rationale for designing scalable GNNs. To this end, we first systematically formulate the representative methods of large-scale graph training into several branches and further establish a fair and consistent benchmark for them by a greedy hyperparameter searching. In addition, regarding efficiency, we theoretically evaluate the time and space complexity of various branches and empirically compare them w.r.t GPU memory usage, throughput, and convergence. Furthermore, We analyze the pros and cons for various branches of scalable GNNs and then present a new ensembling training manner, named EnGCN, to address the existing issues. Our code is available at https://github.com/VITA-Group/Large_Scale_GCN_Benchmarking.

Daochen Zha, Louis Feng, Qiaoyu Tan et al.

We study embedding table placement for distributed recommender systems, which aims to partition and place the tables on multiple hardware devices (e.g., GPUs) to balance the computation and communication costs. Although prior work has explored learning-based approaches for the device placement of computational graphs, embedding table placement remains to be a challenging problem because of 1) the operation fusion of embedding tables, and 2) the generalizability requirement on unseen placement tasks with different numbers of tables and/or devices. To this end, we present DreamShard, a reinforcement learning (RL) approach for embedding table placement. DreamShard achieves the reasoning of operation fusion and generalizability with 1) a cost network to directly predict the costs of the fused operation, and 2) a policy network that is efficiently trained on an estimated Markov decision process (MDP) without real GPU execution, where the states and the rewards are estimated with the cost network. Equipped with sum and max representation reductions, the two networks can directly generalize to any unseen tasks with different numbers of tables and/or devices without fine-tuning. Extensive experiments show that DreamShard substantially outperforms the existing human expert and RNN-based strategies with up to 19% speedup over the strongest baseline on large-scale synthetic tables and our production tables. The code is available at https://github.com/daochenzha/dreamshard

Yao Rong, Guanchu Wang, Qizhang Feng et al.

As Graph Neural Networks (GNNs) have been widely used in real-world applications, model explanations are required not only by users but also by legal regulations. However, simultaneously achieving high fidelity and low computational costs in generating explanations has been a challenge for current methods. In this work, we propose a framework of GNN explanation named LeArn Removal-based Attribution (LARA) to address this problem. Specifically, we introduce removal-based attribution and demonstrate its substantiated link to interpretability fidelity theoretically and experimentally. The explainer in LARA learns to generate removal-based attribution which enables providing explanations with high fidelity. A strategy of subgraph sampling is designed in LARA to improve the scalability of the training process. In the deployment, LARA can efficiently generate the explanation through a feed-forward pass. We benchmark our approach with other state-of-the-art GNN explanation methods on six datasets. Results highlight the effectiveness of our framework regarding both efficiency and fidelity. In particular, LARA is 3.5 times faster and achieves higher fidelity than the state-of-the-art method on the large dataset ogbn-arxiv (more than 160K nodes and 1M edges), showing its great potential in real-world applications. Our source code is available at https://anonymous.4open.science/r/LARA-10D8/README.md.

Jiayi Yuan, Hongyi Liu, Shaochen Zhong et al.

Long context capability is a crucial competency for large language models (LLMs) as it mitigates the human struggle to digest long-form texts. This capability enables complex task-solving scenarios such as book summarization, code assistance, and many more tasks that are traditionally manpower-intensive. However, transformer-based LLMs face significant challenges with long context input due to the growing size of the KV cache and the intrinsic complexity of attending to extended inputs; where multiple schools of efficiency-driven approaches - such as KV cache quantization, token dropping, prompt compression, linear-time sequence models, and hybrid architectures - have been proposed to produce efficient yet long context-capable models. Despite these advancements, no existing work has comprehensively benchmarked these methods in a reasonably aligned environment. In this work, we fill this gap by providing a taxonomy of current methods and evaluating 10+ state-of-the-art approaches across seven categories of long context tasks. Our work reveals numerous previously unknown phenomena and offers insights - as well as a friendly workbench - for the future development of long context-capable LLMs. The source code is available at https://github.com/henryzhongsc/longctx_bench.

Guanchu Wang, Junhao Ran, Ruixiang Tang et al.

Despite the impressive capabilities of Large Language Models (LLMs) in general medical domains, questions remain about their performance in diagnosing rare diseases. To answer this question, we aim to assess the diagnostic performance of LLMs in rare diseases, and explore methods to enhance their effectiveness in this area. In this work, we introduce a rare disease question-answering (ReDis-QA) dataset to evaluate the performance of LLMs in diagnosing rare diseases. Specifically, we collected 1360 high-quality question-answer pairs within the ReDis-QA dataset, covering 205 rare diseases. Additionally, we annotated meta-data for each question, facilitating the extraction of subsets specific to any given disease and its property. Based on the ReDis-QA dataset, we benchmarked several open-source LLMs, revealing that diagnosing rare diseases remains a significant challenge for these models. To facilitate retrieval augmentation generation for rare disease diagnosis, we collect the first rare diseases corpus (ReCOP), sourced from the National Organization for Rare Disorders (NORD) database. Specifically, we split the report of each rare disease into multiple chunks, each representing a different property of the disease, including their overview, symptoms, causes, effects, related disorders, diagnosis, and standard therapies. This structure ensures that the information within each chunk aligns consistently with a question. Experiment results demonstrate that ReCOP can effectively improve the accuracy of LLMs on the ReDis-QA dataset by an average of 8%. Moreover, it significantly guides LLMs to generate trustworthy answers and explanations that can be traced back to existing literature.

Yu-Neng Chuang, Guanchu Wang, Fan Yang et al.

Recently, there has been a growing demand for the deployment of Explainable Artificial Intelligence (XAI) algorithms in real-world applications. However, traditional XAI methods typically suffer from a high computational complexity problem, which discourages the deployment of real-time systems to meet the time-demanding requirements of real-world scenarios. Although many approaches have been proposed to improve the efficiency of XAI methods, a comprehensive understanding of the achievements and challenges is still needed. To this end, in this paper we provide a review of efficient XAI. Specifically, we categorize existing techniques of XAI acceleration into efficient non-amortized and efficient amortized methods. The efficient non-amortized methods focus on data-centric or model-centric acceleration upon each individual instance. In contrast, amortized methods focus on learning a unified distribution of model explanations, following the predictive, generative, or reinforcement frameworks, to rapidly derive multiple model explanations. We also analyze the limitations of an efficient XAI pipeline from the perspectives of the training phase, the deployment phase, and the use scenarios. Finally, we summarize the challenges of deploying XAI acceleration methods to real-world scenarios, overcoming the trade-off between faithfulness and efficiency, and the selection of different acceleration methods.

Xiaotian Han, Zhimeng Jiang, Hongye Jin et al.

Demographic parity is the most widely recognized measure of group fairness in machine learning, which ensures equal treatment of different demographic groups. Numerous works aim to achieve demographic parity by pursuing the commonly used metric $ΔDP$. Unfortunately, in this paper, we reveal that the fairness metric $ΔDP$ can not precisely measure the violation of demographic parity, because it inherently has the following drawbacks: i) zero-value $ΔDP$ does not guarantee zero violation of demographic parity, ii) $ΔDP$ values can vary with different classification thresholds. To this end, we propose two new fairness metrics, Area Between Probability density function Curves (ABPC) and Area Between Cumulative density function Curves (ABCC), to precisely measure the violation of demographic parity at the distribution level. The new fairness metrics directly measure the difference between the distributions of the prediction probability for different demographic groups. Thus our proposed new metrics enjoy: i) zero-value ABCC/ABPC guarantees zero violation of demographic parity; ii) ABCC/ABPC guarantees demographic parity while the classification thresholds are adjusted. We further re-evaluate the existing fair models with our proposed fairness metrics and observe different fairness behaviors of those models under the new metrics. The code is available at https://github.com/ahxt/new_metric_for_demographic_parity

Zirui Liu, Shengyuan Chen, Kaixiong Zhou et al.

The training of graph neural networks (GNNs) is extremely time consuming because sparse graph-based operations are hard to be accelerated by hardware. Prior art explores trading off the computational precision to reduce the time complexity via sampling-based approximation. Based on the idea, previous works successfully accelerate the dense matrix based operations (e.g., convolution and linear) with negligible accuracy drop. However, unlike dense matrices, sparse matrices are stored in the irregular data format such that each row/column may have different number of non-zero entries. Thus, compared to the dense counterpart, approximating sparse operations has two unique challenges (1) we cannot directly control the efficiency of approximated sparse operation since the computation is only executed on non-zero entries; (2) sub-sampling sparse matrices is much more inefficient due to the irregular data format. To address the issues, our key idea is to control the accuracy-efficiency trade off by optimizing computation resource allocation layer-wisely and epoch-wisely. Specifically, for the first challenge, we customize the computation resource to different sparse operations, while limit the total used resource below a certain budget. For the second challenge, we cache previous sampled sparse matrices to reduce the epoch-wise sampling overhead. Finally, we propose a switching mechanisms to improve the generalization of GNNs trained with approximated operations. To this end, we propose Randomized Sparse Computation, which for the first time demonstrate the potential of training GNNs with approximated operations. In practice, rsc can achieve up to $11.6\times$ speedup for a single sparse operation and a $1.6\times$ end-to-end wall-clock time speedup with negligible accuracy drop.

Hailin Zhang, Penghao Zhao, Xupeng Miao et al.

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.

Wenya Xie, Shaochen, Zhong et al.

Large Reasoning Models (LRMs) are often bottlenecked by the high cost of output tokens. We show that a significant portion of these tokens are useless self-repetitions - what we call "word salad" - that exhaust the decoding budget without adding value. Interestingly, we observe that LRMs are self-aware when trapped in these loops: the hidden states of <\n\n> tokens trailing each reasoning chunk exhibit patterns that allow us to detect word salad behavior on-the-fly via a single-layer linear classifier. Once detected, a simple chop appended by a straightforward regeneration prompt yields substantial length savings with minimal quality loss. Our work offers WordSaladChopper (WSC) - a lightweight, turnkey component for LRM that is minimally invasive to its reasoning trajectory by only removing semantically redundant tokens. Given its low overhead, strong savings, and the lack of semantic value of word salad tokens, we believe it is not too far-fetched to argue that WSC - or a similar component - is a must-have for all LRM applications with user experience in mind. Our code is publicly available at https://github.com/wenyaxie023/WordSaladChopper.

Ruixiang Tang, Jiayi Yuan, Yiming Li et al.

In the field of natural language processing, the prevalent approach involves fine-tuning pretrained language models (PLMs) using local samples. Recent research has exposed the susceptibility of PLMs to backdoor attacks, wherein the adversaries can embed malicious prediction behaviors by manipulating a few training samples. In this study, our objective is to develop a backdoor-resistant tuning procedure that yields a backdoor-free model, no matter whether the fine-tuning dataset contains poisoned samples. To this end, we propose and integrate a honeypot module into the original PLM, specifically designed to absorb backdoor information exclusively. Our design is motivated by the observation that lower-layer representations in PLMs carry sufficient backdoor features while carrying minimal information about the original tasks. Consequently, we can impose penalties on the information acquired by the honeypot module to inhibit backdoor creation during the fine-tuning process of the stem network. Comprehensive experiments conducted on benchmark datasets substantiate the effectiveness and robustness of our defensive strategy. Notably, these results indicate a substantial reduction in the attack success rate ranging from 10\% to 40\% when compared to prior state-of-the-art methods.

Guanchu Wang, Zirui Liu, Zhimeng Jiang et al.

Activation compressed training provides a solution towards reducing the memory cost of training deep neural networks~(DNNs). However, state-of-the-art work combines a search of quantization bit-width with the training, which makes the procedure complicated and less transparent. To this end, we propose a simple and effective method to compress DNN training. Our method is motivated by an instructive observation: DNN backward propagation mainly utilizes the low-frequency component (LFC) of the activation maps, while the majority of memory is for caching the high-frequency component (HFC) during the training. This indicates the HFC of activation maps is highly redundant and compressible during DNN training, which inspires our proposed Dual Activation Precision (DIVISION). During the training, DIVISION preserves the high-precision copy of LFC and compresses the HFC into a light-weight copy with low numerical precision. This can significantly reduce the memory cost without negatively affecting the precision of backward propagation such that DIVISION maintains competitive model accuracy. Experiment results show DIVISION has better comprehensive performance than state-of-the-art methods, including over 10x compression of activation maps and competitive training throughput, without loss of model accuracy.

Zirui Liu, Jiayi Yuan, Hongye Jin et al.

Efficiently serving large language models (LLMs) requires batching of many requests to reduce the cost per request. Yet, with larger batch sizes and longer context lengths, the key-value (KV) cache, which stores attention keys and values to avoid re-computations, significantly increases memory demands and becomes the new bottleneck in speed and memory usage. Additionally, the loading of the KV cache causes the computational core to be idle, which limits the inference speed. A straightforward and effective solution to reduce KV cache size is quantization, which decreases the total bytes taken by KV cache. However, there is a lack of in-depth studies that explore the element distribution of KV cache to understand the hardness and limitation of KV cache quantization. To fill the gap, we conducted a comprehensive study on the element distribution in KV cache of popular LLMs. Our findings indicate that the key cache should be quantized per-channel, i.e., group elements along the channel dimension and quantize them together. In contrast, the value cache should be quantized per-token. From this analysis, we developed a tuning-free 2bit KV cache quantization algorithm named KIVI. With hardware-friendly implementation, KIVI can enable Llama, Falcon, and Mistral models to maintain almost the same quality while using $\mathbf{2.6\times}$ less peak memory (including model weight). This reduction in memory usage enables up to $\mathbf{4\times}$ larger batch size, bringing $\mathbf{2.35\times \sim 3.47\times}$ throughput on real LLM inference workload. The source code is available at https://github.com/jy-yuan/KIVI.

Hongye Jin, Xiaotian Han, Jingfeng Yang et al.

It is well known that LLMs cannot generalize well to long contexts whose lengths are larger than the training sequence length. This poses challenges when employing LLMs for processing long input sequences during inference. In this work, we argue that LLMs themselves have inherent capabilities to handle long contexts without fine-tuning. To achieve this goal, we propose SelfExtend to extend the context window of LLMs by constructing bi-level attention information: the grouped attention and the neighbor attention. The grouped attention captures the dependencies among tokens that are far apart, while neighbor attention captures dependencies among adjacent tokens within a specified range. The two-level attentions are computed based on the original model's self-attention mechanism during inference. With minor code modification, our SelfExtend can effortlessly extend existing LLMs' context window without any fine-tuning. We conduct comprehensive experiments on multiple benchmarks and the results show that our SelfExtend can effectively extend existing LLMs' context window length. The code can be found at \url{https://github.com/datamllab/LongLM}.

Rixin Liu, Xingqi Cui, Kaijian Wang et al.

NVIDIA Multi-Process Service (MPS) enables fine-grained GPU sharing by allowing multiple processes to execute concurrently on the same GPU, making it an important mechanism for improving GPU utilization. However, MPS has weak fault resilience: a fault in one process can terminate all co-running processes, limiting its adoption in resilience-critical settings such as multi-tenant GPU clusters. In this work, we design fault-resilient MPS to solve this problem. Our design is guided by insights from a systematic characterization of GPU faults and a deep analysis of their end-to-end processing pipeline. Based on these insights, we design two complementary mechanisms. A fault isolation mechanism for the dominant memory-related faults that can be fully isolated by software intervention in the open GPU driver kernel module. For other faults whose process is within proprietary software, we design a practical mechanism -- fast recovery using virtual memory based GPU-resident state sharing. Our evaluation on different GPUs and workloads shows that these mechanisms can handle corresponding faults effectively with minimal overhead.

Shimao Chen, Zirui Liu, Zhiying Wu et al.

As the foundation of large language models (LLMs), self-attention module faces the challenge of quadratic time and memory complexity with respect to sequence length. FlashAttention accelerates attention computation and reduces its memory usage by leveraging the GPU memory hierarchy. A promising research direction is to integrate FlashAttention with quantization methods. This paper introduces INT-FlashAttention, the first INT8 quantization architecture compatible with the forward workflow of FlashAttention, which significantly improves the inference speed of FlashAttention on Ampere GPUs. We implement our INT-FlashAttention prototype with fully INT8 activations and general matrix-multiplication (GEMM) kernels, making it the first attention operator with fully INT8 input. As a general token-level post-training quantization framework, INT-FlashAttention is also compatible with other data formats like INT4, etc. Experimental results show INT-FlashAttention achieves 72% faster inference speed and 82% smaller quantization error compared to standard FlashAttention with FP16 and FP8 data format.

Yu-Neng Chuang, Leisheng Yu, Guanchu Wang et al.

Large language models (LLMs) are increasingly deployed and democratized on edge devices. To improve the efficiency of on-device deployment, small language models (SLMs) are often adopted due to their efficient decoding latency and reduced energy consumption. However, these SLMs often generate inaccurate responses when handling complex queries. One promising solution is uncertainty-based SLM routing, offloading high-stakes queries to stronger LLMs when resulting in low-confidence responses on SLM. This follows the principle of "If you lack confidence, seek stronger support" to enhance reliability. Relying on more powerful LLMs is yet effective but increases invocation costs. Therefore, striking a routing balance between efficiency and efficacy remains a critical challenge. Additionally, efficiently generalizing the routing strategy to new datasets remains under-explored. In this paper, we conduct a comprehensive investigation into benchmarking and generalization of uncertainty-driven routing strategies from SLMs to LLMs over 1500+ settings. Our findings highlight: First, uncertainty-correctness alignment in different uncertainty quantification (UQ) methods significantly impacts routing performance. Second, uncertainty distributions depend more on both the specific SLM and the chosen UQ method, rather than downstream data. Building on the insight, we propose a calibration data construction instruction pipeline and open-source a constructed hold-out set to enhance routing generalization on new downstream scenarios. The experimental results indicate calibration data effectively bootstraps routing performance without any new data.

Zhimeng Jiang, Xiaotian Han, Chao Fan et al.

There has been significant progress in improving the performance of graph neural networks (GNNs) through enhancements in graph data, model architecture design, and training strategies. For fairness in graphs, recent studies achieve fair representations and predictions through either graph data pre-processing (e.g., node feature masking, and topology rewiring) or fair training strategies (e.g., regularization, adversarial debiasing, and fair contrastive learning). How to achieve fairness in graphs from the model architecture perspective is less explored. More importantly, GNNs exhibit worse fairness performance compared to multilayer perception since their model architecture (i.e., neighbor aggregation) amplifies biases. To this end, we aim to achieve fairness via a new GNN architecture. We propose \textsf{F}air \textsf{M}essage \textsf{P}assing (FMP) designed within a unified optimization framework for GNNs. Notably, FMP \textit{explicitly} renders sensitive attribute usage in \textit{forward propagation} for node classification task using cross-entropy loss without data pre-processing. In FMP, the aggregation is first adopted to utilize neighbors' information and then the bias mitigation step explicitly pushes demographic group node presentation centers together. In this way, FMP scheme can aggregate useful information from neighbors and mitigate bias to achieve better fairness and prediction tradeoff performance. Experiments on node classification tasks demonstrate that the proposed FMP outperforms several baselines in terms of fairness and accuracy on three real-world datasets. The code is available in {\url{https://github.com/zhimengj0326/FMP}}.

Guoqi Yu, Yaoming Li, Xiaoyu Guo et al.

Forecasting models are pivotal in a data-driven world with vast volumes of time series data that appear as a compound of vast Linear and Nonlinear patterns. Recent deep time series forecasting models struggle to utilize seasonal and trend decomposition to separate the entangled components. Such a strategy only explicitly extracts simple linear patterns like trends, leaving the other linear modes and vast unexplored nonlinear patterns to the residual. Their flawed linear and nonlinear feature extraction models and shallow-level decomposition limit their adaptation to the diverse patterns present in real-world scenarios. Given this, we innovate Recursive Residual Decomposition by introducing explicit extraction of both linear and nonlinear patterns. This deeper-level decomposition framework, which is named LiNo, captures linear patterns using a Li block which can be a moving average kernel, and models nonlinear patterns using a No block which can be a Transformer encoder. The extraction of these two patterns is performed alternatively and recursively. To achieve the full potential of LiNo, we develop the current simple linear pattern extractor to a general learnable autoregressive model, and design a novel No block that can handle all essential nonlinear patterns. Remarkably, the proposed LiNo achieves state-of-the-art on thirteen real-world benchmarks under univariate and multivariate forecasting scenarios. Experiments show that current forecasting models can deliver more robust and precise results through this advanced Recursive Residual Decomposition. We hope this work could offer insight into designing more effective forecasting models. Code is available at this Repository: https://github.com/Levi-Ackman/LiNo.

Zhiling Zhou, Zirui Liu, Chengming Xu et al.

While neural networks have made significant strides in many AI tasks, they remain vulnerable to a range of noise types, including natural corruptions, adversarial noise, and low-resolution artifacts. Many existing approaches focus on enhancing robustness against specific noise types, limiting their adaptability to others. Previous studies have addressed general robustness by adopting a spectral perspective, which tends to blur crucial features like texture and object contours. Our proposed solution, however, introduces an inverse scale variational sparsification framework within a time-continuous inverse scale space formulation. This framework progressively learns finer-scale features by discerning variational differences between pixels, ultimately preserving only large-scale features in the smoothed image. Unlike frequency-based methods, our approach not only removes noise by smoothing small-scale features where corruptions often occur but also retains high-contrast details such as textures and object contours. Moreover, our framework offers simplicity and efficiency in implementation. By integrating this algorithm into neural network training, we guide the model to prioritize learning large-scale features. We show the efficacy of our approach through enhanced robustness against various noise types.

Zhimeng Jiang, Zirui Liu, Xiaotian Han et al.

Deep neural networks are ubiquitously adopted in many applications, such as computer vision, natural language processing, and graph analytics. However, well-trained neural networks can make prediction errors after deployment as the world changes. \textit{Model editing} involves updating the base model to correct prediction errors with less accessible training data and computational resources. Despite recent advances in model editors in computer vision and natural language processing, editable training in graph neural networks (GNNs) is rarely explored. The challenge with editable GNN training lies in the inherent information aggregation across neighbors, which can lead model editors to affect the predictions of other nodes unintentionally. In this paper, we first observe the gradient of cross-entropy loss for the target node and training nodes with significant inconsistency, which indicates that directly fine-tuning the base model using the loss on the target node deteriorates the performance on training nodes. Motivated by the gradient inconsistency observation, we propose a simple yet effective \underline{G}radient \underline{R}ewiring method for \underline{E}ditable graph neural network training, named \textbf{GRE}. Specifically, we first store the anchor gradient of the loss on training nodes to preserve the locality. Subsequently, we rewire the gradient of the loss on the target node to preserve performance on the training node using anchor gradient. Experiments demonstrate the effectiveness of GRE on various model architectures and graph datasets in terms of multiple editing situations. The source code is available at \url{https://github.com/zhimengj0326/Gradient_rewiring_editing}

Daochen Zha, Louis Feng, Liang Luo et al.

Sharding a large machine learning model across multiple devices to balance the costs is important in distributed training. This is challenging because partitioning is NP-hard, and estimating the costs accurately and efficiently is difficult. In this work, we explore a "pre-train, and search" paradigm for efficient sharding. The idea is to pre-train a universal and once-for-all neural network to predict the costs of all the possible shards, which serves as an efficient sharding simulator. Built upon this pre-trained cost model, we then perform an online search to identify the best sharding plans given any specific sharding task. We instantiate this idea in deep learning recommendation models (DLRMs) and propose NeuroShard for embedding table sharding. NeuroShard pre-trains neural cost models on augmented tables to cover various sharding scenarios. Then it identifies the best column-wise and table-wise sharding plans with beam search and greedy grid search, respectively. Experiments show that NeuroShard significantly and consistently outperforms the state-of-the-art on the benchmark sharding dataset, achieving up to 23.8% improvement. When deployed in an ultra-large production DLRM with multi-terabyte embedding tables, NeuroShard achieves 11.6% improvement in embedding costs over the state-of-the-art, which translates to 6.6% end-to-end training throughput improvement. To facilitate future research of the "pre-train, and search" paradigm in ML for Systems, we open-source our code at https://github.com/daochenzha/neuroshard

Zirui Liu, Haifeng Jin, Ting-Hsiang Wang et al.

Human-designed data augmentation strategies have been replaced by automatically learned augmentation policy in the past two years. Specifically, recent work has empirically shown that the superior performance of the automated data augmentation methods stems from increasing the diversity of augmented data \cite{autoaug, randaug}. However, two factors regarding the diversity of augmented data are still missing: 1) the explicit definition (and thus measurement) of diversity and 2) the quantifiable relationship between diversity and its regularization effects. To bridge this gap, we propose a diversity measure called Variance Diversity and theoretically show that the regularization effect of data augmentation is promised by Variance Diversity. We validate in experiments that the relative gain from automated data augmentation in test accuracy is highly correlated to Variance Diversity. An unsupervised sampling-based framework, \textbf{DivAug}, is designed to directly maximize Variance Diversity and hence strengthen the regularization effect. Without requiring a separate search process, the performance gain from DivAug is comparable with the state-of-the-art method with better efficiency. Moreover, under the semi-supervised setting, our framework can further improve the performance of semi-supervised learning algorithms compared to RandAugment, making it highly applicable to real-world problems, where labeled data is scarce. The code is available at \texttt{\url{https://github.com/warai-0toko/DivAug}}.

Ting-Hsiang Wang, Qingquan Song, Xiaotian Han et al.

Realistic recommender systems are often required to adapt to ever-changing data and tasks or to explore different models systematically. To address the need, we present AutoRec, an open-source automated machine learning (AutoML) platform extended from the TensorFlow ecosystem and, to our knowledge, the first framework to leverage AutoML for model search and hyperparameter tuning in deep recommendation models. AutoRec also supports a highly flexible pipeline that accommodates both sparse and dense inputs, rating prediction and click-through rate (CTR) prediction tasks, and an array of recommendation models. Lastly, AutoRec provides a simple, user-friendly API. Experiments conducted on the benchmark datasets reveal AutoRec is reliable and can identify models which resemble the best model without prior knowledge.

Ruixiang Tang, Mengnan Du, Yuening Li et al.

Image captioning has made substantial progress with huge supporting image collections sourced from the web. However, recent studies have pointed out that captioning datasets, such as COCO, contain gender bias found in web corpora. As a result, learning models could heavily rely on the learned priors and image context for gender identification, leading to incorrect or even offensive errors. To encourage models to learn correct gender features, we reorganize the COCO dataset and present two new splits COCO-GB V1 and V2 datasets where the train and test sets have different gender-context joint distribution. Models relying on contextual cues will suffer from huge gender prediction errors on the anti-stereotypical test data. Benchmarking experiments reveal that most captioning models learn gender bias, leading to high gender prediction errors, especially for women. To alleviate the unwanted bias, we propose a new Guided Attention Image Captioning model (GAIC) which provides self-guidance on visual attention to encourage the model to capture correct gender visual evidence. Experimental results validate that GAIC can significantly reduce gender prediction errors with a competitive caption quality. Our codes and the designed benchmark datasets are available at https://github.com/datamllab/Mitigating_Gender_Bias_In_Captioning_System.

Zihao Li, Yucheng Shi, Zirui Liu et al.

The development of Large Language Models (LLMs) relies on extensive text corpora, which are often unevenly distributed across languages. This imbalance results in LLMs performing significantly better on high-resource languages like English, German, and French, while their capabilities in low-resource languages remain inadequate. Currently, there is a lack of quantitative methods to evaluate the performance of LLMs in these low-resource languages. To address this gap, we propose the Language Ranker, an intrinsic metric designed to benchmark and rank languages based on LLM performance using internal representations. By comparing the LLM's internal representation of various languages against a baseline derived from English, we can assess the model's multilingual capabilities in a robust and language-agnostic manner. Our analysis reveals that high-resource languages exhibit higher similarity scores with English, demonstrating superior performance, while low-resource languages show lower similarity scores, underscoring the effectiveness of our metric in assessing language-specific capabilities. Besides, the experiments show that there is a strong correlation between the LLM's performance in different languages and the proportion of those languages in its pre-training corpus. These insights underscore the efficacy of the Language Ranker as a tool for evaluating LLM performance across different languages, particularly those with limited resources.

Yu-Neng Chuang, Tianwei Xing, Chia-Yuan Chang et al.

Large language models (LLMs) are great at processing multiple natural language processing tasks, but their abilities are constrained by inferior performance with long context, slow inference speed, and the high cost of computing the results. Deploying LLMs with precise and informative context helps users process large-scale datasets more effectively and cost-efficiently. Existing works rely on compressing long prompt contexts into soft prompts. However, soft prompt compression encounters limitations in transferability across different LLMs, especially API-based LLMs. To this end, this work aims to compress lengthy prompts in the form of natural language with LLM transferability. This poses two challenges: (i) Natural Language (NL) prompts are incompatible with back-propagation, and (ii) NL prompts lack flexibility in imposing length constraints. In this work, we propose a Natural Language Prompt Encapsulation (Nano-Capsulator) framework compressing original prompts into NL formatted Capsule Prompt while maintaining the prompt utility and transferability. Specifically, to tackle the first challenge, the Nano-Capsulator is optimized by a reward function that interacts with the proposed semantics preserving loss. To address the second question, the Nano-Capsulator is optimized by a reward function featuring length constraints. Experimental results demonstrate that the Capsule Prompt can reduce 81.4% of the original length, decrease inference latency up to 4.5x, and save 80.1% of budget overheads while providing transferability across diverse LLMs and different datasets.

Hongyi Liu, Shaochen Zhong, Xintong Sun et al.

Finetuning LLMs with LoRA has gained significant popularity due to its simplicity and effectiveness. Often, users may even find pluggable, community-shared LoRAs to enhance their base models for a specific downstream task of interest; enjoying a powerful, efficient, yet customized LLM experience with negligible investment. However, this convenient share-and-play ecosystem also introduces a new attack surface, where attackers can distribute malicious LoRAs to a community eager to try out shared assets. Despite the high-risk potential, no prior art has comprehensively explored LoRA's attack surface under the downstream-enhancing share-and-play context. In this paper, we investigate how backdoors can be injected into task-enhancing LoRAs and examine the mechanisms of such infections. We find that with a simple, efficient, yet specific recipe, a backdoor LoRA can be trained once and then seamlessly merged (in a training-free fashion) with multiple task-enhancing LoRAs, retaining both its malicious backdoor and benign downstream capabilities. This allows attackers to scale the distribution of compromised LoRAs with minimal effort by leveraging the rich pool of existing shared LoRA assets. We note that such merged LoRAs are particularly infectious -- because their malicious intent is cleverly concealed behind improved downstream capabilities, creating a strong incentive for voluntary download -- and dangerous -- because under local deployment, no safety measures exist to intervene when things go wrong. Our work is among the first to study this new threat model of training-free distribution of downstream-capable-yet-backdoor-injected LoRAs, highlighting the urgent need for heightened security awareness in the LoRA ecosystem. Warning: This paper contains offensive content and involves a real-life tragedy.

Hailin Zhang, Zirui Liu, Boxuan Chen et al.

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.

Zirui Liu, Qingquan Song, Qiang Charles Xiao et al.

The large number of parameters in Pretrained Language Models enhance their performance, but also make them resource-intensive, making it challenging to deploy them on commodity hardware like a single GPU. Due to the memory and power limitations of these devices, model compression techniques are often used to decrease both the model's size and its inference latency. This usually results in a trade-off between model accuracy and efficiency. Therefore, optimizing this balance is essential for effectively deploying LLMs on commodity hardware. A significant portion of the efficiency challenge is the Feed-forward network (FFN) component, which accounts for roughly $\frac{2}{3}$ total parameters and inference latency. In this paper, we first observe that only a few neurons of FFN module have large output norm for any input tokens, a.k.a. heavy hitters, while the others are sparsely triggered by different tokens. Based on this observation, we explicitly split the FFN into two parts according to the heavy hitters. We improve the efficiency-accuracy trade-off of existing compression methods by allocating more resource to FFN parts with heavy hitters. In practice, our method can reduce model size by 43.1\% and bring $1.25\sim1.56\times$ wall clock time speedup on different hardware with negligible accuracy drop.

Haochen Zhang, Junze Yin, Guanchu Wang et al.

Low-rank optimization has emerged as a promising approach to enabling memory-efficient training of large language models (LLMs). Existing low-rank optimization methods typically project gradients onto a low-rank subspace, reducing the memory cost of storing optimizer states. A key challenge in these methods is selecting suitable subspaces to ensure an effective optimization trajectory. Most existing approaches select the dominant subspace to preserve gradient information, as this intuitively provides the best approximation. However, we find that in practice, the dominant subspace stops changing during pretraining, thereby constraining weight updates to similar subspaces. In this paper, we propose importance sampling for low-rank optimization in LLM pretraining with a provable convergence guarantee, which the dominant subspace approach does not have. Empirically, we demonstrate that our method significantly outperforms previous methods in LLM pretraining tasks.

Shuang Zhou, Wenya Xie, Jiaxi Li et al.

As large language models (LLMs) become increasingly integrated into clinical decision-making, ensuring transparent and trustworthy reasoning is essential. However, existing evaluation strategies of LLMs' medical reasoning capability either suffer from unsatisfactory assessment or poor scalability, and a rigorous benchmark remains lacking. To address this, we introduce MedThink-Bench, a benchmark designed for rigorous, explainable, and scalable assessment of LLMs' medical reasoning. MedThink-Bench comprises 500 challenging questions across ten medical domains, each annotated with expert-crafted step-by-step rationales. Building on this, we propose LLM-w-Ref, a novel evaluation framework that leverages fine-grained rationales and LLM-as-a-Judge mechanisms to assess intermediate reasoning with expert-level fidelity while maintaining scalability. Experiments show that LLM-w-Ref exhibits a strong positive correlation with expert judgments. Benchmarking twelve state-of-the-art LLMs, we find that smaller models (e.g., MedGemma-27B) can surpass larger proprietary counterparts (e.g., OpenAI-o3). Overall, MedThink-Bench offers a foundational tool for evaluating LLMs' medical reasoning, advancing their safe and responsible deployment in clinical practice.

Ammar Ahmed, Sheng Di, Franck Cappello et al.

Large language models (LLMs) excel across diverse natural language processing tasks but face resource demands and limited context windows. Although techniques like pruning, quantization, and token dropping can mitigate these issues, their efficacy in long-context scenarios and system evaluation remains underexplored. This paper systematically benchmarks these optimizations, characterizing memory usage, latency, and throughput, and studies how these methods impact the quality of text generation. We first analyze individual optimization methods for two LLM architectures supporting long context and then systematically evaluate combinations of these techniques to assess how this deeper analysis impacts performance metrics. We subsequently study the scalability of individual optimization methods on a larger variant with 70 billion-parameter model. Our novel insights reveal that naive combination inference optimization algorithms can adversely affect larger models due to compounded approximation errors, as compared to their smaller counterparts. Experiments show that relying solely on F1 obscures these effects by hiding precision-recall trade-offs in question answering tasks. By integrating system-level profiling with task-specific insights, this study helps LLM practitioners and researchers explore and balance efficiency, accuracy, and scalability across tasks and hardware configurations.

Zaifu Zhan, Shuang Zhou, Huixue Zhou et al.

Foundation models, including language models, e.g., GPT, and vision models, e.g., CLIP, have significantly advanced numerous biomedical tasks. Despite these advancements, the high inference latency and the "overthinking" issues in model inference impair the efficiency and effectiveness of foundation models, thus limiting their application in real-time clinical settings. To address these challenges, we proposed EPEE (Entropy- and Patience-based Early Exiting), a novel hybrid strategy designed to improve the inference efficiency of foundation models. The core idea was to leverage the strengths of entropy-based and patience-based early exiting methods to overcome their respective weaknesses. To evaluate EPEE, we conducted experiments on three core biomedical tasks-classification, relation extraction, and event extraction-using four foundation models (BERT, ALBERT, GPT-2, and ViT) across twelve datasets, including clinical notes and medical images. The results showed that EPEE significantly reduced inference time while maintaining or improving accuracy, demonstrating its adaptability to diverse datasets and tasks. EPEE addressed critical barriers to deploying foundation models in healthcare by balancing efficiency and effectiveness. It potentially provided a practical solution for real-time clinical decision-making with foundation models, supporting reliable and efficient workflows.

Minrui Luo, Fuhang Kuang, Yu Wang et al.

Parameter-Efficient Fine-Tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), are indispensable for efficiently customizing Large Language Models (LLMs). However, vanilla LoRA suffers from slow convergence speed and knowledge forgetting problems. Recent studies have leveraged the power of designed LoRA initialization, to enhance the fine-tuning efficiency, or to preserve knowledge in the pre-trained LLM. However, none of these works can address the two cases at the same time. To this end, we introduce Subspace-Constrained LoRA (SC-LoRA), a novel LoRA initialization framework engineered to navigate the trade-off between efficient fine-tuning and knowledge preservation. We achieve this by constraining the output of trainable LoRA adapters in a low-rank subspace, where the context information of fine-tuning data is most preserved while the context information of preserved knowledge is least retained, in a balanced way. Such constraint enables the trainable weights to primarily focus on the main features of fine-tuning data while avoiding damaging the preserved knowledge features. We provide theoretical analysis on our method, and conduct extensive experiments including safety preservation and world knowledge preservation, on various downstream tasks. In our experiments, SC-LoRA succeeds in delivering superior fine-tuning performance while markedly diminishing knowledge forgetting, surpassing contemporary LoRA initialization methods.

Wentao Guo, Jikai Long, Yimeng Zeng et al.

Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, the application of ZO fine-tuning in memory-constrained settings such as mobile phones and laptops is still challenging since full precision forward passes are infeasible. In this study, we address this limitation by integrating sparsity and quantization into ZO fine-tuning of LLMs. Specifically, we investigate the feasibility of fine-tuning an extremely small subset of LLM parameters using ZO. This approach allows the majority of un-tuned parameters to be quantized to accommodate the constraint of limited device memory. Our findings reveal that the pre-training process can identify a set of "sensitive parameters" that can guide the ZO fine-tuning of LLMs on downstream tasks. Our results demonstrate that fine-tuning 0.1% sensitive parameters in the LLM with ZO can outperform the full ZO fine-tuning performance, while offering wall-clock time speedup. Additionally, we show that ZO fine-tuning targeting these 0.1% sensitive parameters, combined with 4 bit quantization, enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8 GiB of memory and notably reduced latency.

Guanchu Wang, Yu-Neng Chuang, Fan Yang et al.

Explainable machine learning significantly improves the transparency of deep neural networks. However, existing work is constrained to explaining the behavior of individual model predictions, and lacks the ability to transfer the explanation across various models and tasks. This limitation results in explaining various tasks being time- and resource-consuming. To address this problem, we introduce a Transferable Vision Explainer (TVE) that can effectively explain various vision models in downstream tasks. Specifically, the transferability of TVE is realized through a pre-training process on large-scale datasets towards learning the meta-attribution. This meta-attribution leverages the versatility of generic backbone encoders to comprehensively encode the attribution knowledge for the input instance, which enables TVE to seamlessly transfer to explain various downstream tasks, without the need for training on task-specific data. Empirical studies involve explaining three different architectures of vision models across three diverse downstream datasets. The experimental results indicate TVE is effective in explaining these tasks without the need for additional training on downstream data.

Zirui Liu, Zhimeng Jiang, Shaochen Zhong et al.

Despite Graph Neural Networks (GNNs) have achieved prominent success in many graph-based learning problem, such as credit risk assessment in financial networks and fake news detection in social networks. However, the trained GNNs still make errors and these errors may cause serious negative impact on society. \textit{Model editing}, which corrects the model behavior on wrongly predicted target samples while leaving model predictions unchanged on unrelated samples, has garnered significant interest in the fields of computer vision and natural language processing. However, model editing for graph neural networks (GNNs) is rarely explored, despite GNNs' widespread applicability. To fill the gap, we first observe that existing model editing methods significantly deteriorate prediction accuracy (up to $50\%$ accuracy drop) in GNNs while a slight accuracy drop in multi-layer perception (MLP). The rationale behind this observation is that the node aggregation in GNNs will spread the editing effect throughout the whole graph. This propagation pushes the node representation far from its original one. Motivated by this observation, we propose \underline{E}ditable \underline{G}raph \underline{N}eural \underline{N}etworks (EGNN), a neighbor propagation-free approach to correct the model prediction on misclassified nodes. Specifically, EGNN simply stitches an MLP to the underlying GNNs, where the weights of GNNs are frozen during model editing. In this way, EGNN disables the propagation during editing while still utilizing the neighbor propagation scheme for node prediction to obtain satisfactory results. Experiments demonstrate that EGNN outperforms existing baselines in terms of effectiveness (correcting wrong predictions with lower accuracy drop), generalizability (correcting wrong predictions for other similar nodes), and efficiency (low training time and memory) on various graph datasets.

Zirui Liu, Guanchu Wang, Shaochen Zhong et al.

With the rapid growth in model size, fine-tuning the large pre-trained language model has become increasingly difficult due to its extensive memory usage. Previous works usually focus on reducing the number of trainable parameters in the network. While the model parameters do contribute to memory usage, the primary memory bottleneck during training arises from storing feature maps, also known as activations, as they are crucial for gradient calculation. Notably, neural networks are usually trained using stochastic gradient descent. We argue that in stochastic optimization, models can handle noisy gradients as long as the gradient estimator is unbiased with reasonable variance. Following this motivation, we propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance, which only requires storing the sub-sampled activations for calculating the gradient. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones. By replacing the linear operation with our approximated one in transformers, we can achieve up to 2.7$\times$ peak memory reduction with almost no accuracy drop and enables up to $6.4\times$ larger batch size. Under the same hardware, WTA-CRS enables better down-streaming task performance by applying larger models and/or faster training speed with larger batch sizes.

Zhaozhuo Xu, Zirui Liu, Beidi Chen et al.

While the numerous parameters in Large Language Models (LLMs) contribute to their superior performance, this massive scale makes them inefficient and memory-hungry. Thus, they are hard to deploy on commodity hardware, such as one single GPU. Given the memory and power constraints of such devices, model compression methods are widely employed to reduce both the model size and inference latency, which essentially trades off model quality in return for improved efficiency. Thus, optimizing this accuracy-efficiency trade-off is crucial for the LLM deployment on commodity hardware. In this paper, we introduce a new perspective to optimize this trade-off by prompting compressed models. Specifically, we first observe that for certain questions, the generation quality of a compressed LLM can be significantly improved by adding carefully designed hard prompts, though this isn't the case for all questions. Based on this observation, we propose a soft prompt learning method where we expose the compressed model to the prompt learning process, aiming to enhance the performance of prompts. Our experimental analysis suggests our soft prompt strategy greatly improves the performance of the 8x compressed LLaMA-7B model (with a joint 4-bit quantization and 50% weight pruning compression), allowing them to match their uncompressed counterparts on popular benchmarks. Also, we demonstrate that these learned prompts can be transferred across various datasets, tasks, and compression levels. Hence with this transferability, we can stitch the soft prompt to a newly compressed model to improve the test-time accuracy in an ``in-situ'' way.

Zhimeng Jiang, Xiaotian Han, Chao Fan et al.

Despite recent advances in achieving fair representations and predictions through regularization, adversarial debiasing, and contrastive learning in graph neural networks (GNNs), the working mechanism (i.e., message passing) behind GNNs inducing unfairness issue remains unknown. In this work, we theoretically and experimentally demonstrate that representative aggregation in message-passing schemes accumulates bias in node representation due to topology bias induced by graph topology. Thus, a \textsf{F}air \textsf{M}essage \textsf{P}assing (FMP) scheme is proposed to aggregate useful information from neighbors but minimize the effect of topology bias in a unified framework considering graph smoothness and fairness objectives. The proposed FMP is effective, transparent, and compatible with back-propagation training. An acceleration approach on gradient calculation is also adopted to improve algorithm efficiency. Experiments on node classification tasks demonstrate that the proposed FMP outperforms the state-of-the-art baselines in effectively and efficiently mitigating bias on three real-world datasets.

Kaixiong Zhou, Ninghao Liu, Fan Yang et al.

Graph neural networks (GNNs), which learn the node representations by recursively aggregating information from its neighbors, have become a predominant computational tool in many domains. To handle large-scale graphs, most of the existing methods partition the input graph into multiple sub-graphs (e.g., through node clustering) and apply batch training to save memory cost. However, such batch training will lead to label bias within each batch, and then result in over-confidence in model predictions. Since the connected nodes with positively related labels tend to be assigned together, the traditional cross-entropy minimization process will attend on the predictions of biased classes in the batch, and may intensify the overfitting issue. To overcome the label bias problem, we propose the adaptive label smoothing (ALS) method to replace the one-hot hard labels with smoothed ones, which learns to allocate label confidences from the biased classes to the others. Specifically, ALS propagates node labels to aggregate the neighborhood label distribution in a pre-processing step, and then updates the optimal smoothed labels online to adapt to specific graph structure. Experiments on the real-world datasets demonstrate that ALS can be generally applied to the main scalable learning frameworks to calibrate the biased labels and improve generalization performances.

Zirui Liu, Qingquan Song, Kaixiong Zhou et al.

Detecting statistical interactions between input features is a crucial and challenging task. Recent advances demonstrate that it is possible to extract learned interactions from trained neural networks. It has also been observed that, in neural networks, any interacting features must follow a strongly weighted connection to common hidden units. Motivated by the observation, in this paper, we propose to investigate the interaction detection problem from a novel topological perspective by analyzing the connectivity in neural networks. Specially, we propose a new measure for quantifying interaction strength, based upon the well-received theory of persistent homology. Based on this measure, a Persistence Interaction detection~(PID) algorithm is developed to efficiently detect interactions. Our proposed algorithm is evaluated across a number of interaction detection tasks on several synthetic and real world datasets with different hyperparameters. Experimental results validate that the PID algorithm outperforms the state-of-the-art baselines.