arXiv subjects

Priyadarshini Panda

Publications and source records attributed to Priyadarshini Panda.

At least 19 recordsLinked to original sources

arXiv2026

Celty: SpMSpV GPU Kernel and SIMT Co-Design for Efficient Dual-Sparse LLM Inference

Large Language Models (LLMs) increasingly rely on sparsity to cut inference cost, but most prior work exploits a single sparsity source and targets batched multi-user inference. Dual-sparsity, which pairs unstructured weight pruning with runtime activation sparsity, offers a compelling size-accuracy-latency tradeoff for single-user decoding, but forms a Sparse Matrix-Sparse Vector (SpMSpV) workload that existing GPU kernels handle poorly. We propose Celty, a co-designed sparse format, GPU kernel, and SIMT microarchitecture for SpMSpV in LLM inference. Celty's Run-Length Compressed CSC (RLC-CSC) format enables vectorized loading of compressed weight columns and exploits both sparsity sources to skip memory accesses, accumulating partial products in shared memory. The Celty Sparse SIMT Core then adds a pipelined RLC decoder that removes software index reconstruction and repurposes local register files for conflict-free accumulation, operating on the same compressed representation. The kernel alone achieves up to 2.8x over cuBLAS; with the Sparse SIMT Core, speedup reaches 5.3x over cuBLAS at 70% dual-sparsity.

cs.AR↗

arXiv2026

BRIM: Workload-Balanced Dual-Sided Bit-Serial Sparse Inference Accelerator

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.

cs.AR↗

arXiv2026

KronQ: LLM Quantization via Kronecker-Factored Hessian

Post-training quantization (PTQ) is a widely adopted technique for compressing large language models (LLMs) without retraining. Most existing second-order PTQ methods, including GPTQ, construct quantization objectives from input activation statistics, effectively assuming that all output channels contribute equally to the layer-wise reconstruction objective. We propose KronQ, a PTQ framework that challenges this assumption by introducing the gradient covariance into the quantization pipeline. Under the Kronecker-factored Hessian approximation, the quantization loss depends jointly on both the activation and gradient covariances, and KronQ exploits this at two complementary levels. (1) KronQ introduces bidirectional incoherence processing, extending the existing input-side random rotation to the output dimension using the gradient covariance, reducing weight magnitude variance across both input and output dimensions. (2) KronQ derives a new sensitivity metric for inter-layer mixed-precision allocation, driven by the gradient and activation Hessian traces. Notably, in the case of 2-bit weight-only quantization on LLaMA-3-70B, while GPTQ and GPTAQ diverge or produce degenerate quantizations (>2000 perplexity on WikiText-2), \KronQ{} achieves 7.93 perplexity.

cs.LG↗

arXiv2026

OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers

Diffusion transformers (DiTs) achieve state-of-the-art image and video generation, but their multi-step sampling and growing parameter count make inference expensive. Post-training quantization (PTQ) is the natural remedy, yet DiT activations shift across timesteps, prompts, and guidance branches, forcing prior methods to re-fit calibration data for every new checkpoint or modality. We present OrbitQuant, a data-agnostic weight-activation quantizer that bypasses range estimation by quantizing in a normalized, rotated basis. In this basis, a randomized permuted block-Hadamard (RPBH) rotation concentrates each coordinate around one fixed, known marginal regardless of the input, so a single Lloyd-Max codebook serves all timesteps, prompts, and layers of a given input dimension. We extend the same quantizer to weight rows offline, absorbing the rotation into the weights so that it cancels inside each linear layer and only a forward rotation on the activations remains at runtime. The same recipe transfers from image to video with no per-modality tuning. Across FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX, it sets the state of the art for PTQ at several low-bit settings. It also pushes PTQ of image diffusion transformers to W2A4 with usable generation quality.

cs.CV↗

arXiv2026

Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation

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.

cs.LG↗

arXiv2026

QuRL: Efficient Reinforcement Learning with Quantized Rollout

Reinforcement learning with verifiable rewards (RLVR) has become a trending paradigm for training reasoning large language models (LLMs). However, due to the autoregressive decoding nature of LLMs, the rollout process becomes the efficiency bottleneck of RL training, consisting of up to 70\% of the total training time. In this work, we propose Quantized Reinforcement Learning (QuRL) that uses a quantized actor for accelerating the rollout. We address two challenges in QuRL. First, we propose Adaptive Clipping Range (ACR) that dynamically adjusts the clipping ratio based on the policy ratio between the full-precision actor and the quantized actor, which is essential for mitigating long-term training collapse. Second, we identify the weight update problem, where weight changes between RL steps are extremely small, making it difficult for the quantization operation to capture them effectively. We mitigate this problem through the invariant scaling technique that reduces quantization noise and increases weight update. We evaluate our method with INT8 and FP8 quantization experiments on DeepScaleR and DAPO, and achieve 20% to 80% faster rollout during training.

cs.LG↗

arXiv2025

MD-SNN: Membrane Potential-aware Distillation on Quantized Spiking Neural Network

Spiking Neural Networks (SNNs) offer a promising and energy-efficient alternative to conventional neural networks, thanks to their sparse binary activation. However, they face challenges regarding memory and computation overhead due to complex spatio-temporal dynamics and the necessity for multiple backpropagation computations across timesteps during training. To mitigate this overhead, compression techniques such as quantization are applied to SNNs. Yet, naively applying quantization to SNNs introduces a mismatch in membrane potential, a crucial factor for the firing of spikes, resulting in accuracy degradation. In this paper, we introduce Membrane-aware Distillation on quantized Spiking Neural Network (MD-SNN), which leverages membrane potential to mitigate discrepancies after weight, membrane potential, and batch normalization quantization. To our knowledge, this study represents the first application of membrane potential knowledge distillation in SNNs. We validate our approach on various datasets, including CIFAR10, CIFAR100, N-Caltech101, and TinyImageNet, demonstrating its effectiveness for both static and dynamic data scenarios. Furthermore, for hardware efficiency, we evaluate the MD-SNN with SpikeSim platform, finding that MD-SNNs achieve 14.85X lower energy-delay-area product (EDAP), 2.64X higher TOPS/W, and 6.19X higher TOPS/mm2 compared to floating point SNNs at iso-accuracy on N-Caltech101 dataset.

cs.NE↗

arXiv2025

SpikePool: Event-driven Spiking Transformer with Pooling Attention

Building on the success of transformers, Spiking Neural Networks (SNNs) have increasingly been integrated with transformer architectures, leading to spiking transformers that demonstrate promising performance on event-based vision tasks. However, despite these empirical successes, there remains limited understanding of how spiking transformers fundamentally process event-based data. Current approaches primarily focus on architectural modifications without analyzing the underlying signal processing characteristics. In this work, we analyze spiking transformers through the frequency spectrum domain and discover that they behave as high-pass filters, contrasting with Vision Transformers (ViTs) that act as low-pass filters. This frequency domain analysis reveals why certain designs work well for event-based data, which contains valuable high-frequency information but is also sparse and noisy. Based on this observation, we propose SpikePool, which replaces spike-based self-attention with max pooling attention, a low-pass filtering operation, to create a selective band-pass filtering effect. This design preserves meaningful high-frequency content while capturing critical features and suppressing noise, achieving a better balance for event-based data processing. Our approach demonstrates competitive results on event-based datasets for both classification and object detection tasks while significantly reducing training and inference time by up to 42.5% and 32.8%, respectively.

cs.NE↗

arXiv2025

DiffAxE: Diffusion-driven Hardware Accelerator Generation and Design Space Exploration

Design space exploration (DSE) is critical for developing optimized hardware architectures, especially for AI workloads such as deep neural networks (DNNs) and large language models (LLMs), which require specialized acceleration. As model complexity grows, accelerator design spaces have expanded to O(10^17), becoming highly irregular, non-convex, and exhibiting many-to-one mappings from design configurations to performance metrics. This complexity renders direct inverse derivation infeasible and necessitates heuristic or sampling-based optimization. Conventional methods - including Bayesian optimization, gradient descent, reinforcement learning, and genetic algorithms - depend on iterative sampling, resulting in long runtimes and sensitivity to initialization. Deep learning-based approaches have reframed DSE as classification using recommendation models, but remain limited to small-scale (O(10^3)), less complex design spaces. To overcome these constraints, we propose a generative approach that models hardware design as 1-D image synthesis conditioned on target performance, enabling efficient learning of non-differentiable, non-bijective hardware-performance mappings. Our framework achieves 0.86% lower generation error than Bayesian optimization with a 17000x speedup, and outperforms GANDSE with 30% lower error at only 1.83x slower search. We further extend the method to a structured DSE setting, attaining 9.8% lower energy-delay product (EDP) and 6% higher performance, with up to 145.6x and 1312x faster search compared to existing optimization methods on O(10^17) design spaces. For LLM inference, our method achieves 3.37x and 7.75x lower EDP on a 32nm ASIC and Xilinx Ultrascale+ VPU13 FPGA, respectively, compared to the state-of-the-art DOSA framework.

cs.AR↗

arXiv2025

OpenWorldSAM: Extending SAM2 for Universal Image Segmentation with Language Prompts

The ability to segment objects based on open-ended language prompts remains a critical challenge, requiring models to ground textual semantics into precise spatial masks while handling diverse and unseen categories. We present OpenWorldSAM, a framework that extends the prompt-driven Segment Anything Model v2 (SAM2) to open-vocabulary scenarios by integrating multi-modal embeddings extracted from a lightweight vision-language model (VLM). Our approach is guided by four key principles: i) Unified prompting: OpenWorldSAM supports a diverse range of prompts, including category-level and sentence-level language descriptions, providing a flexible interface for various segmentation tasks. ii) Efficiency: By freezing the pre-trained components of SAM2 and the VLM, we train only 4.5 million parameters on the COCO-stuff dataset, achieving remarkable resource efficiency. iii) Instance Awareness: We enhance the model's spatial understanding through novel positional tie-breaker embeddings and cross-attention layers, enabling effective segmentation of multiple instances. iv) Generalization: OpenWorldSAM exhibits strong zero-shot capabilities, generalizing well on unseen categories and an open vocabulary of concepts without additional training. Extensive experiments demonstrate that OpenWorldSAM achieves state-of-the-art performance in open-vocabulary semantic, instance, and panoptic segmentation across multiple benchmarks. Code is available at https://github.com/GinnyXiao/OpenWorldSAM.

cs.CV↗

arXiv2025

FF-INT8: Efficient Forward-Forward DNN Training on Edge Devices with INT8 Precision

Backpropagation has been the cornerstone of neural network training for decades, yet its inefficiencies in time and energy consumption limit its suitability for resource-constrained edge devices. While low-precision neural network quantization has been extensively researched to speed up model inference, its application in training has been less explored. Recently, the Forward-Forward (FF) algorithm has emerged as a promising alternative to backpropagation, replacing the backward pass with an additional forward pass. By avoiding the need to store intermediate activations for backpropagation, FF can reduce memory footprint, making it well-suited for embedded devices. This paper presents an INT8 quantized training approach that leverages FF's layer-by-layer strategy to stabilize gradient quantization. Furthermore, we propose a novel "look-ahead" scheme to address limitations of FF and improve model accuracy. Experiments conducted on NVIDIA Jetson Orin Nano board demonstrate 4.6% faster training, 8.3% energy savings, and 27.0% reduction in memory usage, while maintaining competitive accuracy compared to the state-of-the-art.

cs.LG↗

arXiv2025

DuoGPT: Training-free Dual Sparsity through Activation-aware Pruning in LLMs

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.

cs.LG↗

arXiv2025

Memba: Membrane-driven Parameter-Efficient Fine-Tuning for Mamba

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.

cs.LG↗

arXiv2025

Assortment of Attention Heads: Accelerating Federated PEFT with Head Pruning and Strategic Client Selection

Parameter Efficient Fine-Tuning (PEFT) has become the de-facto approach in adapting Large Language Models (LLMs) for downstream tasks in Natural Language Processing. However, its adoption in privacy-preserving distributed learning frameworks, such as Federated Learning (FL), remains relatively limited. This is mainly due to challenges specific to FL, such as resource-constrained devices and diverse data distributions among clients. In this paper, we propose an efficient method to perform PEFT within the FL framework for Multi-Head Attention (MHA) based language models. We address the challenges through head pruning, a novel head-specific weighted aggregation mechanism, and a client selection strategy. Head pruning minimizes training complexity within the clients, guided by the importance score computed based on the confidence of the attention head. Weighted aggregation of heads ensures the global model captures crucial updates from diverse clients complementing our client selection strategy. We show results on the MultiNLI benchmark along with 20 Newsgroups, XL-Sum, and E2E NLG datasets. We use the MultiNLI dataset and T5-small model with LoRA as our PEFT method, attaining sparsity levels of up to 90%, resulting in a communication advantage of up to 1.8x and a reduction in training OPs of 3.9x while maintaining the accuracy drop under 2%.

cs.CL↗

arXiv2025

Fast and Cost-effective Speculative Edge-Cloud Decoding with Early Exits

Large Language Models (LLMs) enable various applications on edge devices such as smartphones, wearables, and embodied robots. However, their deployment often depends on expensive cloud-based APIs, creating high operational costs, which limit access for smaller organizations and raise sustainability concerns. Certain LLMs can be deployed on-device, offering a cost-effective solution with reduced latency and improved privacy. Yet, limited computing resources constrain the size and accuracy of models that can be deployed, necessitating a collaborative design between edge and cloud. We propose a fast and cost-effective speculative edge-cloud decoding framework with a large target model on the server and a small draft model on the device. By introducing early exits in the target model, tokens are generated mid-verification, allowing the client to preemptively draft subsequent tokens before final verification, thus utilizing idle time and enhancing parallelism between edge and cloud. Using an NVIDIA Jetson Nano (client) and an A100 GPU (server) with Vicuna-68M (draft) and Llama2-7B (target) models, our method achieves up to a 35% reduction in latency compared to cloud-based autoregressive decoding, with an additional 11% improvement from preemptive drafting. To demonstrate real-world applicability, we deploy our method on the Unitree Go2 quadruped robot using Vision-Language Model (VLM) based control, achieving a 21% speedup over traditional cloud-based autoregressive decoding. These results demonstrate the potential of our framework for real-time LLM and VLM applications on resource-constrained edge devices.

cs.RO↗

arXiv2025

Event2Vec: Processing Neuromorphic Events Directly by Representations in Vector Space

Neuromorphic event cameras possess superior temporal resolution, power efficiency, and dynamic range compared to traditional cameras. However, their asynchronous and sparse data format poses a significant challenge for conventional deep learning methods. Most existing methods either densify events into frames, sacrificing their sparse asynchronous nature, or use irregular models that are less compatible with GPU acceleration. Inspired by word-to-vector models, we propose event2vec, a novel representation that allows Transformers to process events directly. We demonstrate the effectiveness of event2vec on the DVS Gesture, ASL-DVS, and DVS-Lip benchmarks, showing that event2vec is remarkably parameter-efficient, features high throughput and low latency, and achieves high accuracy even with an extremely low number of events or low spatial resolutions. These results show that sparse asynchronous event data can be directly integrated into high-throughput Transformer architectures, offering an efficient paradigm for real-time neuromorphic vision. The code is provided at https://github.com/Intelligent-Computing-Lab-Panda/event2vec.

cs.CV↗

arXiv2025

GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration

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.

cs.LG↗

arXiv2025

PacQ: A SIMT Microarchitecture for Efficient Dataflow in Hyper-asymmetric GEMMs

Weight-only quantization has been widely explored in large language models (LLMs) to reduce memory storage and data loading overhead. During deployment on single-instruction-multiple-threads (SIMT) architectures, weights are stored in low-precision integer (INT) format, while activations remain in full-precision floating-point (FP) format to preserve inference accuracy. Although memory footprint and data loading requirements for weight matrices are reduced, computation performance gains remain limited due to the need to convert weights back to FP format through unpacking and dequantization before GEMM operations. In this work, we investigate methods to accelerate GEMM operations involving packed low-precision INT weights and high-precision FP activations, defining this as the hyper-asymmetric GEMM problem. Our approach co-optimizes tile-level packing and dataflow strategies for INT weight matrices. We further design a specialized FP-INT multiplier unit tailored to our packing and dataflow strategies, enabling parallel processing of multiple INT weights. Finally, we integrate the packing, dataflow, and multiplier unit into PacQ, a SIMT microarchitecture designed to efficiently accelerate hyper-asymmetric GEMMs. We show that PacQ can achieve up to 1.99x speedup and 81.4% reduction in EDP compared to weight-only quantized LLM workloads running on conventional SIMT baselines.

cs.AR↗