PyTorchKR Papers

최신 논문

총 5,357편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2310.03150

Federated Fine-Tuning of LLMs on the Very Edge: The Good, the Bad, the Ugly

Large Language Models (LLM) and foundation models are popular as they offer new opportunities for individuals and businesses to improve natural language processing, interact with data, and retrieve information faster. However, training or fine-tuning LLMs requires a vast amount of data, which can be challenging to access due to legal or technical restrictions and may require private computing resources. Federated Learning (FL) is a solution designed to overcome these challenges and expand data access for deep learning applications. This paper takes a hardware-centric approach to explore how LLMs can be brought to modern edge computing systems. Our study fine-tunes the FLAN-T5 model family, ranging from 80M to 3B parameters, using FL for a text summarization task. We provide a micro-level hardware benchmark, compare the model FLOP utilization to a state-of-the-art data center GPU, and study the network utilization in realistic conditions. Our contribution is twofold: First, we evaluate the current capabilities of edge computing systems and their potential for LLM FL workloads. Second, by comparing these systems with a data-center GPU, we demonstrate the potential for improvement and the next steps toward achieving greater computational efficiency at the edge.

arXiv:2310.03294공식 코드 · Papers with Code

DISTFLASHATTN: Distributed Memory-efficient Attention for Long-context LLMs Training

FlashAttention (Dao, 2023) effectively reduces the quadratic peak memory usage to linear in training transformer-based large language models (LLMs) on a single GPU. In this paper, we introduce DISTFLASHATTN, a distributed memory-efficient attention mechanism optimized for long-context LLMs training. We propose three key techniques: token-level workload balancing, overlapping key-value communication, and a rematerialization-aware gradient checkpointing algorithm. We evaluate DISTFLASHATTN on Llama-7B and variants with sequence lengths from 32K to 512K. DISTFLASHATTN achieves 8x longer sequences, 4.45 - 5.64x speedup compared to Ring Self-Attention, 2 - 8x longer sequences, 1.24 - 2.01x speedup compared to Megatron-LM with FlashAttention. It achieves 1.67x and 1.26 - 1.88x speedup compared to recent Ring Attention and DeepSpeed-Ulysses. Code is available at https://github.com/RulinShao/LightSeq.

arXiv:2310.03003

From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference

Large language models (LLMs) have exploded in popularity due to their new generative capabilities that go far beyond prior state-of-the-art. These technologies are increasingly being leveraged in various domains such as law, finance, and medicine. However, these models carry significant computational challenges, especially the compute and energy costs required for inference. Inference energy costs already receive less attention than the energy costs of training LLMs -- despite how often these large models are called on to conduct inference in reality (e.g., ChatGPT). As these state-of-the-art LLMs see increasing usage and deployment in various domains, a better understanding of their resource utilization is crucial for cost-savings, scaling performance, efficient hardware usage, and optimal inference strategies. In this paper, we describe experiments conducted to study the computational and energy utilization of inference with LLMs. We benchmark and conduct a preliminary analysis of the inference performance and inference energy costs of different sizes of LLaMA -- a recent state-of-the-art LLM -- developed by Meta AI on two generations of popular GPUs (NVIDIA V100 \& A100) and two datasets (Alpaca and GSM8K) to reflect the diverse set of tasks/benchmarks for LLMs in research and practice. We present the results of multi-node, multi-GPU inference using model sharding across up to 32 GPUs. To our knowledge, our work is the one of the first to study LLM inference performance from the perspective of computational and energy resources at this scale.

arXiv:2310.02724공식 코드 · Papers with Code

End-to-End Training of a Neural HMM with Label and Transition Probabilities

We investigate a novel modeling approach for end-to-end neural network training using hidden Markov models (HMM) where the transition probabilities between hidden states are modeled and learned explicitly. Most contemporary sequence-to-sequence models allow for from-scratch training by summing over all possible label segmentations in a given topology. In our approach there are explicit, learnable probabilities for transitions between segments as opposed to a blank label that implicitly encodes duration statistics. We implement a GPU-based forward-backward algorithm that enables the simultaneous training of label and transition probabilities. We investigate recognition results and additionally Viterbi alignments of our models. We find that while the transition model training does not improve recognition performance, it has a positive impact on the alignment quality. The generated alignments are shown to be viable targets in state-of-the-art Viterbi trainings.

arXiv:2310.02410

Mixture of Quantized Experts (MoQE): Complementary Effect of Low-bit Quantization and Robustness

Large Mixture of Experts (MoE) models could achieve state-of-the-art quality on various language tasks, including machine translation task, thanks to the efficient model scaling capability with expert parallelism. However, it has brought a fundamental issue of larger memory consumption and increased memory bandwidth bottleneck at deployment time. In this paper, we propose Mixture of Quantized Experts (MoQE) which is a simple weight-only quantization method applying ultra low-bit down to 2-bit quantizations only to expert weights for mitigating the increased memory and latency issues of MoE models. We show that low-bit quantization together with the MoE architecture delivers a reliable model performance while reducing the memory size significantly even without any additional training in most cases. In particular, expert layers in MoE models are much more robust to the quantization than conventional feedforward networks (FFN) layers. In our comprehensive analysis, we show that MoE models with 2-bit expert weights can deliver better model performance than the dense model trained on the same dataset. As a result of low-bit quantization, we show the model size can be reduced by 79.6% of the original half precision floating point (fp16) MoE model. Combined with an optimized GPU runtime implementation, it also achieves 1.24X speed-up on A100 GPUs.

arXiv:2310.02784

MAD Max Beyond Single-Node: Enabling Large Machine Learning Model Acceleration on Distributed Systems

Training and deploying large-scale machine learning models is time-consuming, requires significant distributed computing infrastructures, and incurs high operational costs. Our analysis, grounded in real-world large model training on datacenter-scale infrastructures, reveals that 14~32% of all GPU hours are spent on communication with no overlapping computation. To minimize this outstanding communication latency and other inherent at-scale inefficiencies, we introduce an agile performance modeling framework, MAD-Max. This framework is designed to optimize parallelization strategies and facilitate hardware-software co-design opportunities. Through the application of MAD-Max to a suite of real-world large-scale ML models on state-of-the-art GPU clusters, we showcase potential throughput enhancements of up to 2.24x for pre-training and up to 5.2x for inference scenarios, respectively.

arXiv:2310.02422

OneAdapt: Fast Configuration Adaptation for Video Analytics Applications via Backpropagation

Deep learning inference on streaming media data, such as object detection in video or LiDAR feeds and text extraction from audio waves, is now ubiquitous. To achieve high inference accuracy, these applications typically require significant network bandwidth to gather high-fidelity data and extensive GPU resources to run deep neural networks (DNNs). While the high demand for network bandwidth and GPU resources could be substantially reduced by optimally adapting the configuration knobs, such as video resolution and frame rate, current adaptation techniques fail to meet three requirements simultaneously: adapt configurations (i) with minimum extra GPU or bandwidth overhead; (ii) to reach near-optimal decisions based on how the data affects the final DNN's accuracy, and (iii) do so for a range of configuration knobs. This paper presents OneAdapt, which meets these requirements by leveraging a gradient-ascent strategy to adapt configuration knobs. The key idea is to embrace DNNs' differentiability to quickly estimate the accuracy's gradient to each configuration knob, called AccGrad. Specifically, OneAdapt estimates AccGrad by multiplying two gradients: InputGrad (i.e. how each configuration knob affects the input to the DNN) and DNNGrad (i.e. how the DNN input affects the DNN inference output). We evaluate OneAdapt across five types of configurations, four analytic tasks, and five types of input data. Compared to state-of-the-art adaptation schemes, OneAdapt cuts bandwidth usage and GPU usage by 15-59% while maintaining comparable accuracy or improves accuracy by 1-5% while using equal or fewer resources.

arXiv:2310.02949

Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models

Warning: This paper contains examples of harmful language, and reader discretion is recommended. The increasing open release of powerful large language models (LLMs) has facilitated the development of downstream applications by reducing the essential cost of data annotation and computation. To ensure AI safety, extensive safety-alignment measures have been conducted to armor these models against malicious use (primarily hard prompt attack). However, beneath the seemingly resilient facade of the armor, there might lurk a shadow. By simply tuning on 100 malicious examples with 1 GPU hour, these safely aligned LLMs can be easily subverted to generate harmful content. Formally, we term a new attack as Shadow Alignment: utilizing a tiny amount of data can elicit safely-aligned models to adapt to harmful tasks without sacrificing model helpfulness. Remarkably, the subverted models retain their capability to respond appropriately to regular inquiries. Experiments across 8 models released by 5 different organizations (LLaMa-2, Falcon, InternLM, BaiChuan2, Vicuna) demonstrate the effectiveness of shadow alignment attack. Besides, the single-turn English-only attack successfully transfers to multi-turn dialogue and other languages. This study serves as a clarion call for a collective effort to overhaul and fortify the safety of open-source LLMs against malicious attackers.

arXiv:2310.01801공식 코드 · Papers with Code

Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs

In this study, we introduce adaptive KV cache compression, a plug-and-play method that reduces the memory footprint of generative inference for Large Language Models (LLMs). Different from the conventional KV cache that retains key and value vectors for all context tokens, we conduct targeted profiling to discern the intrinsic structure of attention modules. Based on the recognized structure, we then construct the KV cache in an adaptive manner: evicting long-range contexts on attention heads emphasizing local contexts, discarding non-special tokens on attention heads centered on special tokens, and only employing the standard KV cache for attention heads that broadly attend to all tokens. Moreover, with the lightweight attention profiling used to guide the construction of the adaptive KV cache, FastGen can be deployed without resource-intensive fine-tuning or re-training. In our experiments across various asks, FastGen demonstrates substantial reduction on GPU memory consumption with negligible generation quality loss. We will release our code and the compatible CUDA kernel for reproducibility.

arXiv:2310.01881

Adaptive Multi-NeRF: Exploit Efficient Parallelism in Adaptive Multiple Scale Neural Radiance Field Rendering

Recent advances in Neural Radiance Fields (NeRF) have demonstrated significant potential for representing 3D scene appearances as implicit neural networks, enabling the synthesis of high-fidelity novel views. However, the lengthy training and rendering process hinders the widespread adoption of this promising technique for real-time rendering applications. To address this issue, we present an effective adaptive multi-NeRF method designed to accelerate the neural rendering process for large scenes with unbalanced workloads due to varying scene complexities. Our method adaptively subdivides scenes into axis-aligned bounding boxes using a tree hierarchy approach, assigning smaller NeRFs to different-sized subspaces based on the complexity of each scene portion. This ensures the underlying neural representation is specific to a particular part of the scene. We optimize scene subdivision by employing a guidance density grid, which balances representation capability for each Multilayer Perceptron (MLP). Consequently, samples generated by each ray can be sorted and collected for parallel inference, achieving a balanced workload suitable for small MLPs with consistent dimensions for regular and GPU-friendly computations. We aosl demonstrated an efficient NeRF sampling strategy that intrinsically adapts to increase parallelism, utilization, and reduce kernel calls, thereby achieving much higher GPU utilization and accelerating the rendering process.

arXiv:2310.02238

Who's Harry Potter? Approximate Unlearning in LLMs

Large language models (LLMs) are trained on massive internet corpora that often contain copyrighted content. This poses legal and ethical challenges for the developers and users of these models, as well as the original authors and publishers. In this paper, we propose a novel technique for unlearning a subset of the training data from a LLM, without having to retrain it from scratch. We evaluate our technique on the task of unlearning the Harry Potter books from the Llama2-7b model (a generative language model recently open-sourced by Meta). While the model took over 184K GPU-hours to pretrain, we show that in about 1 GPU hour of finetuning, we effectively erase the model's ability to generate or recall Harry Potter-related content, while its performance on common benchmarks (such as Winogrande, Hellaswag, arc, boolq and piqa) remains almost unaffected. We make our fine-tuned model publicly available on HuggingFace for community evaluation. To the best of our knowledge, this is the first paper to present an effective technique for unlearning in generative language models. Our technique consists of three main components: First, we use a reinforced model that is further trained on the target data to identify the tokens that are most related to the unlearning target, by comparing its logits with those of a baseline model. Second, we replace idiosyncratic expressions in the target data with generic counterparts, and leverage the model's own predictions to generate alternative labels for every token. These labels aim to approximate the next-token predictions of a model that has not been trained on the target data. Third, we finetune the model on these alternative labels, which effectively erases the original text from the model's memory whenever it is prompted with its context.

arXiv:2310.01775

STAMP: Differentiable Task and Motion Planning via Stein Variational Gradient Descent

Planning for sequential robotics tasks often requires integrated symbolic and geometric reasoning. TAMP algorithms typically solve these problems by performing a tree search over high-level task sequences while checking for kinematic and dynamic feasibility. This can be inefficient because, typically, candidate task plans resulting from the tree search ignore geometric information. This often leads to motion planning failures that require expensive backtracking steps to find alternative task plans. We propose a novel approach to TAMP called Stein Task and Motion Planning (STAMP) that relaxes the hybrid optimization problem into a continuous domain. This allows us to leverage gradients from differentiable physics simulation to fully optimize discrete and continuous plan parameters for TAMP. In particular, we solve the optimization problem using a gradient-based variational inference algorithm called Stein Variational Gradient Descent. This allows us to find a distribution of solutions within a single optimization run. Furthermore, we use an off-the-shelf differentiable physics simulator that is parallelized on the GPU to run parallelized inference over diverse plan parameters. We demonstrate our method on a variety of problems and show that it can find multiple diverse plans in a single optimization run while also being significantly faster than existing approaches.

arXiv:2310.01258

MobileNVC: Real-time 1080p Neural Video Compression on a Mobile Device

Neural video codecs have recently become competitive with standard codecs such as HEVC in the low-delay setting. However, most neural codecs are large floating-point networks that use pixel-dense warping operations for temporal modeling, making them too computationally expensive for deployment on mobile devices. Recent work has demonstrated that running a neural decoder in real time on mobile is feasible, but shows this only for 720p RGB video. This work presents the first neural video codec that decodes 1080p YUV420 video in real time on a mobile device. Our codec relies on two major contributions. First, we design an efficient codec that uses a block-based motion compensation algorithm available on the warping core of the mobile accelerator, and we show how to quantize this model to integer precision. Second, we implement a fast decoder pipeline that concurrently runs neural network components on the neural signal processor, parallel entropy coding on the mobile GPU, and warping on the warping core. Our codec outperforms the previous on-device codec by a large margin with up to 48% BD-rate savings, while reducing the MAC count on the receiver side by $10 \times$. We perform a careful ablation to demonstrate the effect of the introduced motion compensation scheme, and ablate the effect of model quantization.

arXiv:2310.00530

Multi-tiling Neural Radiance Field (NeRF) -- Geometric Assessment on Large-scale Aerial Datasets

Neural Radiance Fields (NeRF) offer the potential to benefit 3D reconstruction tasks, including aerial photogrammetry. However, the scalability and accuracy of the inferred geometry are not well-documented for large-scale aerial assets,since such datasets usually result in very high memory consumption and slow convergence.. In this paper, we aim to scale the NeRF on large-scael aerial datasets and provide a thorough geometry assessment of NeRF. Specifically, we introduce a location-specific sampling technique as well as a multi-camera tiling (MCT) strategy to reduce memory consumption during image loading for RAM, representation training for GPU memory, and increase the convergence rate within tiles. MCT decomposes a large-frame image into multiple tiled images with different camera models, allowing these small-frame images to be fed into the training process as needed for specific locations without a loss of accuracy. We implement our method on a representative approach, Mip-NeRF, and compare its geometry performance with threephotgrammetric MVS pipelines on two typical aerial datasets against LiDAR reference data. Both qualitative and quantitative results suggest that the proposed NeRF approach produces better completeness and object details than traditional approaches, although as of now, it still falls short in terms of accuracy.

arXiv:2310.00426공식 코드 · Papers with Code

PixArt-$α$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis

The most advanced text-to-image (T2I) models require significant training costs (e.g., millions of GPU hours), seriously hindering the fundamental innovation for the AIGC community while increasing CO2 emissions. This paper introduces PIXART-$\alpha$, a Transformer-based T2I diffusion model whose image generation quality is competitive with state-of-the-art image generators (e.g., Imagen, SDXL, and even Midjourney), reaching near-commercial application standards. Additionally, it supports high-resolution image synthesis up to 1024px resolution with low training cost, as shown in Figure 1 and 2. To achieve this goal, three core designs are proposed: (1) Training strategy decomposition: We devise three distinct training steps that separately optimize pixel dependency, text-image alignment, and image aesthetic quality; (2) Efficient T2I Transformer: We incorporate cross-attention modules into Diffusion Transformer (DiT) to inject text conditions and streamline the computation-intensive class-condition branch; (3) High-informative data: We emphasize the significance of concept density in text-image pairs and leverage a large Vision-Language model to auto-label dense pseudo-captions to assist text-image alignment learning. As a result, PIXART-$\alpha$'s training speed markedly surpasses existing large-scale T2I models, e.g., PIXART-$\alpha$ only takes 10.8% of Stable Diffusion v1.5's training time (675 vs. 6,250 A100 GPU days), saving nearly \$300,000 (\$26,000 vs. \$320,000) and reducing 90% CO2 emissions. Moreover, compared with a larger SOTA model, RAPHAEL, our training cost is merely 1%. Extensive experiments demonstrate that PIXART-$\alpha$ excels in image quality, artistry, and semantic control. We hope PIXART-$\alpha$ will provide new insights to the AIGC community and startups to accelerate building their own high-quality yet low-cost generative models from scratch.

arXiv:2310.01208공식 코드 · Papers with Code

Label Supervised LLaMA Finetuning

The recent success of Large Language Models (LLMs) has gained significant attention in both academia and industry. Substantial efforts have been made to enhance the zero- and few-shot generalization capabilities of open-source LLMs through finetuning. Currently, the prevailing approach is instruction-tuning, which trains LLMs to complete real-world tasks by generating responses guided by natural language instructions. It is worth noticing that such an approach may underperform in sequence and token classification tasks. Unlike text generation tasks, classification tasks have a limited label space, where precise label prediction is more appreciated than generating diverse and human-like responses. Prior research has unveiled that instruction-tuned LLMs cannot outperform BERT, prompting us to explore the potential of leveraging latent representations from LLMs for supervised label prediction. In this paper, we introduce a label-supervised adaptation for LLMs, which aims to finetuning the model with discriminant labels. We evaluate this approach with Label Supervised LLaMA (LS-LLaMA), based on LLaMA-2-7B, a relatively small-scale LLM, and can be finetuned on a single GeForce RTX4090 GPU. We extract latent representations from the final LLaMA layer and project them into the label space to compute the cross-entropy loss. The model is finetuned by Low-Rank Adaptation (LoRA) to minimize this loss. Remarkably, without intricate prompt engineering or external knowledge, LS-LLaMA substantially outperforms LLMs ten times its size in scale and demonstrates consistent improvements compared to robust baselines like BERT-Large and RoBERTa-Large in text classification. Moreover, by removing the causal mask from decoders, LS-unLLaMA achieves the state-of-the-art performance in named entity recognition (NER). Our work will shed light on a novel approach to adapting LLMs for various downstream tasks.

arXiv:2309.16952공식 코드 · Papers with Code

Leveraging Optimization for Adaptive Attacks on Image Watermarks

Untrustworthy users can misuse image generators to synthesize high-quality deepfakes and engage in unethical activities. Watermarking deters misuse by marking generated content with a hidden message, enabling its detection using a secret watermarking key. A core security property of watermarking is robustness, which states that an attacker can only evade detection by substantially degrading image quality. Assessing robustness requires designing an adaptive attack for the specific watermarking algorithm. When evaluating watermarking algorithms and their (adaptive) attacks, it is challenging to determine whether an adaptive attack is optimal, i.e., the best possible attack. We solve this problem by defining an objective function and then approach adaptive attacks as an optimization problem. The core idea of our adaptive attacks is to replicate secret watermarking keys locally by creating surrogate keys that are differentiable and can be used to optimize the attack's parameters. We demonstrate for Stable Diffusion models that such an attacker can break all five surveyed watermarking methods at no visible degradation in image quality. Optimizing our attacks is efficient and requires less than 1 GPU hour to reduce the detection accuracy to 6.3% or less. Our findings emphasize the need for more rigorous robustness testing against adaptive, learnable attackers.

arXiv:2309.16809공식 코드 · Papers with Code

GraB-sampler: Optimal Permutation-based SGD Data Sampler for PyTorch

The online Gradient Balancing (GraB) algorithm greedily choosing the examples ordering by solving the herding problem using per-sample gradients is proved to be the theoretically optimal solution that guarantees to outperform Random Reshuffling. However, there is currently no efficient implementation of GraB for the community to easily use it. This work presents an efficient Python library, $\textit{GraB-sampler}$, that allows the community to easily use GraB algorithms and proposes 5 variants of the GraB algorithm. The best performance result of the GraB-sampler reproduces the training loss and test accuracy results while only in the cost of 8.7% training time overhead and 0.85% peak GPU memory usage overhead.

arXiv:2309.16818공식 코드 · Papers with Code

MEM: Multi-Modal Elevation Mapping for Robotics and Learning

Elevation maps are commonly used to represent the environment of mobile robots and are instrumental for locomotion and navigation tasks. However, pure geometric information is insufficient for many field applications that require appearance or semantic information, which limits their applicability to other platforms or domains. In this work, we extend a 2.5D robot-centric elevation mapping framework by fusing multi-modal information from multiple sources into a popular map representation. The framework allows inputting data contained in point clouds or images in a unified manner. To manage the different nature of the data, we also present a set of fusion algorithms that can be selected based on the information type and user requirements. Our system is designed to run on the GPU, making it real-time capable for various robotic and learning tasks. We demonstrate the capabilities of our framework by deploying it on multiple robots with varying sensor configurations and showcasing a range of applications that utilize multi-modal layers, including line detection, human detection, and colorization.

arXiv:2309.16738공식 코드 · Papers with Code

ELIP: Efficient Language-Image Pre-training with Fewer Vision Tokens

Learning a versatile language-image model is computationally prohibitive under a limited computing budget. This paper delves into the \emph{efficient language-image pre-training}, an area that has received relatively little attention despite its importance in reducing computational cost and footprint. To that end, we propose a vision token pruning and merging method ELIP, to remove less influential tokens based on the supervision of language outputs. Our method is designed with several strengths, such as being computation-efficient, memory-efficient, and trainable-parameter-free, and is distinguished from previous vision-only token pruning approaches by its alignment with task objectives. We implement this method in a progressively pruning manner using several sequential blocks. To evaluate its generalization performance, we apply ELIP to three commonly used language-image pre-training models and utilize public image-caption pairs with 4M images for pre-training. Our experiments demonstrate that with the removal of ~30$\%$ vision tokens across 12 ViT layers, ELIP maintains significantly comparable performance with baselines ($\sim$0.32 accuracy drop on average) over various downstream tasks including cross-modal retrieval, VQA, image captioning, \emph{etc}. In addition, the spared GPU resources by our ELIP allow us to scale up with larger batch sizes, thereby accelerating model pre-training and even sometimes enhancing downstream model performance.