PyTorchKR Papers

최신 논문

총 5,357편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2405.04437공식 코드 · Papers with Code

vAttention: Dynamic Memory Management for Serving LLMs without PagedAttention

PagedAttention is a popular approach for dynamic memory allocation in LLM serving systems. It enables on-demand allocation of GPU memory to mitigate KV cache fragmentation -- a phenomenon that crippled the batch size (and consequently throughput) in prior systems. However, in trying to allocate physical memory at runtime, PagedAttention ends up changing the virtual memory layout of the KV cache from contiguous to non-contiguous. Such a design leads to non-trivial programming and performance overheads. We present vAttention -- an approach that mitigates fragmentation in physical memory while retaining the contiguity of KV cache in virtual memory. We achieve this by decoupling the allocation of virtual and physical memory using CUDA virtual memory management APIs. We also introduce various LLM-specific optimizations to address the limitations of CUDA virtual memory support. Overall, vAttention is a simpler, portable, and performant alternative to PagedAttention: it supports various attention kernels out-of-the-box and improves LLM serving throughput by up to 1.23x compared to the use of PagedAttention-based kernels of FlashAttention and FlashInfer.

arXiv:2405.04296공식 코드 · Papers with Code

Open Implementation and Study of BEST-RQ for Speech Processing

Self-Supervised Learning (SSL) has proven to be useful in various speech tasks. However, these methods are generally very demanding in terms of data, memory, and computational resources. BERT-based Speech pre-Training with Random-projection Quantizer (BEST-RQ), is an SSL method that has shown great performance on Automatic Speech Recognition (ASR) while being simpler than other SSL methods, such as wav2vec 2.0. Despite BEST-RQ's great performance, details are lacking in the original paper, such as the amount of GPU/TPU hours used in pre-training, and there is no official easy-to-use open-source implementation. Furthermore, BEST-RQ has not been evaluated on other downstream tasks aside from ASR and speech translation. In this work, we describe a re-implementation of a Random-projection quantizer and perform a preliminary study with a comparison to wav2vec 2.0 on four downstream tasks. We discuss the details and differences of our implementation. We show that a random projection quantizer can achieve similar downstream performance as wav2vec 2.0 while decreasing training time by over a factor of two.

arXiv:2405.03967공식 코드 · Papers with Code

SwiftRL: Towards Efficient Reinforcement Learning on Real Processing-In-Memory Systems

Reinforcement Learning (RL) trains agents to learn optimal behavior by maximizing reward signals from experience datasets. However, RL training often faces memory limitations, leading to execution latencies and prolonged training times. To overcome this, SwiftRL explores Processing-In-Memory (PIM) architectures to accelerate RL workloads. We achieve near-linear performance scaling by implementing RL algorithms like Tabular Q-learning and SARSA on UPMEM PIM systems and optimizing for hardware. Our experiments on OpenAI GYM environments using UPMEM hardware demonstrate superior performance compared to CPU and GPU implementations.

arXiv:2405.03917

KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization

Efficient deployment of Large Language Models (LLMs) requires batching multiple requests together to improve throughput. As the batch size, context length, or model size increases, the size of the key and value (KV) cache can quickly become the main contributor to GPU memory usage and the bottleneck of inference latency. Quantization has emerged as an effective technique for KV cache compression, but existing methods still fail at very low bit widths. We observe that distinct channels of a key/value activation embedding are highly inter-dependent, and the joint entropy of multiple channels grows at a slower rate than the sum of their marginal entropies. Based on this insight, we propose Coupled Quantization (CQ), which couples multiple key/value channels together to exploit their inter-dependency and encode the activations in a more information-efficient manner. Extensive experiments reveal that CQ outperforms or is competitive with existing baselines in preserving model quality. Furthermore, we demonstrate that CQ can preserve model quality with KV cache quantized down to 1-bit.

arXiv:2405.04274

Group-aware Parameter-efficient Updating for Content-Adaptive Neural Video Compression

Content-adaptive compression is crucial for enhancing the adaptability of the pre-trained neural codec for various contents. Although these methods have been very practical in neural image compression (NIC), their application in neural video compression (NVC) is still limited due to two main aspects: 1), video compression relies heavily on temporal redundancy, therefore updating just one or a few frames can lead to significant errors accumulating over time; 2), NVC frameworks are generally more complex, with many large components that are not easy to update quickly during encoding. To address the previously mentioned challenges, we have developed a content-adaptive NVC technique called Group-aware Parameter-Efficient Updating (GPU). Initially, to minimize error accumulation, we adopt a group-aware approach for updating encoder parameters. This involves adopting a patch-based Group of Pictures (GoP) training strategy to segment a video into patch-based GoPs, which will be updated to facilitate a globally optimized domain-transferable solution. Subsequently, we introduce a parameter-efficient delta-tuning strategy, which is achieved by integrating several light-weight adapters into each coding component of the encoding process by both serial and parallel configuration. Such architecture-agnostic modules stimulate the components with large parameters, thereby reducing both the update cost and the encoding time. We incorporate our GPU into the latest NVC framework and conduct comprehensive experiments, whose results showcase outstanding video compression efficiency across four video benchmarks and adaptability of one medical image benchmark.

arXiv:2405.04416

DistGrid: Scalable Scene Reconstruction with Distributed Multi-resolution Hash Grid

Neural Radiance Field~(NeRF) achieves extremely high quality in object-scaled and indoor scene reconstruction. However, there exist some challenges when reconstructing large-scale scenes. MLP-based NeRFs suffer from limited network capacity, while volume-based NeRFs are heavily memory-consuming when the scene resolution increases. Recent approaches propose to geographically partition the scene and learn each sub-region using an individual NeRF. Such partitioning strategies help volume-based NeRF exceed the single GPU memory limit and scale to larger scenes. However, this approach requires multiple background NeRF to handle out-of-partition rays, which leads to redundancy of learning. Inspired by the fact that the background of current partition is the foreground of adjacent partition, we propose a scalable scene reconstruction method based on joint Multi-resolution Hash Grids, named DistGrid. In this method, the scene is divided into multiple closely-paved yet non-overlapped Axis-Aligned Bounding Boxes, and a novel segmented volume rendering method is proposed to handle cross-boundary rays, thereby eliminating the need for background NeRFs. The experiments demonstrate that our method outperforms existing methods on all evaluated large-scale scenes, and provides visually plausible scene reconstruction. The scalability of our method on reconstruction quality is further evaluated qualitatively and quantitatively.

arXiv:2405.03714

UniDEC : Unified Dual Encoder and Classifier Training for Extreme Multi-Label Classification

Extreme Multi-label Classification (XMC) involves predicting a subset of relevant labels from an extremely large label space, given an input query and labels with textual features. Models developed for this problem have conventionally used modular approach with (i) a Dual Encoder (DE) to embed the queries and label texts, (ii) a One-vs-All classifier to rerank the shortlisted labels mined through meta-classifier training. While such methods have shown empirical success, we observe two key uncharted aspects, (i) DE training typically uses only a single positive relation even for datasets which offer more, (ii) existing approaches fixate on using only OvA reduction of the multi-label problem. This work aims to explore these aspects by proposing UniDEC, a novel end-to-end trainable framework which trains the dual encoder and classifier in together in a unified fashion using a multi-class loss. For the choice of multi-class loss, the work proposes a novel pick-some-label (PSL) reduction of the multi-label problem with leverages multiple (in come cases, all) positives. The proposed framework achieves state-of-the-art results on a single GPU, while achieving on par results with respect to multi-GPU SOTA methods on various XML benchmark datasets, all while using 4-16x lesser compute and being practically scalable even beyond million label scale datasets.

arXiv:2405.04305

A New Dataset and Comparative Study for Aphid Cluster Detection and Segmentation in Sorghum Fields

Aphid infestations are one of the primary causes of extensive damage to wheat and sorghum fields and are one of the most common vectors for plant viruses, resulting in significant agricultural yield losses. To address this problem, farmers often employ the inefficient use of harmful chemical pesticides that have negative health and environmental impacts. As a result, a large amount of pesticide is wasted on areas without significant pest infestation. This brings to attention the urgent need for an intelligent autonomous system that can locate and spray sufficiently large infestations selectively within the complex crop canopies. We have developed a large multi-scale dataset for aphid cluster detection and segmentation, collected from actual sorghum fields and meticulously annotated to include clusters of aphids. Our dataset comprises a total of 54,742 image patches, showcasing a variety of viewpoints, diverse lighting conditions, and multiple scales, highlighting its effectiveness for real-world applications. In this study, we trained and evaluated four real-time semantic segmentation models and three object detection models specifically for aphid cluster segmentation and detection. Considering the balance between accuracy and efficiency, Fast-SCNN delivered the most effective segmentation results, achieving 80.46% mean precision, 81.21% mean recall, and 91.66 frames per second (FPS). For object detection, RT-DETR exhibited the best overall performance with a 61.63% mean average precision (mAP), 92.6% mean recall, and 72.55 on an NVIDIA V100 GPU. Our experiments further indicate that aphid cluster segmentation is more suitable for assessing aphid infestations than using detection models.

arXiv:2405.04532공식 코드 · Papers with Code

QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving

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/omniserve.

arXiv:2405.02323

CNN-Based Equalization for Communications: Achieving Gigabit Throughput with a Flexible FPGA Hardware Architecture

To satisfy the growing throughput demand of data-intensive applications, the performance of optical communication systems increased dramatically in recent years. With higher throughput, more advanced equalizers are crucial, to compensate for impairments caused by inter-symbol interference (ISI). The latest research shows that artificial neural network (ANN)-based equalizers are promising candidates to replace traditional algorithms for high-throughput communications. On the other hand, not only throughput but also flexibility is a main objective of beyond-5G and 6G communication systems. A platform that is able to satisfy the strict throughput and flexibility requirements of modern communication systems are field programmable gate arrays (FPGAs). Thus, in this work, we present a high-performance FPGA implementation of an ANN-based equalizer, which meets the throughput requirements of modern optical communication systems. Further, our architecture is highly flexible since it includes a variable degree of parallelism (DOP) and therefore can also be applied to low-cost or low-power applications which is demonstrated for a magnetic recording channel. The implementation is based on a cross-layer design approach featuring optimizations from the algorithm down to the hardware architecture, including a detailed quantization analysis. Moreover, we present a framework to reduce the latency of the ANN-based equalizer under given throughput constraints. As a result, the bit error ratio (BER) of our equalizer for the optical fiber channel is around four times lower than that of a conventional one, while the corresponding FPGA implementation achieves a throughput of more than 40 GBd, outperforming a high-performance graphics processing unit (GPU) by three orders of magnitude for a similar batch size.

arXiv:2405.02817공식 코드 · Papers with Code

Labeling supervised fine-tuning data with the scaling law

This paper introduces a multi-stage manual annotation calibrated by the scaling law, offering a high-quality Supervised Fine-Tuning data acquisition method for environments with constrained resources like GPU poor, limited GPT access, and funding restrictions. We have preprocessed 58k authentic chat data and manually annotated 2.3k questions. After this, we conducted fine-tuning on Qwen models, ranging from 0.5B to 32B parameters. The optimal version improved 29.07 in F1 score. This confirms the viability of fine-tuning Large Language Model (LLM) for downstream Natural Language Processing (NLP) tasks. Our contributions are: 1) Created Supervised Fine-Tuning (SFT) training data in alpaca format, along with a set of Low-Rank Adaptation (LoRA) weights, and 2) Developed a method for acquiring high-quality data leveraging scaling law principle. The script, raw data with alpaca format and experiments track are open-sourced on Github (https://github.com/InternLM/HuixiangDou/tree/main/web/tools), HuggingFace (https://huggingface.co/tpoisonooo) and WandB (https://wandb.ai/tpoisonooo/huixiangdou-cr/table?nw=nwusertpoisonooo). The privacy of the data involved has been authorized by users. SFT data and license comes from ncnn contributors group.

arXiv:2405.03192

QuadraNet V2: Efficient and Sustainable Training of High-Order Neural Networks with Quadratic Adaptation

Machine learning is evolving towards high-order models that necessitate pre-training on extensive datasets, a process associated with significant overheads. Traditional models, despite having pre-trained weights, are becoming obsolete due to architectural differences that obstruct the effective transfer and initialization of these weights. To address these challenges, we introduce a novel framework, QuadraNet V2, which leverages quadratic neural networks to create efficient and sustainable high-order learning models. Our method initializes the primary term of the quadratic neuron using a standard neural network, while the quadratic term is employed to adaptively enhance the learning of data non-linearity or shifts. This integration of pre-trained primary terms with quadratic terms, which possess advanced modeling capabilities, significantly augments the information characterization capacity of the high-order network. By utilizing existing pre-trained weights, QuadraNet V2 reduces the required GPU hours for training by 90\% to 98.4\% compared to training from scratch, demonstrating both efficiency and effectiveness.

arXiv:2405.02774

Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs

This work focuses on leveraging and selecting from vast, unlabeled, open data to pre-fine-tune a pre-trained language model. The goal is to minimize the need for costly domain-specific data for subsequent fine-tuning while achieving desired performance levels. While many data selection algorithms have been designed for small-scale applications, rendering them unsuitable for our context, some emerging methods do cater to language data scales. However, they often prioritize data that aligns with the target distribution. While this strategy may be effective when training a model from scratch, it can yield limited results when the model has already been pre-trained on a different distribution. Differing from prior work, our key idea is to select data that nudges the pre-training distribution closer to the target distribution. We show the optimality of this approach for fine-tuning tasks under certain conditions. We demonstrate the efficacy of our methodology across a diverse array of tasks (NLU, NLG, zero-shot) with models up to 2.7B, showing that it consistently surpasses other selection methods. Moreover, our proposed method is significantly faster than existing techniques, scaling to millions of samples within a single GPU hour. Our code is open-sourced (Code repository: https://anonymous.4open.science/r/DV4LLM-D761/ ). While fine-tuning offers significant potential for enhancing performance across diverse tasks, its associated costs often limit its widespread adoption; with this work, we hope to lay the groundwork for cost-effective fine-tuning, making its benefits more accessible.

arXiv:2405.02019공식 코드 · Papers with Code

Fast Algorithms for Spiking Neural Network Simulation with FPGAs

Using OpenCL-based high-level synthesis, we create a number of spiking neural network (SNN) simulators for the Potjans-Diesmann cortical microcircuit for a high-end Field-Programmable Gate Array (FPGA). Our best simulators simulate the circuit 25\% faster than real-time, require less than 21 nJ per synaptic event, and are bottle-necked by the device's on-chip memory. Speed-wise they compare favorably to the state-of-the-art GPU-based simulators and their energy usage is lower than any other published result. This result is the first for simulating the circuit on a single hardware accelerator. We also extensively analyze the techniques and algorithms we implement our simulators with, many of which can be realized on other types of hardware. Thus, this article is of interest to any researcher or practitioner interested in efficient SNN simulation, whether they target FPGAs or not.

arXiv:2405.01827

SoftMCL: Soft Momentum Contrastive Learning for Fine-grained Sentiment-aware Pre-training

The pre-training for language models captures general language understanding but fails to distinguish the affective impact of a particular context to a specific word. Recent works have sought to introduce contrastive learning (CL) for sentiment-aware pre-training in acquiring affective information. Nevertheless, these methods present two significant limitations. First, the compatibility of the GPU memory often limits the number of negative samples, hindering the opportunities to learn good representations. In addition, using only a few sentiment polarities as hard labels, e.g., positive, neutral, and negative, to supervise CL will force all representations to converge to a few points, leading to the issue of latent space collapse. This study proposes a soft momentum contrastive learning (SoftMCL) for fine-grained sentiment-aware pre-training. Instead of hard labels, we introduce valence ratings as soft-label supervision for CL to fine-grained measure the sentiment similarities between samples. The proposed SoftMCL is conducted on both the word- and sentence-level to enhance the model's ability to learn affective information. A momentum queue was introduced to expand the contrastive samples, allowing storing and involving more negatives to overcome the limitations of hardware platforms. Extensive experiments were conducted on four different sentiment-related tasks, which demonstrates the effectiveness of the proposed SoftMCL method. The code and data of the proposed SoftMCL is available at: https://www.github.com/wangjin0818/SoftMCL/.

arXiv:2405.02267공식 코드 · Papers with Code

Structural Pruning of Pre-trained Language Models via Neural Architecture Search

Pre-trained language models (PLM), for example BERT or RoBERTa, mark the state-of-the-art for natural language understanding task when fine-tuned on labeled data. However, their large size poses challenges in deploying them for inference in real-world applications, due to significant GPU memory requirements and high inference latency. This paper explores neural architecture search (NAS) for structural pruning to find sub-parts of the fine-tuned network that optimally trade-off efficiency, for example in terms of model size or latency, and generalization performance. We also show how we can utilize more recently developed two-stage weight-sharing NAS approaches in this setting to accelerate the search process. Unlike traditional pruning methods with fixed thresholds, we propose to adopt a multi-objective approach that identifies the Pareto optimal set of sub-networks, allowing for a more flexible and automated compression process.

arXiv:2405.00738공식 코드 · Papers with Code

HLSTransform: Energy-Efficient Llama 2 Inference on FPGAs Via High Level Synthesis

Graphics Processing Units (GPUs) have become the leading hardware accelerator for deep learning applications and are used widely in training and inference of transformers; transformers have achieved state-of-the-art performance in many areas of machine learning and are especially used in most modern Large Language Models (LLMs). However, GPUs require large amounts of energy, which poses environmental concerns, demands high operational costs, and causes GPUs to be unsuitable for edge computing. We develop an accelerator for transformers, namely, Llama 2, an open-source state-of-the-art LLM, using high level synthesis (HLS) on Field Programmable Gate Arrays (FPGAs). HLS allows us to rapidly prototype FPGA designs without writing code at the register-transfer level (RTL). We name our method HLSTransform, and the FPGA designs we synthesize with HLS achieve up to a 12.75x reduction and 8.25x reduction in energy used per token on the Xilinx Virtex UltraScale+ VU9P FPGA compared to an Intel Xeon Broadwell E5-2686 v4 CPU and NVIDIA RTX 3090 GPU respectively, while increasing inference speeds by up to 2.46x compared to CPU and maintaining 0.53x the speed of an RTX 3090 GPU despite the GPU's 4 times higher base clock rate. With the lack of existing open-source FPGA accelerators for transformers, we open-source our code and document our steps for synthesis. We hope this work will serve as a step in democratizing the use of FPGAs in transformer inference and inspire research into energy-efficient inference methods as a whole. The code can be found on https://github.com/HLSTransform/submission.

arXiv:2405.01491공식 코드 · Papers with Code

FeNNol: an Efficient and Flexible Library for Building Force-field-enhanced Neural Network Potentials

Neural network interatomic potentials (NNPs) have recently proven to be powerful tools to accurately model complex molecular systems while bypassing the high numerical cost of ab-initio molecular dynamics simulations. In recent years, numerous advances in model architectures as well as the development of hybrid models combining machine-learning (ML) with more traditional, physically-motivated, force-field interactions have considerably increased the design space of ML potentials. In this paper, we present FeNNol, a new library for building, training and running force-field-enhanced neural network potentials. It provides a flexible and modular system for building hybrid models, allowing to easily combine state-of-the-art embeddings with ML-parameterized physical interaction terms without the need for explicit programming. Furthermore, FeNNol leverages the automatic differentiation and just-in-time compilation features of the Jax Python library to enable fast evaluation of NNPs, shrinking the performance gap between ML potentials and standard force-fields. This is demonstrated with the popular ANI-2x model reaching simulation speeds nearly on par with the AMOEBA polarizable force-field on commodity GPUs (GPU=Graphics processing unit). We hope that FeNNol will facilitate the development and application of new hybrid NNP architectures for a wide range of molecular simulation problems.

arXiv:2405.00922

MTDT: A Multi-Task Deep Learning Digital Twin

Traffic congestion has significant impacts on both the economy and the environment. Measures of Effectiveness (MOEs) have long been the standard for evaluating traffic intersections' level of service and operational efficiency. However, the scarcity of traditional high-resolution loop detector data (ATSPM) presents challenges in accurately measuring MOEs or capturing the intricate spatiotemporal characteristics inherent in urban intersection traffic. To address this challenge, we present a comprehensive intersection traffic flow simulation that utilizes a multi-task learning paradigm. This approach combines graph convolutions for primary estimating lane-wise exit and inflow with time series convolutions for secondary assessing multi-directional queue lengths and travel time distribution through any arbitrary urban traffic intersection. Compared to existing deep learning methodologies, the proposed Multi-Task Deep Learning Digital Twin (MTDT) distinguishes itself through its adaptability to local temporal and spatial features, such as signal timing plans, intersection topology, driving behaviors, and turning movement counts. We also show the benefit of multi-task learning in the effectiveness of individual traffic simulation tasks. Furthermore, our approach facilitates sequential computation and provides complete parallelization through GPU implementation. This not only streamlines the computational process but also enhances scalability and performance.

arXiv:2405.01156

Self-Supervised Learning for Interventional Image Analytics: Towards Robust Device Trackers

An accurate detection and tracking of devices such as guiding catheters in live X-ray image acquisitions is an essential prerequisite for endovascular cardiac interventions. This information is leveraged for procedural guidance, e.g., directing stent placements. To ensure procedural safety and efficacy, there is a need for high robustness no failures during tracking. To achieve that, one needs to efficiently tackle challenges, such as: device obscuration by contrast agent or other external devices or wires, changes in field-of-view or acquisition angle, as well as the continuous movement due to cardiac and respiratory motion. To overcome the aforementioned challenges, we propose a novel approach to learn spatio-temporal features from a very large data cohort of over 16 million interventional X-ray frames using self-supervision for image sequence data. Our approach is based on a masked image modeling technique that leverages frame interpolation based reconstruction to learn fine inter-frame temporal correspondences. The features encoded in the resulting model are fine-tuned downstream. Our approach achieves state-of-the-art performance and in particular robustness compared to ultra optimized reference solutions (that use multi-stage feature fusion, multi-task and flow regularization). The experiments show that our method achieves 66.31% reduction in maximum tracking error against reference solutions (23.20% when flow regularization is used); achieving a success score of 97.95% at a 3x faster inference speed of 42 frames-per-second (on GPU). The results encourage the use of our approach in various other tasks within interventional image analytics that require effective understanding of spatio-temporal semantics.