Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories -- Knowledge Editing, Model Compression, and Machine Unlearning. Our results from 310 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions. All of our code is public: https://github.com/hartvigsen-group/composable-interventions.
Increasingly, model compression techniques enable large language models (LLMs) to be deployed in real-world applications. As a result of this momentum towards local deployment, compressed LLMs will interact with a large population. Prior work on compression typically prioritize preserving perplexity, which is directly analogous to training loss. The impact of compression method on other critical aspects of model behavior\, -- \,particularly safety\, -- \,requires systematic assessment. To this end, we investigate the impact of model compression along four dimensions: (1) degeneration harm, i.e., bias and toxicity in generation; (2) representational harm, i.e., biases in discriminative tasks; (3) dialect bias; and(4) language modeling and downstream task performance. We examine a wide spectrum of LLM compression techniques, including unstructured pruning, semi-structured pruning, and quantization. Our analysis reveals that compression can lead to unexpected consequences. Although compression may unintentionally alleviate LLMs' degeneration harm, it can still exacerbate representational harm. Furthermore, increasing compression produces a divergent impact on different protected groups. Finally, different compression methods have drastically different safety impacts: for example, quantization mostly preserves bias while pruning degrades quickly. Our findings underscore the importance of integrating safety assessments into the development of compressed LLMs to ensure their reliability across real-world applications.\footnote{Our implementation and results are available here: \url{https://github.com/zhichaoxu-shufe/Beyond-Perplexity-Compression-Safety-Eval}}
Recent advancements in diffusion models, particularly the architectural transformation from UNet-based models to Diffusion Transformers (DiTs), significantly improve the quality and scalability of image and video generation. However, despite their impressive capabilities, the substantial computational costs of these large-scale models pose significant challenges for real-world deployment. Post-Training Quantization (PTQ) emerges as a promising solution, enabling model compression and accelerated inference for pretrained models, without the costly retraining. However, research on DiT quantization remains sparse, and existing PTQ frameworks, primarily designed for traditional diffusion models, tend to suffer from biased quantization, leading to notable performance degradation. In this work, we identify that DiTs typically exhibit significant spatial variance in both weights and activations, along with temporal variance in activations. To address these issues, we propose Q-DiT, a novel approach that seamlessly integrates two key techniques: automatic quantization granularity allocation to handle the significant variance of weights and activations across input channels, and sample-wise dynamic activation quantization to adaptively capture activation changes across both timesteps and samples. Extensive experiments conducted on ImageNet and VBench demonstrate the effectiveness of the proposed Q-DiT. Specifically, when quantizing DiT-XL/2 to W6A8 on ImageNet ($256 \times 256$), Q-DiT achieves a remarkable reduction in FID by 1.09 compared to the baseline. Under the more challenging W4A8 setting, it maintains high fidelity in image and video generation, establishing a new benchmark for efficient, high-quality quantization in DiTs. Code is available at \href{https://github.com/Juanerx/Q-DiT}{https://github.com/Juanerx/Q-DiT}.
While large language models (LLMs) excel in many domains, their complexity and scale challenge deployment in resource-limited environments. Current compression techniques, such as parameter pruning, often fail to effectively utilize the knowledge from pruned parameters. To address these challenges, we propose Manifold-Based Knowledge Alignment and Layer Merging Compression (MKA), a novel approach that uses manifold learning and the Normalized Pairwise Information Bottleneck (NPIB) measure to merge similar layers, reducing model size while preserving essential performance. We evaluate MKA on multiple benchmark datasets and various LLMs. Our findings show that MKA not only preserves model performance but also achieves substantial compression ratios, outperforming traditional pruning methods. Moreover, when coupled with quantization, MKA delivers even greater compression. Specifically, on the MMLU dataset using the Llama3-8B model, MKA achieves a compression ratio of 43.75% with a minimal performance decrease of only 2.82\%. The proposed MKA method offers a resource-efficient and performance-preserving model compression technique for LLMs.
Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by $3.9\times$ with the same average attention span, boosting retrieval accuracy by $1.5-7.1\times$ over the uniform-attention baseline across Vicuna-{7B,13B}, and Llama3-{8B,70B} models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from $9\%-36\%$ to within $5\%$ across two long-context understanding benchmarks. MoA achieves a $1.2-1.4\times$ GPU memory reduction, boosting decode throughput by $6.6-8.2\times$ and $1.7-1.9\times$ compared to FlashAttention2 and vLLM, with minimal impact on performance. Our code is available at \url{https://github.com/thu-nics/MoA}.
Low-Rank Adaptation (LoRA) methods have gained popularity in efficient parameter fine-tuning of models containing hundreds of billions of parameters. In this work, instead, we demonstrate the application of LoRA methods to train small-vision models in Federated Learning (FL) from scratch. We first propose an aggregation-agnostic method to integrate LoRA within FL, named FLoCoRA, showing that the method is capable of reducing communication costs by 4.8 times, while having less than 1% accuracy degradation, for a CIFAR-10 classification task with a ResNet-8. Next, we show that the same method can be extended with an affine quantization scheme, dividing the communication cost by 18.6 times, while comparing it with the standard method, with still less than 1% of accuracy loss, tested with on a ResNet-18 model. Our formulation represents a strong baseline for message size reduction, even when compared to conventional model compression works, while also reducing the training memory requirements due to the low-rank adaptation.
Large Language Models~(LLMs) have become foundational in the realm of natural language processing, demonstrating performance improvements as model sizes increase. The Mixture-of-Experts~(MoE) approach offers a promising way to scale LLMs more efficiently by using fewer computational FLOPs through sparse activation. However, it suffers from significant memory overheads, necessitating model compression techniques. Post-training quantization, a popular method for model compression, proves less effective when directly applied to MoE models due to MoE's overlooked inherent sparsity. This paper explores several MoE structure-aware quantization heuristics, ranging from coarse to fine granularity, from MoE block to individual linear weight. Our investigations reveal critical principles: different MoE structures (i.e., blocks, experts, linear layers) require varying numbers of weight bits for effective and efficient quantization. Conclusions are supported by extensive benchmarking across two representative MoE models and six tasks. We further introduce novel enhancements to more accurately identify the most critical weights in MoE quantization that necessitate higher bit allocations, including the linear weight outlier scorer and MoE block scorer. Additionally, subsequent experiments validate our findings in the context of both weight and activation quantization.
The ability of machine learning (ML) algorithms to generalize well to unseen data has been studied through the lens of information theory, by bounding the generalization error with the input-output mutual information (MI), i.e., the MI between the training data and the learned hypothesis. Yet, these bounds have limited practicality for modern ML applications (e.g., deep learning), due to the difficulty of evaluating MI in high dimensions. Motivated by recent findings on the compressibility of neural networks, we consider algorithms that operate by slicing the parameter space, i.e., trained on random lower-dimensional subspaces. We introduce new, tighter information-theoretic generalization bounds tailored for such algorithms, demonstrating that slicing improves generalization. Our bounds offer significant computational and statistical advantages over standard MI bounds, as they rely on scalable alternative measures of dependence, i.e., disintegrated mutual information and $k$-sliced mutual information. Then, we extend our analysis to algorithms whose parameters do not need to exactly lie on random subspaces, by leveraging rate-distortion theory. This strategy yields generalization bounds that incorporate a distortion term measuring model compressibility under slicing, thereby tightening existing bounds without compromising performance or requiring model compression. Building on this, we propose a regularization scheme enabling practitioners to control generalization through compressibility. Finally, we empirically validate our results and achieve the computation of non-vacuous information-theoretic generalization bounds for neural networks, a task that was previously out of reach.
Pre-trained large language models (LLMs) based on Transformer have demonstrated striking in-context learning (ICL) abilities. With a few demonstration input-label pairs, they can predict the label for an unseen input without any parameter updates. In this paper, we show an exciting phenomenon that SVD-based weight pruning can enhance ICL performance, and more surprising, pruning weights in deep layers often results in more stable performance improvements than in shallow layers. However, the underlying mechanism of those findings still remains an open question. To reveal those findings, we conduct an in-depth theoretical analysis by presenting the implicit gradient descent (GD) trajectories of ICL and giving the mutual information based generalization bounds of ICL via full implicit GD trajectories. This helps us reasonably explain the surprising experimental findings. Besides, based on all our experimental and theoretical insights, we intuitively propose a simple, model-compression and derivative-free algorithm for downstream tasks in enhancing ICL inference. Experiments on benchmark datasets and open source LLMs display the method effectiveness\footnote{The code is available at \url{https://github.com/chen123CtrlS/EnhancingICL_SVDPruning}.}.
Benefiting from well-trained deep neural networks (DNNs), model compression have captured special attention for computing resource limited equipment, especially edge devices. Knowledge distillation (KD) is one of the widely used compression techniques for edge deployment, by obtaining a lightweight student model from a well-trained teacher model released on public platforms. However, it has been empirically noticed that the backdoor in the teacher model will be transferred to the student model during the process of KD. Although numerous KD methods have been proposed, most of them focus on the distillation of a high-performing student model without robustness consideration. Besides, some research adopts KD techniques as effective backdoor mitigation tools, but they fail to perform model compression at the same time. Consequently, it is still an open problem to well achieve two objectives of robust KD, i.e., student model's performance and backdoor mitigation. To address these issues, we propose RobustKD, a robust knowledge distillation that compresses the model while mitigating backdoor based on feature variance. Specifically, RobustKD distinguishes the previous works in three key aspects: (1) effectiveness: by distilling the feature map of the teacher model after detoxification, the main task performance of the student model is comparable to that of the teacher model; (2) robustness: by reducing the characteristic variance between the teacher model and the student model, it mitigates the backdoor of the student model under backdoored teacher model scenario; (3) generic: RobustKD still has good performance in the face of multiple data models (e.g., WRN 28-4, Pyramid-200) and diverse DNNs (e.g., ResNet50, MobileNet).
Deep learning neural network models must be large enough to adapt to their problem domain, while small enough to avoid overfitting training data during gradient descent. To balance these competing demands, overprovisioned deep learning models such as transformers are trained for a single epoch on large data sets, and hence inefficient with both computing resources and training data. In response to these inefficiencies, we exploit learning theory to derive Occam Gradient Descent, an algorithm that interleaves adaptive reduction of model size to minimize generalization error, with gradient descent on model weights to minimize fitting error. In contrast, traditional gradient descent greedily minimizes fitting error without regard to generalization error. Our algorithm simultaneously descends the space of weights and topological size of any neural network without modification. With respect to loss, compute and model size, our experiments show (a) on image classification benchmarks, linear and convolutional neural networks trained with Occam Gradient Descent outperform traditional gradient descent with or without post-train pruning; (b) on a range of tabular data classification tasks, neural networks trained with Occam Gradient Descent outperform traditional gradient descent, as well as Random Forests; (c) on natural language transformers, Occam Gradient Descent outperforms traditional gradient descent.
Federated Learning (FL) is commonly used in systems with distributed and heterogeneous devices with access to varying amounts of data and diverse computing and storage capacities. FL training process enables such devices to update the weights of a shared model locally using their local data and then a trusted central server combines all of those models to generate a global model. In this way, a global model is generated while the data remains local to devices to preserve privacy. However, training large models such as Deep Neural Networks (DNNs) on resource-constrained devices can take a prohibitively long time and consume a large amount of energy. In the current process, the low-capacity devices are excluded from the training process, although they might have access to unseen data. To overcome this challenge, we propose a model compression approach that enables heterogeneous devices with varying computing capacities to participate in the FL process. In our approach, the server shares a dense model with all devices to train it: Afterwards, the trained model is gradually compressed to obtain submodels with varying levels of sparsity to be used as suitable initial global models for resource-constrained devices that were not capable of train the first dense model. This results in an increased participation rate of resource-constrained devices while the transferred weights from the previous round of training are preserved. Our validation experiments show that despite reaching about 50 per cent global sparsity, generated submodels maintain their accuracy while can be shared to increase participation by around 50 per cent.
Motivated by the huge success of Transformers in the field of natural language processing (NLP), Vision Transformers (ViTs) have been rapidly developed and achieved remarkable performance in various computer vision tasks. However, their huge model sizes and intensive computations hinder ViTs' deployment on embedded devices, calling for effective model compression methods, such as quantization. Unfortunately, due to the existence of hardware-unfriendly and quantization-sensitive non-linear operations, particularly {Softmax}, it is non-trivial to completely quantize all operations in ViTs, yielding either significant accuracy drops or non-negligible hardware costs. In response to challenges associated with \textit{standard ViTs}, we focus our attention towards the quantization and acceleration for \textit{efficient ViTs}, which not only eliminate the troublesome Softmax but also integrate linear attention with low computational complexity, and propose Trio-ViT accordingly. Specifically, at the algorithm level, we develop a {tailored post-training quantization engine} taking the unique activation distributions of Softmax-free efficient ViTs into full consideration, aiming to boost quantization accuracy. Furthermore, at the hardware level, we build an accelerator dedicated to the specific Convolution-Transformer hybrid architecture of efficient ViTs, thereby enhancing hardware efficiency. Extensive experimental results consistently prove the effectiveness of our Trio-ViT framework. {Particularly, we can gain up to $\uparrow$$\mathbf{3.6}\times$, $\uparrow$$\mathbf{5.0}\times$, and $\uparrow$$\mathbf{7.3}\times$ FPS under comparable accuracy over state-of-the-art ViT accelerators, as well as $\uparrow$$\mathbf{6.0}\times$, $\uparrow$$\mathbf{1.5}\times$, and $\uparrow$$\mathbf{2.1}\times$ DSP efficiency.} Codes are available at \url{https://github.com/shihuihong214/Trio-ViT}.
The development of model compression is continuously motivated by the evolution of various neural network accelerators with ASIC or FPGA. On the algorithm side, the ultimate goal of quantization or pruning is accelerating the expensive DNN computations on low-power hardware. However, such a "design-and-deploy" workflow faces under-explored challenges in the current hardware-algorithm co-design community. First, although the state-of-the-art quantization algorithm can achieve low precision with negligible degradation of accuracy, the latest deep learning framework (e.g., PyTorch) can only support non-customizable 8-bit precision, data format, and parameter extraction. Secondly, the objective of quantization is to enable the computation with low-precision data. However, the current SoTA algorithm treats the quantized integer as an intermediate result, while the final output of the quantizer is the "discretized" floating-point values, ignoring the practical needs and adding additional workload to hardware designers for integer parameter extraction and layer fusion. Finally, the compression toolkits designed by the industry are constrained to their in-house product or a handful of algorithms. The limited degree of freedom in the current toolkit and the under-explored customization hinder the prototype ASIC or FPGA-based accelerator design. To resolve these challenges, we propose Torch2Chip, an open-sourced, fully customizable, and high-performance toolkit that supports user-designed compression followed by automatic model fusion and parameter extraction. Torch2Chip incorporates the hierarchical design workflow, and the user-customized compression algorithm will be directly packed into the deployment-ready format for prototype chip verification with either CNN or vision transformer (ViT). The code is available at https://github.com/SeoLabCornell/torch2chip.
Data-free knowledge distillation (DFKD) is a promising approach for addressing issues related to model compression, security privacy, and transmission restrictions. Although the existing methods exploiting DFKD have achieved inspiring achievements in coarse-grained classification, in practical applications involving fine-grained classification tasks that require more detailed distinctions between similar categories, sub-optimal results are obtained. To address this issue, we propose an approach called DFKD-FGVC that extends DFKD to fine-grained visual categorization~(FGVC) tasks. Our approach utilizes an adversarial distillation framework with attention generator, mixed high-order attention distillation, and semantic feature contrast learning. Specifically, we introduce a spatial-wise attention mechanism to the generator to synthesize fine-grained images with more details of discriminative parts. We also utilize the mixed high-order attention mechanism to capture complex interactions among parts and the subtle differences among discriminative features of the fine-grained categories, paying attention to both local features and semantic context relationships. Moreover, we leverage the teacher and student models of the distillation framework to contrast high-level semantic feature maps in the hyperspace, comparing variances of different categories. We evaluate our approach on three widely-used FGVC benchmarks (Aircraft, Cars196, and CUB200) and demonstrate its superior performance.
Recent success of pre-trained foundation vision-language models makes Open-Vocabulary Segmentation (OVS) possible. Despite the promising performance, this approach introduces heavy computational overheads for two challenges: 1) large model sizes of the backbone; 2) expensive costs during the fine-tuning. These challenges hinder this OVS strategy from being widely applicable and affordable in real-world scenarios. Although traditional methods such as model compression and efficient fine-tuning can address these challenges, they often rely on heuristics. This means that their solutions cannot be easily transferred and necessitate re-training on different models, which comes at a cost. In the context of efficient OVS, we target achieving performance that is comparable to or even better than prior OVS works based on large vision-language foundation models, by utilizing smaller models that incur lower training costs. The core strategy is to make our efficiency principled and thus seamlessly transferable from one OVS framework to others without further customization. Comprehensive experiments on diverse OVS benchmarks demonstrate our superior trade-off between segmentation accuracy and computation costs over previous works. Our code is available on https://github.com/Xujxyang/OpenTrans
Large Language Models (LLMs) have ushered in a new era in Natural Language Processing, but their massive size demands effective compression techniques for practicality. Although numerous model compression techniques have been investigated, they typically rely on a calibration set that overlooks the multilingual context and results in significant accuracy degradation for low-resource languages. This paper introduces Multilingual Brain Surgeon (MBS), a novel calibration data sampling method for multilingual LLMs compression. MBS overcomes the English-centric limitations of existing methods by sampling calibration data from various languages proportionally to the language distribution of the model training datasets. Our experiments, conducted on the BLOOM multilingual LLM, demonstrate that MBS improves the performance of existing English-centric compression methods, especially for low-resource languages. We also uncover the dynamics of language interaction during compression, revealing that the larger the proportion of a language in the training set and the more similar the language is to the calibration language, the better performance the language retains after compression. In conclusion, MBS presents an innovative approach to compressing multilingual LLMs, addressing the performance disparities and improving the language inclusivity of existing compression techniques.
This paper introduces LLM-Streamline, a pioneer work on layer pruning for large language models (LLMs). It is based on the observation that different layers have varying impacts on hidden states, enabling the identification of less important layers to be pruned.LLM-Streamline comprises two parts: layer pruning, which removes consecutive layers with the lowest importance based on target sparsity, and layer replacement, a novel module that trains a lightweight network to replace the pruned layers to mitigate performance loss. Additionally, a new metric called stability is proposed to address the limitations of the widely used accuracy metric in evaluating model compression. Experiments show that LLM-Streamline outperforms both previous and concurrent state-of-the-art pruning methods in terms of both performance and training efficiency.Our code is available at https://github.com/RUCKBReasoning/LLM-Streamline
The rise of Modular Deep Learning showcases its potential in various Natural Language Processing applications. Parameter-efficient fine-tuning (PEFT) modularity has been shown to work for various use cases, from domain adaptation to multilingual setups. However, all this work covers the case where the modular components are trained and deployed within one single Pre-trained Language Model (PLM). This model-specific setup is a substantial limitation on the very modularity that modular architectures are trying to achieve. We ask whether current modular approaches are transferable between models and whether we can transfer the modules from more robust and larger PLMs to smaller ones. In this work, we aim to fill this gap via a lens of Knowledge Distillation, commonly used for model compression, and present an extremely straightforward approach to transferring pre-trained, task-specific PEFT modules between same-family PLMs. Moreover, we propose a method that allows the transfer of modules between incompatible PLMs without any change in the inference complexity. The experiments on Named Entity Recognition, Natural Language Inference, and Paraphrase Identification tasks over multiple languages and PEFT methods showcase the initial potential of transferable modularity.
To reduce the inference cost of large language models, model compression is increasingly used to create smaller scalable models. However, little is known about their robustness to minority subgroups defined by the labels and attributes of a dataset. In this paper, we investigate the effects of 18 different compression methods and settings on the subgroup robustness of BERT language models. We show that worst-group performance does not depend on model size alone, but also on the compression method used. Additionally, we find that model compression does not always worsen the performance on minority subgroups. Altogether, our analysis serves to further research into the subgroup robustness of model compression.