Sparsity Acceleration

In this paper, we present DSA, the first automated framework for discovering sparsity allocation schemes for layer-wise pruning in Large Language Models (LLMs). LLMs have become increasingly powerful, but their large parameter counts make them computationally expensive. Existing pruning methods for compressing LLMs primarily focus on evaluating redundancies and removing element-wise weights. However, these methods fail to allocate adaptive layer-wise sparsities, leading to performance degradation in challenging tasks. We observe that per-layer importance statistics can serve as allocation indications, but their effectiveness depends on the allocation function between layers. To address this issue, we develop an expression discovery framework to explore potential allocation strategies. Our allocation functions involve two steps: reducing element-wise metrics to per-layer importance scores, and modelling layer importance to sparsity ratios. To search for the most effective allocation function, we construct a search space consisting of pre-process, reduction, transform, and post-process operations. We leverage an evolutionary algorithm to perform crossover and mutation on superior candidates within the population, guided by performance evaluation. Finally, we seamlessly integrate our discovered functions into various uniform methods, resulting in significant performance improvements. We conduct extensive experiments on multiple challenging tasks such as arithmetic, knowledge reasoning, and multimodal benchmarks spanning GSM8K, MMLU, SQA, and VQA, demonstrating that our DSA method achieves significant performance gains on the LLaMA-1|2|3, Mistral, and OPT models. Notably, the LLaMA-1|2|3 model pruned by our DSA reaches 4.73\%|6.18\%|10.65\% gain over the state-of-the-art techniques (e.g., Wanda and SparseGPT).

Incomplete spatio-temporal data in real-world has spawned many research. However, existing methods often utilize iterative message-passing across temporal and spatial dimensions, resulting in substantial information loss and high computational cost. We provide a theoretical analysis revealing that such iterative models are not only susceptible to data sparsity but also to graph sparsity, causing unstable performances on different datasets. To overcome these limitations, we introduce a novel method named One-step Propagation and Confidence-based Refinement (OPCR). In the first stage, OPCR leverages inherent spatial and temporal relationships by employing sparse attention mechanism. These modules propagate limited observations directly to the global context through one-step imputation, which are theoretically effected only by data sparsity. Following this, we assign confidence levels to the initial imputations by correlating missing data with valid data. This confidence-based propagation refines the seperate spatial and temporal imputation results through spatio-temporal dependencies. We evaluate the proposed model across various downstream tasks involving highly sparse spatio-temporal data. Empirical results indicate that our model outperforms state-of-the-art imputation methods, demonstrating its superior effectiveness and robustness.

Multi-Output Regression (MOR) has been widely used in scientific data analysis for decision-making. Unlike traditional regression models, MOR aims to simultaneously predict multiple real-valued outputs given an input. However, the increasing dimensionality of the outputs poses significant challenges regarding interpretability and computational scalability for modern MOR applications. As a first step to address these challenges, this paper proposes a Sparse \& High-dimensional-Output REgression (SHORE) model by incorporating additional sparsity requirements to resolve the output interpretability, and then designs a computationally efficient two-stage optimization framework capable of solving SHORE with provable accuracy via compression on outputs. Theoretically, we show that the proposed framework is computationally scalable while maintaining the same order of training loss and prediction loss before-and-after compression under arbitrary or relatively weak sample set conditions. Empirically, numerical results further validate the theoretical findings, showcasing the efficiency and accuracy of the proposed framework.

We present a generalization of Nesterov's accelerated gradient descent algorithm. Our algorithm (AGNES) provably achieves acceleration for smooth convex and strongly convex minimization tasks with noisy gradient estimates if the noise intensity is proportional to the magnitude of the gradient at every point. Nesterov's method converges at an accelerated rate if the constant of proportionality is below 1, while AGNES accommodates any signal-to-noise ratio. The noise model is motivated by applications in overparametrized machine learning. AGNES requires only two parameters in convex and three in strongly convex minimization tasks, improving on existing methods. We further provide clear geometric interpretations and heuristics for the choice of parameters.

We study stochastic linear bandits where, in each round, the learner receives a set of actions (i.e., feature vectors), from which it chooses an element and obtains a stochastic reward. The expected reward is a fixed but unknown linear function of the chosen action. We study \emph{sparse} regret bounds, that depend on the number $S$ of non-zero coefficients in the linear reward function. Previous works focused on the case where $S$ is known, or the action sets satisfy additional assumptions. In this work, we obtain the first sparse regret bounds that hold when $S$ is unknown and the action sets are adversarially generated. Our techniques combine online to confidence set conversions with a novel randomized model selection approach over a hierarchy of nested confidence sets. When $S$ is known, our analysis recovers state-of-the-art bounds for adversarial action sets. We also show that a variant of our approach, using Exp3 to dynamically select the confidence sets, can be used to improve the empirical performance of stochastic linear bandits while enjoying a regret bound with optimal dependence on the time horizon.

Large Language Models (LLMs) have achieved remarkable success with their billion-level parameters, yet they incur high inference overheads. The emergence of activation sparsity in LLMs provides a natural approach to reduce this cost by involving only parts of the parameters for inference. However, existing methods only focus on utilizing this naturally formed activation sparsity in a post-training setting, overlooking the potential for further amplifying this inherent sparsity. In this paper, we hypothesize that LLMs can learn to be efficient by achieving more structured activation sparsity. To achieve this, we introduce a novel training algorithm, Learn-To-be-Efficient (LTE), designed to train efficiency-aware LLMs to learn to activate fewer neurons and achieve a better trade-off between sparsity and performance. Furthermore, unlike SOTA MoEfication methods, which mainly focus on ReLU-based models, LTE can also be applied to LLMs like LLaMA using non-ReLU activations. Extensive evaluation on language understanding, language generation, and instruction tuning tasks show that LTE consistently outperforms SOTA baselines. Along with our hardware-aware custom kernel implementation, LTE reduces LLaMA2-7B inference latency by 25% at 50% sparsity.

Considerable research efforts have recently been made to show that a random neural network $N$ contains subnetworks capable of accurately approximating any given neural network that is sufficiently smaller than $N$, without any training. This line of research, known as the Strong Lottery Ticket Hypothesis (SLTH), was originally motivated by the weaker Lottery Ticket Hypothesis, which states that a sufficiently large random neural network $N$ contains sparse subnetworks that can be trained efficiently to achieve performance comparable to that of training the entire network $N$. Despite its original motivation, results on the SLTH have so far not provided any guarantee on the size of subnetworks. Such limitation is due to the nature of the main technical tool leveraged by these results, the Random Subset Sum (RSS) Problem. Informally, the RSS Problem asks how large a random i.i.d. sample $\Omega$ should be so that we are able to approximate any number in $[-1,1]$, up to an error of $ \epsilon$, as the sum of a suitable subset of $\Omega$. We provide the first proof of the SLTH in classical settings, such as dense and equivariant networks, with guarantees on the sparsity of the subnetworks. Central to our results, is the proof of an essentially tight bound on the Random Fixed-Size Subset Sum Problem (RFSS), a variant of the RSS Problem in which we only ask for subsets of a given size, which is of independent interest.

This paper introduces a novel prior called Diversified Block Sparse Prior to characterize the widespread block sparsity phenomenon in real-world data. By allowing diversification on intra-block variance and inter-block correlation matrices, we effectively address the sensitivity issue of existing block sparse learning methods to pre-defined block information, which enables adaptive block estimation while mitigating the risk of overfitting. Based on this, a diversified block sparse Bayesian learning method (DivSBL) is proposed, utilizing EM algorithm and dual ascent method for hyperparameter estimation. Moreover, we establish the global and local optimality theory of our model. Experiments validate the advantages of DivSBL over existing algorithms.

We introduce the CRONOS algorithm for convex optimization of two-layer neural networks. CRONOS is the first algorithm capable of scaling to high-dimensional datasets such as ImageNet, which are ubiquitous in modern deep learning. This significantly improves upon prior work, which has been restricted to downsampled versions of MNIST and CIFAR-10. Taking CRONOS as a primitive, we then develop a new algorithm called CRONOS-AM, which combines CRONOS with alternating minimization, to obtain an algorithm capable of training multi-layer networks with arbitrary architectures. Our theoretical analysis proves that CRONOS converges to the global minimum of the convex reformulation under mild assumptions. In addition, we validate the efficacy of CRONOS and CRONOS-AM through extensive large-scale numerical experiments with GPU acceleration in JAX. Our results show that CRONOS-AM can obtain comparable or better validation accuracy than predominant tuned deep learning optimizers on vision and language tasks with benchmark datasets such as ImageNet and IMDb. To the best of our knowledge, CRONOS is the first algorithm which utilizes the convex reformulation to enhance performance on large-scale learning tasks.

In recent years, Dynamic Sparse Training (DST) has emerged as an alternative to post-training pruning for generating efficient models. In principle, DST allows for a much more memory efficient training process, as it maintains sparsity throughout the entire training run. However, current DST implementations fail to capitalize on this. Because sparse matrix multiplication is much less efficient than dense matrix multiplication on GPUs, most implementations simulate sparsity by masking weights. In this paper, we leverage recent advances in semi-structured sparse training to apply DST in the domain of classification with large output spaces, where memory-efficiency is paramount. With a label space of possibly millions of candidates, the classification layer alone will consume several gigabytes of memory. Switching from a dense to a fixed fan-in sparse layer updated with sparse evolutionary training (SET); however, severely hampers training convergence, especially at the largest label spaces. We find that the gradients fed back from the classifier into the text encoder make it much more difficult to learn good input representations, despite using a dense encoder. By employing an intermediate layer or adding an auxiliary training objective, we recover most of the generalisation performance of the dense model. Overall, we demonstrate the applicability of DST in a challenging domain, characterized by a highly skewed label distribution, that lies outside of DST's typical benchmark datasets, and enable end-to-end training with millions of labels on commodity hardware.

Several challenges make it difficult for sparse neural networks to compete with dense models. First, setting a large fraction of weights to zero impairs forward and gradient signal propagation. Second, sparse studies often need to test multiple sparsity levels, while also introducing new hyperparameters (HPs), leading to prohibitive tuning costs. Indeed, the standard practice is to re-use the learning HPs originally crafted for dense models. Unfortunately, we show sparse and dense networks do not share the same optimal HPs. Without stable dynamics and effective training recipes, it is costly to test sparsity at scale, which is key to surpassing dense networks and making the business case for sparsity acceleration in hardware. A holistic approach is needed to tackle these challenges and we propose S$\textmu$Par as one such approach. For random unstructured static sparsity, S$\textmu$Par ensures activations, gradients, and weight updates all scale independently of sparsity level. Further, by reparameterizing the HPs, S$\textmu$Par enables the same HP values to be optimal as we vary both sparsity level and model width. HPs can be tuned on small dense networks and transferred to large sparse models, greatly reducing tuning costs. On large-scale language modeling, S$\textmu$Par shows increasing improvements over standard parameterization as sparsity increases, leading up to 11.9\% relative loss improvement at 99.2\% sparsity. A minimal implementation of S$\textmu$Par is available at https://github.com/EleutherAI/nanoGPT-mup/tree/supar.

Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable pruning method that establishes Semi-structured (or ``N:M'') Sparsity in LLMs, aimed at reducing computational overhead during inference. Instead of developing a new importance criterion, MaskLLM explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. We assessed MaskLLM using 2:4 sparsity on various LLMs, including LLaMA-2, Nemotron-4, and GPT-3, with sizes ranging from 843M to 15B parameters, and our empirical results show substantial improvements over state-of-the-art methods. For instance, leading approaches achieve a perplexity (PPL) of 10 or greater on Wikitext compared to the dense model's 5.12 PPL, but MaskLLM achieves a significantly lower 6.72 PPL solely by learning the masks with frozen weights. Furthermore, MaskLLM's learnable nature allows customized masks for lossless application of 2:4 sparsity to downstream tasks or domains. Code is available at https://github.com/NVlabs/MaskLLM.

Large Language Models (LLMs) have revolutionized the field of natural language processing with their impressive capabilities. However, their enormous size presents challenges for deploying them in real-world applications. Traditional compression techniques, like pruning, often lead to suboptimal performance due to their uniform pruning ratios and lack of consideration for the varying importance of features across different layers. To address these limitations, we present a novel Adaptive Layer Sparsity (ALS) approach to optimize LLMs. Our approach consists of two key steps. Firstly, we estimate the correlation matrix between intermediate layers by leveraging the concept of information orthogonality. This novel perspective allows for a precise measurement of the importance of each layer across the model. Secondly, we employ a linear optimization algorithm to develop an adaptive sparse allocation strategy based on evaluating the correlation matrix. This strategy enables us to selectively prune features in intermediate layers, achieving fine-grained optimization of the LLM model. Considering the varying importance across different layers, we can significantly reduce the model size without sacrificing performance. We conduct extensive experiments on publicly available language processing datasets, including the LLaMA-V1|V2|V3 family and OPT, covering various benchmarks. Our experimental results validate the effectiveness of our ALS method, showcasing its superiority over previous approaches. The performance gains demonstrate its potential for enhancing LLMs' efficiency and resource utilization. Notably, our approach surpasses the state-of-the-art models Wanda and SparseGPT, showcasing its ability to excel even under high sparsity levels. Codes at: https://github.com/lliai/ALS.

Network pruning is a commonly used measure to alleviate the storage and computational burden of deep neural networks. However, the fundamental limit of network pruning is still lacking. To close the gap, in this work we'll take a first-principles approach, i.e. we'll directly impose the sparsity constraint on the loss function and leverage the framework of statistical dimension in convex geometry, thus enabling us to characterize the sharp phase transition point, which can be regarded as the fundamental limit of the pruning ratio. Through this limit, we're able to identify two key factors that determine the pruning ratio limit, namely, weight magnitude and network sharpness. Generally speaking, the flatter the loss landscape or the smaller the weight magnitude, the smaller pruning ratio. Moreover, we provide efficient countermeasures to address the challenges in the computation of the pruning limit, which mainly involves the accurate spectrum estimation of a large-scale and non-positive Hessian matrix. Moreover, through the lens of the pruning ratio threshold, we can also provide rigorous interpretations on several heuristics in existing pruning algorithms. Extensive experiments are performed which demonstrate that our theoretical pruning ratio threshold coincides very well with the experiments. All codes are available at: https://github.com/QiaozheZhang/Global-One-shot-Pruning

Sparsity is a central aspect of interpretability in machine learning. Typically, sparsity is measured in terms of the size of a model globally, such as the number of variables it uses. However, this notion of sparsity is not particularly relevant for decision making; someone subjected to a decision does not care about variables that do not contribute to the decision. In this work, we dramatically expand a notion of *decision sparsity* called the *Sparse Explanation Value* (SEV) so that its explanations are more meaningful. SEV considers movement along a hypercube towards a reference point. By allowing flexibility in that reference and by considering how distances along the hypercube translate to distances in feature space, we can derive sparser and more meaningful explanations for various types of function classes. We present cluster-based SEV and its variant tree-based SEV, introduce a method that improves credibility of explanations, and propose algorithms that optimize decision sparsity in machine learning models.

Training deep neural networks (DNNs) is costly. Fortunately, Nvidia Ampere and Hopper GPUs can accelerate matrix multiplications twice as fast as a dense equivalent by implementing 2:4 sparsity. However, previous STE-based 2:4 pre-training methods (\eg~STE with hard-thresholding, SR-STE) suffer from optimization difficulties because of discontinuous pruning function. In this study, we comprehensively analyse the bottleneck of traditional N:M sparse training and recognize three drawbacks with discontinuity: incorrect descending direction, inability to predict the amount of descent and sparse mask oscillation. In the light of this statement, we propose S-STE, a simple yet powerful 2:4 training method that contains two parts: to continuously project weights to be 2:4 sparse, and to rescale sparse weights with a per-tensor fixed scaling factor. Besides, we adopt minimum-variance unbiased estimation for activation gradient and FP8 quantization for whole process. Results show that our method surpass previous 2:4 pre-training recipes and is comparable even with full parameter models.

Diffusion models have garnered significant interest from the community for their great generative ability across various applications. However, their typical multi-step sequential-denoising nature gives rise to high cumulative latency, thereby precluding the possibilities of parallel computation. To address this, we introduce AsyncDiff, a universal and plug-and-play acceleration scheme that enables model parallelism across multiple devices. Our approach divides the cumbersome noise prediction model into multiple components, assigning each to a different device. To break the dependency chain between these components, it transforms the conventional sequential denoising into an asynchronous process by exploiting the high similarity between hidden states in consecutive diffusion steps. Consequently, each component is facilitated to compute in parallel on separate devices. The proposed strategy significantly reduces inference latency while minimally impacting the generative quality. Specifically, for the Stable Diffusion v2.1, AsyncDiff achieves a 2.7x speedup with negligible degradation and a 4.0x speedup with only a slight reduction of 0.38 in CLIP Score, on four NVIDIA A5000 GPUs. Our experiments also demonstrate AsyncDiff can be readily applied to video diffusion models with encouraging performances.

We accelerate the iterative hard thresholding (IHT) method, which finds \(k\) important elements from a parameter vector in a linear regression model. Although the plain IHT repeatedly updates the parameter vector during the optimization, computing gradients is the main bottleneck. Our method safely prunes unnecessary gradient computations to reduce the processing time.The main idea is to efficiently construct a candidate set, which contains \(k\) important elements in the parameter vector, for each iteration. Specifically, before computing the gradients, we prune unnecessary elements in the parameter vector for the candidate set by utilizing upper bounds on absolute values of the parameters. Our method guarantees the same optimization results as the plain IHT because our pruning is safe. Experiments show that our method is up to 73 times faster than the plain IHT without degrading accuracy.