Large Language Models (LLMs) have achieved significant success across various NLP tasks. However, their massive computational costs limit their widespread use, particularly in real-time applications. Structured pruning offers an effective solution by compressing models and directly providing end-to-end speed improvements, regardless of the hardware environment. Meanwhile, different components of the model exhibit varying sensitivities towards pruning, calling for \emph{non-uniform} model compression. However, a pruning method should not only identify a capable substructure, but also account for post-compression training. To this end, we propose \sysname, a method for \emph{training-aware} structured pruning. \sysname builds upon an evolutionary search process, generating multiple offspring models in each generation through mutation, and selecting the fittest for survival. To assess the effect of post-training, we incorporate a lightweight, multistep training process within the offspring population, progressively increasing the number of tokens and eliminating poorly performing models in each selection stage. We validate our method through extensive experiments on Llama-2-7B, Llama-3.1-8B and Qwen-2.5-14B-Instruct, achieving state-of-the-art performance for structured pruning. For instance, \sysname surpasses ShearedLlama while requiring $5\times$ less training data during post-compression training.
Model merging, a method that combines the parameters and embeddings of multiple fine-tuned large language models (LLMs), offers a promising approach to enhance model performance across various tasks while maintaining computational efficiency. This paper introduces Activation-Informed Merging (AIM), a technique that integrates the information from the activation space of LLMs into the merging process to improve performance and robustness. AIM is designed as a flexible, complementary solution that is applicable to any existing merging method. It aims to preserve critical weights from the base model, drawing on principles from continual learning~(CL) and model compression. Utilizing a task-agnostic calibration set, AIM selectively prioritizes essential weights during merging. We empirically demonstrate that AIM significantly enhances the performance of merged models across multiple benchmarks. Our findings suggest that considering the activation-space information can provide substantial advancements in the model merging strategies for LLMs with up to 40\% increase in benchmark performance.
The neurons of artificial neural networks were originally invented when much less was known about biological neurons than is known today. Our work explores a modification to the core neuron unit to make it more parallel to a biological neuron. The modification is made with the knowledge that biological dendrites are not simply passive activation funnels, but also compute complex non-linear functions as they transmit activation to the cell body. The paper explores a novel system of "Perforated" backpropagation empowering the artificial neurons of deep neural networks to achieve better performance coding for the same features they coded for in the original architecture. After an initial network training phase, additional "Dendrite Nodes" are added to the network and separately trained with a different objective: to correlate their output with the remaining error of the original neurons. The trained Dendrite Nodes are then frozen, and the original neurons are further trained, now taking into account the additional error signals provided by the Dendrite Nodes. The cycle of training the original neurons and then adding and training Dendrite Nodes can be repeated several times until satisfactory performance is achieved. Our algorithm was successfully added to modern state-of-the-art PyTorch networks across multiple domains, improving upon original accuracies and allowing for significant model compression without a loss in accuracy.
This paper presents SANA-1.5, a linear Diffusion Transformer for efficient scaling in text-to-image generation. Building upon SANA-1.0, we introduce three key innovations: (1) Efficient Training Scaling: A depth-growth paradigm that enables scaling from 1.6B to 4.8B parameters with significantly reduced computational resources, combined with a memory-efficient 8-bit optimizer. (2) Model Depth Pruning: A block importance analysis technique for efficient model compression to arbitrary sizes with minimal quality loss. (3) Inference-time Scaling: A repeated sampling strategy that trades computation for model capacity, enabling smaller models to match larger model quality at inference time. Through these strategies, SANA-1.5 achieves a text-image alignment score of 0.81 on GenEval, which can be further improved to 0.96 through inference scaling with VILA-Judge, establishing a new SoTA on GenEval benchmark. These innovations enable efficient model scaling across different compute budgets while maintaining high quality, making high-quality image generation more accessible. Our code and pre-trained models are released.
Recently, state-of-the-art approaches for pruning large pre-trained models (LPMs) have demonstrated that the training-free removal of non-critical residual blocks in Transformers is viable for reducing model size, achieving results that outperform previous training-free pruning approaches. Motivated by these findings, we extend BlockPruner (Zhong et al., 2024) and propose MultiPruner, a pruning approach that surpasses recent training-free pruning methods by adopting a multidimensional, iterative, fine-grained pruning strategy. In MultiPruner, multidimensional pruning reinstates the structural balance in block-pruned models by sequentially compressing along three dimensions: i) residual blocks, ii) channels of multilayer perceptrons (MLP), and iii) attention heads. This solution enhances zero-shot accuracy on downstream tasks compared to other techniques while improving model compression ratios, producing compressed models with fewer computing and memory requirements. Extensive experiments demonstrate the advantages of the proposed method across various large pre-trained models. The code and pruning configurations are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
We present a tensorization algorithm for constructing tensor train representations of functions, drawing on sketching and cross interpolation ideas. The method only requires black-box access to the target function and a small set of sample points defining the domain of interest. Thus, it is particularly well-suited for machine learning models, where the domain of interest is naturally defined by the training dataset. We show that this approach can be used to enhance the privacy and interpretability of neural network models. Specifically, we apply our decomposition to (i) obfuscate neural networks whose parameters encode patterns tied to the training data distribution, and (ii) estimate topological phases of matter that are easily accessible from the tensor train representation. Additionally, we show that this tensorization can serve as an efficient initialization method for optimizing tensor trains in general settings, and that, for model compression, our algorithm achieves a superior trade-off between memory and time complexity compared to conventional tensorization methods of neural networks.
With the rise and ubiquity of larger deep learning models, the need for high-quality compression techniques is growing in order to deploy these models widely. The sheer parameter count of these models makes it difficult to fit them into the memory constraints of different hardware. In this work, we present a novel approach to model compression by merging similar parameter groups within a model, rather than pruning away less important parameters. Specifically, we select, align, and merge separate feed-forward sublayers in Transformer models, and test our method on language modeling, image classification, and machine translation. With our method, we demonstrate performance comparable to the original models while combining more than a third of model feed-forward sublayers, and demonstrate improved performance over a strong layer-pruning baseline. For instance, we can remove over 21% of total parameters from a Vision Transformer, while maintaining 99% of its original performance. Additionally, we observe that some groups of feed-forward sublayers exhibit high activation similarity, which may help explain their surprising mergeability.
We develop a pipeline to streamline neural architecture codesign for physics applications to reduce the need for ML expertise when designing models for novel tasks. Our method employs neural architecture search and network compression in a two-stage approach to discover hardware efficient models. This approach consists of a global search stage that explores a wide range of architectures while considering hardware constraints, followed by a local search stage that fine-tunes and compresses the most promising candidates. We exceed performance on various tasks and show further speedup through model compression techniques such as quantization-aware-training and neural network pruning. We synthesize the optimal models to high level synthesis code for FPGA deployment with the hls4ml library. Additionally, our hierarchical search space provides greater flexibility in optimization, which can easily extend to other tasks and domains. We demonstrate this with two case studies: Bragg peak finding in materials science and jet classification in high energy physics, achieving models with improved accuracy, smaller latencies, or reduced resource utilization relative to the baseline models.
Graph neural networks (GNNs) have demonstrated superior performance in collaborative recommendation through their ability to conduct high-order representation smoothing, effectively capturing structural information within users' interaction patterns. However, existing GNN paradigms face significant challenges in scalability and robustness when handling large-scale, noisy, and real-world datasets. To address these challenges, we present LightGNN, a lightweight and distillation-based GNN pruning framework designed to substantially reduce model complexity while preserving essential collaboration modeling capabilities. Our LightGNN framework introduces a computationally efficient pruning module that adaptively identifies and removes redundant edges and embedding entries for model compression. The framework is guided by a resource-friendly hierarchical knowledge distillation objective, whose intermediate layer augments the observed graph to maintain performance, particularly in high-rate compression scenarios. Extensive experiments on public datasets demonstrate LightGNN's effectiveness, significantly improving both computational efficiency and recommendation accuracy. Notably, LightGNN achieves an 80% reduction in edge count and 90% reduction in embedding entries while maintaining performance comparable to more complex state-of-the-art baselines. The implementation of our LightGNN framework is available at the github repository: https://github.com/HKUDS/LightGNN.
While pruning methods effectively maintain model performance without extra training costs, they often focus solely on preserving crucial connections, overlooking the impact of pruned weights on subsequent fine-tuning or distillation, leading to inefficiencies. Moreover, most compression techniques for generative models have been developed primarily for GANs, tailored to specific architectures like StyleGAN, and research into compressing Diffusion models has just begun. Even more, these methods are often applicable only to GANs or Diffusion models, highlighting the need for approaches that work across both model types. In this paper, we introduce Singular Value Scaling (SVS), a versatile technique for refining pruned weights, applicable to both model types. Our analysis reveals that pruned weights often exhibit dominant singular vectors, hindering fine-tuning efficiency and leading to suboptimal performance compared to random initialization. Our method enhances weight initialization by minimizing the disparities between singular values of pruned weights, thereby improving the fine-tuning process. This approach not only guides the compressed model toward superior solutions but also significantly speeds up fine-tuning. Extensive experiments on StyleGAN2, StyleGAN3 and DDPM demonstrate that SVS improves compression performance across model types without additional training costs. Our code is available at: https://github.com/LAIT-CVLab/Singular-Value-Scaling.
Large Language Models (LLMs) have achieved remarkable success, yet recent findings reveal that their deeper layers often contribute minimally and can be pruned without affecting overall performance. While some view this as an opportunity for model compression, we identify it as a training shortfall rooted in the widespread use of Pre-Layer Normalization (Pre-LN). We demonstrate that Pre-LN, commonly employed in models like GPT and LLaMA, leads to diminished gradient norms in its deeper layers, reducing their effectiveness. In contrast, Post-Layer Normalization (Post-LN) preserves larger gradient norms in deeper layers but suffers from vanishing gradients in earlier layers. To address this, we introduce Mix-LN, a novel normalization technique that combines the strengths of Pre-LN and Post-LN within the same model. Mix-LN applies Post-LN to the earlier layers and Pre-LN to the deeper layers, ensuring more uniform gradients across layers. This allows all parts of the network--both shallow and deep layers--to contribute effectively to training. Extensive experiments with various model sizes from 70M to 7B demonstrate that Mix-LN consistently outperforms both Pre-LN and Post-LN, promoting more balanced, healthier gradient norms throughout the network, and enhancing the overall quality of LLM pre-training. Furthermore, we demonstrate that models pre-trained with Mix-LN learn better compared to those using Pre-LN or Post-LN during supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), highlighting the critical importance of high-quality deep layers. By effectively addressing the inefficiencies of deep layers in current LLMs, Mix-LN unlocks their potential, enhancing model capacity without increasing model size. Our code is available at https://github.com/pixeli99/MixLN.
Since high resolution remote sensing image classification often requires a relatively high computation complexity, lightweight models tend to be practical and efficient. Model pruning is an effective method for model compression. However, existing methods rarely take into account the specificity of remote sensing images, resulting in significant accuracy loss after pruning. To this end, we propose an effective structural pruning approach for remote sensing image classification. Specifically, a pruning strategy that amplifies the differences in channel importance of the model is introduced. Then an adaptive mining loss function is designed for the fine-tuning process of the pruned model. Finally, we conducted experiments on two remote sensing classification datasets. The experimental results demonstrate that our method achieves minimal accuracy loss after compressing remote sensing classification models, achieving state-of-the-art (SoTA) performance.
Knowledge distillation approaches are model compression techniques, with the goal of training a highly performant student model by using a teacher network that is larger or contains a different inductive bias. These approaches are particularly useful when applied to large computer vision foundation models, which can be compressed into smaller variants that retain desirable properties such as improved robustness. This paper presents a label-free knowledge distillation approach called Teacher in the Middle (TinTeM), which improves on previous methods by learning an approximately orthogonal mapping from the latent space of the teacher to the student network. This produces a more faithful student, which better replicates the behavior of the teacher network across a range of benchmarks testing model robustness, generalisability and out-of-distribution detection. It is further shown that knowledge distillation with TinTeM on task specific datasets leads to more accurate models with greater generalisability and OOD detection performance, and that this technique provides a competitive pathway for training highly performant lightweight models on small datasets.
This study examines quantisation and pruning strategies to reduce energy consumption in code Large Language Models (LLMs) inference. Using StarCoder2, we observe increased energy demands with quantization due to lower throughput and some accuracy losses. Conversely, pruning reduces energy usage but impairs performance. The results highlight challenges and trade-offs in LLM model compression. We suggest future work on hardware-optimized quantization to enhance efficiency with minimal loss in accuracy.
The integration of Large Language Models (LLMs) into autonomous driving systems offers promising enhancements in environmental understanding and decision-making. However, the substantial computational demands of deploying LLMs locally on vehicles render this approach unfeasible for real-world automotive applications. To address this challenge, we introduce OWLed, the Outlier-Weighed Layerwise Pruning for Efficient Autonomous Driving Framework that leverages outlier-weighted layerwise sparsity for model compression. Our method assigns non-uniform sparsity ratios to different layers based on the distribution of outlier features, significantly reducing the model size without the need for fine-tuning. To ensure the compressed model adapts well to autonomous driving tasks, we incorporate driving environment data into both the calibration and pruning processes. Our empirical studies reveal that the encoder component is more sensitive to pruning than the LLM, highlighting its critical role in the system. Experimental results demonstrate that OWLed outperforms existing methods in perception, action prediction, and language understanding while substantially lowering computational requirements. These findings underscore the potential of combining advanced pruning techniques with LLMs to develop efficient and robust autonomous driving systems capable of handling complex scenarios. Code will be made publicly available.
With the growth of model sizes and the scale of their deployment, their sheer size burdens the infrastructure requiring more network and more storage to accommodate these. While there is a vast model compression literature deleting parts of the model weights for faster inference, we investigate a more traditional type of compression - one that represents the model in a compact form and is coupled with a decompression algorithm that returns it to its original form and size - namely lossless compression. We present ZipNN a lossless compression tailored to neural networks. Somewhat surprisingly, we show that specific lossless compression can gain significant network and storage reduction on popular models, often saving 33% and at times reducing over 50% of the model size. We investigate the source of model compressibility and introduce specialized compression variants tailored for models that further increase the effectiveness of compression. On popular models (e.g. Llama 3) ZipNN shows space savings that are over 17% better than vanilla compression while also improving compression and decompression speeds by 62%. We estimate that these methods could save over an ExaByte per month of network traffic downloaded from a large model hub like Hugging Face.
This paper introduces a novel model compression approach through dynamic layer-specific pruning in Large Language Models (LLMs), enhancing the traditional methodology established by SliceGPT. By transitioning from constant to dynamic slicing, our method leverages the newly proposed Layer Redundancy (LR) score, which assesses how much change each layer changes its input by measuring the cosine similarity of the input to the output of the layer. We use this score to prune parts of individual layers based on redundancy in such a way that the average pruned percentage for all layers is a fixed value. We conducted extensive experiments using models like Llama3-8B and Mistral-7B on multiple datasets, evaluating different slicing bases and percentages to determine optimal configurations that balance efficiency and performance. Our findings show that our dynamic slicing approach not only maintains but, in many cases, enhances model performance compared to the baseline established by constant slicing methods. For instance, in several settings, we see performance improvements of up to 5% over the SliceGPT baseline. Additionally, a perplexity decrease by as much as 7% was observed across multiple benchmarks, validating the effectiveness of our method. The code, model weights, and datasets are open-sourced at https://github.com/RazvanDu/DynamicSlicing.
Artificial intelligence agents are increasingly capable of performing complex tasks across various domains. As these agents advance, there is a growing need to accurately measure and benchmark their capabilities, particularly in accelerating AI research and development. Current benchmarks focus on general machine learning tasks, but lack comprehensive evaluation methods for assessing AI agents' abilities in tackling research-level problems and competition-level challenges in the field of AI. We present the ML Research Benchmark (MLRB), comprising 7 competition-level tasks derived from recent machine learning conference tracks. These tasks span activities typically undertaken by AI researchers, including model training efficiency, pretraining on limited data, domain specific fine-tuning, and model compression. This paper introduces a novel benchmark and evaluates it using agent scaffolds powered by frontier models, including Claude-3 and GPT-4o. The results indicate that the Claude-3.5 Sonnet agent performs best across our benchmark, excelling in planning and developing machine learning models. However, both tested agents struggled to perform non-trivial research iterations. We observed significant performance variations across tasks, highlighting the complexity of AI development and the challenges in creating versatile agent scaffolds. While current AI agents can successfully navigate complex instructions and produce baseline results, they fall short of the capabilities required for advanced AI research. The ML Research Benchmark provides a valuable framework for assessing and comparing AI agents on tasks mirroring real-world AI research challenges.
Although large language models (LLMs) have demonstrated their strong intelligence ability, the high demand for computation and storage hinders their practical application. To this end, many model compression techniques are proposed to increase the efficiency of LLMs. However, current researches only validate their methods on limited models, datasets, metrics, etc, and still lack a comprehensive evaluation under more general scenarios. So it is still a question of which model compression approach we should use under a specific case. To mitigate this gap, we present the Large Language Model Compression Benchmark (LLMCBench), a rigorously designed benchmark with an in-depth analysis for LLM compression algorithms. We first analyze the actual model production requirements and carefully design evaluation tracks and metrics. Then, we conduct extensive experiments and comparison using multiple mainstream LLM compression approaches. Finally, we perform an in-depth analysis based on the evaluation and provide useful insight for LLM compression design. We hope our LLMCBench can contribute insightful suggestions for LLM compression algorithm design and serve as a foundation for future research. Our code is available at https://github.com/AboveParadise/LLMCBench.
The high computational costs of large language models (LLMs) have led to a flurry of research on LLM compression, via methods such as quantization, sparsification, or structured pruning. A new frontier in this area is given by \emph{dynamic, non-uniform} compression methods, which adjust the compression levels (e.g., sparsity) per-block or even per-layer in order to minimize accuracy loss, while guaranteeing a global compression threshold. Yet, current methods rely on heuristics for identifying the "importance" of a given layer towards the loss, based on assumptions such as \emph{error monotonicity}, i.e. that the end-to-end model compression error is proportional to the sum of layer-wise errors. In this paper, we revisit this area, and propose a new and general approach for dynamic compression that is provably optimal in a given input range. We begin from the motivating observation that, in general, \emph{error monotonicity does not hold for LLMs}: compressed models with lower sum of per-layer errors can perform \emph{worse} than models with higher error sums. To address this, we propose a new general evolutionary framework for dynamic LLM compression called EvoPress, which has provable convergence, and low sample and evaluation complexity. We show that these theoretical guarantees lead to highly competitive practical performance for dynamic compression of Llama, Mistral and Phi models. Via EvoPress, we set new state-of-the-art results across all compression approaches: structural pruning (block/layer dropping), unstructured sparsity, as well as quantization with dynamic bitwidths. Our code is available at https://github.com/IST-DASLab/EvoPress.