Inference with transformer-based large language models (LLMs) is often limited by the memory-bound KV cache and quadratic attention cost. State-space models (SSMs) mitigate this through linear attention and fixed-size recurrent states, but their large dense linear projections remain computationally expensive even after quantization. We introduce a method that induces sparse neural activity in heavily quantized linear-attention models with minimal performance loss. Activations below a per-projection trainable threshold ($\pm Δ$) are nullified while preserving crucial outliers, achieving comparable performance to dense models with up to 4$\times$ fewer effective arithmetic operations. Targeting a multi-core, multi-chip neuromorphic platform, where event-driven execution converts unstructured sparsity into throughput at both the compute and communication levels, a capability GPU architectures fundamentally lack, we project up to 37$\times$ higher throughput and 16$\times$ lower power versus edge GPU inference of a comparable transformer-based model, and up to 5.4$\times$ improvements over the non-sparsified baseline. These results position sparse, quantized linear-attention models as a natural fit for deploying LLMs on event-driven multi-core platforms.
Prohibitive computational and environmental costs impede the scalable deployment of Large Language Models (LLMs). Traditional compression techniques (sparsity, quantization, low-rank approximations) are typically applied in isolation, and each hits an accuracy-efficiency wall. This thesis proposes the "Compression Trinity," a unified framework that applies the three pillars jointly: sparsity to reduce computation, quantization to minimize memory bandwidth, and low-rank approximations to recover accuracy. To accelerate pretraining, we apply the Trinity to the optimizer and model architecture. MKOR approximates curvature via block-diagonal sparsity and low-rank inversion, maintaining numerical stability for quantized states; it reduces curvature update complexity from $O(d^3)$ to $O(d^2)$ and accelerates convergence by up to 1.85x over KFAC. SLoPe accelerates training by up to 1.25x via a double-pruned backward pass for N:M sparsity, using low-rank "lazy" adapters in the final 1% of training to recover accuracy. For post-training compression, OPTIMA stabilizes static masks in a zero-training regime by formulating weight reconstruction as globally optimal column-wise quadratic programs, improving zero-shot accuracy by up to 3.97%. Given a fine-tuning budget, PATCH breaks the ceiling of static masks by learning a dynamic hybrid sparsity ratio between 0% and 50%, yielding up to 1.38x speedups. Finally, SLiM realizes the full Compression Trinity in one shot, using mathematically derived low-rank adapters to recover information lost to quantization and sparsity, improving accuracy by up to 5.66% over state-of-the-art methods and outperforming uncompressed dense models at equal parameter budgets by 0.6%. Together, these results show that jointly applying the Compression Trinity is essential for efficient, scalable, high-performance LLMs.
Pruning reduces the inference cost of large language models, but existing criteria primarily preserve large activations or reconstruct layer outputs. We argue that this overlooks a key computation performed by particularly sparsity-sensitive neurons in the MLP up and gate projections: separating similar inputs into dissimilar outputs. This suggests that effective pruning should preserve not only activations, but also the differences between outputs more broadly. We introduce a family of difference-informed pruning methods built upon this principle. Wisp is a first-order, update-free method that scores weights using input-difference norms, and Wisp+ refines this score neuronwise using the input pairs each neuron separates most strongly. Finally, Whisper is a second-order method that uses a lightly regularized difference Hessian as its reconstruction objective. Across Llama 2 and 3.1 models from 7B to 405B parameters, our second-order variant consistently improves over strong reconstruction-based baselines, while our update-free variants improve over activation-aware baselines, especially in constrained settings. The improvements over Wanda and SparseGPT extend to structured sparsity, downstream evaluations, and other model families. Augmenting stronger techniques such as RIA and ALPS with our difference-informed criteria yields further improvements, shifting the overall accuracy-runtime frontier outward at negligible additional cost. These results suggest that preserving output differences is a broadly useful and composable signal for post-training LLM sparsification.
Multi-task inference models share a single backbone across diverse tasks, yet execute identical computation regardless of which task is active - wasting energy and cycles on task-irrelevant operations. We observe that the task command, typically available before inference begins, provides a free signal that can be exploited to skip unnecessary computation at the hardware level. We present a HW/SW co-designed approach in which a lightweight gating network, trained jointly with the backbone, predicts per-tile binary execution masks conditioned on the task input. Each tile corresponds to a fixed group of output channels (the native scheduling granularity of the accelerator), enabling masked tiles to be skipped with zero overhead. This yields a task-dependent reduction in compute, where each command activates only the subset of the network it requires, without changes to the model architecture or inference pipeline. We co-design the full system stack: a command-conditioned training procedure that learns hardware-aligned tile masks under a sparsity objective; an instruction set architecture whose instructions carry per-tile bitmask fields, allowing the hardware to skip masked tiles without software intervention; and a tiled inference accelerator with configurable parallelism, double-buffered memory, and INT8 datapath that natively supports sparse tile execution. We prototype on an AMD/Xilinx Alveo U50 FPGA and evaluate on a closed-loop visuomotor driving task in CARLA autonomous driving simulator. Task-conditional sparsity reduces FLOPs by 66-76% while maintaining driving quality. On-device latency decreases by 51-59%, from 9.12 ms to 3.74-4.44 ms (2.1-2.4x speedup), with energy per inference dropping from 263 to 108-128mJ.
Varun Manjunath, Ruokai Yin, Donghyun Lee +2cs.AR cs.AI
Bit-serial accelerators exploit bit-level sparsity to reduce DNN inference cost, but existing designs exploit sparsity on only one operand, bounding the speedup. Extending sparsity exploitation to both operands simultaneously yields compounding reductions in partial products but introduces a critical new bottleneck: workload imbalance. Because each concurrent weight - activation pair's execution cost depends on the product of two independently varying operand non-zero bit counts, pairs that must complete together finish at vastly different times, leaving faster computations idle. We show this limits PE utilization to 56 - 64% in existing dual-sided designs. We present BRIM, a hardware - software co-designed dual-sided bit-serial sparse accelerator that directly targets this bottleneck. BRIM combines two integrated mechanisms: 1) Cyclic-Balanced Pruning (CBP), a post-training weight optimization that reshapes weight representations based on profiled activation statistics to equalize expected workloads across concurrently processed pairs offline; and 2) Pairwise Slot Donation, a lightweight hardware mechanism that absorbs residual runtime imbalance with negligible area overhead. Evaluated across CNNs, ViTs, and LLMs under iso-area constraints, BRIM achieves over 90% PE utilization, up to 2.37x speedup, and up to 1.63x energy efficiency improvement over prior dual-sided designs.
Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints. Since the actual output length is unknown, output sequences are padded to a predefined maximum length, resulting in substantial redundant computation over unnecessary [EOS] tokens. In this work, we discover that DMLLMs implicitly reveal their valid semantic boundary at the very first denoising step through a distinct shift in MLP activation sparsity. Leveraging this observation, we propose Seer, a training-free framework that detects this boundary using a Signal-to-Noise Ratio (SNR)-based criterion and performs one-shot truncation of the redundant suffix for all subsequent computations. To preserve these theoretical gains during batched serving, Seer incorporates a hybrid execution strategy that maximizes throughput while seamlessly accommodating dynamic sequence lengths. Experimental results demonstrate that Seer effectively eliminates padding waste, accelerating throughput by up to $\sim$31$\times$. Across 9 benchmarks, Seer robustly maintains overall performance and even improves accuracy on complex visual tasks by mitigating noise leakage (e.g., DocVQA score increases from 63.52 to 63.66), offering a highly efficient, plug-and-play solution for DMLLM acceleration.
Elijah Cadenhead, Cristian McGee, Xin Li +2cs.LG cs.AI cs.IT
Low-rank adaptation (LoRA) and its variants provide a memory- and compute-efficient alternative to full fine-tuning of pre-trained models. However, questions remain about the comparative generalizability of these approaches and how the structural restrictions on low-rank updates preserve effective adaptation performance. We present a historical framing, covering the past (full fine-tuning and original LoRA), the present (different variants of LoRA), and propose simpler, cheaper, parameter-efficient extensions by inducing sparsity within existing LoRA variants: Cheap LoRA (cLA), training a single low-rank factor with the other fixed (deterministically or, in its randomized variant, stochastically), and the chained circulant variant, ${c}^3$LA. We frame cLA as a structured instance of asymmetric LoRA, serving as a controlled column-subspace restriction of full fine-tuning. We derive information-theoretic generalization error bounds for these variants, marking one of the first endeavors in this area. Empirically, we evaluate 11 fine-tuning methods across 10 pre-trained models and 14 datasets, analyzing the fine-tuned models' performance and generalization using tools such as loss landscapes and spectral analysis. Despite the sensitivity of fine-tuned models to the pre-trained model, datasets, and other factors, our study suggests that restricting LoRA-based PEFT methods' adaptation to a sparse, structured column space remains competitive across tasks with their parameter-matched baselines while reducing up to 10% training time and peak GPU memory up to 15%, even with a naïve, non-optimized, sparse implementation. Our theoretical and empirical generalization measures provide a more consistent and principled approach to their cost-effective adaptation than commonly used analytical tools. Overview and code are available at: https://elicaden.github.io/Beyond_LoRA/.
Semi-structured 2:4 sparsity is widely supported by modern accelerators, providing up to a 2x theoretical speedup. However, its strict 50% sparsity constraint often causes non-negligible accuracy degradation under post-training pruning. Meanwhile, existing relaxed sparsity formats either require specialized compiler support or introduce runtime overheads that limit end-to-end speedup. We propose Spense, a practical hybrid sparse-dense format that splits each weight matrix into a 2:4 sparse region and a dense region. This design relaxes the effective sparsity constraint while remaining compatible with existing high-performance sparse and dense GEMM libraries, avoiding both custom compiler support and input activation expansion. Building on this format, we introduce SpenseGPT, a one-shot post-training pruning method that produces sparse and dense regions. Notably, we show that selecting the right dense regions is important, and we devise two different strategies to choose them. Experiments on Qwen3-32B and Seed-OSS-36B demonstrate that our method achieves up to 1.2x end-to-end decoding speedup on B200 GPUs with FP8 precision, while preserving accuracy. To the best of our knowledge, this is the first one-shot pruning demonstration of real-world end-to-end LLM decoding speedup from semi-structured sparse tensor cores on recent GPUs such as B200s, while maintaining model quality.
Calibration data are often treated as a minor implementation detail in post-training LLM pruning because averaged evaluations suggest only modest effects. We show that this conclusion is an averaging artifact: at 60\% SparseGPT sparsity, calibration strategies separated by only 2.85 points in averaged commonsense accuracy differ by 51.9 points in Code retention. Across 15 sources, capability-decomposed analysis reveals an opposing pattern: calibration perplexity is positively associated with General retention but negatively associated with Math or Code retention, leaving no evaluated single source uniformly strong across capabilities. This finding motivates capability-balanced multi-source calibration. Under the same calibration budget, a balanced real-data mixture outperforms every evaluated single source on LLaMA-3.1-8B, beating C4 by 18.8 points; the advantage grows with sparsity and persists on LLaMA-3.1-70B. Because the original pretraining data of advanced LLMs are often inaccessible, we further introduce Information-Guided Self-Calibration for Pruning (IGSP). Using only the base model and evaluation taxonomy, IGSP generates capability-stratified pools and selects low-redundancy samples within capability-specific perplexity ranges, outperforming Self-Cal and SGS by up to 4.8 points. Together, these results recast calibration as a capability-coverage problem and identify multi-source design as a practical principle for preserving capabilities in high-sparsity LLM pruning.