Accurate INT8 Training Through Dynamic Block-Level Fallback

Transformer models have achieved remarkable success across various AI applications but face significant training costs. Low-bit training, such as INT8 training, can leverage computational units with higher throughput, and has already demonstrated its effectiveness on GPT2 models with block-level quantization. However, it struggles with modern Transformer variants incorporating GLU units. This is because those variants demonstrate complex distributions of activation outliers. To address the challenge, we propose Fallback Quantization, implementing mixed-precision GEMM that dynamically falls back 8-bit to 16-bit for activation blocks containing outliers. Experiments show that our approach is robustly competent in both fine-tuning and pretraining settings. Moreover, our method achieves a 1.57x end-to-end training speedup on RTX4090 GPUs.

SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference

An efficient attention implementation is essential for large models due to its quadratic time complexity. Fortunately, attention commonly exhibits sparsity, i.e., many values in the attention map are near zero, allowing for the omission of corresponding computations. Many studies have utilized the sparse pattern to accelerate attention. However, most existing works focus on optimizing attention within specific models by exploiting certain sparse patterns of the attention map. A universal sparse attention that guarantees both the speedup and end-to-end performance of diverse models remains elusive. In this paper, we propose SpargeAttn, a universal sparse and quantized attention for any model. Our method uses a two-stage online filter: in the first stage, we rapidly and accurately predict the attention map, enabling the skip of some matrix multiplications in attention. In the second stage, we design an online softmax-aware filter that incurs no extra overhead and further skips some matrix multiplications. Experiments show that our method significantly accelerates diverse models, including language, image, and video generation, without sacrificing end-to-end metrics. The codes are available at https://github.com/thu-ml/SpargeAttn.

DeFusion: An Effective Decoupling Fusion Network for Multi-Modal Pregnancy Prediction

Temporal embryo images and parental fertility table indicators are both valuable for pregnancy prediction in \textbf{in vitro fertilization embryo transfer} (IVF-ET). However, current machine learning models cannot make full use of the complementary information between the two modalities to improve pregnancy prediction performance. In this paper, we propose a Decoupling Fusion Network called DeFusion to effectively integrate the multi-modal information for IVF-ET pregnancy prediction. Specifically, we propose a decoupling fusion module that decouples the information from the different modalities into related and unrelated information, thereby achieving a more delicate fusion. And we fuse temporal embryo images with a spatial-temporal position encoding, and extract fertility table indicator information with a table transformer. To evaluate the effectiveness of our model, we use a new dataset including 4046 cases collected from Southern Medical University. The experiments show that our model outperforms state-of-the-art methods. Meanwhile, the performance on the eye disease prediction dataset reflects the model's good generalization. Our code and dataset are available at https://github.com/Ou-Young-1999/DFNet.

SageAttention2 Technical Report: Accurate 4 Bit Attention for Plug-and-play Inference Acceleration

Although quantization for linear layers has been widely used, its application to accelerate the attention process remains limited. SageAttention utilizes 8-bit matrix multiplication, 16-bit matrix multiplication with 16-bit accumulator, and precision-enhancing methods, implementing an accurate and 2x speedup kernel compared to FlashAttention2. To further enhance the efficiency of attention computation while maintaining precision, we propose SageAttention2, which utilizes significantly faster 4-bit matrix multiplication (Matmul) alongside additional precision-enhancing techniques. First, we propose to quantize matrixes $(Q, K)$ to INT4 in a warp-level granularity and quantize matrixes $(\widetilde P, V)$ to FP8. Second, we propose a method to smooth $Q$ and $V$, enhancing the accuracy of attention with INT4 $QK$ and FP8 $PV$. Third, we analyze the quantization accuracy across timesteps and layers, then propose an adaptive quantization method to ensure the end-to-end metrics over various models. The operations per second (OPS) of SageAttention2 surpass FlashAttention2 and xformers by about 3x and 5x on RTX4090, respectively. Comprehensive experiments confirm that our approach incurs negligible end-to-end metrics loss across diverse models, including those for large language processing, image generation, and video generation. The codes are available at https://github.com/thu-ml/SageAttention.

SAM-Swin: SAM-Driven Dual-Swin Transformers with Adaptive Lesion Enhancement for Laryngo-Pharyngeal Tumor Detection

Laryngo-pharyngeal cancer (LPC) is a highly lethal malignancy in the head and neck region. Recent advancements in tumor detection, particularly through dual-branch network architectures, have significantly improved diagnostic accuracy by integrating global and local feature extraction. However, challenges remain in accurately localizing lesions and fully capitalizing on the complementary nature of features within these branches. To address these issues, we propose SAM-Swin, an innovative SAM-driven Dual-Swin Transformer for laryngo-pharyngeal tumor detection. This model leverages the robust segmentation capabilities of the Segment Anything Model 2 (SAM2) to achieve precise lesion segmentation. Meanwhile, we present a multi-scale lesion-aware enhancement module (MS-LAEM) designed to adaptively enhance the learning of nuanced complementary features across various scales, improving the quality of feature extraction and representation. Furthermore, we implement a multi-scale class-aware guidance (CAG) loss that delivers multi-scale targeted supervision, thereby enhancing the model's capacity to extract class-specific features. To validate our approach, we compiled three LPC datasets from the First Affiliated Hospital (FAHSYSU), the Sixth Affiliated Hospital (SAHSYSU) of Sun Yat-sen University, and Nanfang Hospital of Southern Medical University (NHSMU). The FAHSYSU dataset is utilized for internal training, while the SAHSYSU and NHSMU datasets serve for external evaluation. Extensive experiments demonstrate that SAM-Swin outperforms state-of-the-art methods, showcasing its potential for advancing LPC detection and improving patient outcomes. The source code of SAM-Swin is available at the URL of \href{https://github.com/VVJia/SAM-Swin}{https://github.com/VVJia/SAM-Swin}.

Multi-Scale Diffusion: Enhancing Spatial Layout in High-Resolution Panoramic Image Generation

Diffusion models have recently gained recognition for generating diverse and high-quality content, especially in the domain of image synthesis. These models excel not only in creating fixed-size images but also in producing panoramic images. However, existing methods often struggle with spatial layout consistency when producing high-resolution panoramas, due to the lack of guidance of the global image layout. In this paper, we introduce the Multi-Scale Diffusion (MSD) framework, a plug-and-play module that extends the existing panoramic image generation framework to multiple resolution levels. By utilizing gradient descent techniques, our method effectively incorporates structural information from low-resolution images into high-resolution outputs. A comprehensive evaluation of the proposed method was conducted, comparing it with the prior works in qualitative and quantitative dimensions. The evaluation results demonstrate that our method significantly outperforms others in generating coherent high-resolution panoramas.

SAM-FNet: SAM-Guided Fusion Network for Laryngo-Pharyngeal Tumor Detection

Laryngo-pharyngeal cancer (LPC) is a highly fatal malignant disease affecting the head and neck region. Previous studies on endoscopic tumor detection, particularly those leveraging dual-branch network architectures, have shown significant advancements in tumor detection. These studies highlight the potential of dual-branch networks in improving diagnostic accuracy by effectively integrating global and local (lesion) feature extraction. However, they are still limited in their capabilities to accurately locate the lesion region and capture the discriminative feature information between the global and local branches. To address these issues, we propose a novel SAM-guided fusion network (SAM-FNet), a dual-branch network for laryngo-pharyngeal tumor detection. By leveraging the powerful object segmentation capabilities of the Segment Anything Model (SAM), we introduce the SAM into the SAM-FNet to accurately segment the lesion region. Furthermore, we propose a GAN-like feature optimization (GFO) module to capture the discriminative features between the global and local branches, enhancing the fusion feature complementarity. Additionally, we collect two LPC datasets from the First Affiliated Hospital (FAHSYSU) and the Sixth Affiliated Hospital (SAHSYSU) of Sun Yat-sen University. The FAHSYSU dataset is used as the internal dataset for training the model, while the SAHSYSU dataset is used as the external dataset for evaluating the model's performance. Extensive experiments on both datasets of FAHSYSU and SAHSYSU demonstrate that the SAM-FNet can achieve competitive results, outperforming the state-of-the-art counterparts. The source code of SAM-FNet is available at the URL of https://github.com/VVJia/SAM-FNet.

Unsupervised Tumor-Aware Distillation for Multi-Modal Brain Image Translation

Multi-modal brain images from MRI scans are widely used in clinical diagnosis to provide complementary information from different modalities. However, obtaining fully paired multi-modal images in practice is challenging due to various factors, such as time, cost, and artifacts, resulting in modality-missing brain images. To address this problem, unsupervised multi-modal brain image translation has been extensively studied. Existing methods suffer from the problem of brain tumor deformation during translation, as they fail to focus on the tumor areas when translating the whole images. In this paper, we propose an unsupervised tumor-aware distillation teacher-student network called UTAD-Net, which is capable of perceiving and translating tumor areas precisely. Specifically, our model consists of two parts: a teacher network and a student network. The teacher network learns an end-to-end mapping from source to target modality using unpaired images and corresponding tumor masks first. Then, the translation knowledge is distilled into the student network, enabling it to generate more realistic tumor areas and whole images without masks. Experiments show that our model achieves competitive performance on both quantitative and qualitative evaluations of image quality compared with state-of-the-art methods. Furthermore, we demonstrate the effectiveness of the generated images on downstream segmentation tasks. Our code is available at https://github.com/scut-HC/UTAD-Net.

BEND: Bagging Deep Learning Training Based on Efficient Neural Network Diffusion

Bagging has achieved great success in the field of machine learning by integrating multiple base classifiers to build a single strong classifier to reduce model variance. The performance improvement of bagging mainly relies on the number and diversity of base classifiers. However, traditional deep learning model training methods are expensive to train individually and difficult to train multiple models with low similarity in a restricted dataset. Recently, diffusion models, which have been tremendously successful in the fields of imaging and vision, have been found to be effective in generating neural network model weights and biases with diversity. We creatively propose a Bagging deep learning training algorithm based on Efficient Neural network Diffusion (BEND). The originality of BEND comes from the first use of a neural network diffusion model to efficiently build base classifiers for bagging. Our approach is simple but effective, first using multiple trained model weights and biases as inputs to train autoencoder and latent diffusion model to realize a diffusion model from noise to valid neural network parameters. Subsequently, we generate several base classifiers using the trained diffusion model. Finally, we integrate these ba se classifiers for various inference tasks using the Bagging method. Resulting experiments on multiple models and datasets show that our proposed BEND algorithm can consistently outperform the mean and median accuracies of both the original trained model and the diffused model. At the same time, new models diffused using the diffusion model have higher diversity and lower cost than multiple models trained using traditional methods. The BEND approach successfully introduces diffusion models into the new deep learning training domain and provides a new paradigm for future deep learning training and inference.

Learning Multi-Modal Brain Tumor Segmentation from Privileged Semi-Paired MRI Images with Curriculum Disentanglement Learning

Due to the difficulties of obtaining multimodal paired images in clinical practice, recent studies propose to train brain tumor segmentation models with unpaired images and capture complementary information through modality translation. However, these models cannot fully exploit the complementary information from different modalities. In this work, we thus present a novel two-step (intra-modality and inter-modality) curriculum disentanglement learning framework to effectively utilize privileged semi-paired images, i.e. limited paired images that are only available in training, for brain tumor segmentation. Specifically, in the first step, we propose to conduct reconstruction and segmentation with augmented intra-modality style-consistent images. In the second step, the model jointly performs reconstruction, unsupervised/supervised translation, and segmentation for both unpaired and paired inter-modality images. A content consistency loss and a supervised translation loss are proposed to leverage complementary information from different modalities in this step. Through these two steps, our method effectively extracts modality-specific style codes describing the attenuation of tissue features and image contrast, and modality-invariant content codes containing anatomical and functional information from the input images. Experiments on three brain tumor segmentation tasks show that our model outperforms competing segmentation models based on unpaired images.