Chayne Thrash

Ali Abbasi, Chayne Thrash, Haoran Qin et al.

Large language models deliver strong performance across language and reasoning tasks, but their storage and compute costs remain major barriers to deployment in resource-constrained and latency-sensitive settings. SVD-based post-training compression offers a hardware-agnostic way to reduce model size and improve inference efficiency through low-rank factorization. However, existing methods often rely on input-only whitening spaces, homogeneous rank allocation, or loss-agnostic allocation heuristics, limiting their ability to preserve model quality under aggressive compression. We propose Input-Output Whitened SVD (IO-SVD), a post-training compression method that forms a KL-aware double-sided whitening space for model weights. Using a second-order expansion of the KL loss over the top-K token probabilities, IO-SVD constructs an output-side metric that captures predictive sensitivity, while input whitening captures activation statistics. We further introduce an efficient heterogeneous rank-allocation strategy that scores whitened singular components using first-order calibration loss estimates and prunes the least sensitive components under a global budget. Inspired by prior work that combines SVD truncation with quantization, we improve hybrid SVD-quantization compression through loss-aware remapping, which selects low-rank factor rows for 8-bit quantization based on the predicted loss change incurred by quantizing them. Extensive experiments across diverse LLM and VLM families, and inference-time analysis shows that IO-SVD compresses LLMs with minimal performance degradation while delivering practical inference speedups. Code is available at https://github.com/mint-vu/IO-SVD.git

Ali Abbasi, Chayne Thrash, Elaheh Akbari et al.

The rapid progress of AI, combined with its unprecedented public adoption and the propensity of large neural networks to memorize training data, has given rise to significant data privacy concerns. To address these concerns, machine unlearning has emerged as an essential technique to selectively remove the influence of specific training data points on trained models. In this paper, we approach the machine unlearning problem through the lens of continual learning. Given a trained model and a subset of training data designated to be forgotten (i.e., the "forget set"), we introduce a three-step process, named CovarNav, to facilitate this forgetting. Firstly, we derive a proxy for the model's training data using a model inversion attack. Secondly, we mislabel the forget set by selecting the most probable class that deviates from the actual ground truth. Lastly, we deploy a gradient projection method to minimize the cross-entropy loss on the modified forget set (i.e., learn incorrect labels for this set) while preventing forgetting of the inverted samples. We rigorously evaluate CovarNav on the CIFAR-10 and Vggface2 datasets, comparing our results with recent benchmarks in the field and demonstrating the efficacy of our proposed approach.

Haoran Qin, Shansita Sharma, Ali Abbasi et al.

Low-rank approximation methods such as singular value decomposition (SVD) and its variants (e.g., Fisher-weighted SVD, Activation SVD) have recently emerged as effective tools for neural network compression. In this setting, decomposition acts as a "surgical" intervention, followed by fine-tuning that serves as "rehab" to recover accuracy. Inspired by prehabilitation in surgery, we introduce a pre-compression fine-tuning stage, Low-Rank Prehab, that explicitly encourages low-rank structure in weight matrices while preserving task performance. By conditioning the model before SVD, Prehab steers weights toward spectrally compact regions of the parameter space, enabling smoother low-rank approximation and improved recovery. Experiments on large language models (LLMs) and other Transformer-based architectures, including Vision Transformers (ViTs), show that Prehab substantially reduces the immediate accuracy drop after compression and consistently improves post-finetuning performance. Across a wide range of compression ratios, our method outperforms state-of-the-art SVD-based techniques such as SVD-LLM, highlighting the importance of preparing models for compression rather than only improving the compression and recovery stages. Source code is available at https://github.com/niqretnuh/PREHAB-SVD

Ali Abbasi, Chayne Thrash, Haoran Qin et al.

Advances in large language models have driven strong performance across many tasks, but their memory and compute costs still hinder deployment. SVD-based compression reduces storage and can speed up inference via low-rank factors, yet performance depends on how rank is allocated under a global compression ratio. Prior methods often use homogeneous ranks for similarly sized matrices, despite large differences in loss sensitivity, or rely on expensive iterative pre-truncation optimization to determine per matrix ranks. We propose \textbf{Zero Sum SVD} (\textbf{ZS-SVD}), a post-training method that performs \emph{global} singular component selection using activation whitening and first-order calibration loss estimates in whitened coordinates. \textbf{ZS-SVD} prunes components across the whole model with a \textbf{zero sum} rule that keeps the cumulative predicted loss change near zero, automatically yielding heterogeneous ranks without solving a rank allocation optimization. Motivated by evidence that gradients near pretrained solutions exhibit low rank structure, we also introduce an optional lightweight correction that applies a \textbf{single} projected gradient update after truncation, followed by re-truncation. Extensive experiments across multiple LLM architectures show consistent gains across diverse benchmarks and compression ratios. Code is available at https://github.com/mint-vu/Zero-Sum-SVD

Chayne Thrash, Ali Abbasi, Soheil Kolouri

Large language models (LLMs) are costly to deploy due to their large memory footprint and high inference cost. Weight-activation quantization can reduce these costs, but low-bit activation quantization remains difficult because activation outliers induce large quantization error. Recent rotation-based methods address this by applying orthogonal transformations that redistribute activation magnitude across dimensions, but existing approaches either require expensive end-to-end rotation training or rely on stored activation corpora, introducing significant compute or storage overhead. We propose a lightweight post-training rotation calibration method for LLM activation quantization. Our method learns orthogonal rotations that align normalized activations with the corners of an inscribed hypercube, encouraging activation energy to be distributed more evenly across dimensions. This objective admits an efficient closed-form update via the orthogonal Procrustes problem, avoiding gradient-based optimization over the orthogonal group. We further introduce an online calibration procedure that updates rotations as calibration samples are processed, eliminating the need to store activations on disk and allowing rotations to adapt to quantized activation distributions during calibration. Experiments on Llama-2 and Llama-3 models from 3B to 70B parameters show that our method achieves competitive or improved performance across perplexity benchmarks and common sense reasoning tasks while avoiding both costly end-to-end training and large offline activation storage.

Chayne Thrash, Ali Abbasi, Reed Andreas et al.

The outstanding performance of large foundational models across diverse tasks, from computer vision to speech and natural language processing, has significantly increased their demand. However, storing and transmitting these models poses significant challenges due to their massive size (e.g., 750GB for Llama 3.1 405B). Recent literature has focused on compressing the original weights or reducing the number of parameters required for fine-tuning these models. These compression methods generally constrain the parameter space, for example, through low-rank reparametrization (e.g., LoRA), pruning, or quantization (e.g., QLoRA) during or after the model training. In this paper, we present a novel model compression method, which we term Manifold-Constrained Neural Compression (MCNC). This method constrains the parameter space to low-dimensional pre-defined and frozen nonlinear manifolds, which effectively cover this space. Given the prevalence of good solutions in over-parameterized deep neural networks, we show that by constraining the parameter space to our proposed manifold, we can identify high-quality solutions while achieving unprecedented compression rates across a wide variety of tasks and architectures. Through extensive experiments in computer vision and natural language processing tasks, we demonstrate that our method significantly outperforms state-of-the-art baselines in terms of compression, accuracy, and/or model reconstruction time. Our code is publicly available at https://github.com/mint-vu/MCNC.

Ashkan Shahbazi, Chayne Thrash, Yikun Bai et al.

Transformers have proven highly effective across a wide range of modalities. However, the quadratic complexity of the standard softmax attention mechanism poses a fundamental barrier to scaling them to long context windows. A large body of work addresses this with linear attention, which reformulates attention as a kernel function and approximates it with finite feature maps to achieve linear-time computation. Orthogonal to computational scaling, most attention mechanisms -- both quadratic and linear -- produce row-normalized maps that can over-focus on a few tokens, degrading robustness and information flow. Enforcing doubly-stochastic attention alleviates this by balancing token participation across rows and columns, but existing doubly-stochastic attention mechanisms typically introduce substantial overhead, undermining scalability. We propose LOTFormer, a principled attention mechanism that is simultaneously linear-time and doubly-stochastic. Our approach exploits the connection between attention maps and transportation plans between query and key measures. The central idea is to constrain the transport plan to be low-rank by conditioning it on a learnable pivot measure with small support. Concretely, we solve two entropic optimal transport problems (queries $\to$ pivot and pivot $\to$ keys) and compose them into a conditional (glued) coupling. This yields an attention matrix that is provably doubly-stochastic, has rank at most $r \ll n$, and applies to values in $O(nr)$ time without forming the full $n \times n$ map. The pivot locations and masses are learned end-to-end. Empirically, LOTFormer achieves state-of-the-art results on the Long Range Arena benchmark, surpassing prior linear and transport-based attention methods in both accuracy and efficiency.