Souvik Kundu

Amey Agrawal, Anmol Agarwal, Nitin Kedia et al. · gatech

Serving large language models (LLMs) in production can incur substantial costs, which has prompted recent advances in inference system optimizations. Today, these systems are evaluated against conventional latency and throughput metrics (eg. TTFT, TBT, Normalised Latency and TPOT). However, these metrics fail to fully capture the nuances of LLM inference, leading to an incomplete assessment of user-facing performance crucial for real-time applications such as chat and translation. In this paper, we first identify the pitfalls of current performance metrics in evaluating LLM inference systems. We then propose Etalon, a comprehensive performance evaluation framework that includes fluidity-index -- a novel metric designed to reflect the intricacies of the LLM inference process and its impact on real-time user experience. Finally, we evaluate various existing open-source platforms and model-as-a-service offerings using Etalon, discussing their strengths and weaknesses. Etalon is available at https://github.com/project-etalon/etalon.

Aradhana Mohan Parvathy, Soumendu Kumar Ghosh, Shamik Kundu et al.

The rapid growth in sizes of Large language models (LLMs) results in high compute and memory costs during inference. Quantization has been a significant pathway to addressing this challenge. In the quest to push the limits of quantization, weights, which are static, can often be quantized aggressively (e.g. 4 bits), while activations often require higher precision (e.g., 8 bits) to preserve accuracy, forcing hardware to operate with higher-precision datapaths. We leverage the statistical property that a significant fraction of activations are concentrated around zero, resulting in sparsity in the higher-order bits. Our proposal, SPARQLe, is a hardware-software co-design framework that exploits this sub-precision redundancy in any given quantized model. SPARQLe represents each 2k-bit activation tensor as a dense k-bit LSB tensor and a sparse k-bit MSB tensor compressed with a precision bitmap, and proposes a lightweight algorithm to increase MSB sparsity. SPARQLe reduces activation memory traffic and enables efficient computation on k-bit datapaths while preserving 2k-bit activation accuracy. SPARQLe includes an accelerator that operates directly on this hybrid format with minimal control overheads. Across the BitNet 3B, Llama2 7B, and Llama3 8B models, SPARQLe reduces prefill latency by 16-24.3% and decode latency by 13.5-23.4%, with 17-26.7% and 6.5-14.2% lower prefill and decode energy, respectively. SPARQLe demonstrates that sub-precision activation sparsity offers an effective and complementary pathway towards efficient LLM inference.

Sara Babakniya, Souvik Kundu, Saurav Prakash et al.

Edge devices can benefit remarkably from federated learning due to their distributed nature; however, their limited resource and computing power poses limitations in deployment. A possible solution to this problem is to utilize off-the-shelf sparse learning algorithms at the clients to meet their resource budget. However, such naive deployment in the clients causes significant accuracy degradation, especially for highly resource-constrained clients. In particular, our investigations reveal that the lack of consensus in the sparsity masks among the clients may potentially slow down the convergence of the global model and cause a substantial accuracy drop. With these observations, we present \textit{federated lottery aware sparsity hunting} (FLASH), a unified sparse learning framework for training a sparse sub-model that maintains the performance under ultra-low parameter density while yielding proportional communication benefits. Moreover, given that different clients may have different resource budgets, we present \textit{hetero-FLASH} where clients can take different density budgets based on their device resource limitations instead of supporting only one target parameter density. Experimental analysis on diverse models and datasets shows the superiority of FLASH in closing the gap with an unpruned baseline while yielding up to $\mathord{\sim}10.1\%$ improved accuracy with $\mathord{\sim}10.26\times$ fewer communication, compared to existing alternatives, at similar hyperparameter settings. Code is available at \url{https://github.com/SaraBabakN/flash_fl}.

Hongyi Wang, Felipe Maia Polo, Yuekai Sun et al.

Training AI models that generalize across tasks and domains has long been among the open problems driving AI research. The emergence of Foundation Models made it easier to obtain expert models for a given task, but the heterogeneity of data that may be encountered at test time often means that any single expert is insufficient. We consider the Fusion of Experts (FoE) problem of fusing outputs of expert models with complementary knowledge of the data distribution and formulate it as an instance of supervised learning. Our method is applicable to both discriminative and generative tasks and leads to significant performance improvements in image and text classification, text summarization, multiple-choice QA, and automatic evaluation of generated text. We also extend our method to the "frugal" setting where it is desired to reduce the number of expert model evaluations at test time. Our implementation is publicly available at https://github.com/hwang595/FoE-ICLR2024.

Yonggan Fu, Ye Yuan, Souvik Kundu et al.

Generalizable Neural Radiance Fields (GNeRF) are one of the most promising real-world solutions for novel view synthesis, thanks to their cross-scene generalization capability and thus the possibility of instant rendering on new scenes. While adversarial robustness is essential for real-world applications, little study has been devoted to understanding its implication on GNeRF. We hypothesize that because GNeRF is implemented by conditioning on the source views from new scenes, which are often acquired from the Internet or third-party providers, there are potential new security concerns regarding its real-world applications. Meanwhile, existing understanding and solutions for neural networks' adversarial robustness may not be applicable to GNeRF, due to its 3D nature and uniquely diverse operations. To this end, we present NeRFool, which to the best of our knowledge is the first work that sets out to understand the adversarial robustness of GNeRF. Specifically, NeRFool unveils the vulnerability patterns and important insights regarding GNeRF's adversarial robustness. Built upon the above insights gained from NeRFool, we further develop NeRFool+, which integrates two techniques capable of effectively attacking GNeRF across a wide range of target views, and provide guidelines for defending against our proposed attacks. We believe that our NeRFool/NeRFool+ lays the initial foundation for future innovations in developing robust real-world GNeRF solutions. Our codes are available at: https://github.com/GATECH-EIC/NeRFool.

Lu Yin, Ajay Jaiswal, Shiwei Liu et al.

We present Junk DNA Hypothesis by adopting a novel task-centric angle for the pre-trained weights of large language models (LLMs). It has been believed that weights in LLMs contain significant redundancy, leading to the conception that a considerable chunk of the parameters can be removed by pruning without compromising performance. Contrary to this belief, this paper presents a counter-argument: small-magnitude weights of pre-trained model weights encode vital knowledge essential for tackling difficult downstream tasks - manifested as the monotonic relationship between the performance drop of downstream tasks across the difficulty spectrum, as we prune more pre-trained weights by magnitude. Moreover, we reveal that these seemingly inconsequential weights can result in irreparable loss of knowledge and performance degradation in difficult tasks, even when downstream continual training is allowed. Interestingly, our evaluations show that the other popular compression, namely quantization, fails to exhibit similar monotonic effect and does not as convincingly disentangle this task-difficulty information. To study formally, we introduce several quantifiable metrics to gauge the downstream task difficulty: (1) within the same task category, and (2) across different task categories. Our extensive experiments substantiate the Junk DNA Hypothesis across a diverse range of model sizes, tasks, datasets, and even pruning methods. Codes are available at: https://github.com/VITA-Group/Junk_DNA_Hypothesis.git.

Fang Chen, Gourav Datta, Souvik Kundu et al. · tsinghua

Efficient custom pooling techniques that can aggressively trim the dimensions of a feature map and thereby reduce inference compute and memory footprint for resource-constrained computer vision applications have recently gained significant traction. However, prior pooling works extract only the local context of the activation maps, limiting their effectiveness. In contrast, we propose a novel non-local self-attentive pooling method that can be used as a drop-in replacement to the standard pooling layers, such as max/average pooling or strided convolution. The proposed self-attention module uses patch embedding, multi-head self-attention, and spatial-channel restoration, followed by sigmoid activation and exponential soft-max. This self-attention mechanism efficiently aggregates dependencies between non-local activation patches during down-sampling. Extensive experiments on standard object classification and detection tasks with various convolutional neural network (CNN) architectures demonstrate the superiority of our proposed mechanism over the state-of-the-art (SOTA) pooling techniques. In particular, we surpass the test accuracy of existing pooling techniques on different variants of MobileNet-V2 on ImageNet by an average of 1.2%. With the aggressive down-sampling of the activation maps in the initial layers (providing up to 22x reduction in memory consumption), our approach achieves 1.43% higher test accuracy compared to SOTA techniques with iso-memory footprints. This enables the deployment of our models in memory-constrained devices, such as micro-controllers (without losing significant accuracy), because the initial activation maps consume a significant amount of on-chip memory for high-resolution images required for complex vision tasks. Our proposed pooling method also leverages the idea of channel pruning to further reduce memory footprints.

Saptarshi Mitra, Yifan Zhang, Rachid Karami et al.

Mixture-of-Agents (MoA) systems improve reasoning accuracy by routing each query to multiple expert LLMs and aggregating their outputs. Efficiently executing this workload on limited GPU resources has bottlenecks. Skill-based routing creates skewed expert demand, and combining instruction-tuned LLMs with long-reasoning models results in extreme variability in generation lengths. Consequently, traditional scheduling strategies suffer from significant GPU idling and throughput collapse due to load imbalances. We present MOSAIC, a scheduling framework to accelerate MoA workloads. First, we formulate an Integer Linear Program (ILP) based scheduler that jointly optimizes expert placement and per-worker prompt assignment from offline-profiled costs, replicating reasoning experts across workers while pinning lightweight ones. Second, MOSAIC uses confidence-aware adaptive aggregation, leveraging inter-expert agreement to bypass the heavy final aggregator LLM for consensus queries. In our 4-GPU system, MOSAIC achieves up to 2.5x expert-stage, 4.23x aggregator-stage and 1.7~2.3x end-to-end speedups over the baseline scheduler, while matching accuracy within 0.1pp.

Akshat Ramachandran, Souvik Kundu, Tushar Krishna

We present CLAMP-ViT, a data-free post-training quantization method for vision transformers (ViTs). We identify the limitations of recent techniques, notably their inability to leverage meaningful inter-patch relationships, leading to the generation of simplistic and semantically vague data, impacting quantization accuracy. CLAMP-ViT employs a two-stage approach, cyclically adapting between data generation and model quantization. Specifically, we incorporate a patch-level contrastive learning scheme to generate richer, semantically meaningful data. Furthermore, we leverage contrastive learning in layer-wise evolutionary search for fixed- and mixed-precision quantization to identify optimal quantization parameters while mitigating the effects of a non-smooth loss landscape. Extensive evaluations across various vision tasks demonstrate the superiority of CLAMP-ViT, with performance improvements of up to 3% in top-1 accuracy for classification, 0.6 mAP for object detection, and 1.5 mIoU for segmentation at similar or better compression ratio over existing alternatives. Code is available at https://github.com/georgia-tech-synergy-lab/CLAMP-ViT.git

Gourav Datta, Souvik Kundu, Zihan Yin et al.

The demand to process vast amounts of data generated from state-of-the-art high resolution cameras has motivated novel energy-efficient on-device AI solutions. Visual data in such cameras are usually captured in the form of analog voltages by a sensor pixel array, and then converted to the digital domain for subsequent AI processing using analog-to-digital converters (ADC). Recent research has tried to take advantage of massively parallel low-power analog/digital computing in the form of near- and in-sensor processing, in which the AI computation is performed partly in the periphery of the pixel array and partly in a separate on-board CPU/accelerator. Unfortunately, high-resolution input images still need to be streamed between the camera and the AI processing unit, frame by frame, causing energy, bandwidth, and security bottlenecks. To mitigate this problem, we propose a novel Processing-in-Pixel-in-memory (P2M) paradigm, that customizes the pixel array by adding support for analog multi-channel, multi-bit convolution, batch normalization, and ReLU (Rectified Linear Units). Our solution includes a holistic algorithm-circuit co-design approach and the resulting P2M paradigm can be used as a drop-in replacement for embedding memory-intensive first few layers of convolutional neural network (CNN) models within foundry-manufacturable CMOS image sensor platforms. Our experimental results indicate that P2M reduces data transfer bandwidth from sensors and analog to digital conversions by ~21x, and the energy-delay product (EDP) incurred in processing a MobileNetV2 model on a TinyML use case for visual wake words dataset (VWW) by up to ~11x compared to standard near-processing or in-sensor implementations, without any significant drop in test accuracy.

Shashank Nag, Gourav Datta, Souvik Kundu et al.

Vision Transformer models, such as ViT, Swin Transformer, and Transformer-in-Transformer, have recently gained significant traction in computer vision tasks due to their ability to capture the global relation between features which leads to superior performance. However, they are compute-heavy and difficult to deploy in resource-constrained edge devices. Existing hardware accelerators, including those for the closely-related BERT transformer models, do not target highly resource-constrained environments. In this paper, we address this gap and propose ViTA - a configurable hardware accelerator for inference of vision transformer models, targeting resource-constrained edge computing devices and avoiding repeated off-chip memory accesses. We employ a head-level pipeline and inter-layer MLP optimizations, and can support several commonly used vision transformer models with changes solely in our control logic. We achieve nearly 90% hardware utilization efficiency on most vision transformer models, report a power of 0.88W when synthesised with a clock of 150 MHz, and get reasonable frame rates - all of which makes ViTA suitable for edge applications.

Souvik Kundu, Shunlin Lu, Yuke Zhang et al.

The large number of ReLU non-linearity operations in existing deep neural networks makes them ill-suited for latency-efficient private inference (PI). Existing techniques to reduce ReLU operations often involve manual effort and sacrifice significant accuracy. In this paper, we first present a novel measure of non-linearity layers' ReLU sensitivity, enabling mitigation of the time-consuming manual efforts in identifying the same. Based on this sensitivity, we then present SENet, a three-stage training method that for a given ReLU budget, automatically assigns per-layer ReLU counts, decides the ReLU locations for each layer's activation map, and trains a model with significantly fewer ReLUs to potentially yield latency and communication efficient PI. Experimental evaluations with multiple models on various datasets show SENet's superior performance both in terms of reduced ReLUs and improved classification accuracy compared to existing alternatives. In particular, SENet can yield models that require up to ~2x fewer ReLUs while yielding similar accuracy. For a similar ReLU budget SENet can yield models with ~2.32% improved classification accuracy, evaluated on CIFAR-100.

Vladimir Araujo, Andrés Carvallo, Souvik Kundu et al.

Due to the success of pre-trained language models, versions of languages other than English have been released in recent years. This fact implies the need for resources to evaluate these models. In the case of Spanish, there are few ways to systematically assess the models' quality. In this paper, we narrow the gap by building two evaluation benchmarks. Inspired by previous work (Conneau and Kiela, 2018; Chen et al., 2019), we introduce Spanish SentEval and Spanish DiscoEval, aiming to assess the capabilities of stand-alone and discourse-aware sentence representations, respectively. Our benchmarks include considerable pre-existing and newly constructed datasets that address different tasks from various domains. In addition, we evaluate and analyze the most recent pre-trained Spanish language models to exhibit their capabilities and limitations. As an example, we discover that for the case of discourse evaluation tasks, mBERT, a language model trained on multiple languages, usually provides a richer latent representation than models trained only with documents in Spanish. We hope our contribution will motivate a fairer, more comparable, and less cumbersome way to evaluate future Spanish language models.

Yue Niu, Saurav Prakash, Souvik Kundu et al.

Federated Learning (FL) is emerging as a popular, promising decentralized learning framework that enables collaborative training among clients, with no need to share private data between them or to a centralized server. However, considering many edge clients do not have sufficient computing, memory, or communication capabilities, federated learning of large models still faces significant bottlenecks. To keep such weak but crucial clients in the loop, prior works either consider a heterogeneous-client setting where clients train models with different sizes; or offload training to the server. However, the heterogeneous-client setting requires some clients to train full model, which is not aligned with the resource-constrained setting; while the latter ones break privacy promises in FL when sharing intermediate representations or labels with the server. To overcome these limitations, in this work, we formulate a realistic, but much less explored, cross-device FL setting in which no client can train a full large model nor is willing to share any intermediate information with the remote server. Under such a formulation, we develop a principal sub-model (PriSM) training methodology to collaboratively train a full large model, while assigning each client a small sub-model that is a probabilistic low-rank approximation to the full server model. When creating sub-models, PriSM first performs a principal kernel analysis in the orthogonal kernel space to obtain importance of each kernel. Then, PriSM adopts a novel importance-aware sampling process to select a subset of kernels (i.e., a kernel with high importance is assigned with a higher sampling probability). This sampling process ensures each sub-model is still a low-rank approximation to the full model, while all sub-models together achieve nearly full coverage on the principal kernels.

Souvik Kundu, Yuke Zhang, Dake Chen et al.

Large number of ReLU and MAC operations of Deep neural networks make them ill-suited for latency and compute-efficient private inference. In this paper, we present a model optimization method that allows a model to learn to be shallow. In particular, we leverage the ReLU sensitivity of a convolutional block to remove a ReLU layer and merge its succeeding and preceding convolution layers to a shallow block. Unlike existing ReLU reduction methods, our joint reduction method can yield models with improved reduction of both ReLUs and linear operations by up to 1.73x and 1.47x, respectively, evaluated with ResNet18 on CIFAR-100 without any significant accuracy-drop.

Gourav Datta, Zeyu Liu, Md Abdullah-Al Kaiser et al.

Due to the high activation sparsity and use of accumulates (AC) instead of expensive multiply-and-accumulates (MAC), neuromorphic spiking neural networks (SNNs) have emerged as a promising low-power alternative to traditional DNNs for several computer vision (CV) applications. However, most existing SNNs require multiple time steps for acceptable inference accuracy, hindering real-time deployment and increasing spiking activity and, consequently, energy consumption. Recent works proposed direct encoding that directly feeds the analog pixel values in the first layer of the SNN in order to significantly reduce the number of time steps. Although the overhead for the first layer MACs with direct encoding is negligible for deep SNNs and the CV processing is efficient using SNNs, the data transfer between the image sensors and the downstream processing costs significant bandwidth and may dominate the total energy. To mitigate this concern, we propose an in-sensor computing hardware-software co-design framework for SNNs targeting image recognition tasks. Our approach reduces the bandwidth between sensing and processing by 12-96x and the resulting total energy by 2.32x compared to traditional CV processing, with a 3.8% reduction in accuracy on ImageNet.

Pavan Kumar Reddy Boppidi, Victor Jeffry Louis, Arvind Subramaniam et al.

Independent component analysis is an unsupervised learning approach for computing the independent components (ICs) from the multivariate signals or data matrix. The ICs are evaluated based on the multiplication of the weight matrix with the multivariate data matrix. This study proposes a novel memristor crossbar array for the implementation of both ACY ICA and Fast ICA for blind source separation. The data input was applied in the form of pulse width modulated voltages to the crossbar array and the weight of the implemented neural network is stored in the memristor. The output charges from the memristor columns are used to calculate the weight update, which is executed through the voltages kept higher than the memristor Set/Reset voltages. In order to demonstrate its potential application, the proposed memristor crossbar arrays based fast ICA architecture is employed for image source separation problem. The experimental results demonstrate that the proposed approach is very effective to separate image sources, and also the contrast of the images are improved with an improvement factor in terms of percentage of structural similarity as 67.27% when compared with the software-based implementation of conventional ACY ICA and Fast ICA algorithms.

Sreetama Sarkar, Gourav Datta, Souvik Kundu et al.

Video tasks are compute-heavy and thus pose a challenge when deploying in real-time applications, particularly for tasks that require state-of-the-art Vision Transformers (ViTs). Several research efforts have tried to address this challenge by leveraging the fact that large portions of the video undergo very little change across frames, leading to redundant computations in frame-based video processing. In particular, some works leverage pixel or semantic differences across frames, however, this yields limited latency benefits with significantly increased memory overhead. This paper, in contrast, presents a strategy for masking regions in video frames that leverages the semantic information in images and the temporal correlation between frames to significantly reduce FLOPs and latency with little to no penalty in performance over baseline models. In particular, we demonstrate that by leveraging extracted features from previous frames, ViT backbones directly benefit from region masking, skipping up to 80% of input regions, improving FLOPs and latency by 3.14x and 1.5x. We improve memory and latency over the state-of-the-art (SOTA) by 2.3x and 1.14x, while maintaining similar detection performance. Additionally, our approach demonstrates promising results on convolutional neural networks (CNNs) and provides latency improvements over the SOTA up to 1.3x using specialized computational kernels.

Yue Hu, Xinan Ye, Yifei Liu et al.

This paper presents "FireFly", a synthetic dataset for ember detection created using Unreal Engine 4 (UE4), designed to overcome the current lack of ember-specific training resources. To create the dataset, we present a tool that allows the automated generation of the synthetic labeled dataset with adjustable parameters, enabling data diversity from various environmental conditions, making the dataset both diverse and customizable based on user requirements. We generated a total of 19,273 frames that have been used to evaluate FireFly on four popular object detection models. Further to minimize human intervention, we leveraged a trained model to create a semi-automatic labeling process for real-life ember frames. Moreover, we demonstrated an up to 8.57% improvement in mean Average Precision (mAP) in real-world wildfire scenarios compared to models trained exclusively on a small real dataset.

Souvik Kundu, Sairam Sundaresan, Sharath Nittur Sridhar et al.

Existing deep neural networks (DNNs) that achieve state-of-the-art (SOTA) performance on both clean and adversarially-perturbed images rely on either activation or weight conditioned convolution operations. However, such conditional learning costs additional multiply-accumulate (MAC) or addition operations, increasing inference memory and compute costs. To that end, we present a sparse mixture once for all adversarial training (SMART), that allows a model to train once and then in-situ trade-off between accuracy and robustness, that too at a reduced compute and parameter overhead. In particular, SMART develops two expert paths, for clean and adversarial images, respectively, that are then conditionally trained via respective dedicated sets of binary sparsity masks. Extensive evaluations on multiple image classification datasets across different models show SMART to have up to 2.72x fewer non-zero parameters costing proportional reduction in compute overhead, while yielding SOTA accuracy-robustness trade-off. Additionally, we present insightful observations in designing sparse masks to successfully condition on both clean and perturbed images.

Yuchen Xia, Jiho Kim, Yuhan Chen et al.

Due to the cost-prohibitive nature of training Large Language Models (LLMs), fine-tuning has emerged as an attractive alternative for specializing LLMs for specific tasks using limited compute resources in a cost-effective manner. In this paper, we characterize sparse Mixture of Experts (MoE) based LLM fine-tuning to understand their accuracy and runtime performance on a single GPU. Our evaluation provides unique insights into the training efficacy of sparse and dense versions of MoE models, as well as their runtime characteristics, including maximum batch size, execution time breakdown, end-to-end throughput, GPU hardware utilization, and load distribution. Our study identifies the optimization of the MoE layer as crucial for further improving the performance of LLM fine-tuning. Using our profiling results, we also develop and validate an analytical model to estimate the cost of LLM fine-tuning on the cloud. This model, based on parameters of the model and GPU architecture, estimates LLM throughput and the cost of training, aiding practitioners in industry and academia to budget the cost of fine-tuning a specific model.

Hanjiang Wu, Abhimanyu Rajeshkumar Bambhaniya, Sarbartha Banerjee et al.

Modern large language model (LLM) inference has progressively disaggregated to keep pace with growing model sizes and tight TTFT and TPOT service-level objectives: from chunked-prefill aggregation, to prefill-decode (P/D) disaggregation, and most recently to operator-level Attention-FFN Disaggregation (AFD). This trend is especially important for mixture-of-experts (MoE) models, where memory-bound attention, compute-intensive expert FFNs, and MoE dispatch/combine communication create distinct resource demands. AFD further exposes this heterogeneity by placing attention and MoE-FFN execution on separate GPU groups. Each level of disaggregation deepens the scheduling design space across workload characteristics, resource allocation, and interconnect topology, raising the central question: when does each level actually pay off? We systematically characterize this trade-off for MoE inference across realistic workloads spanning input/output sequence lengths, prefix-KV reuse, and per-user latency constraints. Using chunked-prefill and P/D disaggregation as baselines, we study the benefits and limits of AFD at scale through a framework that fuses on-device kernel measurements with high-fidelity network simulation. Under strict TTFT/TPOT SLOs, AFD sustains around 4k tokens/s of system throughput on DeepSeek-V3.2 across chat, coding, and agentic-coding workloads, where non-AFD deployments are infeasible. We distill concrete takeaways for jointly optimizing throughput and interactivity, including how to partition attention and FFN across GPUs as a function of workload and model architecture, providing design principles for current rack- and cluster-scale deployments as well as future disaggregated AI infrastructure.

Erfan Baghaei Potraghloo, Seyedarmin Azizi, Souvik Kundu et al.

Instruction-tuned large language models produce helpful, structured responses, but how robust is this helpfulness when trivially constrained? We show that simple lexical constraints (banning a single punctuation character or common word) cause instruction-tuned LLMs to collapse their responses, losing 14--48% of comprehensiveness in pairwise evaluation across three open-weight model families and one closed-weight model (GPT-4o-mini). The baseline response is preferred in 77--100% of 1,920 pairwise comparisons judged by GPT-4o-mini and GPT-4o. Notably, GPT-4o-mini suffers 31% comprehensiveness loss (99% baseline win rate), demonstrating that the fragility extends to commercially deployed closed-weight models, contrary to prior findings on format-level constraints. Through mechanistic analysis, we identify this as a planning failure: two-pass generation (free generation followed by constrained rewriting) recovers 59--96% of response length, and linear probes on prompt representations predict response length with $R^2 = 0.51$--$0.93$ before generation begins, with $R^2$ tracking collapse severity across models. The same probes yield negative $R^2$ on base models, confirming that instruction tuning creates the representational structure encoding the collapse decision. Crucially, base models show no systematic collapse under identical constraints, with effects that are small, noisy, and bidirectional, demonstrating that instruction tuning creates this fragility by coupling task competence to narrow surface-form templates. The effect replicates on MT-Bench across all eight task categories. We further show that standard independent LLM-as-judge evaluation detects only a 3.5% average quality drop where pairwise evaluation reveals 23%, exposing a methodological blind spot in how constrained generation is assessed.

Souvik Kundu, Sairam Sundaresan, Massoud Pedram et al.

Existing models that achieve state-of-the-art (SOTA) performance on both clean and adversarially-perturbed images rely on convolution operations conditioned with feature-wise linear modulation (FiLM) layers. These layers require many new parameters and are hyperparameter sensitive. They significantly increase training time, memory cost, and potential latency which can prove costly for resource-limited or real-time applications. In this paper, we present a fast learnable once-for-all adversarial training (FLOAT) algorithm, which instead of the existing FiLM-based conditioning, presents a unique weight conditioned learning that requires no additional layer, thereby incurring no significant increase in parameter count, training time, or network latency compared to standard adversarial training. In particular, we add configurable scaled noise to the weight tensors that enables a trade-off between clean and adversarial performance. Extensive experiments show that FLOAT can yield SOTA performance improving both clean and perturbed image classification by up to ~6% and ~10%, respectively. Moreover, real hardware measurement shows that FLOAT can reduce the training time by up to 1.43x with fewer model parameters of up to 1.47x on iso-hyperparameter settings compared to the FiLM-based alternatives. Additionally, to further improve memory efficiency we introduce FLOAT sparse (FLOATS), a form of non-iterative model pruning and provide detailed empirical analysis to provide a three way accuracy-robustness-complexity trade-off for these new class of pruned conditionally trained models.

Yichao Yuan, Ankita Nayak, Souvik Kundu et al.

Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. This paper characterizes ReAct-style agents from both the LLM-serving and tool-execution perspectives using an end-to-end tracing infrastructure across reasoning and non-reasoning Gemma and Qwen configurations on five agentic benchmarks. Our study shows that agentic workloads are not simply long-prompt workloads: with effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. We also find that tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later. These results show that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.

Hao Kang, Qingru Zhang, Souvik Kundu et al. · gatech

Key-value (KV) caching has become the de-facto to accelerate generation speed for large language models (LLMs) inference. However, the growing cache demand with increasing sequence length has transformed LLM inference to be a memory bound problem, significantly constraining the system throughput. Existing methods rely on dropping unimportant tokens or quantizing all entries uniformly. Such methods, however, often incur high approximation errors to represent the compressed matrices. The autoregressive decoding process further compounds the error of each step, resulting in critical deviation in model generation and deterioration of performance. To tackle this challenge, we propose GEAR, an efficient KV cache compression framework that achieves near-lossless high-ratio compression. GEAR first applies quantization to majority of entries of similar magnitudes to ultra-low precision. It then employs a low rank matrix to approximate the quantization error, and a sparse matrix to remedy individual errors from outlier entries. By adeptly integrating three techniques, GEAR is able to fully exploit their synergistic potentials. Our experiments demonstrate that compared to alternatives, GEAR achieves near-lossless 4-bit KV cache compression with up to 2.38x throughput improvement, while reducing peak-memory size up to 2.29x. Our code is publicly available at https://github.com/HaoKang-Timmy/GEAR.

Sharath Nittur Sridhar, Souvik Kundu, Sairam Sundaresan et al.

One-Shot Neural Architecture Search (NAS) algorithms often rely on training a hardware agnostic super-network for a domain specific task. Optimal sub-networks are then extracted from the trained super-network for different hardware platforms. However, training super-networks from scratch can be extremely time consuming and compute intensive especially for large models that rely on a two-stage training process of pre-training and fine-tuning. State of the art pre-trained models are available for a wide range of tasks, but their large sizes significantly limits their applicability on various hardware platforms. We propose InstaTune, a method that leverages off-the-shelf pre-trained weights for large models and generates a super-network during the fine-tuning stage. InstaTune has multiple benefits. Firstly, since the process happens during fine-tuning, it minimizes the overall time and compute resources required for NAS. Secondly, the sub-networks extracted are optimized for the target task, unlike prior work that optimizes on the pre-training objective. Finally, InstaTune is easy to "plug and play" in existing frameworks. By using multi-objective evolutionary search algorithms along with lightly trained predictors, we find Pareto-optimal sub-networks that outperform their respective baselines across different performance objectives such as accuracy and MACs. Specifically, we demonstrate that our approach performs well across both unimodal (ViT and BERT) and multi-modal (BEiT-3) transformer based architectures.

Jiayi Tian, Seyedarmin Azizi, Yequan Zhao et al.

Large reasoning models (LRMs) often cost significant key-value (KV) cache overhead, due to their linear growth with the verbose chain-of-thought (CoT) reasoning process. This costs both memory and throughput bottleneck limiting their efficient deployment. Towards reducing KV cache size during inference, we first investigate the effectiveness of existing KV cache eviction methods for CoT reasoning. Interestingly, we find that due to unstable token-wise scoring and the reduced effective KV budget caused by padding tokens, state-of-the-art (SoTA) eviction methods fail to maintain accuracy in the multi-batch setting. Additionally, these methods often generate longer sequences than the original model, as semantic-unaware token-wise eviction leads to repeated revalidation during reasoning. To address these issues, we present \textbf{SkipKV}, a \textbf{\textit{training-free}} KV compression method for selective \textit{eviction} and \textit{generation} operating at a coarse-grained sentence-level sequence removal for efficient CoT reasoning. In specific, it introduces a \textit{sentence-scoring metric} to identify and remove highly similar sentences while maintaining semantic coherence. To suppress redundant generation, SkipKV dynamically adjusts a steering vector to update the hidden activation states during inference enforcing the LRM to generate concise response. Extensive evaluations on multiple reasoning benchmarks demonstrate the effectiveness of SkipKV in maintaining up to $\mathbf{26.7}\%$ improved accuracy compared to the alternatives, at a similar compression budget. Additionally, compared to SoTA, SkipKV yields up to $\mathbf{1.6}\times$ fewer generation length while improving throughput up to $\mathbf{1.7}\times$.

Souvik Kundu, Sharath Nittur Sridhar, Maciej Szankin et al.

Large pre-trained language models have recently gained significant traction due to their improved performance on various down-stream tasks like text classification and question answering, requiring only few epochs of fine-tuning. However, their large model sizes often prohibit their applications on resource-constrained edge devices. Existing solutions of yielding parameter-efficient BERT models largely rely on compute-exhaustive training and fine-tuning. Moreover, they often rely on additional compute heavy models to mitigate the performance gap. In this paper, we present Sensi-BERT, a sensitivity driven efficient fine-tuning of BERT models that can take an off-the-shelf pre-trained BERT model and yield highly parameter-efficient models for downstream tasks. In particular, we perform sensitivity analysis to rank each individual parameter tensor, that then is used to trim them accordingly during fine-tuning for a given parameter or FLOPs budget. Our experiments show the efficacy of Sensi-BERT across different downstream tasks including MNLI, QQP, QNLI, SST-2 and SQuAD, showing better performance at similar or smaller parameter budget compared to various alternatives.

Runjin Chen, Zhenyu Zhang, Junyuan Hong et al.

Large Language Models (LLMs), such as OpenAI's o1-series have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism. However, recent studies reveal substantial redundancy in the CoT reasoning traces, which not only increases inference latency but also negatively impacts model performance by diverting attention to unnecessary reasoning paths. To address this issue, we investigate the internal reasoning structures of LLMs and categorize them into three primary thought types: execution, reflection, and transition thoughts. Moreover, our analysis reveals that excessive reflection and transition thoughts are strongly correlated with failure cases and these thought categories exhibit clear separation in the latent space. Based on these, we introduce SEAL (Steerable reasoning calibration), a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains. SEAL consists of an offline stage for extracting the reasoning steering vector in the latent space, followed by an on-the-fly calibration of the reasoning trace through representation intervention using the steering vector. Notably, the steering vector exhibits strong transferability across various tasks. Extensive experiments across multiple models (DeepSeek-R1-Distill and QwQ-32B-Preview) and benchmarks (Math500, GSM8K, LiveCodeBench) validate the effectiveness of SEAL, up to a 11% improvement in accuracy while reducing reasoning tokens by 11.8% to 50.4%. Our code is publicly available at https://github.com/VITA-Group/SEAL.

Wenhao Zheng, Yixiao Chen, Weitong Zhang et al.

Large language models have achieved remarkable success in various tasks but suffer from high computational costs during inference, limiting their deployment in resource-constrained applications. To address this issue, we propose a novel Collaborative Inference with Token-lEvel Routing (CITER) framework that enables efficient collaboration between small and large language models (SLMs \& LLMs) through a token-level routing strategy. Specifically, CITER routes non-critical tokens to an SLM for efficiency and routes critical tokens to an LLM for generalization quality. We formulate router training as a policy optimization, where the router receives rewards based on both the quality of predictions and the inference costs of generation. This allows the router to learn to predict token-level routing scores and make routing decisions based on both the current token and the future impact of its decisions. To further accelerate the reward evaluation process, we introduce a shortcut which significantly reduces the costs of the reward estimation and improving the practicality of our approach. Extensive experiments on five benchmark datasets demonstrate that CITER reduces the inference costs while preserving high-quality generation, offering a promising solution for real-time and resource-constrained applications. Our data and code are available at https://github.com/aiming-lab/CITER.

Akshat Ramachandran, Souvik Kundu, Tushar Krishna

Quantization of foundational models (FMs) is significantly more challenging than traditional DNNs due to the emergence of large magnitude values called outliers. Existing outlier-aware algorithm-architecture co-design techniques either use mixed-precision, retaining outliers at high precision but compromise hardware efficiency, or quantize inliers and outliers at the same precision, improving hardware efficiency at the cost of accuracy. To address this mutual exclusivity, we propose MicroScopiQ, a novel co-design technique that leverages pruning to complement outlier-aware quantization. MicroScopiQ retains outliers at higher precision while pruning a certain fraction of least important weights to distribute the additional outlier bits; ensuring high accuracy, aligned memory and hardware efficiency. We design a high-throughput, low overhead accelerator architecture composed of multi-precision INT processing elements and a network-on-chip called ReCoN that efficiently abstracts the complexity of supporting high-precision outliers. Additionally, unlike prior techniques, MicroScopiQ does not assume any locality of outlier weights, enabling applicability to a broad range of FMs. Extensive experiments across diverse quantization settings demonstrate that MicroScopiQ achieves state-of-the-art quantization accuracy, while delivering up to 3x faster inference and 2x lower energy consumption compared to existing alternatives. Code is available at: https://github.com/georgia-tech-synergy-lab/MicroScopiQ-LLM-Quantization

Yuhang Li, Youngeun Kim, Donghyun Lee et al.

In the realm of deep neural network deployment, low-bit quantization presents a promising avenue for enhancing computational efficiency. However, it often hinges on the availability of training data to mitigate quantization errors, a significant challenge when data availability is scarce or restricted due to privacy or copyright concerns. Addressing this, we introduce GenQ, a novel approach employing an advanced Generative AI model to generate photorealistic, high-resolution synthetic data, overcoming the limitations of traditional methods that struggle to accurately mimic complex objects in extensive datasets like ImageNet. Our methodology is underscored by two robust filtering mechanisms designed to ensure the synthetic data closely aligns with the intrinsic characteristics of the actual training data. In case of limited data availability, the actual data is used to guide the synthetic data generation process, enhancing fidelity through the inversion of learnable token embeddings. Through rigorous experimentation, GenQ establishes new benchmarks in data-free and data-scarce quantization, significantly outperforming existing methods in accuracy and efficiency, thereby setting a new standard for quantization in low data regimes. Code is released at \url{https://github.com/Intelligent-Computing-Lab-Yale/GenQ}.

Sreetama Sarkar, Yue Che, Alex Gavin et al.

Despite their remarkable progress in multimodal understanding tasks, large vision language models (LVLMs) often suffer from "hallucinations", generating texts misaligned with the visual context. Existing methods aimed at reducing hallucinations through inference time intervention incur a significant increase in latency. To mitigate this, we present SPIN, a task-agnostic attention-guided head suppression strategy that can be seamlessly integrated during inference, without incurring any significant compute or latency overhead. We investigate whether hallucination in LVLMs can be linked to specific model components. Our analysis suggests that hallucinations can be attributed to a dynamic subset of attention heads in each layer. Leveraging this insight, for each text query token, we selectively suppress attention heads that exhibit low attention to image tokens, keeping the top-K attention heads intact. Extensive evaluations on visual question answering and image description tasks demonstrate the efficacy of SPIN in reducing hallucination scores up to 2.7x while maintaining F1, and improving throughput by 1.8x compared to existing alternatives. Code is available at https://github.com/YUECHE77/SPIN.

Souvik Kundu, Anahita Bhiwandiwalla, Sungduk Yu et al.

Despite recent efforts in understanding the compression impact on large language models (LLMs) in terms of their downstream task performance and trustworthiness on relatively simpler uni-modal benchmarks (for example, question answering, common sense reasoning), their detailed study on multi-modal Large Vision-Language Models (LVLMs) is yet to be unveiled. Towards mitigating this gap, we present LVLM-Compress-Bench, a framework to first thoroughly study the broad impact of compression on the generative performance of LVLMs with multi-modal input driven tasks. In specific, we consider two major classes of compression for autoregressive models, namely KV cache and weight compression, for the dynamically growing intermediate cache and static weights, respectively. We use four LVLM variants of the popular LLaVA framework to present our analysis via integrating various state-of-the-art KV and weight compression methods including uniform, outlier-reduced, and group quantization for the KV cache and weights. With this framework we demonstrate on ten different multi-modal datasets with different capabilities including recognition, knowledge, language generation, spatial awareness, visual reasoning, hallucination and visual illusion identification, toxicity, stereotypes and bias. In specific, our framework demonstrates the compression impact on both general and ethically critical metrics leveraging a combination of real world and synthetic datasets to encompass diverse societal intersectional attributes. Extensive experimental evaluations yield diverse and intriguing observations on the behavior of LVLMs at different quantization budget of KV and weights, in both maintaining and losing performance as compared to the baseline model with FP16 data format. Code will be open-sourced at https://github.com/opengear-project/LVLM-compress-bench.

Sihwan Park, Doohyuk Jang, Sungyub Kim et al.

Speculative decoding has been widely used to accelerate auto-regressive (AR) text generation. However, its effectiveness for visual AR models remains limited due to token selection ambiguity, where multiple tokens share similarly low probabilities and thus reduce acceptance rates. Recently, relaxed speculative decoding with dynamic tree drafting was proposed to mitigate this ambiguity, demonstrating promising results in accelerating visual AR models. However, we observe that token selection ambiguity still negatively affects dynamic tree drafting, resulting in shallow draft trees and limited acceleration. To overcome this issue, we introduce LANTERN++, a refined framework that integrates static tree drafting with a tailored relaxed acceptance condition, allowing drafts to be selected independently of low-confidence predictions. This enables the acceptance of deeper sequences, improving decoding efficiency while preserving image quality. Extensive experiments on state-of-the-art visual AR models demonstrate that LANTERN++ significantly accelerates inference, achieving up to $\mathbf{\times 2.56}$ speedup over standard AR decoding while maintaining high image quality. The code is publicly available at https://github.com/jadohu/LANTERN.

Anand Menon, Samit S Miftah, Shamik Kundu et al.

Hardware verification is crucial in modern SoC design, consuming around 70% of development time. SystemVerilog assertions ensure correct functionality. However, existing industrial practices rely on manual efforts for assertion generation, which becomes increasingly untenable as hardware systems become complex. Recent research shows that Large Language Models (LLMs) can automate this process. However, proprietary SOTA models like GPT-4o often generate inaccurate assertions and require expensive licenses, while smaller open-source LLMs need fine-tuning to manage HDL code complexities. To address these issues, we introduce **VERT**, an open-source dataset designed to enhance SystemVerilog assertion generation using LLMs. VERT enables researchers in academia and industry to fine-tune open-source models, outperforming larger proprietary ones in both accuracy and efficiency while ensuring data privacy through local fine-tuning and eliminating costly licenses. The dataset is curated by systematically augmenting variables from open-source HDL repositories to generate synthetic code snippets paired with corresponding assertions. Experimental results demonstrate that fine-tuned models like Deepseek Coder 6.7B and Llama 3.1 8B outperform GPT-4o, achieving up to 96.88% improvement over base models and 24.14% over GPT-4o on platforms including OpenTitan, CVA6, OpenPiton and Pulpissimo. VERT is available at https://github.com/AnandMenon12/VERT.

Erfan Baghaei Potraghloo, Seyedarmin Azizi, Souvik Kundu et al.

Large language models (LLMs), despite their impressive performance across a wide range of tasks, often struggle to balance two competing objectives in open-ended text generation: fostering diversity and creativity while preserving logical coherence. Existing truncated sampling techniques, including temperature scaling, top-\$p\$ (nucleus) sampling, and min-\$p\$ sampling, aim to manage this trade-off. However, they exhibit limitations, particularly in the effective incorporation of the confidence of the model into the corresponding sampling strategy. For example, min-\$p\$ sampling relies on a single top token as a heuristic for confidence, eventually underutilizing the information of the probability distribution. Toward effective incorporation of the confidence of the model, in this paper, we present **top-H** decoding. We first establish the theoretical foundation of the interplay between creativity and coherence in truncated sampling by formulating an **entropy-constrained minimum divergence** problem. We then prove this minimization problem to be equivalent to an **entropy-constrained mass maximization** (ECMM) problem, which is NP-hard. Finally, we present top-H decoding, a computationally efficient greedy algorithm to solve the ECMM problem. Extensive empirical evaluations demonstrate that top-H outperforms the state-of-the-art (SoTA) alternative of min-\$p\$ sampling by up to **25.63%** on creative writing benchmarks, while maintaining robustness on question-answering datasets such as GPQA, GSM8K, and MT-Bench. Additionally, an *LLM-as-judge* evaluation confirms that top-H indeed produces coherent outputs even at higher temperatures, where creativity is especially critical. In summary, top-H advances SoTA in open-ended text generation and can be *easily integrated* into creative writing applications. The code is available at https://github.com/ErfanBaghaei/Top-H-Decoding.

Yeonju Ro, Zhenyu Zhang, Souvik Kundu et al.

Large language models (LLMs) excel at capturing global token dependencies via self-attention but face prohibitive compute and memory costs on lengthy inputs. While sub-quadratic methods (e.g., linear attention) can reduce these costs, they often degrade accuracy due to overemphasizing recent tokens. In this work, we first propose dual-state linear attention (DSLA), a novel design that maintains two specialized hidden states-one for preserving historical context and one for tracking recency-thereby mitigating the short-range bias typical of linear-attention architectures. To further balance efficiency and accuracy under dynamic workload conditions, we introduce DSLA-Serve, an online adaptive distillation framework that progressively replaces Transformer layers with DSLA layers at inference time, guided by a sensitivity-based layer ordering. DSLA-Serve uses a chained fine-tuning strategy to ensure that each newly converted DSLA layer remains consistent with previously replaced layers, preserving the overall quality. Extensive evaluations on commonsense reasoning, long-context QA, and text summarization demonstrate that DSLA-Serve yields 2.3x faster inference than Llama2-7B and 3.0x faster than the hybrid Zamba-7B, while retaining comparable performance across downstream tasks. Our ablation studies show that DSLA's dual states capture both global and local dependencies, addressing the historical-token underrepresentation seen in prior linear attentions. Codes are available at https://github.com/utnslab/DSLA-Serve.

Akshat Ramachandran, Souvik Kundu, Arnab Raha et al.

Large language model (LLM) pruning with fixed N:M structured sparsity significantly limits the expressivity of the sparse model, yielding sub-optimal performance. In contrast, supporting multiple N:M patterns to provide sparse representational freedom introduces costly overhead in hardware. To address these challenges for LLMs, we first present a flexible layer-wise outlier-density-aware N:M sparsity (FLOW) selection method. FLOW enables the identification of optimal layer-wise N and M values (from a given range) by simultaneously accounting for the presence and distribution of outliers, allowing a higher degree of representational freedom. To deploy sparse models with such N:M flexibility, we then introduce a flexible, low-overhead digital compute-in-memory architecture (FlexCiM). FlexCiM supports diverse sparsity patterns by partitioning a digital CiM (DCiM) macro into smaller sub-macros, which are adaptively aggregated and disaggregated through distribution and merging mechanisms for different N and M values. Extensive experiments on both transformer-based and recurrence-based state space foundation models (SSMs) demonstrate that FLOW outperforms existing alternatives with an accuracy improvement of up to 36%, while FlexCiM achieves up to 1.75x lower inference latency and 1.5x lower energy consumption compared to existing sparse accelerators. Code is available at: https://github.com/FLOW-open-project/FLOW

Jiayi Tian, Yupeng Su, Ryan Solgi et al.

Large reasoning models (LRMs) enhance problem-solving capabilities by generating explicit multi-step chains of thought (CoT) reasoning; however, they incur substantial inference latency and computational overhead. To mitigate this issue, recent works have explored model collaboration paradigms, where small reasoning models (SRMs) generate intermediate reasoning steps to achieve a better accuracy--latency trade-off. Despite recent progress, effectively and efficiently detecting and mitigating SRM failures in collaborative systems remains a key challenge. To address this issue, we analyze SRM inference in both the generated text and hidden-state spaces, and identify three types of failure modes: \textit{overconfidence}, \textit{uncertainty}, and \textit{heavy revalidation}. Building on these insights, we propose \textbf{RankGuide}, a framework that improves the efficiency and effectiveness of SRM--LRM collaboration through tensor-rank-guided routing and steering. Specifically, RankGuide leverages a routing signal that incorporates tensor-rank signals derived from consecutive hidden states to detect when SRMs are likely to fail and selectively invoke LRMs. In addition, we introduce a tensor-rank-filtered steering vector extraction method to modulate the reasoning trajectory of SRMs, thereby improving their generation quality. By improving both routing and steering through tensor-rank signals, RankGuide enables SRM--LRM collaborative systems to achieve more efficient reasoning with fewer steps and improved accuracy. Experiments on multiple reasoning benchmarks demonstrate the efficacy of RankGuide in reducing latency by up to $1.75\times$ compared to LRM, while maintaining competitive accuracy relative to prior methods.

Jucheng Shen, Gaurav Sarkar, Yeonju Ro et al.

We present CadLLM, a training-free method to accelerate the inference throughput of diffusion-based LLMs (dLLMs). We first investigate the dynamic nature of token unmasking confidence across blocks and steps. Based on this observation, we present a lightweight adaptive approach that controls the generation block size, step size, and threshold based on the average confidence of unmasked tokens. We further reduce softmax overhead by dynamically leveraging a subset of the vocabulary to regulate sampling breadth. CadLLM is a plug-and-play, model-agnostic method compatible with KV-cache-based dLLMs. Extensive experiments on four popular tasks demonstrate that CadLLM yields up to 2.28x throughput improvement over the state-of-the-art baseline with competitive accuracy.

Zeyu Liu, Souvik Kundu, Lianghao Jiang et al.

Although transformer architectures have achieved state-of-the-art performance across diverse domains, their quadratic computational complexity with respect to sequence length remains a significant bottleneck, particularly for latency-sensitive long-context applications. While recent linear-complexity alternatives are increasingly powerful, effectively training them from scratch is still resource-intensive. To overcome these limitations, we propose LAWCAT (Linear Attention with Convolution Across Time), a novel linearization framework designed to efficiently transfer the capabilities of pre-trained transformers into a performant linear attention architecture. LAWCAT integrates causal Conv1D layers to enhance local dependency modeling and employs normalized gated linear attention to improve generalization across varying context lengths. Our comprehensive evaluations demonstrate that, distilling Mistral-7B with only 1K-length sequences yields over 90\% passkey retrieval accuracy up to 22K tokens, significantly extending its effective context window. Similarly, Llama3.2-1B LAWCAT variant achieves competitive performance on S-NIAH 1\&2\&3 tasks (1K-8K context length) and BABILong benchmark (QA2\&QA3, 0K-16K context length), requiring less than 0.1\% pre-training tokens compared with pre-training models. Furthermore, LAWCAT exhibits faster prefill speeds than FlashAttention-2 for sequences exceeding 8K tokens. LAWCAT thus provides an efficient pathway to high-performance, long-context linear models suitable for edge deployment, reducing reliance on extensive long-sequence training data and computational resources. Code is released at: https://github.com/zeyuliu1037/LAWCAT

Haoran You, Yipin Guo, Yichao Fu et al.

Large language models (LLMs) have shown impressive performance on language tasks but face challenges when deployed on resource-constrained devices due to their extensive parameters and reliance on dense multiplications, resulting in high memory demands and latency bottlenecks. Shift-and-add reparameterization offers a promising solution by replacing costly multiplications with hardware-friendly primitives in both the attention and multi-layer perceptron (MLP) layers of an LLM. However, current reparameterization techniques require training from scratch or full parameter fine-tuning to restore accuracy, which is resource-intensive for LLMs. To address this, we propose accelerating pretrained LLMs through post-training shift-and-add reparameterization, creating efficient multiplication-free models, dubbed ShiftAddLLM. Specifically, we quantize each weight matrix into binary matrices paired with group-wise scaling factors. The associated multiplications are reparameterized into (1) shifts between activations and scaling factors and (2) queries and adds according to the binary matrices. To reduce accuracy loss, we present a multi-objective optimization method to minimize both weight and output activation reparameterization errors. Additionally, based on varying sensitivity across layers to reparameterization, we develop an automated bit allocation strategy to further reduce memory usage and latency. Experiments on five LLM families and eight tasks consistently validate the effectiveness of ShiftAddLLM, achieving average perplexity improvements of 5.6 and 22.7 points at comparable or lower latency compared to the most competitive quantized LLMs at 3 and 2 bits, respectively, and more than 80% memory and energy reductions over the original LLMs. Codes and models are available at https://github.com/GATECH-EIC/ShiftAddLLM.

Abhimanyu Bambhaniya, Ritik Raj, Geonhwa Jeong et al.

Large language models (LLMs) have shown remarkable performance across a wide range of applications, often outperforming human experts. However, deploying these gigantic models efficiently for diverse inference use cases requires carefully designed hardware platforms with ample computing, memory, and network resources. With constant innovation in LLM serving optimizations and model architecture evolving at breakneck speed, the hardware requirements to meet Service Level Objectives (SLOs) remain an open research question. To answer the question, we present an analytical tool, GenZ, to efficiently navigate the relationship between diverse LLM model architectures(Dense, GQA, MoE, Mamba), LLM serving optimizations(Chunking, Speculative decoding, quanitization), and AI platform design parameters. Our tool estimates LLM inference performance metrics for the given scenario. We have validated against real hardware platforms running various different LLM models, achieving a max geomean error of 5.82.We use GenZ to identify compute, memory capacity, memory bandwidth, network latency, and network bandwidth requirements across diverse LLM inference use cases. We also study diverse architectural choices in use today (inspired by LLM serving platforms from several vendors) to help inform computer architects designing next-generation AI hardware accelerators and platforms. The trends and insights derived from GenZ can guide AI engineers deploying LLMs as well as computer architects designing next-generation hardware accelerators and platforms. Ultimately, this work sheds light on the platform design considerations for unlocking the full potential of large language models across a spectrum of applications. The source code is available at https://github.com/abhibambhaniya/GenZ-LLM-Analyzer . Users can also be tried it on at https://genz-llm-analyzer.streamlit.app/ without any setup on your web browser.

Zeyu Liu, Souvik Kundu, Anni Li et al.

We present a novel Parameter-Efficient Fine-Tuning (PEFT) method, dubbed as Adaptive Freezing of Low Rank Adaptation (AFLoRA). Specifically, for each pre-trained frozen weight tensor, we add a parallel path of trainable low-rank matrices, namely a down-projection and an up-projection matrix, each of which is followed by a feature transformation vector. Based on a novel freezing score, we the incrementally freeze these projection matrices during fine-tuning to reduce the computation and alleviate over-fitting. Our experimental results demonstrate that we can achieve state-of-the-art performance with an average improvement of up to $0.85\%$ as evaluated on GLUE benchmark while yeilding up to $9.5\times$ fewer average trainable parameters. While compared in terms of runtime, AFLoRA can yield up to $1.86\times$ improvement as opposed to similar PEFT alternatives. Besides the practical utility of our approach, we provide insights on the trainability requirements of LoRA paths at different modules and the freezing schedule for the different projection matrices. Code will be released.

Yeshwanth Venkatesha, Souvik Kundu, Priyadarshini Panda

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.

Sreetama Sarkar, Souvik Kundu, Kai Zheng et al.

The ubiquity of vision transformers (ViTs) for various edge applications, including personalized learning, has created the demand for on-device fine-tuning. However, training with the limited memory and computation power of edge devices remains a significant challenge. In particular, the memory required for training is much higher than that needed for inference, primarily due to the need to store activations across all layers in order to compute the gradients needed for weight updates. Previous works have explored reducing this memory requirement via frozen-weight training as well storing the activations in a compressed format. However, these methods are deemed inefficient due to their inability to provide training or inference speedup. In this paper, we first investigate the limitations of existing on-device training methods aimed at reducing memory and compute requirements. We then present block selective reprogramming (BSR) in which we fine-tune only a fraction of total blocks of a pre-trained model and selectively drop tokens based on self-attention scores of the frozen layers. To show the efficacy of BSR, we present extensive evaluations on ViT-B and DeiT-S with five different datasets. Compared to the existing alternatives, our approach simultaneously reduces training memory by up to 1.4x and compute cost by up to 2x while maintaining similar accuracy. We also showcase results for Mixture-of-Expert (MoE) models, demonstrating the effectiveness of our approach in multitask learning scenarios.

Amey Agrawal, Nitin Kedia, Anmol Agarwal et al. · gatech

The rapid evolution of Large Language Model (LLM) inference systems has yielded significant efficiency improvements. However, our systematic analysis reveals that current evaluation methodologies frequently exhibit fundamental flaws, often manifesting as common evaluation anti-patterns that obscure true performance characteristics and impede scientific progress. Through a comprehensive examination of recent systems, we identify recurring anti-patterns across three key dimensions: Baseline Fairness, Evaluation Setup, and Metric Design. These anti-patterns are uniquely problematic for LLM inference due to its dual-phase nature combining distinct prefill and decode operations, its handling of highly heterogeneous workloads, and its strict temporal requirements for interactive use. We demonstrate how common anti-patterns -- such as inadequate baseline comparisons that conflate engineering effort with algorithmic novelty, workload selections that fail to represent production scenarios, and metric normalizations that hide substantial performance variability like generation stalls-lead to misleading conclusions. To address these challenges, we provide a comprehensive checklist derived from our analysis, establishing a framework for recognizing and avoiding these anti-patterns in favor of robust LLM inference evaluation. To demonstrate the practical application of our framework, we present a case study analyzing speculative decoding, a technique whose bursty, non-uniform token generation is easily misinterpreted when evaluated using approaches characteristic of these anti-patterns. Our work establishes a rigorous foundation for evaluation methodology, enabling meaningful comparisons, ensuring reproducible results, and ultimately accelerating genuine progress in LLM inference systems by moving beyond common anti-patterns to align evaluation with real-world requirements.

Sanjay Das, Swastik Bhattacharya, Souvik Kundu et al.

Large Language Models (LLMs) have revolutionized natural language processing (NLP), excelling in tasks like text generation and summarization. However, their increasing adoption in mission-critical applications raises concerns about hardware-based threats, particularly bit-flip attacks (BFAs). BFAs, enabled by fault injection methods such as Rowhammer, target model parameters in memory, compromising both integrity and performance. Identifying critical parameters for BFAs in the vast parameter space of LLMs poses significant challenges. While prior research suggests transformer-based architectures are inherently more robust to BFAs compared to traditional deep neural networks, we challenge this assumption. For the first time, we demonstrate that as few as three bit-flips can cause catastrophic performance degradation in an LLM with billions of parameters. Current BFA techniques are inadequate for exploiting this vulnerability due to the difficulty of efficiently identifying critical parameters within the immense parameter space. To address this, we propose AttentionBreaker, a novel framework tailored for LLMs that enables efficient traversal of the parameter space to identify critical parameters. Additionally, we introduce GenBFA, an evolutionary optimization strategy designed to refine the search further, isolating the most critical bits for an efficient and effective attack. Empirical results reveal the profound vulnerability of LLMs to AttentionBreaker. For example, merely three bit-flips (4.129 x 10^-9% of total parameters) in the LLaMA3-8B-Instruct 8-bit quantized (W8) model result in a complete performance collapse: accuracy on MMLU tasks drops from 67.3% to 0%, and Wikitext perplexity skyrockets from 12.6 to 4.72 x 10^5. These findings underscore the effectiveness of AttentionBreaker in uncovering and exploiting critical vulnerabilities within LLM architectures.