Attention Head Pruning

Model merging integrates task-specific fine-tuned models into a single multi-task model, but often suffers from parameter interference caused by conflicting task-vector updates. Existing methods typically mitigate conflicts by pruning task vectors based on weight magnitude or random heuristics, treating Transformers as unstructured ``bags of parameters'' and overlooking their inherent modularity. In this paper, we propose \textbf{C}ontribution-\textbf{A}ware \textbf{S}tructured \textbf{S}parsity (CASS), a unified framework that reduces parameter interference by identifying and preserving task-specific components. At the core of CASS is a contribution-aware structured mask that identifies task-relevant attention heads and FFN neurons. We instantiate this mask in two settings: CASS-Merging, the primary post-hoc setting where masks serve as a plug-and-play denoising filter for existing merging operators, and CASS-Tuning, an extension for scenarios with fine-tuning access where masks constrain gradients to reduce structural overlap between task vectors. Our analysis shows that task-relevant components are sparse and partially disjoint, supporting structured component-level filtering as an effective way to reduce merging interference. Extensive experiments across vision (ViT, 20 tasks) and language (RoBERTa, 8 tasks; Qwen2.5, 4 tasks) benchmarks demonstrate that CASS improves a range of representative merging baselines.

Vision foundation models, such as DINOv2, learn highly expressive representations but rely on massive, opaque architectures that demand substantial computational power and memory. To provide an interpretable-guided and efficient solution to this issue, we first propose a spectral analysis and new visualization technique for individual attention heads based on the Laplacian eigenvectors of their attention maps. Building upon recent observations regarding the block structure of Vision Transformers, we perform semantic clustering of attention heads and identify functional redundancies. Leveraging these insights, we introduce SAPER (Soft Attention PrunER), an end-to-end differentiable pruning framework based on the LapSum Soft Top-K approach. Extensive experiments on ImageNet-1K demonstrate that SAPER achieves a highly favorable accuracy-efficiency trade-off, outperforming the competitive RAPTOR baseline in FLOPs reduction while preserving strong classification performance.

Pruning is a promising approach for improving the efficiency of LLMs. Existing static structured pruning methods are hardware-friendly and can deliver practical throughput gains, but their input-agnostic computation allocation often causes substantial accuracy degradation under aggressive sparsity. Recent dynamic sparsity methods improve quality retention by adapting computation to individual inputs, yet they remain largely limited to coarse-grained structural decisions and their practical acceleration under real-world inference scenarios remains challenging. To address these challenges, we present WIDE, the first end-to-end differentiable token-level dynamic width pruning framework designed for both prefill and decode scenarios. WIDE enables fine-grained computation allocation by allowing each token to dynamically select attention-head groups and FFN-channel groups, extending dynamic pruning beyond layer-level decisions to neuron-block-level granularity. Through a two-stage training pipeline, WIDE learns effective token-wise sparse execution patterns and achieves substantially better quality retention than existing approaches. To make such fine-grained dynamic pruning practical, we further propose a pruning--kernel co-design framework that decomposes dynamic sparsity acceleration into mask reordering, hardware-agnostic block-level skipping, and hardware-dependent intra-block skipping, enabling efficient execution across different granularities. At 50% sparsity, WIDE provides 55.1% performance boost when compared to the state-of-the-art dynamic depth pruning under calibration-only settings. Under prefill and decoding inference workloads, WIDE achieves close-to-theoretical kernel-level speedups of up to 1.98x for prefill and 4.95x for decoding, as well as 1.68x and 1.55x end-to-end acceleration. Our code is available at https://github.com/EIT-NLP/LLM-Pruning/tree/main/WIDE.

Modern text-to-image diffusion transformers (DiTs) generate images through joint attention, in which text and image tokens interact directly within a single sequence. In large-scale DiTs, the conditioning input contains not only the user prompt but also chat-template tokens introduced by LLM-based text encoders. Yet how these tokens participate in the denoising computation remains poorly understood. To probe this, we introduce a causal interpretability framework. Using it to separate prompt-content tokens from chat-template tokens, we find that the template tokens carry little prompt-specific information at the encoder output. Yet surprisingly, they emerge as dominant image-to-text attention sinks and causally maintain object identity inside the DiT, acting as implicit semantic registers. We show that they acquire this identity indirectly. Rather than reading the prompt tokens, they draw the identity from the image latents into which the prompt semantics have already been injected at the very first layer. We further reveal a division of labor across heads and depth in DiTs, where distinct heads route semantics or render visual structure, and identity is committed in early blocks, carried by middle blocks, and refined in late ones. As a practical payoff, this analysis yields a training-free pruning rule that removes the causally inert prompt-reading heads and cuts $20\%$ of joint-attention FLOPs at a $1.4$-point cost in GenEval accuracy. Overall, our work not only reveals that the tokens encoding semantics at the input need not be those that maintain them during generation, but also provides a causal view of internal mechanisms in diffusion transformers.

Structured pruning compresses large language models (LLMs) by removing whole computational units, such as attention heads and feed-forward (FFN) channel groups. Most training-free methods, however, rank these units independently, implicitly treating the loss from pruning a set as the sum of its individual losses. This view fails for Transformers, whose sublayers are coupled through a shared residual stream. Two individually weak units can thus be jointly indispensable, yet independent scoring is blind to such dependence and removes them together. We introduce CoCurve (Cross-Module Co-Pruning Curvature), a calibration-only, fine-tuning-free method that prunes attention and FFN units jointly. A second-order Taylor expansion of the token-level KL between the frozen model and its masked copy yields a single Fisher matrix whose diagonal is classical node saliency and whose off-diagonal entries are co-pruning curvature edges: the extra damage of removing two units together. Under a single-ablation additivity approximation this matrix reduces to a Gram product of single-unit ablation features, so the full M x M interaction is recovered from M forward passes, with no pairwise sweeps or gradients. Pruning then reduces to one budgeted quadratic program, solved in a single shot under a shared attention--FFN budget, with no labels, fine-tuning, or recovery.

Recent advances in video diffusion have been driven by scaling transformer-based architectures to billions of parameters, substantially improving visual fidelity and motion coherence. In contrast, existing mobile video diffusion models remain limited to relatively small parameter budgets, typically 0.4-1.8B, restricting generation quality. In this work, we show that high-quality mobile video generation does not require small models. Instead, we demonstrate that a server-scale 5B-parameter video diffusion transformer can be deployed efficiently on memory-constrained mobile hardware through recurrent reformulation and structured compression. Starting from Wan2.2-5B, we rely on a recurrence distillation framework that converts video generation into a chunk-wise autoregressive process with constant-memory attention computation. Combined with causal linear attention, the model operates as an RNN at inference time while preserving temporal coherence across chunks. We further propose a learnable attention head pruning method based on binary per-head gates optimized end-to-end using a noise-biased sparsity objective and distillation-based finetuning. Together with sampling-step distillation and memory-optimized VAE decoding, MobileWan becomes the first 5B-scale video diffusion model deployable on a commercial mobile device. Our system generates 5-second 480x832 videos at 16 FPS in 20 seconds end-to-end latency, achieving a VBench score of 83.79 and establishing a new state of the art in mobile video generation. Project page: https://qualcomm-ai-research.github.io/mobilewan

The remarkable success of Transformer-based models in natural language processing stems from architectural scaling, which leads to a large number of parameters and hinders deployment in resource-constrained environments. While structured pruning offers a pathway to compression, existing state-of-the-art methods often rely on gradient-based importance ranking or stochastic gating, which suffer from instability, structural degeneration, and the need for extensive manual hyperparameter tuning. In this paper, we introduce CAHP (Complementary Attention Head Pruning), a novel post-hoc framework that redefines head selection as a global graph-theoretical problem. Rather than evaluating heads in isolation, CAHP utilizes graph-based clustering combined with information-theoretic distance measures to identify and preserve a topologically diverse subset of complementary attention heads. Without requiring a predefined sparsity level or pruning ratio, the framework automatically determines the number of selected attention heads across layers by identifying a diminishing marginal performance curve, where pruning additional heads leads to a sharp degradation in performance, as determined by the chosen polynomial degree. Extensive evaluations on the SST-5 and MNLI benchmarks, across different Transformer model scales, demonstrate that CAHP consistently outperforms competitive baselines, particularly in high-compression regimes. Furthermore, our structural analysis shows that CAHP avoids the "proximity bias" of gradient-based pruning methods, which tend to preserve heads mainly in layers close to the output, and instead retains a functionally critical set of attention heads in the model's intermediate layers.

Adapting Large Language Models (LLMs) to specialized domains typically incurs high data and computational overhead. While prior efficiency efforts have largely treated data selection and parameter-efficient fine-tuning as isolated processes, our empirical analysis suggests they may be intrinsically coupled. We posit the Strong Map Hypothesis: a sparse subset of attention heads plays a dominant role in task-specific adaptation, acting as keys that unlock specific data patterns. Building on this observation, we propose From Parameters to Data (P2D), a unified framework that leverages these task-sensitive attention heads as a dual compass for both sample mining and structural pruning. To rigorously quantify the total pipeline cost, we introduce the Alignment Efficiency Ratio (AER) metric for both selection latency and training time. Mechanistically, P2D identifies critical heads via a lightweight proxy and uses them as a functional filter to curate high-affinity data, establishing a synergistic pipeline. Empirically, by updating merely 10% of attention heads on 10% of the data, P2D achieves an 8.3 pp performance gain over strong baselines and delivers a 7.0x end-to-end time speedup. These results validate that precise parameter-data synchronization eliminates redundancy, offering a new paradigm for efficient alignment.

Large Language Models (LLMs) have experienced significant growth and development in recent years. However, performing inference on LLMs remains costly, especially for long-context inference or in resource-constrained devices. This motivates the development of new post-training pruning (PTP) methods. These methods reduce LLMs' requirements by removing a substantial part of the model's parameters. The discarded weights are selected depending on their impact on the models performance. Current PTP methods prune the models by removing the less informative hidden nodes from the FFN layers, and the least important attention layers. We propose Putri, a PTP method that introduces three changes to the State-of-the-art. First, we update the un-pruned weights of the FFN to compensate for the introduced pruning error. Second, the FFN layers are pruned sequentially, taking into account the updates done to the previous layers. Third, instead of removing full attention layers, we remove individual attention-heads. We extend this method such that it can also address Grouped-Query Attention. In summary, Putri is a structure pruning method which remains simple while showing SOTA performance. Pruning experiments on multiple models with a wide variety of sparsity ranges and on different datasets, validate the generality of Putri. Notably, we demonstrate that, unlike previous methods, Putri can prune LLMs on extreme sparsity ratios. The code is available at: https://github.com/Coello-dev/Putri.

Large language models (LLMs) excel at multi-step reasoning but incur substantial inference cost. We introduce Causal Attribution Pruning (CAP), a training-free method that identifies critical attention heads by measuring their causal impact on reasoning tasks and uses these head-level scores to guide fine-grained weight pruning. For each attention head, CAP estimates the expected performance degradation when the head is masked during forward passes on a small calibration set of reasoning problems. These causal scores are then converted into weight-level importance values for the corresponding projection matrices. Unlike magnitude-only or activation-based criteria, CAP's interventional measurement directly captures each head's functional contribution, yielding relative accuracy gains of up to 61% over Wanda on ARC-Challenge at 20% sparsity. We evaluate CAP on GSM8K, StrategyQA, and ARC-Challenge using Llama-3-8B-Instruct and Mistral-7B-Instruct at 10%, 20%, and 50% sparsity. At moderate sparsity (10-20%), CAP improves over Wanda in most model-benchmark configurations. with especially large gains on ARC-Challenge for Llama-3. Our results suggest that attention-head-level causal attribution can better preserve reasoning performance on downstream benchmarks than correlational pruning criteria at equivalent sparsity, while remaining limited by coarse MLP attribution at 50% sparsity.

Vision Transformers require significant computational resources and memory bandwidth, severely limiting their deployment on resource-constraint hardware. Most structured pruning methods reduce theoretical cost effectively, yet they typically operate at a single structural granularity and depend on multi-stage pipelines with importance ranking, auxiliary solvers or post-hoc magnitude thresholding, followed by a separate fine-tuning phase to recover accuracy. We propose Hierarchical Auto-Pruning (HiAP), which casts ViT pruning as a single budget-aware learning problem and jointly allocates sparsity across four granularities in one end-to-end phase. HiAP introduces stochastic Gumbel-Sigmoid gates at macro level (attention heads and FFN blocks) and micro level (intra-head dimensions and FFN neurons), and optimizes them against the task loss together with an analytical MAC cost term. The budget coefficient steers the network to a target compute level while the gates gradually harden into a dense, smaller sub-network at convergence. It does not require importance heuristics, ranking metrics, auxillary solvers or secondary fine-tuning. On ImageNet with DeiT small, HiAP automatically discovers hetergenous architectures, pruning depths, heads, and width by different amount across layers, and reaches competitive accuracies against substantially more complex pruning pipelines at comparable compute from a single training run.

Large Language Models (LLMs) are known to contain significant redundancy, yet a systematic explanation for why certain components, particularly in higher layers, are more redundant has remained elusive. In this work, we identify the BOS sink phenomenon as a key mechanism driving this layer-wise sensitivity. We show that attention heads with high BOS sink scores are strongly associated with functional redundancy: such heads, especially in deeper layers, contribute little to predictive performance and effectively serve as dumping grounds for superfluous attention weights. Leveraging this insight, we introduce a simple pruning strategy that removes high-BOS sink heads. Experiments on Gemma-3, Llama-3.1, and Qwen3 demonstrate that this approach identifies redundant transformer components more reliably than weight- and activation-based criteria in terms of downstream task retention, remaining close to dense baselines at low-to-moderate pruning ratios. We further find that high-scoring sink heads sustain their focus on BOS as context length grows. Overall, our results suggest that structural properties of attention offer a more direct basis for model compression than magnitude-based methods.

We study continual pre-training (CPT) as a mechanism for adapting general-purpose large language models to specialized domains: mathematics, instruction, code, and natural text. Using singular value decomposition of weight matrices, we find that CPT leaves singular value spectra largely invariant, with adaptation driven mainly by changes in singular vectors. An analysis of attention-head projection matrices reveals strong, domain-dependent head heterogeneity, which we exploit to define a head importance criterion: up to 60% of head updates can be removed without measurable quality loss. Selectively rewinding low-importance heads to their pre-trained state improves benchmark accuracy by up to 4% versus the fully trained baseline. Finally, we identify domain connectivity - linear interpolation between CPT checkpoints yields smooth domain-quality interpolation without notable degradation on either domain - and release Diffract, an open-source toolkit for scalable spectral analysis of billion-parameter models.