PyTorchKR Papers

최신 논문

총 5,357편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2407.03168공식 코드 · Papers with Code

LivePortrait: Efficient Portrait Animation with Stitching and Retargeting Control

Portrait Animation aims to synthesize a lifelike video from a single source image, using it as an appearance reference, with motion (i.e., facial expressions and head pose) derived from a driving video, audio, text, or generation. Instead of following mainstream diffusion-based methods, we explore and extend the potential of the implicit-keypoint-based framework, which effectively balances computational efficiency and controllability. Building upon this, we develop a video-driven portrait animation framework named LivePortrait with a focus on better generalization, controllability, and efficiency for practical usage. To enhance the generation quality and generalization ability, we scale up the training data to about 69 million high-quality frames, adopt a mixed image-video training strategy, upgrade the network architecture, and design better motion transformation and optimization objectives. Additionally, we discover that compact implicit keypoints can effectively represent a kind of blendshapes and meticulously propose a stitching and two retargeting modules, which utilize a small MLP with negligible computational overhead, to enhance the controllability. Experimental results demonstrate the efficacy of our framework even compared to diffusion-based methods. The generation speed remarkably reaches 12.8ms on an RTX 4090 GPU with PyTorch. The inference code and models are available at https://github.com/KwaiVGI/LivePortrait

arXiv:2407.00603

Hierarchical Memory for Long Video QA

This paper describes our champion solution to the LOVEU Challenge @ CVPR'24, Track 1 (Long Video VQA). Processing long sequences of visual tokens is computationally expensive and memory-intensive, making long video question-answering a challenging task. The key is to compress visual tokens effectively, reducing memory footprint and decoding latency, while preserving the essential information for accurate question-answering. We adopt a hierarchical memory mechanism named STAR Memory, proposed in Flash-VStream, that is capable of processing long videos with limited GPU memory (VRAM). We further utilize the video and audio data of MovieChat-1K training set to fine-tune the pretrained weight released by Flash-VStream, achieving 1st place in the challenge. Code is available at project homepage https://invinciblewyq.github.io/vstream-page .

arXiv:2407.00599공식 코드 · Papers with Code

Parm: Efficient Training of Large Sparsely-Activated Models with Dedicated Schedules

Sparsely-activated Mixture-of-Expert (MoE) layers have found practical applications in enlarging the model size of large-scale foundation models, with only a sub-linear increase in computation demands. Despite the wide adoption of hybrid parallel paradigms like model parallelism, expert parallelism, and expert-sharding parallelism (i.e., MP+EP+ESP) to support MoE model training on GPU clusters, the training efficiency is hindered by communication costs introduced by these parallel paradigms. To address this limitation, we propose Parm, a system that accelerates MP+EP+ESP training by designing two dedicated schedules for placing communication tasks. The proposed schedules eliminate redundant computations and communications and enable overlaps between intra-node and inter-node communications, ultimately reducing the overall training time. As the two schedules are not mutually exclusive, we provide comprehensive theoretical analyses and derive an automatic and accurate solution to determine which schedule should be applied in different scenarios. Experimental results on an 8-GPU server and a 32-GPU cluster demonstrate that Parm outperforms the state-of-the-art MoE training system, DeepSpeed-MoE, achieving 1.13$\times$ to 5.77$\times$ speedup on 1296 manually configured MoE layers and approximately 3$\times$ improvement on two real-world MoE models based on BERT and GPT-2.

arXiv:2407.00945공식 코드 · Papers with Code

Efficient Expert Pruning for Sparse Mixture-of-Experts Language Models: Enhancing Performance and Reducing Inference Costs

The rapid advancement of large language models (LLMs) has led to architectures with billions to trillions of parameters, posing significant deployment challenges due to their substantial demands on memory, processing power, and energy consumption. Sparse Mixture-of-Experts (SMoE) architectures have emerged as a solution, activating only a subset of parameters per token, thereby achieving faster inference while maintaining performance. However, SMoE models still face limitations in broader deployment due to their large parameter counts and significant GPU memory requirements. In this work, we introduce a gradient-free evolutionary strategy named EEP (Efficient Expert P}runing) to enhance the pruning of experts in SMoE models. EEP relies solely on model inference (i.e., no gradient computation) and achieves greater sparsity while maintaining or even improving performance on downstream tasks. EEP can be used to reduce both the total number of experts (thus saving GPU memory) and the number of active experts (thus accelerating inference). For example, we demonstrate that pruning up to 75% of experts in Mixtral $8\times7$B-Instruct results in a substantial reduction in parameters with minimal performance loss. Remarkably, we observe improved performance on certain tasks, such as a significant increase in accuracy on the SQuAD dataset (from 53.4% to 75.4%), when pruning half of the experts. With these results, EEP not only lowers the barrier to deploying SMoE models,but also challenges the conventional understanding of model pruning by showing that fewer experts can lead to better task-specific performance without any fine-tuning. Code is available at https://github.com/imagination-research/EEP.

arXiv:2407.02327공식 코드 · Papers with Code

QSync: Quantization-Minimized Synchronous Distributed Training Across Hybrid Devices

A number of production deep learning clusters have attempted to explore inference hardware for DNN training, at the off-peak serving hours with many inference GPUs idling. Conducting DNN training with a combination of heterogeneous training and inference GPUs, known as hybrid device training, presents considerable challenges due to disparities in compute capability and significant differences in memory capacity. We propose QSync, a training system that enables efficient synchronous data-parallel DNN training over hybrid devices by strategically exploiting quantized operators. According to each device's available resource capacity, QSync selects a quantization-minimized setting for operators in the distributed DNN training graph, minimizing model accuracy degradation but keeping the training efficiency brought by quantization. We carefully design a predictor with a bi-directional mixed-precision indicator to reflect the sensitivity of DNN layers on fixed-point and floating-point low-precision operators, a replayer with a neighborhood-aware cost mapper to accurately estimate the latency of distributed hybrid mixed-precision training, and then an allocator that efficiently synchronizes workers with minimized model accuracy degradation. QSync bridges the computational graph on PyTorch to an optimized backend for quantization kernel performance and flexible support for various GPU architectures. Extensive experiments show that QSync's predictor can accurately simulate distributed mixed-precision training with <5% error, with a consistent 0.27-1.03% accuracy improvement over the from-scratch training tasks compared to uniform precision.

arXiv:2407.01376

Badllama 3: removing safety finetuning from Llama 3 in minutes

We show that extensive LLM safety fine-tuning is easily subverted when an attacker has access to model weights. We evaluate three state-of-the-art fine-tuning methods-QLoRA, ReFT, and Ortho-and show how algorithmic advances enable constant jailbreaking performance with cuts in FLOPs and optimisation power. We strip safety fine-tuning from Llama 3 8B in one minute and Llama 3 70B in 30 minutes on a single GPU, and sketch ways to reduce this further.

arXiv:2407.01781

fVDB: A Deep-Learning Framework for Sparse, Large-Scale, and High-Performance Spatial Intelligence

We present fVDB, a novel GPU-optimized framework for deep learning on large-scale 3D data. fVDB provides a complete set of differentiable primitives to build deep learning architectures for common tasks in 3D learning such as convolution, pooling, attention, ray-tracing, meshing, etc. fVDB simultaneously provides a much larger feature set (primitives and operators) than established frameworks with no loss in efficiency: our operators match or exceed the performance of other frameworks with narrower scope. Furthermore, fVDB can process datasets with much larger footprint and spatial resolution than prior works, while providing a competitive memory footprint on small inputs. To achieve this combination of versatility and performance, fVDB relies on a single novel VDB index grid acceleration structure paired with several key innovations including GPU accelerated sparse grid construction, convolution using tensorcores, fast ray tracing kernels using a Hierarchical Digital Differential Analyzer algorithm (HDDA), and jagged tensors. Our framework is fully integrated with PyTorch enabling interoperability with existing pipelines, and we demonstrate its effectiveness on a number of representative tasks such as large-scale point-cloud segmentation, high resolution 3D generative modeling, unbounded scale Neural Radiance Fields, and large-scale point cloud reconstruction.

arXiv:2407.01131

M$^2$IST: Multi-Modal Interactive Side-Tuning for Efficient Referring Expression Comprehension

Referring expression comprehension (REC) is a vision-language task to locate a target object in an image based on a language expression. Fully fine-tuning general-purpose pre-trained vision-language foundation models for REC yields impressive performance but becomes increasingly costly. Parameter-efficient transfer learning (PETL) methods have shown strong performance with fewer tunable parameters. However, directly applying PETL to REC faces two challenges: (1) insufficient multi-modal interaction between pre-trained vision-language foundation models, and (2) high GPU memory usage due to gradients passing through the heavy vision-language foundation models. To this end, we present M$^2$IST: Multi-Modal Interactive Side-Tuning with M$^3$ISAs: Mixture of Multi-Modal Interactive Side-Adapters. During fine-tuning, we keep the pre-trained uni-modal encoders fixed, updating M$^3$ISAs on side networks to progressively connect them, enabling more comprehensive vision-language alignment and efficient tuning for REC. Empirical results reveal that M$^2$IST achieves an optimal balance between performance and efficiency compared to most full fine-tuning and other PETL methods. With M$^2$IST, standard transformer-based REC methods present competitive or even superior performance compared to full fine-tuning, while utilizing only 2.11\% of the tunable parameters, 39.61\% of the GPU memory, and 63.46\% of the fine-tuning time required for full fine-tuning.

arXiv:2407.00949

SpectralKAN: Kolmogorov-Arnold Network for Hyperspectral Images Change Detection

It has been verified that deep learning methods, including convolutional neural networks (CNNs), graph neural networks (GNNs), and transformers, can accurately extract features from hyperspectral images (HSIs). These algorithms perform exceptionally well on HSIs change detection (HSIs-CD). However, the downside of these impressive results is the enormous number of parameters, FLOPs, GPU memory, training and test times required. In this paper, we propose an spectral Kolmogorov-Arnold Network for HSIs-CD (SpectralKAN). SpectralKAN represent a multivariate continuous function with a composition of activation functions to extract HSIs feature and classification. These activation functions are b-spline functions with different parameters that can simulate various functions. In SpectralKAN, a KAN encoder is proposed to enhance computational efficiency for HSIs. And a spatial-spectral KAN encoder is introduced, where the spatial KAN encoder extracts spatial features and compresses the spatial dimensions from patch size to one. The spectral KAN encoder then extracts spectral features and classifies them into changed and unchanged categories. We use five HSIs-CD datasets to verify the effectiveness of SpectralKAN. Experimental verification has shown that SpectralKAN maintains high HSIs-CD accuracy while requiring fewer parameters, FLOPs, GPU memory, training and testing times, thereby increasing the efficiency of HSIs-CD. The code will be available at https://github.com/yanhengwang-heu/SpectralKAN.

arXiv:2407.01638

LASSI: An LLM-based Automated Self-Correcting Pipeline for Translating Parallel Scientific Codes

This paper addresses the problem of providing a novel approach to sourcing significant training data for LLMs focused on science and engineering. In particular, a crucial challenge is sourcing parallel scientific codes in the ranges of millions to billions of codes. To tackle this problem, we propose an automated pipeline framework called LASSI, designed to translate between parallel programming languages by bootstrapping existing closed- or open-source LLMs. LASSI incorporates autonomous enhancement through self-correcting loops where errors encountered during the compilation and execution of generated code are fed back to the LLM through guided prompting for debugging and refactoring. We highlight the bi-directional translation of existing GPU benchmarks between OpenMP target offload and CUDA to validate LASSI. The results of evaluating LASSI with different application codes across four LLMs demonstrate the effectiveness of LASSI for generating executable parallel codes, with 80% of OpenMP to CUDA translations and 85% of CUDA to OpenMP translations producing the expected output. We also observe approximately 78% of OpenMP to CUDA translations and 62% of CUDA to OpenMP translations execute within 10% of or at a faster runtime than the original benchmark code in the same language.

arXiv:2407.01437

Needle in the Haystack for Memory Based Large Language Models

Current large language models (LLMs) often perform poorly on simple fact retrieval tasks. Here we investigate if coupling a dynamically adaptable external memory to a LLM can alleviate this problem. For this purpose, we test Larimar, a recently proposed language model architecture which uses an external associative memory, on long-context recall tasks including passkey and needle-in-the-haystack tests. We demonstrate that the external memory of Larimar, which allows fast write and read of an episode of text samples, can be used at test time to handle contexts much longer than those seen during training. We further show that the latent readouts from the memory (to which long contexts are written) control the decoder towards generating correct outputs, with the memory stored off of the GPU. Compared to existing transformer-based LLM architectures for long-context recall tasks that use larger parameter counts or modified attention mechanisms, a relatively smaller size Larimar is able to maintain strong performance without any task-specific training or training on longer contexts.

arXiv:2407.02109공식 코드 · Papers with Code

HRSAM: Efficient Interactive Segmentation in High-Resolution Images

The Segment Anything Model (SAM) has advanced interactive segmentation but is limited by the high computational cost on high-resolution images. This requires downsampling to meet GPU constraints, sacrificing the fine-grained details needed for high-precision interactive segmentation. To address SAM's limitations, we focus on visual length extrapolation and propose a lightweight model named HRSAM. The extrapolation enables HRSAM trained on low resolutions to generalize to high resolutions. We begin by finding the link between the extrapolation and attention scores, which leads us to base HRSAM on Swin attention. We then introduce the Flexible Local Attention (FLA) framework, using CUDA-optimized Efficient Memory Attention to accelerate HRSAM. Within FLA, we implement Flash Swin attention, achieving over a 35% speedup compared to traditional Swin attention, and propose a KV-only padding mechanism to enhance extrapolation. We also develop the Cycle-scan module that uses State Space models to efficiently expand HRSAM's receptive field. We further develop the HRSAM++ within FLA by adding an anchor map, providing multi-scale data augmentation for the extrapolation and a larger receptive field at slight computational cost. Experiments show that, under standard training, HRSAMs surpass the previous SOTA with only 38% of the latency. With SAM-distillation, the extrapolation enables HRSAMs to outperform the teacher model at lower latency. Further finetuning achieves performance significantly exceeding the previous SOTA.

arXiv:2407.00079공식 코드 · Papers with Code

Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. It features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache. The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs). Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake's innovative architecture enables Kimi to handle 75% more requests.

arXiv:2407.02490공식 코드 · Papers with Code

MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention

The computational challenges of Large Language Model (LLM) inference remain a significant barrier to their widespread deployment, especially as prompt lengths continue to increase. Due to the quadratic complexity of the attention computation, it takes 30 minutes for an 8B LLM to process a prompt of 1M tokens (i.e., the pre-filling stage) on a single A100 GPU. Existing methods for speeding up prefilling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs. To address this gap, we introduce MInference (Milliontokens Inference), a sparse calculation method designed to accelerate pre-filling of long-sequence processing. Specifically, we identify three unique patterns in long-context attention matrices-the A-shape, Vertical-Slash, and Block-Sparsethat can be leveraged for efficient sparse computation on GPUs. We determine the optimal pattern for each attention head offline and dynamically build sparse indices based on the assigned pattern during inference. With the pattern and sparse indices, we perform efficient sparse attention calculations via our optimized GPU kernels to significantly reduce the latency in the pre-filling stage of long-context LLMs. Our proposed technique can be directly applied to existing LLMs without any modifications to the pre-training setup or additional fine-tuning. By evaluating on a wide range of downstream tasks, including InfiniteBench, RULER, PG-19, and Needle In A Haystack, and models including LLaMA-3-1M, GLM4-1M, Yi-200K, Phi-3-128K, and Qwen2-128K, we demonstrate that MInference effectively reduces inference latency by up to 10x for pre-filling on an A100, while maintaining accuracy. Our code is available at https://aka.ms/MInference.

arXiv:2407.02159공식 코드 · Papers with Code

SparseSSP: 3D Subcellular Structure Prediction from Sparse-View Transmitted Light Images

Traditional fluorescence staining is phototoxic to live cells, slow, and expensive; thus, the subcellular structure prediction (SSP) from transmitted light (TL) images is emerging as a label-free, faster, low-cost alternative. However, existing approaches utilize 3D networks for one-to-one voxel level dense prediction, which necessitates a frequent and time-consuming Z-axis imaging process. Moreover, 3D convolutions inevitably lead to significant computation and GPU memory overhead. Therefore, we propose an efficient framework, SparseSSP, predicting fluorescent intensities within the target voxel grid in an efficient paradigm instead of relying entirely on 3D topologies. In particular, SparseSSP makes two pivotal improvements to prior works. First, SparseSSP introduces a one-to-many voxel mapping paradigm, which permits the sparse TL slices to reconstruct the subcellular structure. Secondly, we propose a hybrid dimensions topology, which folds the Z-axis information into channel features, enabling the 2D network layers to tackle SSP under low computational cost. We conduct extensive experiments to validate the effectiveness and advantages of SparseSSP on diverse sparse imaging ratios, and our approach achieves a leading performance compared to pure 3D topologies. SparseSSP reduces imaging frequencies compared to previous dense-view SSP (i.e., the number of imaging is reduced up to 87.5% at most), which is significant in visualizing rapid biological dynamics on low-cost devices and samples.

arXiv:2407.01873

Automated Text Scoring in the Age of Generative AI for the GPU-poor

Current research on generative language models (GLMs) for automated text scoring (ATS) has focused almost exclusively on querying proprietary models via Application Programming Interfaces (APIs). Yet such practices raise issues around transparency and security, and these methods offer little in the way of efficiency or customizability. With the recent proliferation of smaller, open-source models, there is the option to explore GLMs with computers equipped with modest, consumer-grade hardware, that is, for the "GPU poor." In this study, we analyze the performance and efficiency of open-source, small-scale GLMs for ATS. Results show that GLMs can be fine-tuned to achieve adequate, though not state-of-the-art, performance. In addition to ATS, we take small steps towards analyzing models' capacity for generating feedback by prompting GLMs to explain their scores. Model-generated feedback shows promise, but requires more rigorous evaluation focused on targeted use cases.

arXiv:2406.19657공식 코드 · Papers with Code

LLMEasyQuant: Scalable Quantization for Parallel and Distributed LLM Inference

As large language models (LLMs) grow in size and deployment scale, quantization has become an essential technique for reducing memory footprint and improving inference efficiency. However, existing quantization toolkits often lack transparency, flexibility, and system-level scalability across GPUs and distributed environments. We present \textbf{LLMEasyQuant}, a modular, system-aware quantization framework designed for efficient, low-bit inference of LLMs on single-node multi-GPU, multi-node, and edge hardware. LLMEasyQuant supports a wide range of quantization methods -- including Symmetric Quantization, ZeroQuant, SmoothQuant, and SimQuant -- with unified interfaces for per-layer calibration, bitwidth assignment, and runtime adaptation. It integrates fused CUDA kernels with NCCL-based distributed synchronization and supports both static and online quantization. Empirical results show that LLMEasyQuant can achieve substantial speedups in GEMM execution, HBM load time, and near-linear multi-GPU scaling. Ablation studies further validate its ability to balance latency, memory, and accuracy under diverse deployment conditions. LLMEasyQuant offers a practical quantization serving system for scalable, hardware-optimized LLM inference.

arXiv:2406.20037공식 코드 · Papers with Code

Explore as a Storm, Exploit as a Raindrop: On the Benefit of Fine-Tuning Kernel Schedulers with Coordinate Descent

Machine-learning models consist of kernels, which are algorithms applying operations on tensors -- data indexed by a linear combination of natural numbers. Examples of kernels include convolutions, transpositions, and vectorial products. There are many ways to implement a kernel. These implementations form the kernel's optimization space. Kernel scheduling is the problem of finding the best implementation, given an objective function -- typically execution speed. Kernel optimizers such as Ansor, Halide, and AutoTVM solve this problem via search heuristics, which combine two phases: exploration and exploitation. The first step evaluates many different kernel optimization spaces. The latter tries to improve the best implementations by investigating a kernel within the same space. For example, Ansor combines kernel generation through sketches for exploration and leverages an evolutionary algorithm to exploit the best sketches. In this work, we demonstrate the potential to reduce Ansor's search time while enhancing kernel quality by incorporating Droplet Search, an AutoTVM algorithm, into Ansor's exploration phase. The approach involves limiting the number of samples explored by Ansor, selecting the best, and exploiting it with a coordinate descent algorithm. By applying this approach to the first 300 kernels that Ansor generates, we usually obtain better kernels in less time than if we let Ansor analyze 10,000 kernels. This result has been replicated in 20 well-known deep-learning models (AlexNet, ResNet, VGG, DenseNet, etc.) running on four architectures: an AMD Ryzen 7 (x86), an NVIDIA A100 tensor core, an NVIDIA RTX 3080 GPU, and an ARM A64FX. A patch with this combined approach was approved in Ansor in February 2024. As evidence of the generality of this search methodology, a similar patch, achieving equally good results, was submitted to TVM's MetaSchedule in June 2024.

arXiv:2404.00234

Grid Diffusion Models for Text-to-Video Generation

Recent advances in the diffusion models have significantly improved text-to-image generation. However, generating videos from text is a more challenging task than generating images from text, due to the much larger dataset and higher computational cost required. Most existing video generation methods use either a 3D U-Net architecture that considers the temporal dimension or autoregressive generation. These methods require large datasets and are limited in terms of computational costs compared to text-to-image generation. To tackle these challenges, we propose a simple but effective novel grid diffusion for text-to-video generation without temporal dimension in architecture and a large text-video paired dataset. We can generate a high-quality video using a fixed amount of GPU memory regardless of the number of frames by representing the video as a grid image. Additionally, since our method reduces the dimensions of the video to the dimensions of the image, various image-based methods can be applied to videos, such as text-guided video manipulation from image manipulation. Our proposed method outperforms the existing methods in both quantitative and qualitative evaluations, demonstrating the suitability of our model for real-world video generation.

arXiv:2406.18790

MUMU: Bootstrapping Multimodal Image Generation from Text-to-Image Data

We train a model to generate images from multimodal prompts of interleaved text and images such as "a <picture of a man> man and his <picture of a dog> dog in an <picture of a cartoon> animated style." We bootstrap a multimodal dataset by extracting semantically meaningful image crops corresponding to words in the image captions of synthetically generated and publicly available text-image data. Our model, MUMU, is composed of a vision-language model encoder with a diffusion decoder and is trained on a single 8xH100 GPU node. Despite being only trained on crops from the same image, MUMU learns to compose inputs from different images into a coherent output. For example, an input of a realistic person and a cartoon will output the same person in the cartoon style, and an input of a standing subject and a scooter will output the subject riding the scooter. As a result, our model generalizes to tasks such as style transfer and character consistency. Our results show the promise of using multimodal models as general purpose controllers for image generation.