Ruokai Yin

Youngeun Kim, Yuhang Li, Hyoungseob Park et al.

Spiking Neural Networks (SNNs) have recently emerged as a new generation of low-power deep neural networks, which is suitable to be implemented on low-power mobile/edge devices. As such devices have limited memory storage, neural pruning on SNNs has been widely explored in recent years. Most existing SNN pruning works focus on shallow SNNs (2~6 layers), however, deeper SNNs (>16 layers) are proposed by state-of-the-art SNN works, which is difficult to be compatible with the current SNN pruning work. To scale up a pruning technique towards deep SNNs, we investigate Lottery Ticket Hypothesis (LTH) which states that dense networks contain smaller subnetworks (i.e., winning tickets) that achieve comparable performance to the dense networks. Our studies on LTH reveal that the winning tickets consistently exist in deep SNNs across various datasets and architectures, providing up to 97% sparsity without huge performance degradation. However, the iterative searching process of LTH brings a huge training computational cost when combined with the multiple timesteps of SNNs. To alleviate such heavy searching cost, we propose Early-Time (ET) ticket where we find the important weight connectivity from a smaller number of timesteps. The proposed ET ticket can be seamlessly combined with a common pruning techniques for finding winning tickets, such as Iterative Magnitude Pruning (IMP) and Early-Bird (EB) tickets. Our experiment results show that the proposed ET ticket reduces search time by up to 38% compared to IMP or EB methods. Code is available at Github.

Yuhang Li, Ruokai Yin, Hyoungseob Park et al.

We study the Human Activity Recognition (HAR) task, which predicts user daily activity based on time series data from wearable sensors. Recently, researchers use end-to-end Artificial Neural Networks (ANNs) to extract the features and perform classification in HAR. However, ANNs pose a huge computation burden on wearable devices and lack temporal feature extraction. In this work, we leverage Spiking Neural Networks (SNNs)--an architecture inspired by biological neurons--to HAR tasks. SNNs allow spatio-temporal extraction of features and enjoy low-power computation with binary spikes. We conduct extensive experiments on three HAR datasets with SNNs, demonstrating that SNNs are on par with ANNs in terms of accuracy while reducing up to 94% energy consumption. The code is publicly available in https://github.com/Intelligent-Computing-Lab-Yale/SNN_HAR

Ruokai Yin, Youngeun Kim, Di Wu et al.

Spiking Neural Networks (SNNs) have gained significant research attention in the last decade due to their potential to drive resource-constrained edge devices. Though existing SNN accelerators offer high efficiency in processing sparse spikes with dense weights, opportunities are less explored in SNNs with sparse weights, i.e., dual-sparsity. In this work, we study the acceleration of dual-sparse SNNs, focusing on their core operation, sparse-matrix-sparse-matrix multiplication (spMspM). We observe that naively running a dual-sparse SNN on existing spMspM accelerators designed for dual-sparse Artificial Neural Networks (ANNs) exhibits sub-optimal efficiency. The main challenge is that processing timesteps, a natural property of SNNs, introduces an extra loop to ANN spMspM, leading to longer latency and more memory traffic. To address the problem, we propose a fully temporal-parallel (FTP) dataflow, which minimizes both data movement across timesteps and the end-to-end latency of dual-sparse SNNs. To maximize the efficiency of FTP dataflow, we propose an FTP-friendly spike compression mechanism that efficiently compresses single-bit spikes and ensures contiguous memory access. We further propose an FTP-friendly inner-join circuit that can lower the cost of the expensive prefix-sum circuits with almost no throughput penalty. All the above techniques for FTP dataflow are encapsulated in LoAS, a Low-latency inference Accelerator for dual-sparse SNNs. With FTP dataflow, compression, and inner-join, running dual-sparse SNN workloads on LoAS demonstrates significant speedup (up to $8.51\times$) and energy reduction (up to $3.68\times$) compared to running it on prior dual-sparse accelerators.

Yuhang Li, Ruokai Yin, Donghyun Lee et al.

We introduce GPTAQ, a novel finetuning-free quantization method for compressing large-scale transformer architectures. Unlike the previous GPTQ method, which independently calibrates each layer, we always match the quantized layer's output to the exact output in the full-precision model, resulting in a scheme that we call asymmetric calibration. Such a scheme can effectively reduce the quantization error accumulated in previous layers. We analyze this problem using optimal brain compression to derive a close-formed solution. The new solution explicitly minimizes the quantization error as well as the accumulated asymmetry error. Furthermore, we utilize various techniques to parallelize the solution calculation, including channel parallelization, neuron decomposition, and Cholesky reformulation for matrix fusion. As a result, GPTAQ is easy to implement, simply using 20 more lines of code than GPTQ but improving its performance under low-bit quantization. Remarkably, on a single GPU, we quantize a 405B language transformer as well as EVA-02, the rank first vision transformer that achieves 90% pretraining Imagenet accuracy. Code is available at Github.

Ruokai Yin, Yuhang Li, Donghyun Lee et al.

Large language models (LLMs) deliver strong performance but are difficult to deploy due to high memory and compute costs. While pruning reduces these demands, most methods ignore activation sparsity observed at runtime. We reinterpret activation sparsity as dynamic structured weight sparsity and propose DuoGPT, a unified framework that constructs dual-sparse (spMspV) workloads by combining unstructured weight pruning with activation sparsity. To preserve accuracy, we extend the Optimal Brain Compression (OBC) framework with activation-aware calibration and introduce output residuals from the dense model as correction terms. We further optimize the solution for efficient GPU execution, enabling scalability to billion-parameter LLMs. Evaluations on LLaMA-2 and LLaMA-3 show that DuoGPT outperforms state-of-the-art structured pruning methods by up to 9.17% accuracy at an iso-speedup of 1.39$\times$ compared to the baseline dense model. Code is available at Github.

Donghyun Lee, Yuhang Li, Ruokai Yin et al.

State Space Models (SSMs) have emerged as powerful alternatives to attention-based Transformers, with Mamba demonstrating impressive efficiency and scalability. As these models grow increasingly larger, the need for Parameter-Efficient Fine-Tuning (PEFT) methods becomes critical to adapt pre-trained Mamba to downstream tasks without prohibitive computational costs. However, previous approaches simply apply traditional Transformer-tailored PEFT methods without addressing the unique temporal processing dynamics of SSMs. To address this limitation, we propose Memba, a membrane-driven PEFT approach specifically designed for Mamba. Memba introduces Leaky Integrate Membrane (LIM) neurons as bio-inspired gating mechanisms that naturally accumulate membrane potentials over time, enhancing selective information retention. By strategically combining LIM neurons with Low-Rank Adaptations (LoRA) and cross-layer membrane transfer, our approach significantly improves Mamba's temporal modeling capabilities. Extensive experiments across language and vision tasks demonstrate that Memba achieves substantial improvements over existing PEFT methods. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/Memba.

Yuhang Li, Donghyun Lee, Ruokai Yin et al.

Low-rank decomposition has emerged as an important problem in Large Language Model (LLM) fine-tuning and inference. Through Singular Value Decomposition (SVD), the weight matrix can be factorized into low-rank spaces optimally. Previously, a common practice was to decompose the weight in the activation-whitened space, and then achieve satisfying results. In this work, we propose Optimal Brain Decomposition LLM (OBD-LLM), which studies the decomposition problem in the model space by utilizing second-order Hessian information. Through a rigorous Kronecker-factorization of the Hessian, we show that the decomposition needs to consider both input and output information of the layer, and achieves much better decomposition results compared to input only method. Our loss-aware decomposition method involves a bi-directional whitening on the weight matrix. As a result, OBD-LLM is a closed-form solution for the optimal decomposition of weights in the language model. Remarkably, we achieve ~20-40\% better results than previous state-of-the-art decomposition methods, the SVD-LLM.

Ruokai Yin, Sattwik Deb Mishra, Xuan Zuo et al.

Distributed LLM inference requires careful coordination of parallelization strategies across hundreds to thousands of NPUs to meet production SLOs. Current systems like Megatron-LM rely on static heuristics that separately configure parallelism degrees and per-operator sharding dimensions, leaving significant performance on the table as models scale and hardware topologies diversify. We introduce Learn to Shard, to our knowledge, the first RL-based approach to co-optimize both coarse-grained parallelism degrees and fine-grained per-operator sharding dimensions for distributed LLM inference. Our method employs an attention-based policy over an elite history that learns from high-performing strategies to efficiently navigate the vast combinatorial search space. Evaluated on H100 clusters with MoE models up to 1.6T parameters, Learn to Shard achieves up to 3.5x throughput improvement over metaheuristic baselines and 1.06x over Megatron heuristics.

Youngeun Kim, Yuhang Li, Abhishek Moitra et al.

Due to increasing interest in adapting models on resource-constrained edges, parameter-efficient transfer learning has been widely explored. Among various methods, Visual Prompt Tuning (VPT), prepending learnable prompts to input space, shows competitive fine-tuning performance compared to training of full network parameters. However, VPT increases the number of input tokens, resulting in additional computational overhead. In this paper, we analyze the impact of the number of prompts on fine-tuning performance and self-attention operation in a vision transformer architecture. Through theoretical and empirical analysis we show that adding more prompts does not lead to linear performance improvement. Further, we propose a Prompt Condensation (PC) technique that aims to prevent performance degradation from using a small number of prompts. We validate our methods on FGVC and VTAB-1k tasks and show that our approach reduces the number of prompts by ~70% while maintaining accuracy.