Search

Long-context capability is critical for multi-modal foundation models, especially for long video understanding. We introduce LongVILA, a full-stack solution for long-context visual-language models by co-designing the algorithm and system. For model training, we upgrade existing VLMs to support long video understanding by incorporating two additional stages, i.e., long context extension and long supervised fine-tuning. However, training on long video is computationally and memory intensive. We introduce the long-context Multi-Modal Sequence Parallelism (MM-SP) system that efficiently parallelizes long video training and inference, enabling 2M context length training on 256 GPUs without any gradient checkpointing. LongVILA efficiently extends the number of video frames of VILA from 8 to 1024, improving the long video captioning score from 2.00 to 3.26 (out of 5), achieving 99.5% accuracy in 1400-frame (274k context length) video needle-in-a-haystack. LongVILA-8B demonstrates consistent accuracy improvements on long videos in the VideoMME benchmark as the number of frames increases. Besides, MM-SP is 2.1x - 5.7x faster than ring sequence parallelism and 1.1x - 1.4x faster than Megatron with context parallelism + tensor parallelism. Moreover, it seamlessly integrates with Hugging Face Transformers., Comment: Code and models are available at https://github.com/NVlabs/VILA/blob/main/LongVILA.md

Semantic segmentation empowers numerous real-world applications, such as autonomous driving and augmented/mixed reality. These applications often operate on high-resolution images (e.g., 8 megapixels) to capture the fine details. However, this comes at the cost of considerable computational complexity, hindering the deployment in latency-sensitive scenarios. In this paper, we introduce SparseRefine, a novel approach that enhances dense low-resolution predictions with sparse high-resolution refinements. Based on coarse low-resolution outputs, SparseRefine first uses an entropy selector to identify a sparse set of pixels with high entropy. It then employs a sparse feature extractor to efficiently generate the refinements for those pixels of interest. Finally, it leverages a gated ensembler to apply these sparse refinements to the initial coarse predictions. SparseRefine can be seamlessly integrated into any existing semantic segmentation model, regardless of CNN- or ViT-based. SparseRefine achieves significant speedup: 1.5 to 3.7 times when applied to HRNet-W48, SegFormer-B5, Mask2Former-T/L and SegNeXt-L on Cityscapes, with negligible to no loss of accuracy. Our "dense+sparse" paradigm paves the way for efficient high-resolution visual computing., Comment: ECCV 2024. The first two authors contributed equally to this work. Project page: https://sparserefine.mit.edu

Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/qserve., Comment: The first three authors contribute equally to this project and are listed in the alphabetical order. Yujun Lin leads the quantization algorithm, Haotian Tang and Shang Yang lead the GPU kernels and the serving system. Code is available at https://github.com/mit-han-lab/qserve

Data acquisition, image processing, and image quality are the long-lasting issues for terahertz (THz) 3D reconstructed imaging. Existing methods are primarily designed for 2D scenarios, given the challenges associated with obtaining super-resolution (SR) data and the absence of an efficient SR 3D reconstruction framework in conventional computed tomography (CT). Here, we demonstrate BLIss, a new approach for THz SR 3D reconstruction with sparse 2D data input. BLIss seamlessly integrates conventional CT techniques and variational framework with the core of the adapted Euler-Elastica-based model. The quantitative 3D image evaluation metrics, including the standard deviation of Gaussian, mean curvatures, and the multi-scale structural similarity index measure (MS-SSIM), validate the superior smoothness and fidelity achieved with our variational framework approach compared with conventional THz CT modal. Beyond its contributions to advancing THz SR 3D reconstruction, BLIss demonstrates potential applicability in other imaging modalities, such as X-ray and MRI. This suggests extensive impacts on the broader field of imaging applications., Comment: 15 pages, 7 figures. Supplemental Document: https://doi.org/10.6084/m9.figshare.24455206

Sub-terahertz (Sub-THz) waves possess exceptional attributes, capable of penetrating non-metallic and non-polarized materials while ensuring bio-safety. However, their practicality in imaging is marred by the emergence of troublesome speckle artifacts, primarily due to diffraction effects caused by wavelengths comparable to object dimensions. In addressing this limitation, we present the Diffuser-aided sub-THz Imaging System (DaISy), which utilizes a diffuser and a focusing lens to convert coherent waves into incoherent counterparts. The cornerstone of our progress lies in a coherence theory-based theoretical framework, pivotal for designing and validating the THz diffuser, and systematically evaluating speckle phenomena. Our experimental results utilizing DaISy reveal substantial improvements in imaging quality and nearly diffraction-limited spatial resolution. Moreover, we demonstrate a tangible application of DaISy in the scenario of security scanning, highlighting the versatile potential of sub-THz waves in miscellaneous fields., Comment: These authors (Shao-Hsuan Wu and Yiyao Zhang) contributed equally to this work. 15 pages, 7 figures. Supplemental Document: https://doi.org/10.6084/m9.figshare.25328746

Sparse convolution plays a pivotal role in emerging workloads, including point cloud processing in AR/VR, autonomous driving, and graph understanding in recommendation systems. Since the computation pattern is sparse and irregular, specialized high-performance kernels are required. Existing GPU libraries offer two dataflow types for sparse convolution. The gather-GEMM-scatter dataflow is easy to implement but not optimal in performance, while the dataflows with overlapped computation and memory access (e.g.implicit GEMM) are highly performant but have very high engineering costs. In this paper, we introduce TorchSparse++, a new GPU library that achieves the best of both worlds. We create a highly efficient Sparse Kernel Generator that generates performant sparse convolution kernels at less than one-tenth of the engineering cost of the current state-of-the-art system. On top of this, we design the Sparse Autotuner, which extends the design space of existing sparse convolution libraries and searches for the best dataflow configurations for training and inference workloads. Consequently, TorchSparse++ achieves 2.9x, 3.3x, 2.2x and 1.7x measured end-to-end speedup on an NVIDIA A100 GPU over state-of-the-art MinkowskiEngine, SpConv 1.2, TorchSparse and SpConv v2 in inference; and is 1.2-1.3x faster than SpConv v2 in mixed precision training across seven representative autonomous driving benchmarks. It also seamlessly supports graph convolutions, achieving 2.6-7.6x faster inference speed compared with state-of-the-art graph deep learning libraries., Comment: MICRO 2023; Haotian Tang and Shang Yang contributed equally to this project

In many imaging applications where segmented features (e.g. blood vessels) are further used for other numerical simulations (e.g. finite element analysis), the obtained surfaces do not have fine resolutions suitable for the task. Increasing the resolution of such surfaces becomes crucial. This paper proposes a new variational model for solving this problem, based on an Euler-Elastica-based regulariser. Further, we propose and implement two numerical algorithms for solving the model, a projected gradient descent method and the alternating direction method of multipliers. Numerical experiments using real-life examples (including two from outputs of another variational model) have been illustrated for effectiveness. The advantages of the new model are shown through quantitative comparisons by the standard deviation of Gaussian curvatures and mean curvatures from the viewpoint of discrete geometry., Comment: 33 pages, 25 figures

Clinical guidelines underscore the importance of regularly monitoring and surveilling arteriovenous fistula (AVF) access in hemodialysis patients to promptly detect any dysfunction. Although phono-angiography/sound analysis overcomes the limitations of standardized AVF stenosis diagnosis tool, prior studies have depended on conventional feature extraction methods, restricting their applicability in diverse contexts. In contrast, representation learning captures fundamental underlying factors that can be readily transferred across different contexts. We propose an approach based on deep denoising autoencoders (DAEs) that perform dimensionality reduction and reconstruction tasks using the waveform obtained through one-level discrete wavelet transform, utilizing representation learning. Our results demonstrate that the latent representation generated by the DAE surpasses expectations with an accuracy of 0.93. The incorporation of noise-mixing and the utilization of a noise-to-clean scheme effectively enhance the discriminative capabilities of the latent representation. Moreover, when employed to identify patient-specific characteristics, the latent representation exhibited performance by surpassing an accuracy of 0.92. Appropriate light-weighted methods can restore the detection performance of the excessively reduced dimensionality version and enable operation on less computational devices. Our findings suggest that representation learning is a more feasible approach for extracting auscultation features in AVF, leading to improved generalization and applicability across multiple tasks. The manipulation of latent representations holds immense potential for future advancements. Further investigations in this area are promising and warrant continued exploration.

Large language models (LLMs) have transformed numerous AI applications. On-device LLM is becoming increasingly important: running LLMs locally on edge devices can reduce the cloud computing cost and protect users' privacy. However, the astronomical model size and the limited hardware resource pose significant deployment challenges. We propose Activation-aware Weight Quantization (AWQ), a hardware-friendly approach for LLM low-bit weight-only quantization. AWQ finds that not all weights in an LLM are equally important. Protecting only 1% salient weights can greatly reduce quantization error. To identify salient weight channels, we should refer to the activation distribution, not weights. To avoid the hardware-inefficient mix-precision quantization, we mathematically derive that scaling up the salient channels can reduce the quantization error. AWQ employs an equivalent transformation to scale the salient weight channels to protect them. The scale is determined by collecting the activation statistics offline. AWQ does not rely on any backpropagation or reconstruction, so it generalizes to different domains and modalities without overfitting the calibration set. AWQ outperforms existing work on various language modeling and domain-specific benchmarks (coding and math). Thanks to better generalization, it achieves excellent quantization performance for instruction-tuned LMs and, for the first time, multi-modal LMs. Alongside AWQ, we implement TinyChat, an efficient and flexible inference framework tailored for 4-bit on-device LLM/VLMs. With kernel fusion and platform-aware weight packing, TinyChat offers more than 3x speedup over the Huggingface FP16 implementation on both desktop and mobile GPUs. It also democratizes the deployment of the 70B Llama-2 model on mobile GPUs., Comment: MLSys 2024 Best Paper Award. Code available at: https://github.com/mit-han-lab/llm-awq

Terahertz (THz) tomographic imaging has recently attracted significant attention thanks to its non-invasive, non-destructive, non-ionizing, material-classification, and ultra-fast nature for object exploration and inspection. However, its strong water absorption nature and low noise tolerance lead to undesired blurs and distortions of reconstructed THz images. The diffraction-limited THz signals highly constrain the performances of existing restoration methods. To address the problem, we propose a novel multi-view Subspace-Attention-guided Restoration Network (SARNet) that fuses multi-view and multi-spectral features of THz images for effective image restoration and 3D tomographic reconstruction. To this end, SARNet uses multi-scale branches to extract intra-view spatio-spectral amplitude and phase features and fuse them via shared subspace projection and self-attention guidance. We then perform inter-view fusion to further improve the restoration of individual views by leveraging the redundancies between neighboring views. Here, we experimentally construct a THz time-domain spectroscopy (THz-TDS) system covering a broad frequency range from 0.1 THz to 4 THz for building up a temporal/spectral/spatial/ material THz database of hidden 3D objects. Complementary to a quantitative evaluation, we demonstrate the effectiveness of our SARNet model on 3D THz tomographic reconstruction applications., Comment: 34 pages, 13 figures

Transformer, as an alternative to CNN, has been proven effective in many modalities (e.g., texts and images). For 3D point cloud transformers, existing efforts focus primarily on pushing their accuracy to the state-of-the-art level. However, their latency lags behind sparse convolution-based models (3x slower), hindering their usage in resource-constrained, latency-sensitive applications (such as autonomous driving). This inefficiency comes from point clouds' sparse and irregular nature, whereas transformers are designed for dense, regular workloads. This paper presents FlatFormer to close this latency gap by trading spatial proximity for better computational regularity. We first flatten the point cloud with window-based sorting and partition points into groups of equal sizes rather than windows of equal shapes. This effectively avoids expensive structuring and padding overheads. We then apply self-attention within groups to extract local features, alternate sorting axis to gather features from different directions, and shift windows to exchange features across groups. FlatFormer delivers state-of-the-art accuracy on Waymo Open Dataset with 4.6x speedup over (transformer-based) SST and 1.4x speedup over (sparse convolutional) CenterPoint. This is the first point cloud transformer that achieves real-time performance on edge GPUs and is faster than sparse convolutional methods while achieving on-par or even superior accuracy on large-scale benchmarks., Comment: CVPR 2023. The first two authors contributed equally to this work. Project page: https://flatformer.mit.edu

Visualizing information inside objects is an ever-lasting need to bridge the world from physics, chemistry, biology to computation. Among all tomographic techniques, terahertz (THz) computational imaging has demonstrated its unique sensing features to digitalize multi-dimensional object information in a non-destructive, non-ionizing, and non-invasive way. Applying modern signal processing and physics-guided modalities, THz computational imaging systems are now launched in various application fields in industrial inspection, security screening, chemical inspection and non-destructive evaluation. In this article, we overview recent advances in THz computational imaging modalities in the aspects of system configuration, wave propagation and interaction models, physics-guided algorithm for digitalizing interior information of imaged objects. Several image restoration and reconstruction issues based on multi-dimensional THz signals are further discussed, which provides a crosslink between material digitalization, functional property extraction, and multi-dimensional imager utilization from a signal processing perspective.

Terahertz computed tomography (THz CT) has drawn significant attention because of its unique capability to bring multi-dimensional object information from invisible to visible. However, current physics-model-based THz CT modalities present low data use efficiency on time-resolved THz signals and low model fusion extensibility, limiting their application fields' practical use. In this paper, we propose a supervised THz deep learning computed tomography (THz DL-CT) framework based on time-domain information. THz DL-CT restores superior THz tomographic images of 3D objects by extracting features from spatio-temporal THz signals without any prior material information. Compared with conventional and machine learning based methods, THz DL-CT delivers at least 50.2%, and 52.6% superior in root mean square error (RMSE) and structural similarity index (SSIM), respectively. Additionally, we have experimentally demonstrated that the pretrained THz DL-CT model can generalize to reconstruct multi-material systems with no prerequisite information. THz CT through the DL data fusion approach provides a new pathway for non-invasive functional imaging in object investigation.

Face anti-spoofing (FAS) plays a critical role in securing face recognition systems from different presentation attacks. Previous works leverage auxiliary pixel-level supervision and domain generalization approaches to address unseen spoof types. However, the local characteristics of image captures, i.e., capturing devices and presenting materials, are ignored in existing works and we argue that such information is required for networks to discriminate between live and spoof images. In this work, we propose PatchNet which reformulates face anti-spoofing as a fine-grained patch-type recognition problem. To be specific, our framework recognizes the combination of capturing devices and presenting materials based on the patches cropped from non-distorted face images. This reformulation can largely improve the data variation and enforce the network to learn discriminative feature from local capture patterns. In addition, to further improve the generalization ability of the spoof feature, we propose the novel Asymmetric Margin-based Classification Loss and Self-supervised Similarity Loss to regularize the patch embedding space. Our experimental results verify our assumption and show that the model is capable of recognizing unseen spoof types robustly by only looking at local regions. Moreover, the fine-grained and patch-level reformulation of FAS outperforms the existing approaches on intra-dataset, cross-dataset, and domain generalization benchmarks. Furthermore, our PatchNet framework can enable practical applications like Few-Shot Reference-based FAS and facilitate future exploration of spoof-related intrinsic cues., Comment: CVPR 2022

Sparse Matrix-Matrix Multiplication (SpMM) has served as fundamental components in various domains. Many previous studies exploit GPUs for SpMM acceleration because GPUs provide high bandwidth and parallelism. We point out that a static design does not always improve the performance of SpMM on different input data (e.g., >85\% performance loss with a single algorithm). In this paper, we consider the challenge of input dynamics from a novel auto-tuning perspective, while following issues remain to be solved: (1) Orthogonal design principles considering sparsity. Orthogonal design principles for such a sparse problem should be extracted to form different algorithms, and further used for performance tuning. (2) Nontrivial implementations in the algorithm space. Combining orthogonal design principles to create new algorithms needs to tackle with new challenges like thread race handling. (3) Heuristic adaptability to input dynamics. The heuristic adaptability is required to dynamically optimize code for input dynamics. To tackle these challenges, we first propose a novel three-loop model to extract orthogonal design principles for SpMM on GPUs. The model not only covers previous SpMM designs, but also comes up with new designs absent from previous studies. We propose techniques like conditional reduction to implement algorithms missing in previous studies. We further propose DA-SpMM, a Data-Aware heuristic GPU kernel for SpMM. DA-SpMM adaptively optimizes code considering input dynamics. Extensive experimental results show that, DA-SpMM achieves 1.26x~1.37x speedup compared with the best NVIDIA cuSPARSE algorithm on average, and brings up to 5.59x end-to-end speedup to applications like Graph Neural Networks., Comment: Accepted by Design Automation Conference (DAC) 2022

We propose a new architecture, double-pass multiple-plate continuum (DPMPC), for nonlinear pulse compression. In addition to smaller footprint, a double-pass configuration is designed to achieve substantial bandwidth broadening without incurring noticeable higher-order dispersion, thus improving the temporal contrast over those of traditional single-pass geometry when only quadratic spectral phase can be compensated. In our proof-of-concept experiment, 187~$\mu$J, 190-fs Yb-based laser pulse is compressed to 20~fs with high throughput (75%), high Strehl ratio (0.76) and excellent beam homogeneity by using DPMPC. Subsequently generated octave-spanning spectrum exhibits a significantly raised blue tail compared with that driven by pulses from a single-pass counterpart.

Terahertz (THz) technology has been a great candidate for applications, including pharmaceutic analysis, chemical identification, and remote sensing and imaging due to its non-invasive and non-destructive properties. Among those applications, penetrating-type hyperspectral THz signals, which provide crucial material information, normally involve a noisy, complex mixture system. Additionally, the measured THz signals could be ill-conditioned due to the overlap of the material absorption peak in the measured bands. To address those issues, we consider penetrating-type signal mixtures and aim to develop a blind hyperspectral unmixing (HU) method without requiring any information from a prebuilt database. The proposed HYperspectral Penetrating-type Ellipsoidal ReconstructION (HYPERION) algorithm is unsupervised, not relying on collecting extensive data or sophisticated model training. Instead, it is developed based on elegant ellipsoidal geometry under a very mild requirement on data purity, whose excellent efficacy is experimentally demonstrated., Comment: 51 pages, 15 figures

Optical pump-probe spectroscopy is a powerful tool for the study of non-equilibrium electronic dynamics and finds wide applications across a range of fields, from physics and chemistry to material science and biology. However, a shortcoming of conventional pump-probe spectroscopy is that photoinduced changes in transmission, reflection and scattering can simultaneously contribute to the measured differential spectra, leading to ambiguities in assigning the origin of spectral signatures and ruling out quantitative interpretation of the spectra. Ideally, these methods would measure the underlying dielectric function (or the complex refractive index) which would then directly provide quantitative information on the transient excited state dynamics free of these ambiguities. Here we present and test a model independent route to transform differential transmission or reflection spectra, measured via conventional optical pump-probe spectroscopy, to changes in the quantitative transient dielectric function. We benchmark this method against changes in the real refractive index measured using time-resolved Frequency Domain Interferometry in prototypical inorganic and organic semiconductor films. Our methodology can be applied to existing and future pump-probe data sets, allowing for an unambiguous and quantitative characterisation of the transient photoexcited spectra of materials. This in turn will accelerate the adoption of pump-probe spectroscopy as a facile and robust materials characterisation and screening tool.

Terahertz (THz) imaging has recently attracted significant attention thanks to its non-invasive, non-destructive, non-ionizing, material-classification, and ultra-fast nature for object exploration and inspection. However, its strong water absorption nature and low noise tolerance lead to undesired blurs and distortions of reconstructed THz images. The performances of existing restoration methods are highly constrained by the diffraction-limited THz signals. To address the problem, we propose a novel Subspace-and-Attention-guided Restoration Network (SARNet) that fuses multi-spectral features of a THz image for effective restoration. To this end, SARNet uses multi-scale branches to extract spatio-spectral features of amplitude and phase which are then fused via shared subspace projection and attention guidance. Here, we experimentally construct ultra-fast THz time-domain spectroscopy system covering a broad frequency range from 0.1 THz to 4 THz for building up temporal/spectral/spatial/phase/material THz database of hidden 3D objects. Complementary to a quantitative evaluation, we demonstrate the effectiveness of our SARNet model on 3D THz tomographic reconstruction, Comment: 10 pages, 9 figures

Action-constrained reinforcement learning (RL) is a widely-used approach in various real-world applications, such as scheduling in networked systems with resource constraints and control of a robot with kinematic constraints. While the existing projection-based approaches ensure zero constraint violation, they could suffer from the zero-gradient problem due to the tight coupling of the policy gradient and the projection, which results in sample-inefficient training and slow convergence. To tackle this issue, we propose a learning algorithm that decouples the action constraints from the policy parameter update by leveraging state-wise Frank-Wolfe and a regression-based policy update scheme. Moreover, we show that the proposed algorithm enjoys convergence and policy improvement properties in the tabular case as well as generalizes the popular DDPG algorithm for action-constrained RL in the general case. Through experiments, we demonstrate that the proposed algorithm significantly outperforms the benchmark methods on a variety of control tasks., Comment: Published in UAI 2021

We propose a unified representation learning framework to address the Cross Model Compatibility (CMC) problem in the context of visual search applications. Cross compatibility between different embedding models enables the visual search systems to correctly recognize and retrieve identities without re-encoding user images, which are usually not available due to privacy concerns. While there are existing approaches to address CMC in face identification, they fail to work in a more challenging setting where the distributions of embedding models shift drastically. The proposed solution improves CMC performance by introducing a light-weight Residual Bottleneck Transformation (RBT) module and a new training scheme to optimize the embedding spaces. Extensive experiments demonstrate that our proposed solution outperforms previous approaches by a large margin for various challenging visual search scenarios of face recognition and person re-identification., Comment: To appear in British Machine Vision Conference (BMVC) 2020