We provide an optimized implementation of the forward pass of FlashAttention-2, a popular memory-aware scaled dot-product attention algorithm, as a custom fused CUDA kernel targeting NVIDIA Hopper architecture and written using the open-source CUTLASS library. In doing so, we explain the challenges and techniques involved in fusing online-softmax with back-to-back GEMM kernels, utilizing the Hopper-specific Tensor Memory Accelerator (TMA) and Warpgroup Matrix-Multiply-Accumulate (WGMMA) instructions, defining and transforming CUTLASS Layouts and Tensors, overlapping copy and GEMM operations, and choosing optimal tile sizes for the Q, K and V attention matrices while balancing the register pressure and shared memory utilization. In head-to-head benchmarks on a single H100 PCIe GPU for some common choices of hyperparameters, we observe 20-50% higher FLOPs/s over a version of FlashAttention-2 optimized for last-generation NVIDIA Ampere architecture.
Multi-camera perception tasks have gained significant attention in the field of autonomous driving. However, existing frameworks based on Lift-Splat-Shoot (LSS) in the multi-camera setting cannot produce suitable dense 3D features due to the projection nature and uncontrollable densification process. To resolve this problem, we propose to regulate intermediate dense 3D features with the help of volume rendering. Specifically, we employ volume rendering to process the dense 3D features to obtain corresponding 2D features (e.g., depth maps, semantic maps), which are supervised by associated labels in the training. This manner regulates the generation of dense 3D features on the feature level, providing appropriate dense and unified features for multiple perception tasks. Therefore, our approach is termed Vampire, stands for "Volume rendering As Multi-camera Perception Intermediate feature REgulator". Experimental results on the Occ3D and nuScenes datasets demonstrate that Vampire facilitates fine-grained and appropriate extraction of dense 3D features, and is competitive with existing SOTA methods across diverse downstream perception tasks like 3D occupancy prediction, LiDAR segmentation and 3D objection detection, while utilizing moderate GPU resources. We provide a video demonstration in the supplementary materials and Codes are available at github.com/cskkxjk/Vampire.
Super-resolution (SR) techniques have recently been proposed to upscale the outputs of neural radiance fields (NeRF) and generate high-quality images with enhanced inference speeds. However, existing NeRF+SR methods increase training overhead by using extra input features, loss functions, and/or expensive training procedures such as knowledge distillation. In this paper, we aim to leverage SR for efficiency gains without costly training or architectural changes. Specifically, we build a simple NeRF+SR pipeline that directly combines existing modules, and we propose a lightweight augmentation technique, random patch sampling, for training. Compared to existing NeRF+SR methods, our pipeline mitigates the SR computing overhead and can be trained up to 23x faster, making it feasible to run on consumer devices such as the Apple MacBook. Experiments show our pipeline can upscale NeRF outputs by 2-4x while maintaining high quality, increasing inference speeds by up to 18x on an NVIDIA V100 GPU and 12.8x on an M1 Pro chip. We conclude that SR can be a simple but effective technique for improving the efficiency of NeRF models for consumer devices.
We propose a method that achieves state-of-the-art rendering quality and efficiency on monocular dynamic scene reconstruction using deformable 3D Gaussians. Implicit deformable representations commonly model motion with a canonical space and time-dependent backward-warping deformation field. Our method, GauFRe, uses a forward-warping deformation to explicitly model non-rigid transformations of scene geometry. Specifically, we propose a template set of 3D Gaussians residing in a canonical space, and a time-dependent forward-warping deformation field to model dynamic objects. Additionally, we tailor a 3D Gaussian-specific static component supported by an inductive bias-aware initialization approach which allows the deformation field to focus on moving scene regions, improving the rendering of complex real-world motion. The differentiable pipeline is optimized end-to-end with a self-supervised rendering loss. Experiments show our method achieves competitive results and higher efficiency than both previous state-of-the-art NeRF and Gaussian-based methods. For real-world scenes, GauFRe can train in ~20 mins and offer 96 FPS real-time rendering on an RTX 3090 GPU. Project website: https://lynl7130.github.io/gaufre/index.html
This paper introduces an efficient strategy to transform Large Language Models (LLMs) into Multi-Modal Large Language Models (MLLMs). By conceptualizing this transformation as a domain adaptation process, i.e., transitioning from text understanding to embracing multiple modalities, we intriguingly note that, within each attention block, tuning LayerNorm suffices to yield strong performance. Moreover, when benchmarked against other tuning approaches like full parameter finetuning or LoRA, its benefits on efficiency are substantial. For example, when compared to LoRA on a 13B model scale, performance can be enhanced by an average of over 20% across five multi-modal tasks, and meanwhile, results in a significant reduction of trainable parameters by 41.9% and a decrease in GPU memory usage by 17.6%. On top of this LayerNorm strategy, we showcase that selectively tuning only with conversational data can improve efficiency further. Beyond these empirical outcomes, we provide a comprehensive analysis to explore the role of LayerNorm in adapting LLMs to the multi-modal domain and improving the expressive power of the model.
The Poisson-Boltzmann (PB) model governs the electrostatics of solvated biomolecules, i.e., potential, field, energy, and force. These quantities can provide useful information about protein properties, functions, and dynamics. By considering the advantages of current algorithms and computer hardware, we focus on the parallelization of the treecode-accelerated boundary integral (TABI) PB solver using the Message Passing Interface (MPI) on CPUs and the direct-sum boundary integral (DSBI) PB solver using KOKKOS on GPUs. We provide optimization guidance for users when the DSBI solver on GPU or the TABI solver with MPI on CPU should be used depending on the size of the problem. Specifically, when the number of unknowns is smaller than a predetermined threshold, the GPU-accelerated DSBI solver converges rapidly thus has the potential to perform PB model-based molecular dynamics or Monte Carlo simulation. As practical appliations, our parallelized boundary integral PB solvers are used to solve electrostatics on selected proteins that play significant roles in the spread, treatment, and prevention of COVID-19 virus diseases. For each selected protein, the simulation produces the electrostatic solvation energy as a global measurement and electrostatic surface potential for local details.
The Cellular Potts Model (CPM) is a widely used simulation paradigm for systems of interacting cells that has been used to study scenarios ranging from plant development to morphogenesis, tumour growth and cell migration. Despite their wide use, CPM simulations are considered too computationally intensive for three-dimensional (3D) models at organ scale. CPMs have been difficult to parallelise because of their inherently sequential update scheme. Here, we present a Graphical Processing Unit (GPU)-based parallelisation scheme that preserves local update statistics and is up to 3-4 orders of magnitude faster than serial implementations. We show several examples where our scheme preserves simulation behaviors that are drastically altered by existing parallelisation methods. We use our framework to construct tissue-scale models of liver and lymph node environments containing millions of cells that are directly based on microscopy-imaged tissue structures. Thus, our GPU-based CPM framework enables in silico studies of multicellular systems of unprecedented scale.
Retail product or packaged grocery goods images need to classified in various computer vision applications like self checkout stores, supply chain automation and retail execution evaluation. Previous works explore ways to finetune deep models for this purpose. But because of the fact that finetuning a large model or even linear layer for a pretrained backbone requires to run at least a few epochs of gradient descent for every new retail product added in classification range, frequent retrainings are needed in a real world scenario. In this work, we propose finetuning the vision encoder of a CLIP model in a way that its embeddings can be easily used for nearest neighbor based classification, while also getting accuracy close to or exceeding full finetuning. A nearest neighbor based classifier needs no incremental training for new products, thus saving resources and wait time.
GPUs have become the \emph{defacto} hardware devices for accelerating Deep Neural Network (DNN) inference workloads. However, the conventional \emph{sequential execution mode of DNN operators} in mainstream deep learning frameworks cannot fully utilize GPU resources, even with the operator fusion enabled, due to the increasing complexity of model structures and a greater diversity of operators. Moreover, the \emph{inadequate operator launch order} in parallelized execution scenarios can lead to GPU resource wastage and unexpected performance interference among operators. In this paper, we propose \emph{Opara}, a resource- and interference-aware DNN \underline{Op}erator \underline{para}llel scheduling framework to accelerate DNN inference on GPUs. Specifically, \emph{Opara} first employs \texttt{CUDA Streams} and \texttt{CUDA Graph} to \emph{parallelize} the execution of multiple operators automatically. To further expedite DNN inference, \emph{Opara} leverages the resource demands of operators to judiciously adjust the operator launch order on GPUs, overlapping the execution of compute-intensive and memory-intensive operators. We implement and open source a prototype of \emph{Opara} based on PyTorch in a \emph{non-intrusive} manner. Extensive prototype experiments with representative DNN and Transformer-based models demonstrate that \emph{Opara} outperforms the default sequential \texttt{CUDA Graph} in PyTorch and the state-of-the-art operator parallelism systems by up to $1.68\times$ and $1.29\times$, respectively, yet with acceptable runtime overhead.
The goal of multimodal alignment is to learn a single latent space that is shared between multimodal inputs. The most powerful models in this space have been trained using massive datasets of paired inputs and large-scale computational resources, making them prohibitively expensive to train in many practical scenarios. We surmise that existing unimodal encoders pre-trained on large amounts of unimodal data should provide an effective bootstrap to create multimodal models from unimodal ones at much lower costs. We therefore propose FuseMix, a multimodal augmentation scheme that operates on the latent spaces of arbitrary pre-trained unimodal encoders. Using FuseMix for multimodal alignment, we achieve competitive performance -- and in certain cases outperform state-of-the art methods -- in both image-text and audio-text retrieval, with orders of magnitude less compute and data: for example, we outperform CLIP on the Flickr30K text-to-image retrieval task with $\sim \! 600\times$ fewer GPU days and $\sim \! 80\times$ fewer image-text pairs. Additionally, we show how our method can be applied to convert pre-trained text-to-image generative models into audio-to-image ones. Code is available at: https://github.com/layer6ai-labs/fusemix.
Binary code summarization, while invaluable for understanding code semantics, is challenging due to its labor-intensive nature. This study delves into the potential of large language models (LLMs) for binary code comprehension. To this end, we present BinSum, a comprehensive benchmark and dataset of over 557K binary functions and introduce a novel method for prompt synthesis and optimization. To more accurately gauge LLM performance, we also propose a new semantic similarity metric that surpasses traditional exact-match approaches. Our extensive evaluation of prominent LLMs, including ChatGPT, GPT-4, Llama 2, and Code Llama, reveals 10 pivotal insights. This evaluation generates 4 billion inference tokens, incurred a total expense of 11,418 US dollars and 873 NVIDIA A100 GPU hours. Our findings highlight both the transformative potential of LLMs in this field and the challenges yet to be overcome.
This paper proposes a novel, resource-efficient approach to Visual Speech Recognition (VSR) leveraging speech representations produced by any trained Automatic Speech Recognition (ASR) model. Moving away from the resource-intensive trends prevalent in recent literature, our method distills knowledge from a trained Conformer-based ASR model, achieving competitive performance on standard VSR benchmarks with significantly less resource utilization. Using unlabeled audio-visual data only, our baseline model achieves a word error rate (WER) of 47.4% and 54.7% on the LRS2 and LRS3 test benchmarks, respectively. After fine-tuning the model with limited labeled data, the word error rate reduces to 35% (LRS2) and 45.7% (LRS3). Our model can be trained on a single consumer-grade GPU within a few days and is capable of performing real-time end-to-end VSR on dated hardware, suggesting a path towards more accessible and resource-efficient VSR methodologies.
This paper introduces a novel Parameter-Efficient Fine-Tuning (PEFT) framework for multi-modal, multi-task transfer learning with pre-trained language models. PEFT techniques such as LoRA, BitFit and IA3 have demonstrated comparable performance to full fine-tuning of pre-trained models for specific downstream tasks, all while demanding significantly fewer trainable parameters and reduced GPU memory consumption. However, in the context of multi-modal fine-tuning, the need for architectural modifications or full fine-tuning often becomes apparent. To address this we propose Context-PEFT, which learns different groups of adaptor parameters based on the token's domain. This approach enables LoRA-like weight injection without requiring additional architectural changes. Our method is evaluated on the COCO captioning task, where it outperforms full fine-tuning under similar data constraints while simultaneously offering a substantially more parameter-efficient and computationally economical solution.
Diffusion models have revolutionized text-driven video editing. However, applying these methods to real-world editing encounters two significant challenges: (1) the rapid increase in GPU memory demand as the number of frames grows, and (2) the inter-frame inconsistency in edited videos. To this end, we propose NVEdit, a novel text-driven video editing framework designed to mitigate memory overhead and improve consistent editing for real-world long videos. Specifically, we construct a neural video field, powered by tri-plane and sparse grid, to enable encoding long videos with hundreds of frames in a memory-efficient manner. Next, we update the video field through off-the-shelf Text-to-Image (T2I) models to impart text-driven editing effects. A progressive optimization strategy is developed to preserve original temporal priors. Importantly, both the neural video field and T2I model are adaptable and replaceable, thus inspiring future research. Experiments demonstrate the ability of our approach to edit hundreds of frames with impressive inter-frame consistency. Our project is available at: https://nvedit.github.io/.
The advancement of text-driven 3D content editing has been blessed by the progress from 2D generative diffusion models. However, a major obstacle hindering the widespread adoption of 3D content editing is its time-intensive processing. This challenge arises from the iterative and refining steps required to achieve consistent 3D outputs from 2D image-based generative models. Recent state-of-the-art methods typically require optimization time ranging from tens of minutes to several hours to edit a 3D scene using a single GPU. In this work, we propose that by incorporating correspondence regularization into diffusion models, the process of 3D editing can be significantly accelerated. This approach is inspired by the notion that the estimated samples during diffusion should be multiview-consistent during the diffusion generation process. By leveraging this multiview consistency, we can edit 3D content at a much faster speed. In most scenarios, our proposed technique brings a 10$\times$ speed-up compared to the baseline method and completes the editing of a 3D scene in 2 minutes with comparable quality.
Deep graph convolution networks (GCNs) have recently shown excellent performance in traffic prediction tasks. However, they face some challenges. First, few existing models consider the influence of auxiliary information, i.e., weather and holidays, which may result in a poor grasp of spatial-temporal dynamics of traffic data. Second, both the construction of a dynamic adjacent matrix and regular graph convolution operations have quadratic computation complexity, which restricts the scalability of GCN-based models. To address such challenges, this work proposes a deep encoder-decoder model entitled AIMSAN. It contains an auxiliary information-aware module (AIM) and sparse cross attention-based graph convolution network (SAN). The former learns multi-attribute auxiliary information and obtains its embedded presentation of different time-window sizes. The latter uses a cross-attention mechanism to construct dynamic adjacent matrices by fusing traffic data and embedded auxiliary data. Then, SAN applies diffusion GCN on traffic data to mine rich spatial-temporal dynamics. Furthermore, AIMSAN considers and uses the spatial sparseness of traffic nodes to reduce the quadratic computation complexity. Experimental results on three public traffic datasets demonstrate that the proposed method outperforms other counterparts in terms of various performance indices. Specifically, the proposed method has competitive performance with the state-of-the-art algorithms but saves 35.74% of GPU memory usage, 42.25% of training time, and 45.51% of validation time on average.
In the acceleration of deep neural network training, the GPU has become the mainstream platform. GPUs face substantial challenges on GNNs, such as workload imbalance and memory access irregularities, leading to underutilized hardware. Existing solutions such as PyG, DGL with cuSPARSE, and GNNAdvisor frameworks partially address these challenges but memory traffic is still significant. We argue that drastic performance improvements can only be achieved by the vertical optimization of algorithm and system innovations, rather than treating the speedup optimization as an "after-thought" (i.e., (i) given a GNN algorithm, designing an accelerator, or (ii) given hardware, mainly optimizing the GNN algorithm). In this paper, we present MaxK-GNN, an advanced high-performance GPU training system integrating algorithm and system innovation. (i) We introduce the MaxK nonlinearity and provide a theoretical analysis of MaxK nonlinearity as a universal approximator, and present the Compressed Balanced Sparse Row (CBSR) format, designed to store the data and index of the feature matrix after nonlinearity; (ii) We design a coalescing enhanced forward computation with row-wise product-based SpGEMM Kernel using CBSR for input feature matrix fetching and strategic placement of a sparse output accumulation buffer in shared memory; (iii) We develop an optimized backward computation with outer product-based and SSpMM Kernel. We conduct extensive evaluations of MaxK-GNN and report the end-to-end system run-time. Experiments show that MaxK-GNN system could approach the theoretical speedup limit according to Amdahl's law. We achieve comparable accuracy to SOTA GNNs, but at a significantly increased speed: 3.22/4.24 times speedup (vs. theoretical limits, 5.52/7.27 times) on Reddit compared to DGL and GNNAdvisor implementations.
On-device memory concerns in distributed deep learning have become severe due to (i) the growth of model size in multi-GPU training, and (ii) the wide adoption of deep neural networks for federated learning on IoT devices which have limited storage. In such settings, communication efficient optimization methods are attractive alternatives, however they still struggle with memory issues. To tackle these challenges, we propose an communication efficient method called contractive error feedback (ConEF). As opposed to SGD with error-feedback (EFSGD) that inefficiently manages memory, ConEF obtains the sweet spot of convergence and memory usage, and achieves communication efficiency by leveraging biased and all-reducable gradient compression. We empirically validate ConEF on various learning tasks that include image classification, language modeling, and machine translation and observe that ConEF saves 80\% - 90\% of the extra memory in EFSGD with almost no loss on test performance, while also achieving 1.3x - 5x speedup of SGD. Through our work, we also demonstrate the feasibility and convergence of ConEF to clear up the theoretical barrier of integrating ConEF to popular memory efficient frameworks such as ZeRO-3.
NN-Descent is a classic k-NN graph construction approach. It is still widely employed in machine learning, computer vision, and information retrieval tasks due to its efficiency and genericness. However, the current design only works well on CPU. In this paper, NN-Descent has been redesigned to adapt to the GPU architecture. A new graph update strategy called selective update is proposed. It reduces the data exchange between GPU cores and GPU global memory significantly, which is the processing bottleneck under GPU computation architecture. This redesign leads to full exploitation of the parallelism of the GPU hardware. In the meantime, the genericness, as well as the simplicity of NN-Descent, are well-preserved. Moreover, a procedure that allows to k-NN graph to be merged efficiently on GPU is proposed. It makes the construction of high-quality k-NN graphs for out-of-GPU-memory datasets tractable. Our approach is 100-250× faster than the single-thread NN-Descent and is 2.5-5× faster than the existing GPU-based approaches as we tested on million as well as billion scale datasets.
Dataset distillation is the technique of synthesizing smaller condensed datasets from large original datasets while retaining necessary information to persist the effect. In this paper, we approach the dataset distillation problem from a novel perspective: we regard minimizing the prediction discrepancy on the real data distribution between models, which are respectively trained on the large original dataset and on the small distilled dataset, as a conduit for condensing information from the raw data into the distilled version. An adversarial framework is proposed to solve the problem efficiently. In contrast to existing distillation methods involving nested optimization or long-range gradient unrolling, our approach hinges on single-level optimization. This ensures the memory efficiency of our method and provides a flexible tradeoff between time and memory budgets, allowing us to distil ImageNet-1K using a minimum of only 6.5GB of GPU memory. Under the optimal tradeoff strategy, it requires only 2.5$\times$ less memory and 5$\times$ less runtime compared to the state-of-the-art. Empirically, our method can produce synthetic datasets just 10% the size of the original, yet achieve, on average, 94% of the test accuracy of models trained on the full original datasets including ImageNet-1K, significantly surpassing state-of-the-art. Additionally, extensive tests reveal that our distilled datasets excel in cross-architecture generalization capabilities.