Recent large language model applications, such as Retrieval-Augmented Generation and chatbots, have led to an increased need to process longer input contexts. However, this requirement is hampered by inherent limitations. Architecturally, models are constrained by a context window defined during training. Additionally, processing extensive texts requires substantial GPU memory. We propose a novel approach, Finch, to compress the input context by leveraging the pre-trained model weights of the self-attention. Given a prompt and a long text, Finch iteratively identifies the most relevant Key (K) and Value (V) pairs over chunks of the text conditioned on the prompt. Only such pairs are stored in the KV cache, which, within the space constrained by the context window, ultimately contains a compressed version of the long text. Our proposal enables models to consume large inputs even with high compression (up to 93x) while preserving semantic integrity without the need for fine-tuning.
Multiphysics problems that are characterized by complex interactions among fluid dynamics, heat transfer, structural mechanics, and electromagnetics, are inherently challenging due to their coupled nature. While experimental data on certain state variables may be available, integrating these data with numerical solvers remains a significant challenge. Physics-informed neural networks (PINNs) have shown promising results in various engineering disciplines, particularly in handling noisy data and solving inverse problems in partial differential equations (PDEs). However, their effectiveness in forecasting nonlinear phenomena in multiphysics regimes, particularly involving turbulence, is yet to be fully established. This study introduces NeuroSEM, a hybrid framework integrating PINNs with the high-fidelity Spectral Element Method (SEM) solver, Nektar++. NeuroSEM leverages the strengths of both PINNs and SEM, providing robust solutions for multiphysics problems. PINNs are trained to assimilate data and model physical phenomena in specific subdomains, which are then integrated into the Nektar++ solver. We demonstrate the efficiency and accuracy of NeuroSEM for thermal convection in cavity flow and flow past a cylinder. We applied NeuroSEM to the Rayleigh-B\'enard convection system, including cases with missing thermal boundary conditions and noisy datasets, and to real particle image velocimetry (PIV) data to capture flow patterns characterized by horseshoe vortical structures. The framework's plug-and-play nature facilitates its extension to other multiphysics or multiscale problems. Furthermore, NeuroSEM is optimized for efficient execution on emerging integrated GPU-CPU architectures. This hybrid approach enhances the accuracy and efficiency of simulations, making it a powerful tool for tackling complex engineering challenges in various scientific domains.
Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tensors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) low-rank-aware quantization compatibility enhancements, and (4) optimized GPU kernels with operators fusion. Extensive experiments with popular LLMs show that Palu compresses KV-Cache by 50% while maintaining strong accuracy and delivering up to 1.89x on the RoPE-based attention module. When combined with quantization, Palu's inherent quantization-friendly design yields small to negligible extra accuracy degradation while saving additional memory than quantization-only methods and achieving up to 2.91x speedup for the RoPE-based attention. Moreover, it maintains comparable or even better accuracy (up to 1.19 lower perplexity) compared to quantization-only methods. These results demonstrate Palu's superior capability to effectively address the efficiency and memory challenges of LLM inference posed by KV-Cache. Our code is publicly available at: https://github.com/shadowpa0327/Palu
Keyphrase Recommendation has been a pivotal problem in advertising and e-commerce where advertisers/sellers are recommended keyphrases (search queries) to bid on to increase their sales. It is a challenging task due to the plethora of items shown on online platforms and various possible queries that users search while showing varying interest in the displayed items. Moreover, query/keyphrase recommendations need to be made in real-time and in a resource-constrained environment. This problem can be framed as an Extreme Multi-label (XML) Short text classification by tagging the input text with keywords as labels. Traditional neural network models are either infeasible or have slower inference latency due to large label spaces. We present Graphite, a graph-based classifier model that provides real-time keyphrase recommendations that are on par with standard text classification models. Furthermore, it doesn't utilize GPU resources, which can be limited in production environments. Due to its lightweight nature and smaller footprint, it can train on very large datasets, where state-of-the-art XML models fail due to extreme resource requirements. Graphite is deterministic, transparent, and intrinsically more interpretable than neural network-based models. We present a comprehensive analysis of our model's performance across forty categories spanning eBay's English-speaking sites.
N:M sparsity pruning is a powerful technique for compressing deep neural networks, utilizing NVIDIA's Sparse Tensor Core technology. This method benefits from hardware support for sparse indexing, enabling the adoption of fine-grained sparsity to maintain model accuracy while minimizing the overhead typically associated with irregular data access. Although restricted to a fixed level of sparsity due to its reliance on hardware, N:M sparsity can be combined with coarser sparsity techniques to achieve diverse compression ratios. Initially, column-wise vector sparsity is applied to a dense model, followed by row-wise N:M sparsity on the preserved column vectors. We call this multi-level approach as hierarchical N:M (HiNM) sparsity. Similar to earlier single-level sparsity techniques, HiNM sparsity necessitates an effective channel permutation strategy to maximize the accuracy of the compressed networks. However, it introduces further complexities by requiring the rearrangement of both input and output channels, addressing challenges such as permutation sequence, HiNM-sparsity-aware permutation, and maintaining consistency in channel ordering across layers. In this paper, we introduce a channel permutation method designed specifically for HiNM sparsity, named gyro-permutation. This method is crafted to exploit the unique characteristics of HiNM pruning, incorporating a strategic policy in each permutation phase, including channel sampling, clustering, and assignment, to circumvent local minima. Additionally, we have developed a GPU kernel that facilitates independent layer permutation during the execution of HiNM sparse networks. Our extensive experimental evaluations on various DNN models demonstrate that our gyro-permutation significantly enhances the accuracy of HiNM sparse networks, allowing them to reach performance levels comparable to those of unstructured sparse networks.
Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications. However, their increased computational and memory demands present significant challenges, especially when handling long sequences. This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. Unlike existing approaches that optimize the memory based on the sequence length, we identify substantial redundancy in the channel dimension of the KV cache, as indicated by an uneven magnitude distribution and a low-rank structure in the attention weights. In response, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in KV cache memory costs by over 20% compared with vanilla KV cache eviction and quantization methods. For instance, ThinK integrated with KIVI can achieve a 2.8x reduction in peak memory usage while maintaining nearly the same quality, enabling up to a 5x increase in batch size when using a single GPU. Extensive evaluations on the LLaMA and Mistral models across various long-sequence datasets verified the efficiency of ThinK, establishing a new baseline algorithm for efficient LLM deployment without compromising performance.
Recent advancements in Contrastive Language-Image Pre-training (CLIP) have demonstrated notable success in self-supervised representation learning across various tasks. However, the existing CLIP-like approaches often demand extensive GPU resources and prolonged training times due to the considerable size of the model and dataset, making them poor for medical applications, in which large datasets are not always common. Meanwhile, the language model prompts are mainly manually derived from labels tied to images, potentially overlooking the richness of information within training samples. We introduce a novel language-image Contrastive Learning method with an Efficient large language model and prompt Fine-Tuning (CLEFT) that harnesses the strengths of the extensive pre-trained language and visual models. Furthermore, we present an efficient strategy for learning context-based prompts that mitigates the gap between informative clinical diagnostic data and simple class labels. Our method demonstrates state-of-the-art performance on multiple chest X-ray and mammography datasets compared with various baselines. The proposed parameter efficient framework can reduce the total trainable model size by 39% and reduce the trainable language model to only 4% compared with the current BERT encoder.
Recently, vision-language instruct-tuning models have made significant progress due to their more comprehensive understanding of the world. In this work, we discovered that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. We rebalanced the computational loads from data, model, and memory perspectives to address this issue, achieving more balanced computation across devices. These three components are not independent but are closely connected, forming an omniverse balanced training framework. Specifically, for the data, we grouped instances into new balanced mini-batches within and across devices. For the model, we employed a search-based method to achieve a more balanced partitioning. For memory optimization, we adaptively adjusted the re-computation strategy for each partition to utilize the available memory fully. We conducted extensive experiments to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, we significantly reduced GPU days, achieving about 1.8x speed-up. Our method's efficacy and generalizability were further demonstrated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.
The remarkable success of Large Language Models (LLMs) relies heavily on their substantial scale, which poses significant challenges during model deployment in terms of latency and memory consumption. Recently, numerous studies have attempted to compress LLMs using one-shot pruning methods. However, these methods often suffer from considerable performance degradation on complex language understanding tasks, raising concerns about the feasibility of pruning in LLMs. To address this issue, we propose Adaptive Sparse Trainer (AST), a novel and efficient retraining framework tailored for semi-structured sparse models. AST enables models to learn optimal masks during the weight update process without incurring additional computational overhead. Furthermore, we demonstrate that incorporating knowledge distillation significantly improves retraining efficiency and enhances model performance under fixed computational constraints. Additionally, a supplementary set of well-initialized parameters is integrated to further augment the model's efficacy. AST achieves state-of-the-art performance with minimal training cost. When applied to the LLaMA2-7B model, AST reduces the perplexity and zero-shot accuracy gap between dense and 2:4 semi-structured sparse models to 0.6 and 1.16%, respectively, utilizing less than 0.4% of the pretraining tokens and GPU hours. Our work demonstrates the feasibility of deploying semi-structured sparse LLMs and offers a promising alternative for achieving highly compressed models when combined with existing quantization techniques.
This paper presents a new open-source high-fidelity dataset for Machine Learning (ML) containing 355 geometric variants of the Windsor body, to help the development and testing of ML surrogate models for external automotive aerodynamics. Each Computational Fluid Dynamics (CFD) simulation was run with a GPU-native high-fidelity Wall-Modeled Large-Eddy Simulations (WMLES) using a Cartesian immersed-boundary method using more than 280M cells to ensure the greatest possible accuracy. The dataset contains geometry variants that exhibits a wide range of flow characteristics that are representative of those observed on road-cars. The dataset itself contains the 3D time-averaged volume & boundary data as well as the geometry and force & moment coefficients. This paper discusses the validation of the underlying CFD methods as well as contents and structure of the dataset. To the authors knowledge, this represents the first, large-scale high-fidelity CFD dataset for the Windsor body with a permissive open-source license (CC-BY-SA).
Exploring scientific datasets with billions of samples in real-time visualization presents a challenge - balancing high-fidelity rendering with speed. This work introduces a novel renderer - Neural Accelerated Renderer (NAR), that uses the neural deferred rendering framework to visualize large-scale scientific point cloud data. NAR augments a real-time point cloud rendering pipeline with high-quality neural post-processing, making the approach ideal for interactive visualization at scale. Specifically, we train a neural network to learn the point cloud geometry from a high-performance multi-stream rasterizer and capture the desired postprocessing effects from a conventional high-quality renderer. We demonstrate the effectiveness of NAR by visualizing complex multidimensional Lagrangian flow fields and photometric scans of a large terrain and compare the renderings against the state-of-the-art high-quality renderers. Through extensive evaluation, we demonstrate that NAR prioritizes speed and scalability while retaining high visual fidelity. We achieve competitive frame rates of $>$ 126 fps for interactive rendering of $>$ 350M points (i.e., an effective throughput of $>$ 44 billion points per second) using $\sim$12 GB of memory on RTX 2080 Ti GPU. Furthermore, we show that NAR is generalizable across different point clouds with similar visualization needs and the desired post-processing effects could be obtained with substantial high quality even at lower resolutions of the original point cloud, further reducing the memory requirements.
Despite extreme sample inefficiency, on-policy reinforcement learning, aka policy gradients, has become a fundamental tool in decision-making problems. With the recent advances in GPU-driven simulation, the ability to collect large amounts of data for RL training has scaled exponentially. However, we show that current RL methods, e.g. PPO, fail to ingest the benefit of parallelized environments beyond a certain point and their performance saturates. To address this, we propose a new on-policy RL algorithm that can effectively leverage large-scale environments by splitting them into chunks and fusing them back together via importance sampling. Our algorithm, termed SAPG, shows significantly higher performance across a variety of challenging environments where vanilla PPO and other strong baselines fail to achieve high performance. Website at https://sapg-rl.github.io/
Checkpointing to preserve training states is crucial during the development of Large Foundation Models (LFMs), for training resumption upon various failures or changes in GPU resources and parallelism configurations. In addition, saved checkpoints are dispatched to evaluation tasks or transferred across different training stages (e.g., from pre-training to post-training). All these scenarios require resharding distributed checkpoints from one parallelism to another. In production, different LFMs are trained with various frameworks and storage backends, depending on model sizes and training scales. A high-performance checkpointing system is needed to enable efficient checkpoint management at scale. This paper presents ByteCheckpoint, an industrial-grade checkpointing system for large-scale LFM training. ByteCheckpoint employs a parallelism-agnostic checkpoint representation that enables efficient load-time checkpoint resharding. ByteCheckpoint advocates a generic checkpoint saving/loading workflow to accommodate multiple training frameworks and support different storage backends. To ensure high I/O efficiency, we take a full-stack approach to optimize saving/loading plan generation, critical stages of checkpointing pipelines, and irregular tensor processing required by resharding. To guarantee the scalability of ByteCheckpoint in large-scale training, we enhance the storage system to efficiently handle high volumes of checkpointing I/O requests, devise communication optimizations within the checkpointing workflow, and introduce a suite of monitoring tools to analyze performance and detect bottlenecks. Compared to existing open-source checkpointing systems [40, 46], ByteCheckpoint significantly reduces runtime checkpoint stalls, achieving an average reduction of 54.20x. For saving and loading times, ByteCheckpoint achieves improvements of up to 9.96x and 8.80x, respectively.
Nearest neighbor machine translation is a successful approach for fast domain adaption, which interpolates the pre-trained transformers with domain-specific token-level k-nearest-neighbor (kNN) retrieval without retraining. Despite kNN MT's success, searching large reference corpus and fixed interpolation between the kNN and pre-trained model led to computational complexity and translation quality challenges. Among other papers, Dai et al. proposed methods to obtain a small number of reference samples dynamically for which they introduced a distance-aware interpolation method using an equation that includes free parameters. This paper proposes a simply trainable nearest neighbor machine translation and carry out inference experiments on GPU. Similar to Dai et al., we first adaptively construct a small datastore for each input sentence. Second, we train a single-layer network for the interpolation coefficient between the knnMT and pre-trained result to automatically interpolate in different domains. Experimental results on different domains show that our proposed method either improves or sometimes maintain the translation quality of methods in Dai et al. while being automatic. In addition, our GPU inference results demonstrate that knnMT can be integrated into GPUs with a drop of only 5% in terms of speed.
Compared with still image object detection, video object detection (VOD) needs to particularly concern the high across-frame variation in object appearance, and the diverse deterioration in some frames. In principle, the detection in a certain frame of a video can benefit from information in other frames. Thus, how to effectively aggregate features across different frames is key to the target problem. Most of contemporary aggregation methods are tailored for two-stage detectors, suffering from high computational costs due to the dual-stage nature. On the other hand, although one-stage detectors have made continuous progress in handling static images, their applicability to VOD lacks sufficient exploration. To tackle the above issues, this study invents a very simple yet potent strategy of feature selection and aggregation, gaining significant accuracy at marginal computational expense. Concretely, for cutting the massive computation and memory consumption from the dense prediction characteristic of one-stage object detectors, we first condense candidate features from dense prediction maps. Then, the relationship between a target frame and its reference frames is evaluated to guide the aggregation. Comprehensive experiments and ablation studies are conducted to validate the efficacy of our design, and showcase its advantage over other cutting-edge VOD methods in both effectiveness and efficiency. Notably, our model reaches \emph{a new record performance, i.e., 92.9\% AP50 at over 30 FPS on the ImageNet VID dataset on a single 3090 GPU}, making it a compelling option for large-scale or real-time applications. The implementation is simple, and accessible at \url{https://github.com/YuHengsss/YOLOV}.
Training with larger mini-batches improves the convergence rate and can yield superior performance. However, training with large mini-batches becomes prohibitive for Large Language Models (LLMs), due to the large GPU memory requirement. To address this problem, an effective approach is finding small mini-batch coresets that closely match the gradient of larger mini-batches. However, this approach becomes infeasible and ineffective for LLMs, due to the highly imbalanced nature of the sources in language data, use of the Adam optimizer, and the very large gradient dimensionality of LLMs. In this work, we address the above challenges by proposing Coresets for Training LLMs (CoLM). First, we show that mini-batch coresets found by gradient matching do not contain representative examples of the small sources w.h.p., and thus including all examples of the small sources in the mini-batch coresets is crucial for optimal performance. Second, we normalize the gradients by their historical exponential to find mini-batch coresets for training with Adam. Finally, we leverage zeroth-order methods to find smooth gradient of the last V -projection matrix and sparsify it to keep the dimensions with the largest normalized gradient magnitude. We apply CoLM to fine-tuning Phi-2, Phi-3, and Zephyr with LoRA on MathInstruct and SuperGLUE benchmark. Remarkably, CoLM reduces the memory requirement of fine-tuning by 2x and even outperforms training with 4x larger mini-batches. Notably, CoLM easily stack with existing memory-efficient training methods, such as LoRA.
In this study, state-of-the-art unsupervised detection models were evaluated for the purpose of automated anomaly inspection of wool carpets. A custom dataset of four unique types of carpet textures was created to thoroughly test the models and their robustness in detecting subtle anomalies in complex textures. Due to the requirements of an inline inspection system in a manufacturing use case, the metrics of importance in this study were accuracy in detecting anomalous areas, the number of false detections, and the inference times of each model for real-time performance. Of the evaluated models, the student-teacher network based methods were found on average to yield the highest detection accuracy and lowest false detection rates. When trained on a multi-class dataset the models were found to yield comparable if not better results than single-class training. Finally, in terms of detection speed, with exception to the generative model, all other evaluated models were found to have comparable inference times on a GPU, with an average of 0.16s per image. On a CPU, most of these models typically produced results between 1.5 to 2 times the respective GPU inference times.
We propose HYBRIDDEPTH, a robust depth estimation pipeline that addresses key challenges in depth estimation,including scale ambiguity, hardware heterogeneity, and generalizability. HYBRIDDEPTH leverages focal stack, data conveniently accessible in common mobile devices, to produce accurate metric depth maps. By incorporating depth priors afforded by recent advances in singleimage depth estimation, our model achieves a higher level of structural detail compared to existing methods. We test our pipeline as an end-to-end system, with a newly developed mobile client to capture focal stacks, which are then sent to a GPU-powered server for depth estimation. Comprehensive quantitative and qualitative analyses demonstrate that HYBRIDDEPTH outperforms state-of-the-art(SOTA) models on common datasets such as DDFF12 and NYU Depth V2. HYBRIDDEPTH also shows strong zero-shot generalization. When trained on NYU Depth V2, HYBRIDDEPTH surpasses SOTA models in zero-shot performance on ARKitScenes and delivers more structurally accurate depth maps on Mobile Depth. The code is available at https://github.com/cake-lab/HybridDepth/.
High-refresh rate displays have become very popular in recent years due to the need for superior visual quality in gaming, professional displays and specialized applications like medical imaging. However, high-refresh rate displays alone do not guarantee a superior visual experience; the GPU needs to render frames at a matching rate. Otherwise, we observe disconcerting visual artifacts such as screen tearing and stuttering. Temporal supersampling is an effective technique to increase frame rates by predicting new frames from other rendered frames. There are two methods in this space: interpolation and extrapolation. Interpolation-based methods provide good image quality at the cost of a higher latency because they also require the next rendered frame. On the other hand, extrapolation methods are much faster at the cost of quality. This paper introduces PatchEX, a novel frame extrapolation method that aims to provide the quality of interpolation at the speed of extrapolation. It smartly partitions the extrapolation task into sub-tasks and executes them in parallel to improve both quality and latency. It then uses a patch-based inpainting method and a custom shadow prediction approach to fuse the generated sub-frames. This approach significantly reduces the overall latency while maintaining the quality of the output. Our results demonstrate that PatchEX achieves a 65.29% and 48.46% improvement in PSNR over the latest extrapolation methods ExtraNet and ExtraSS, respectively, while being 6x and 2x faster, respectively.
Large Language Models (LLMs), epitomized by ChatGPT's release in late 2022, have revolutionized various industries with their advanced language comprehension. However, their efficiency is challenged by the Transformer architecture's struggle with handling long texts. KV Cache has emerged as a pivotal solution to this issue, converting the time complexity of token generation from quadratic to linear, albeit with increased GPU memory overhead proportional to conversation length. With the development of the LLM community and academia, various KV Cache compression methods have been proposed. In this review, we dissect the various properties of KV Cache and elaborate on various methods currently used to optimize the KV Cache space usage of LLMs. These methods span the pre-training phase, deployment phase, and inference phase, and we summarize the commonalities and differences among these methods. Additionally, we list some metrics for evaluating the long-text capabilities of large language models, from both efficiency and capability perspectives. Our review thus sheds light on the evolving landscape of LLM optimization, offering insights into future advancements in this dynamic field. Links to the papers mentioned in this review can be found in our Github Repo https://github.com/zcli-charlie/Awesome-KV-Cache.