Semi-structured N:M sparsity has emerged as a practical direction for accelerating large language models (LLMs). However, existing learnable-mask approaches incur substantial parameter and memory overhead, limiting their scalability to large models and aggressive sparsity regimes. In this work, we revisit semi-structured pruning from a perspective that reconciles efficiency with scalability. We propose Reservoir of Importance (RoI), a lightweight semi-structured pruning framework that learns sparsity masks through differentiable subset sampling. Unlike prior methods that model full categorical distributions over all feasible N:M patterns, RoI introduces a compact-logit parameterization for sparsity mask learning and performs sampling without replacement to select masks, thereby reducing trainable parameters from combinatorial complexity to O(M). As a result, RoI requires 1.5-8.75× fewer learnable parameters and significantly lower memory cost, while remaining fully aligned with hardware-friendly sparsity patterns. Extensive evaluations across multiple scales of the Qwen2.5 LLM family (0.5-7B parameters) demonstrate that RoI achieves competitive performance with strong memory efficiency, stability, and scalability to more aggressive N:M sparsity patterns, offering a practical path toward efficient LLM deployment.
Figures & tables
Figure 1: Learnable semi-structured N : M sparsity methods: a) modeling the mask selection process using a categorical distribution over feasible masks, and b) our proposed method by learning to sample subsets without replacement of model parameters. The proposed method is more memory efficient than previous works for most practical N : M sparsity patterns. The memory advantage becomes more pronounced as M increases or when N is around M/2 .
Figure 2: Overview of the RoI framework for semi-structure sparsity learning via differentiable subset sampling, illustrating the training and inference phases.
Figure 3: RoI demonstrates better scalability than MaskLLM and ProxSparse, with WikiText-2 perplexity (on the Qwen2.5-1.5B model) improving more consistently as the number of training tokens increases.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Mask difference comparison between RoI and previous works on the Qwen2.5-7B model. Besides the name of each baseline, place an overlapping percentage indicating the similarity of the produced masks between that baseline and RoI. Bold pixels indicate masks corresponding to retained weights.
Semi-structured pruning compresses large language models (LLMs) while keeping a regular sparse structure, but the prevailing N:M pattern fixes the same local sparsity ratio in every layer. Layer-adaptive sparsity allocation improves unstructured pruning, yet it has been reported to be less effective under N:M sparsity, leaving open whether adaptive allocation is of limited value for semi-structured pruning in general or only under the fine-grained N:M pattern. We examine this question with group-level sparsity, which partitions each weight matrix into regular groups, retains or prunes each group as a whole, and allows each layer's sparsity ratio to vary under a global budget. We propose GroupMask, which generates the group selectors of all layers with a lightweight hypernetwork, relaxes them with a Gumbel-Sigmoid parameterization and a straight-through estimator, and learns them through sparsity-budget regularization and self-distillation while keeping the pretrained weights frozen. On LLaMA-2-7B at 50% sparsity with the same 1×256 group size, learned layer-adaptive allocation reduces WikiText-2 perplexity from 10.02 to 8.30 and raises the average zero-shot accuracy from 0.455 to 0.496 relative to a uniform per-layer ratio. GroupMask obtains the lowest WikiText-2 perplexity on LLaMA-2-7B and the highest average zero-shot accuracy with Alpaca calibration among the evaluated baselines on five LLaMA and Qwen models. Our code is available at https://github.com/ZhengaoLi/GroupMask.
Zhengao Li, Shuoqiu Li, Xiaofang Zhang +7
Florida State University · University of Connecticut · College of William & Mary +1
Semi-structured sparsity provides a practical path to accelerate large language models (LLMs) with native hardware support, but post-training semi-structured pruning often suffers from substantial quality degradation due to strong structural coupling. Existing methods rely on large-scale sparse retraining to recover accuracy, resulting in high computational cost. We propose SparseForge, a post-training framework that improves recovery efficiency by directly optimizing the sparsity mask rather than scaling up retraining tokens. SparseForge combines Hessian-aware importance estimation with progressive annealing of soft masks into hardware-executable structured sparsity, enabling stable and efficient sparse recovery. On LLaMA-2-7B under 2:4 sparsity, SparseForge achieves 57.27% average zero-shot accuracy with only 5B retraining tokens, surpassing the dense model's 56.43% accuracy and approaching the 57.52% result of a state-of-the-art method using 40B tokens. Such improvements on the accuracy-efficiency trade-off from SparseForge are shown to be consistent across model families.
Structured pruning reduces LLM inference cost by removing low-importance architectural components. This can be viewed as learning a multiplicative gate for each component under an l0 sparsity constraint. Due to the discreteness of the l0 norm, prior work typically adopts stochastic hard-concrete relaxations to enable differentiable optimization; however, this stochasticity can introduce a train--test mismatch when sampled masks are discretized for deployment and restricts masks to a bounded, near-binary range. To address this, we propose Deterministic Differentiable Pruning (DDP), a mask-only optimization method that eliminates stochasticity by directly optimizing a deterministic soft surrogate of the discrete l0 objective. Compared with prior approaches, DDP offers greater expressiveness, reduced train--test mismatch, and faster convergence. We apply our method to several dense and MoE models, including Qwen3-32B and Qwen3-30B-A3B, achieving a performance loss as small as 1% on downstream tasks while outperforming previous methods at 20% sparsity. We further demonstrate end-to-end inference speedups in realistic deployment settings with vLLM.
Weiyu Huang, Pengle Zhang, Xiaolu Zhang +3
Department of Computer Science and Technology, Tsinghua University, Beijing, China · Ant Group, Beijing, China