Zhixu Du

Zhixu Du, Shiyu Li, Yuhao Wu et al.

Mixture-of-Experts (MoE) has emerged as a favorable architecture in the era of large models due to its inherent advantage, i.e., enlarging model capacity without incurring notable computational overhead. Yet, the realization of such benefits often results in ineffective GPU memory utilization, as large portions of the model parameters remain dormant during inference. Moreover, the memory demands of large models consistently outpace the memory capacity of contemporary GPUs. Addressing this, we introduce SiDA-MoE ($\textbf{S}$parsity-$\textbf{i}$nspired $\textbf{D}$ata-$\textbf{A}$ware), an efficient inference approach tailored for large MoE models. SiDA-MoE judiciously exploits both the system's main memory, which is now abundant and readily scalable, and GPU memory by capitalizing on the inherent sparsity on expert activation in MoE models. By adopting a data-aware perspective, SiDA-MoE achieves enhanced model efficiency with a neglectable performance drop. Specifically, SiDA-MoE attains a remarkable speedup in MoE inference with up to $3.93\times$ throughput increasing, up to $72\%$ latency reduction, and up to $80\%$ GPU memory saving with down to $1\%$ performance drop. This work paves the way for scalable and efficient deployment of large MoE models, even with constrained resources. Code is available at: https://github.com/timlee0212/SiDA-MoE.

Zhixu Du, Jingwei Sun, Ang Li et al.

Federated learning (FL) is a popular distributed learning framework that can reduce privacy risks by not explicitly sharing private data. In this work, we explicitly uncover external covariate shift problem in FL, which is caused by the independent local training processes on different devices. We demonstrate that external covariate shifts will lead to the obliteration of some devices' contributions to the global model. Further, we show that normalization layers are indispensable in FL since their inherited properties can alleviate the problem of obliterating some devices' contributions. However, recent works have shown that batch normalization, which is one of the standard components in many deep neural networks, will incur accuracy drop of the global model in FL. The essential reason for the failure of batch normalization in FL is poorly studied. We unveil that external covariate shift is the key reason why batch normalization is ineffective in FL. We also show that layer normalization is a better choice in FL which can mitigate the external covariate shift and improve the performance of the global model. We conduct experiments on CIFAR10 under non-IID settings. The results demonstrate that models with layer normalization converge fastest and achieve the best or comparable accuracy for three different model architectures.

Jingwei Sun, Zhixu Du, Anna Dai et al.

Vertical federated learning (VFL) enables a service provider (i.e., active party) who owns labeled features to collaborate with passive parties who possess auxiliary features to improve model performance. Existing VFL approaches, however, have two major vulnerabilities when passive parties unexpectedly quit in the deployment phase of VFL - severe performance degradation and intellectual property (IP) leakage of the active party's labels. In this paper, we propose \textbf{Party-wise Dropout} to improve the VFL model's robustness against the unexpected exit of passive parties and a defense method called \textbf{DIMIP} to protect the active party's IP in the deployment phase. We evaluate our proposed methods on multiple datasets against different inference attacks. The results show that Party-wise Dropout effectively maintains model performance after the passive party quits, and DIMIP successfully disguises label information from the passive party's feature extractor, thereby mitigating IP leakage.

Zishan Shao, Yixiao Wang, Qinsi Wang et al.

Singular Value Decomposition (SVD) has recently seen a surge of interest as a simple yet powerful tool for large language models (LLMs) compression, with a growing number of works demonstrating 20-80% parameter reductions at minimal accuracy loss. Previous SVD-based approaches have focused primarily on reducing the memory footprint of model weights, largely overlooking the additional activation memory overhead incurred during inference when applying truncated factors via standard dense CUDA kernels. Our experiments demonstrate that this activation overhead, scaling with sequence length and hidden dimension, prevents current SVD compression techniques from achieving any reduction in peak inference memory, thereby limiting their viability for real-world, on-device deployments. We introduce FlashSVD, a novel, end-to-end rank-aware streaming inference framework specifically designed for SVD-compressed large language models. FlashSVD can be seamlessly integrated with any model that employs SVD-based methods for parameter reduction. By fusing low-rank projection kernels directly into both the self-attention and feed-forward network (FFN) pipelines, FlashSVD avoid materializing full-size activation buffers. Instead, small tiles of the truncated factors are loaded into on-chip SRAM, multiplied and reduced on the fly, and immediately evicted, preserving high GPU occupancy and adding no extra latency. On standard encoder benchmarks (e.g., BERT-Base), FlashSVD cuts peak activation memory by up to 70.2% and intermediate transient memory by 75%, all while incur no accuracy loss with upstreaming compression methods, offering a practical path toward memory-constrained deployment of low-rank LLMs.

Haiyue Ma, Zhixu Du, Yiran Chen

In multi-GPU Mixture-of-Experts (MoE) network, experts are distributed across different GPUs, which creates load imbalance as each expert processes different number of tokens. Recent works improve MoE inference load balance by dynamically duplicating popular experts to more GPUs to process excessive tokens, which requires predicting the distribution before routing. In this paper, we discuss the tradeoff of prediction strategies, accuracies, overhead, and end-to-end system performance. We propose MoE-GPS, a framework that guides the selection of the optimal predictor design under various system configurations, by quantifying the performance impact to system-level model runtime. Specifically, we advocate for Distribution-Only Prediction, a prediction strategy that only predicts overall token distribution which significantly reduces overhead compared to the traditional Token-to-Expert Prediction. On Mixtral 8x7B MMLU dataset, MoE-GPS suggests Distribution-Only Prediction which improves end-to-end inference performance by more than 23% compared with Token-to-Expert Prediction.

Tuan Dinh, Daewon Seo, Zhixu Du et al.

We study the GAN conditioning problem, whose goal is to convert a pretrained unconditional GAN into a conditional GAN using labeled data. We first identify and analyze three approaches to this problem -- conditional GAN training from scratch, fine-tuning, and input reprogramming. Our analysis reveals that when the amount of labeled data is small, input reprogramming performs the best. Motivated by real-world scenarios with scarce labeled data, we focus on the input reprogramming approach and carefully analyze the existing algorithm. After identifying a few critical issues of the previous input reprogramming approach, we propose a new algorithm called InRep+. Our algorithm InRep+ addresses the existing issues with the novel uses of invertible neural networks and Positive-Unlabeled (PU) learning. Via extensive experiments, we show that InRep+ outperforms all existing methods, particularly when label information is scarce, noisy, and/or imbalanced. For instance, for the task of conditioning a CIFAR10 GAN with 1% labeled data, InRep+ achieves an average Intra-FID of 76.24, whereas the second-best method achieves 114.51.