Search

We have observed a distinctive quantization-related behavior in the LLaMA3/3.1-70B models that is absent in both the LLaMA2-70B and LLaMA3/3.1/3.2-1B/3B/8B/405B models. Quantization is a crucial technique for deploying large language models (LLMs) efficiently. The impact of W8A8 post-training quantization on model accuracy, especially on the recently released LLaMA3/3.1 model series, remains contentious. In this paper, we explore three key questions: What makes the LLaMA3-70B model series uniquely vulnerable to quantization? Why is this the case? And how can the issue be addressed? We empirically investigate multiple LLMs featured on an open LLM leaderboard, discovering that the LLaMA3-70B model series have a unique accuracy degradation behavior with W8A8 per-channel post-training quantization. In contrast, other model series such as LLaMA2, LLaMA3/3.1-8B, LLaMA3.2, Qwen, Mixtral, Mistral, Phi-3, and Falcon demonstrate robust performance with W8A8. Contrary to previous assertions attributing degradation to the large dynamic range of activations, our findings indicate that the weight distribution of the LLaMA3-70B is the primary factor behind the vulnerability. By meticulously analyzing the distinct characteristics of weight distributions across Transformer blocks, we propose two solutions that make different tradeoffs in hardware/software overhead. First, we propose a mixed strategy where less than 3\% of the layers employ finer per-group W8A8 quantization granularity. Second, we introduce a bi-smoothing strategy that balances quantization errors between weights and activations while maintaining per-channel quantization throughout. Experimental results demonstrate that both strategies effectively preserve the accuracy of the entire LLaMA3-70B model series under W8A8 quantization, achieving performance on par with their FP16 counterparts., Comment: 27 pages, 41 figures

Deep neural networks (DNNs) are frequently employed in a variety of computer vision applications. Nowadays, an emerging trend in the current video distribution system is to take advantage of DNN's overfitting properties to perform video resolution upscaling. By splitting videos into chunks and applying a super-resolution (SR) model to overfit each chunk, this scheme of SR models plus video chunks is able to replace traditional video transmission to enhance video quality and transmission efficiency. However, many models and chunks are needed to guarantee high performance, which leads to tremendous overhead on model switching and memory footprints at the user end. To resolve such problems, we propose a Dynamic Deep neural network assisted by a Content-Aware data processing pipeline to reduce the model number down to one (Dy-DCA), which helps promote performance while conserving computational resources. Additionally, to achieve real acceleration on the user end, we designed a framework that optimizes dynamic features (e.g., dynamic shapes, sizes, and control flow) in Dy-DCA to enable a series of compilation optimizations, including fused code generation, static execution planning, etc. By employing such techniques, our method achieves better PSNR and real-time performance (33 FPS) on an off-the-shelf mobile phone. Meanwhile, assisted by our compilation optimization, we achieve a 1.7$\times$ speedup while saving up to 1.61$\times$ memory consumption. Code available in https://github.com/coulsonlee/Dy-DCA-ECCV2024., Comment: ECCV2024

As deep convolutional neural networks (DNNs) are widely used in various fields of computer vision, leveraging the overfitting ability of the DNN to achieve video resolution upscaling has become a new trend in the modern video delivery system. By dividing videos into chunks and overfitting each chunk with a super-resolution model, the server encodes videos before transmitting them to the clients, thus achieving better video quality and transmission efficiency. However, a large number of chunks are expected to ensure good overfitting quality, which substantially increases the storage and consumes more bandwidth resources for data transmission. On the other hand, decreasing the number of chunks through training optimization techniques usually requires high model capacity, which significantly slows down execution speed. To reconcile such, we propose a novel method for high-quality and efficient video resolution upscaling tasks, which leverages the spatial-temporal information to accurately divide video into chunks, thus keeping the number of chunks as well as the model size to minimum. Additionally, we advance our method into a single overfitting model by a data-aware joint training technique, which further reduces the storage requirement with negligible quality drop. We deploy our models on an off-the-shelf mobile phone, and experimental results show that our method achieves real-time video super-resolution with high video quality. Compared with the state-of-the-art, our method achieves 28 fps streaming speed with 41.6 PSNR, which is 14$\times$ faster and 2.29 dB better in the live video resolution upscaling tasks. Code available in https://github.com/coulsonlee/STDO-CVPR2023.git, Comment: CVPR 2023 Highlight Paper

Deep neural networks (DNNs) have great potential to solve many real-world problems, but they usually require an extensive amount of computation and memory. It is of great difficulty to deploy a large DNN model to a single resource-limited device with small memory capacity. Distributed computing is a common approach to reduce single-node memory consumption and to accelerate the inference of DNN models. In this paper, we explore the "within-layer model parallelism", which distributes the inference of each layer into multiple nodes. In this way, the memory requirement can be distributed to many nodes, making it possible to use several edge devices to infer a large DNN model. Due to the dependency within each layer, data communications between nodes during this parallel inference can be a bottleneck when the communication bandwidth is limited. We propose a framework to train DNN models for Distributed Inference with Sparse Communications (DISCO). We convert the problem of selecting which subset of data to transmit between nodes into a model optimization problem, and derive models with both computation and communication reduction when each layer is inferred on multiple nodes. We show the benefit of the DISCO framework on a variety of CV tasks such as image classification, object detection, semantic segmentation, and image super resolution. The corresponding models include important DNN building blocks such as convolutions and transformers. For example, each layer of a ResNet-50 model can be distributively inferred across two nodes with five times less data communications, almost half overall computations and half memory requirement for a single node, and achieve comparable accuracy to the original ResNet-50 model. This also results in 4.7 times overall inference speedup.

During the deployment of deep neural networks (DNNs) on edge devices, many research efforts are devoted to the limited hardware resource. However, little attention is paid to the influence of dynamic power management. As edge devices typically only have a budget of energy with batteries (rather than almost unlimited energy support on servers or workstations), their dynamic power management often changes the execution frequency as in the widely-used dynamic voltage and frequency scaling (DVFS) technique. This leads to highly unstable inference speed performance, especially for computation-intensive DNN models, which can harm user experience and waste hardware resources. We firstly identify this problem and then propose All-in-One, a highly representative pruning framework to work with dynamic power management using DVFS. The framework can use only one set of model weights and soft masks (together with other auxiliary parameters of negligible storage) to represent multiple models of various pruning ratios. By re-configuring the model to the corresponding pruning ratio for a specific execution frequency (and voltage), we are able to achieve stable inference speed, i.e., keeping the difference in speed performance under various execution frequencies as small as possible. Our experiments demonstrate that our method not only achieves high accuracy for multiple models of different pruning ratios, but also reduces their variance of inference latency for various frequencies, with minimal memory consumption of only one model and one soft mask.

As data becomes increasingly vital, a company would be very cautious about releasing data, because the competitors could use it to train high-performance models, thereby posing a tremendous threat to the company's commercial competence. To prevent training good models on the data, we could add imperceptible perturbations to it. Since such perturbations aim at hurting the entire training process, they should reflect the vulnerability of DNN training, rather than that of a single model. Based on this new idea, we seek perturbed examples that are always unrecognized (never correctly classified) in training. In this paper, we uncover them by model checkpoints' gradients, forming the proposed self-ensemble protection (SEP), which is very effective because (1) learning on examples ignored during normal training tends to yield DNNs ignoring normal examples; (2) checkpoints' cross-model gradients are close to orthogonal, meaning that they are as diverse as DNNs with different architectures. That is, our amazing performance of ensemble only requires the computation of training one model. By extensive experiments with 9 baselines on 3 datasets and 5 architectures, SEP is verified to be a new state-of-the-art, e.g., our small $\ell_\infty=2/255$ perturbations reduce the accuracy of a CIFAR-10 ResNet18 from 94.56% to 14.68%, compared to 41.35% by the best-known method. Code is available at https://github.com/Sizhe-Chen/SEP., Comment: ICLR 2023

Vision transformers (ViTs) have recently obtained success in many applications, but their intensive computation and heavy memory usage at both training and inference time limit their generalization. Previous compression algorithms usually start from the pre-trained dense models and only focus on efficient inference, while time-consuming training is still unavoidable. In contrast, this paper points out that the million-scale training data is redundant, which is the fundamental reason for the tedious training. To address the issue, this paper aims to introduce sparsity into data and proposes an end-to-end efficient training framework from three sparse perspectives, dubbed Tri-Level E-ViT. Specifically, we leverage a hierarchical data redundancy reduction scheme, by exploring the sparsity under three levels: number of training examples in the dataset, number of patches (tokens) in each example, and number of connections between tokens that lie in attention weights. With extensive experiments, we demonstrate that our proposed technique can noticeably accelerate training for various ViT architectures while maintaining accuracy. Remarkably, under certain ratios, we are able to improve the ViT accuracy rather than compromising it. For example, we can achieve 15.2% speedup with 72.6% (+0.4) Top-1 accuracy on Deit-T, and 15.7% speedup with 79.9% (+0.1) Top-1 accuracy on Deit-S. This proves the existence of data redundancy in ViT., Comment: AAAI 2023

The conventional lottery ticket hypothesis (LTH) claims that there exists a sparse subnetwork within a dense neural network and a proper random initialization method called the winning ticket, such that it can be trained from scratch to almost as good as the dense counterpart. Meanwhile, the research of LTH in vision transformers (ViTs) is scarcely evaluated. In this paper, we first show that the conventional winning ticket is hard to find at the weight level of ViTs by existing methods. Then, we generalize the LTH for ViTs to input data consisting of image patches inspired by the input dependence of ViTs. That is, there exists a subset of input image patches such that a ViT can be trained from scratch by using only this subset of patches and achieve similar accuracy to the ViTs trained by using all image patches. We call this subset of input patches the em winning tickets, which represent a significant amount of information in the input data. We use a ticket selector to generate the winning tickets based on the informativeness of patches for various types of ViT, including DeiT, LV-ViT, and Swin Transformers. The experiments show that there is a clear difference between the performance of models trained with winning tickets and randomly selected subsets, which verifies our proposed theory. We elaborate on the analogical similarity between our proposed Data-LTH-ViTs and the conventional LTH to further verify the integrity of our theory. The Source codes are available at https://github.com/shawnricecake/vit-lottery-ticket-input., Comment: Accepted by IJCAI 2023

Deep learning-based super-resolution (SR) has gained tremendous popularity in recent years because of its high image quality performance and wide application scenarios. However, prior methods typically suffer from large amounts of computations and huge power consumption, causing difficulties for real-time inference, especially on resource-limited platforms such as mobile devices. To mitigate this, we propose a compiler-aware SR neural architecture search (NAS) framework that conducts depth search and per-layer width search with adaptive SR blocks. The inference speed is directly taken into the optimization along with the SR loss to derive SR models with high image quality while satisfying the real-time inference requirement. Instead of measuring the speed on mobile devices at each iteration during the search process, a speed model incorporated with compiler optimizations is leveraged to predict the inference latency of the SR block with various width configurations for faster convergence. With the proposed framework, we achieve real-time SR inference for implementing 720p resolution with competitive SR performance (in terms of PSNR and SSIM) on GPU/DSP of mobile platforms (Samsung Galaxy S21).

Channel pruning has been broadly recognized as an effective technique to reduce the computation and memory cost of deep convolutional neural networks. However, conventional pruning methods have limitations in that: they are restricted to pruning process only, and they require a fully pre-trained large model. Such limitations may lead to sub-optimal model quality as well as excessive memory and training cost. In this paper, we propose a novel Channel Exploration methodology, dubbed as CHEX, to rectify these problems. As opposed to pruning-only strategy, we propose to repeatedly prune and regrow the channels throughout the training process, which reduces the risk of pruning important channels prematurely. More exactly: From intra-layer's aspect, we tackle the channel pruning problem via a well known column subset selection (CSS) formulation. From inter-layer's aspect, our regrowing stages open a path for dynamically re-allocating the number of channels across all the layers under a global channel sparsity constraint. In addition, all the exploration process is done in a single training from scratch without the need of a pre-trained large model. Experimental results demonstrate that CHEX can effectively reduce the FLOPs of diverse CNN architectures on a variety of computer vision tasks, including image classification, object detection, instance segmentation, and 3D vision. For example, our compressed ResNet-50 model on ImageNet dataset achieves 76% top1 accuracy with only 25% FLOPs of the original ResNet-50 model, outperforming previous state-of-the-art channel pruning methods. The checkpoints and code are available at here ., Comment: Accepted to CVPR 2022

Weight pruning in deep neural networks (DNNs) can reduce storage and computation cost, but struggles to bring practical speedup to the model inference time. Tensor-cores can significantly boost the throughput of GPUs on dense computation, but exploiting tensor-cores for sparse DNNs is very challenging. Compared to existing CUDA-cores, tensor-cores require higher data reuse and matrix-shaped instruction granularity, both difficult to yield from sparse DNN kernels. Existing pruning approaches fail to balance the demands of accuracy and efficiency: random sparsity preserves the model quality well but prohibits tensor-core acceleration, while highly-structured block-wise sparsity can exploit tensor-cores but suffers from severe accuracy loss. In this work, we propose a novel sparse pattern, Shuffled Block-wise sparsity (Shfl-BW), designed to efficiently utilize tensor-cores while minimizing the constraints on the weight structure. Our insight is that row- and column-wise permutation provides abundant flexibility for the weight structure, while introduces negligible overheads using our GPU kernel designs. We optimize the GPU kernels for Shfl-BW in linear and convolution layers. Evaluations show that our techniques can achieve the state-of-the-art speed-accuracy trade-offs on GPUs. For example, with small accuracy loss, we can accelerate the computation-intensive layers of Transformer by 1.81, 4.18 and 1.90 times on NVIDIA V100, T4 and A100 GPUs respectively at 75% sparsity., Comment: To-appear in Design Automation Conference (DAC), July 2022

A primary source of increased read time on NAND flash comes from the fact that in the presence of noise, the flash medium must be read several times using different read threshold voltages for the decoder to succeed. This paper proposes an algorithm that uses a limited number of re-reads to characterize the noise distribution and recover the stored information. Both hard and soft decoding are considered. For hard decoding, the paper attempts to find a read threshold minimizing bit-error-rate (BER) and derives an expression for the resulting codeword-error-rate. For soft decoding, it shows that minimizing BER and minimizing codeword-error-rate are competing objectives in the presence of a limited number of allowed re-reads, and proposes a trade-off between the two. The proposed method does not require any prior knowledge about the noise distribution, but can take advantage of such information when it is available. Each read threshold is chosen based on the results of previous reads, following an optimal policy derived through a dynamic programming backward recursion. The method and results are studied from the perspective of an SLC Flash memory with Gaussian noise for each level but the paper explains how the method could be extended to other scenarios.

Recently, Vision Transformer (ViT) has continuously established new milestones in the computer vision field, while the high computation and memory cost makes its propagation in industrial production difficult. Pruning, a traditional model compression paradigm for hardware efficiency, has been widely applied in various DNN structures. Nevertheless, it stays ambiguous on how to perform exclusive pruning on the ViT structure. Considering three key points: the structural characteristics, the internal data pattern of ViTs, and the related edge device deployment, we leverage the input token sparsity and propose a computation-aware soft pruning framework, which can be set up on vanilla Transformers of both flatten and CNN-type structures, such as Pooling-based ViT (PiT). More concretely, we design a dynamic attention-based multi-head token selector, which is a lightweight module for adaptive instance-wise token selection. We further introduce a soft pruning technique, which integrates the less informative tokens generated by the selector module into a package token that will participate in subsequent calculations rather than being completely discarded. Our framework is bound to the trade-off between accuracy and computation constraints of specific edge devices through our proposed computation-aware training strategy. Experimental results show that our framework significantly reduces the computation cost of ViTs while maintaining comparable performance on image classification. Moreover, our framework can guarantee the identified model to meet resource specifications of mobile devices and FPGA, and even achieve the real-time execution of DeiT-T on mobile platforms. For example, our method reduces the latency of DeiT-T to 26 ms (26%$\sim $41% superior to existing works) on the mobile device with 0.25%$\sim $4% higher top-1 accuracy on ImageNet., Comment: ECCV 2022

Deep neural networks (DNNs) have been proven to be effective in solving many real-life problems, but its high computation cost prohibits those models from being deployed to edge devices. Pruning, as a method to introduce zeros to model weights, has shown to be an effective method to provide good trade-offs between model accuracy and computation efficiency, and is a widely-used method to generate compressed models. However, the granularity of pruning makes important trade-offs. At the same sparsity level, a coarse-grained structured sparse pattern is more efficient on conventional hardware but results in worse accuracy, while a fine-grained unstructured sparse pattern can achieve better accuracy but is inefficient on existing hardware. On the other hand, some modern processors are equipped with fast on-chip scratchpad memories and gather/scatter engines that perform indirect load and store operations on such memories. In this work, we propose a set of novel sparse patterns, named gather-scatter (GS) patterns, to utilize the scratchpad memories and gather/scatter engines to speed up neural network inferences. Correspondingly, we present a compact sparse format. The proposed set of sparse patterns, along with a novel pruning methodology, address the load imbalance issue and result in models with quality close to unstructured sparse models and computation efficiency close to structured sparse models. Our experiments show that GS patterns consistently make better trade-offs between accuracy and computation efficiency compared to conventional structured sparse patterns. GS patterns can reduce the runtime of the DNN components by two to three times at the same accuracy levels. This is confirmed on three different deep learning tasks and popular models, namely, GNMT for machine translation, ResNet50 for image recognition, and Japser for acoustic speech recognition.

Deep neural networks (DNNs) have shown to provide superb performance in many real life applications, but their large computation cost and storage requirement have prevented them from being deployed to many edge and internet-of-things (IoT) devices. Sparse deep neural networks, whose majority weight parameters are zeros, can substantially reduce the computation complexity and memory consumption of the models. In real-use scenarios, devices may suffer from large fluctuations of the available computation and memory resources under different environment, and the quality of service (QoS) is difficult to maintain due to the long tail inferences with large latency. Facing the real-life challenges, we propose to train a sparse model that supports multiple sparse levels. That is, a hierarchical structure of weights are satisfied such that the locations and the values of the non-zero parameters of the more-sparse sub-model area subset of the less-sparse sub-model. In this way, one can dynamically select the appropriate sparsity level during inference, while the storage cost is capped by the least sparse sub-model. We have verified our methodologies on a variety of DNN models and tasks, including the ResNet-50, PointNet++, GNMT, and graph attention networks. We obtain sparse sub-models with an average of 13.38% weights and 14.97% FLOPs, while the accuracies are as good as their dense counterparts. More-sparse sub-models with 5.38% weights and 4.47% of FLOPs, which are subsets of the less-sparse ones, can be obtained with only 3.25% relative accuracy loss.

Recently, a new trend of exploring sparsity for accelerating neural network training has emerged, embracing the paradigm of training on the edge. This paper proposes a novel Memory-Economic Sparse Training (MEST) framework targeting for accurate and fast execution on edge devices. The proposed MEST framework consists of enhancements by Elastic Mutation (EM) and Soft Memory Bound (&S) that ensure superior accuracy at high sparsity ratios. Different from the existing works for sparse training, this current work reveals the importance of sparsity schemes on the performance of sparse training in terms of accuracy as well as training speed on real edge devices. On top of that, the paper proposes to employ data efficiency for further acceleration of sparse training. Our results suggest that unforgettable examples can be identified in-situ even during the dynamic exploration of sparsity masks in the sparse training process, and therefore can be removed for further training speedup on edge devices. Comparing with state-of-the-art (SOTA) works on accuracy, our MEST increases Top-1 accuracy significantly on ImageNet when using the same unstructured sparsity scheme. Systematical evaluation on accuracy, training speed, and memory footprint are conducted, where the proposed MEST framework consistently outperforms representative SOTA works. A reviewer strongly against our work based on his false assumptions and misunderstandings. On top of the previous submission, we employ data efficiency for further acceleration of sparse training. And we explore the impact of model sparsity, sparsity schemes, and sparse training algorithms on the number of removable training examples. Our codes are publicly available at: https://github.com/boone891214/MEST., Comment: NeurIPS 2021 Spotlight Paper