arXiv subjects

Grace Li Zhang

Publications and source records attributed to Grace Li Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2025

CompressKV: Semantic Retrieval Heads Know What Tokens are Not Important Before Generation

Recent advances in large language models (LLMs) have significantly boosted long-context processing. However, the increasing key-value (KV) cache size poses critical challenges to memory and execution efficiency. Most KV cache compression methods rely on heuristic token eviction using all attention heads in Grouped Query Attention (GQA)-based LLMs. This method ignores the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrades the performance of LLMs. To address the issue above, instead of using all the attention heads in GQA-based LLMs to determine important tokens as in the previous work, we first identify the attention heads in each layer that are not only capable of retrieving the initial and final tokens of a prompt, but also capable of retrieving important tokens within the text and attending to their surrounding semantic context. Afterwards, we exploit such heads to determine the important tokens and retain their corresponding KV cache pairs. Furthermore, we analyze the cache eviction error of each layer individually and introduce a layer-adaptive KV cache allocation strategy. Experimental results demonstrate the proposed CompressKV consistently outperforms state-of-the-art approaches under various memory budgets on LongBench and Needle-in-a-Haystack benchmarks. Our code is publicly available at: https://github.com/TUDa-HWAI/CompressKV.git.

cs.CL↗

arXiv2025

Deep Joint Source-Channel Coding for Small Satellite Applications

Small satellites used for Earth observation generate vast amounts of high-dimensional data, but their operation in low Earth orbit creates a significant communication bottleneck due to limited contact times and harsh, varying channel conditions. While deep joint source-channel coding (DJSCC) has emerged as a promising technique, its practical application to the complex satellite environment remains an open question. This paper presents a comprehensive DJSCC framework tailored for satellite communications. We first establish a basic system, DJSCC-SAT, and integrate a realistic, multi-state statistical channel model to guide its training and evaluation. To overcome the impracticality of using separate models for every channel condition, we then introduce an adaptable architecture, ADJSCC-SAT, which leverages attention modules to allow a single neural network to adjust to a wide range of channel states with minimal overhead. Through extensive evaluation on Sentinel-2 multi-spectral data, we demonstrate that our adaptable approach achieves performance comparable to using multiple specialized networks while significantly reducing model storage requirements. Furthermore, the adaptable model shows enhanced robustness to channel estimation errors, outperforming the non-adaptable baseline. The proposed framework is a practical and efficient step toward deploying robust, adaptive DJSCC systems for real-world satellite missions.

cs.NI↗

arXiv2025

HLSTester: Efficient Testing of Behavioral Discrepancies with LLMs for High-Level Synthesis

In high-level synthesis (HLS), C/C++ programs with synthesis directives are used to generate circuits for FPGA implementations. However, hardware-specific and platform-dependent characteristics in these implementations can introduce behavioral discrepancies between the original C/C++ programs and the circuits after high-level synthesis. Existing methods for testing behavioral discrepancies in HLS are still immature, and the testing workflow requires significant human efforts. To address this challenge, we propose HLSTester, a large language model (LLM) aided testing framework that efficiently detects behavioral discrepancies in HLS. To mitigate hallucinations in LLMs and enhance prompt quality, the testbenches for original C/C++ programs are leveraged to guide LLMs in generating HLS-compatible testbenches, effectively eliminating certain traditional C/C++ constructs that are incompatible with HLS tools. Key variables are pinpointed through a backward slicing technique in both C/C++ and HLS programs to monitor their runtime spectra, enabling an in-depth analysis of the discrepancy symptoms. To reduce test time, a testing input generation mechanism is introduced to integrate dynamic mutation with insights from an LLM-based progressive reasoning chain. In addition, repetitive hardware testing is skipped by a redundancy-aware filtering technique for the generated test inputs. Experimental results demonstrate that the proposed LLM-aided testing framework significantly accelerates the testing workflow while achieving higher testbench simulation pass rates compared with the traditional method and the direct use of LLMs on the same HLS programs.

cs.SE↗

arXiv2025

Paradigm-Based Automatic HDL Code Generation Using LLMs

While large language models (LLMs) have demonstrated the ability to generate hardware description language (HDL) code for digital circuits, they still face the hallucination problem, which can result in the generation of incorrect HDL code or misinterpretation of specifications. In this work, we introduce a human-expert-inspired method to mitigate the hallucination of LLMs and enhance their performance in HDL code generation. We begin by constructing specialized paradigm blocks that consist of several steps designed to divide and conquer generation tasks, mirroring the design methodology of human experts. These steps include information extraction, human-like design flows, and the integration of external tools. LLMs are then instructed to classify the type of circuit in order to match it with the appropriate paradigm block and execute the block to generate the HDL codes. Additionally, we propose a two-phase workflow for multi-round generation, aimed at effectively improving the testbench pass rate of the generated HDL codes within a limited number of generation and verification rounds. Experimental results demonstrate that our method significantly enhances the functional correctness of the generated Verilog code

cs.PL↗

arXiv2025

VRank: Enhancing Verilog Code Generation from Large Language Models via Self-Consistency

Large Language Models (LLMs) have demonstrated promising capabilities in generating Verilog code from module specifications. To improve the quality of such generated Verilog codes, previous methods require either time-consuming manual inspection or generation of multiple Verilog codes, from which the one with the highest quality is selected with manually designed testbenches. To enhance the generation efficiency while maintaining the quality of the generated codes, we propose VRank, an automatic framework that generates Verilog codes with LLMs. In our framework, multiple code candidates are generated with LLMs by leveraging their probabilistic nature. Afterwards, we group Verilog code candidates into clusters based on identical outputs when tested against the same testbench, which is also generated by LLMs. Clusters are ranked based on the consistency they show on testbench. To determine the best candidate, Chain-of-Thought is further applied to select the best candidate from the top-ranked clusters. By systematically analyzing diverse outputs of generated codes, VRank reduces errors and enhances the overall quality of the generated Verilog code. Experimental results on the VerilogEval-Human benchmark demonstrate a significant 10.5% average increase in functional correctness (passl1) across multiple LLMs, demonstrating VRank's effectiveness in improving the accuracy of automated hardware description language generation for complex design tasks.

cs.AR↗

arXiv2024

CorrectBench: Automatic Testbench Generation with Functional Self-Correction using LLMs for HDL Design

Functional simulation is an essential step in digital hardware design. Recently, there has been a growing interest in leveraging Large Language Models (LLMs) for hardware testbench generation tasks. However, the inherent instability associated with LLMs often leads to functional errors in the generated testbenches. Previous methods do not incorporate automatic functional correction mechanisms without human intervention and still suffer from low success rates, especially for sequential tasks. To address this issue, we propose CorrectBench, an automatic testbench generation framework with functional self-validation and self-correction. Utilizing only the RTL specification in natural language, the proposed approach can validate the correctness of the generated testbenches with a success rate of 88.85%. Furthermore, the proposed LLM-based corrector employs bug information obtained during the self-validation process to perform functional self-correction on the generated testbenches. The comparative analysis demonstrates that our method achieves a pass ratio of 70.13% across all evaluated tasks, compared with the previous LLM-based testbench generation framework's 52.18% and a direct LLM-based generation method's 33.33%. Specifically in sequential circuits, our work's performance is 62.18% higher than previous work in sequential tasks and almost 5 times the pass ratio of the direct method. The codes and experimental results are open-sourced at the link: https://github.com/AutoBench/CorrectBench

cs.SE↗

arXiv2024

EncodingNet: A Novel Encoding-based MAC Design for Efficient Neural Network Acceleration

Deep neural networks (DNNs) have achieved great breakthroughs in many fields such as image classification and natural language processing. However, the execution of DNNs needs to conduct massive numbers of multiply-accumulate (MAC) operations on hardware and thus incurs a large power consumption. To address this challenge, we propose a novel digital MAC design based on encoding. In this new design, the multipliers are replaced by simple logic gates to represent the results with a wide bit representation. The outputs of the new multipliers are added by bit-wise weighted accumulation and the accumulation results are compatible with existing computing platforms accelerating neural networks. Since the multiplication function is replaced by a simple logic representation, the critical paths in the resulting circuits become much shorter. Correspondingly, pipelining stages and intermediate registers used to store partial sums in the MAC array can be reduced, leading to a significantly smaller area as well as better power efficiency. The proposed design has been synthesized and verified by ResNet18- Cifar10, ResNet20-Cifar100, ResNet50-ImageNet, MobileNetV2-Cifar10, MobileNetV2-Cifar100, and EfficientNetB0-ImageNet. The experimental results confirmed the reduction of circuit area by up to 48.79% and the reduction of power consumption of executing DNNs by up to 64.41%, while the accuracy of the neural networks can still be well maintained. The open source code of this work can be found on GitHub with link https://github.com/Bo-Liu-TUM/EncodingNet/.

cs.AR↗

arXiv2024

LiveMind: Low-latency Large Language Models with Simultaneous Inference

In this paper, we introduce LiveMind, a novel low-latency inference framework for large language model (LLM) inference which enables LLMs to perform inferences with incomplete user input. By reallocating computational processes to the input phase, a substantial reduction in latency is achieved, thereby significantly enhancing the interactive experience for users of LLMs. The framework adeptly manages the visibility of the streaming input to the model, allowing it to infer from incomplete user input or await additional content. Compared with traditional inference methods on complete user input, our approach demonstrates an average reduction in response latency of 84.0% on the MMLU dataset and 71.6% on the MMLU-Pro dataset, while maintaining comparable accuracy. Additionally, our framework facilitates collaborative inference and output across different models. By employing an large LLM for inference and a small LLM for output, we achieve an average 37% reduction in response latency, alongside a 4.30% improvement in accuracy on the MMLU-Pro dataset compared with the baseline. The proposed LiveMind framework advances the field of human-AI interaction by enabling more responsive and efficient communication between users and AI systems.

cs.AI↗

arXiv2024

LLM-Aided Efficient Hardware Design Automation

With the rapidly increasing complexity of modern chips, hardware engineers are required to invest more effort in tasks such as circuit design, verification, and physical implementation. These workflows often involve continuous modifications, which are labor-intensive and prone to errors. Therefore, there is an increasing need for more efficient and cost-effective Electronic Design Automation (EDA) solutions to accelerate new hardware development. Recently, large language models (LLMs) have made significant advancements in contextual understanding, logical reasoning, and response generation. Since hardware designs and intermediate scripts can be expressed in text format, it is reasonable to explore whether integrating LLMs into EDA could simplify and fully automate the entire workflow. Accordingly, this paper discusses such possibilities in several aspects, covering hardware description language (HDL) generation, code debugging, design verification, and physical implementation. Two case studies, along with their future outlook, are introduced to highlight the capabilities of LLMs in code repair and testbench generation. Finally, future directions and challenges are highlighted to further explore the potential of LLMs in shaping the next-generation EDA

eess.SY↗

arXiv2024

Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression

Large Language Models (LLMs) have achieved remarkable breakthroughs. However, the huge number of parameters in LLMs require significant amount of memory storage in inference, which prevents their practical deployment in many applications. To reduce memory storage of LLMs, singular value decomposition (SVD) provides a promising solution to approximate weight matrices for compressing LLMs. In this paper, we take a step further to explore parameter sharing across different layers with SVD to achieve more effective compression for LLMs. Specifically, weight matrices in different layers are decomposed and represented as a linear combination of a set of shared basis vectors and unique coefficients. The types of weight matrices and the layer selection for basis sharing are examined when compressing LLMs to maintain the performance. Comprehensive experiments demonstrate that Basis Sharing outperforms state-of-the-art SVD-based compression approaches and parameter sharing techniques, especially under large compression ratios. Code is available at: https://github.com/TUDa-HWAI/Basis_Sharing

cs.CL↗

arXiv2024

An Efficient General-Purpose Optical Accelerator for Neural Networks

General-purpose optical accelerators (GOAs) have emerged as a promising platform to accelerate deep neural networks (DNNs) due to their low latency and energy consumption. Such an accelerator is usually composed of a given number of interleaving Mach-Zehnder- Interferometers (MZIs). This interleaving architecture, however, has a low efficiency when accelerating neural networks of various sizes due to the mismatch between weight matrices and the GOA architecture. In this work, a hybrid GOA architecture is proposed to enhance the mapping efficiency of neural networks onto the GOA. In this architecture, independent MZI modules are connected with microring resonators (MRRs), so that they can be combined to process large neural networks efficiently. Each of these modules implements a unitary matrix with inputs adjusted by tunable coefficients. The parameters of the proposed architecture are searched using genetic algorithm. To enhance the accuracy of neural networks, selected weight matrices are expanded to multiple unitary matrices applying singular value decomposition (SVD). The kernels in neural networks are also adjusted to use up the on-chip computational resources. Experimental results show that with a given number of MZIs, the mapping efficiency of neural networks on the proposed architecture can be enhanced by 21.87%, 21.20%, 24.69%, and 25.52% for VGG16 and Resnet18 on datasets Cifar10 and Cifar100, respectively. The energy consumption and computation latency can also be reduced by over 67% and 21%, respectively.

cs.NE↗

arXiv2024

AutoBench: Automatic Testbench Generation and Evaluation Using LLMs for HDL Design

In digital circuit design, testbenches constitute the cornerstone of simulation-based hardware verification. Traditional methodologies for testbench generation during simulation-based hardware verification still remain partially manual, resulting in inefficiencies in testing various scenarios and requiring expensive time from designers. Large Language Models (LLMs) have demonstrated their potential in automating the circuit design flow. However, directly applying LLMs to generate testbenches suffers from a low pass rate. To address this challenge, we introduce AutoBench, the first LLM-based testbench generator for digital circuit design, which requires only the description of the design under test (DUT) to automatically generate comprehensive testbenches. In AutoBench, a hybrid testbench structure and a self-checking system are realized using LLMs. To validate the generated testbenches, we also introduce an automated testbench evaluation framework to evaluate the quality of generated testbenches from multiple perspectives. Experimental results demonstrate that AutoBench achieves a 57% improvement in the testbench pass@1 ratio compared with the baseline that directly generates testbenches using LLMs. For 75 sequential circuits, AutoBench successfully has a 3.36 times testbench pass@1 ratio compared with the baseline. The source codes and experimental results are open-sourced at this link: https://github.com/AutoBench/AutoBench

cs.SE↗

arXiv2024

BasisN: Reprogramming-Free RRAM-Based In-Memory-Computing by Basis Combination for Deep Neural Networks

Deep neural networks (DNNs) have made breakthroughs in various fields including image recognition and language processing. DNNs execute hundreds of millions of multiply-and-accumulate (MAC) operations. To efficiently accelerate such computations, analog in-memory-computing platforms have emerged leveraging emerging devices such as resistive RAM (RRAM). However, such accelerators face the hurdle of being required to have sufficient on-chip crossbars to hold all the weights of a DNN. Otherwise, RRAM cells in the crossbars need to be reprogramed to process further layers, which causes huge time/energy overhead due to the extremely slow writing and verification of the RRAM cells. As a result, it is still not possible to deploy such accelerators to process large-scale DNNs in industry. To address this problem, we propose the BasisN framework to accelerate DNNs on any number of available crossbars without reprogramming. BasisN introduces a novel representation of the kernels in DNN layers as combinations of global basis vectors shared between all layers with quantized coefficients. These basis vectors are written to crossbars only once and used for the computations of all layers with marginal hardware modification. BasisN also provides a novel training approach to enhance computation parallelization with the global basis vectors and optimize the coefficients to construct the kernels. Experimental results demonstrate that cycles per inference and energy-delay product were reduced to below 1% compared with applying reprogramming on crossbars in processing large-scale DNNs such as DenseNet and ResNet on ImageNet and CIFAR100 datasets, while the training and hardware costs are negligible.

eess.SY↗

arXiv2024

Automated C/C++ Program Repair for High-Level Synthesis via Large Language Models

In High-Level Synthesis (HLS), converting a regular C/C++ program into its HLS-compatible counterpart (HLS-C) still requires tremendous manual effort. Various program scripts have been introduced to automate this process. But the resulting codes usually contain many issues that should be manually repaired by developers. Since Large Language Models (LLMs) have the ability to automate code generation, they can also be used for automated program repair in HLS. However, due to the limited training of LLMs considering hardware and software simultaneously, hallucinations may occur during program repair using LLMs, leading to compilation failures. Besides, using LLMs for iterative repair also incurs a high cost. To address these challenges, we propose an LLM-driven program repair framework that takes regular C/C++ code as input and automatically generates its corresponding HLS-C code for synthesis while minimizing human repair effort. To mitigate the hallucinations in LLMs and enhance the prompt quality, a Retrieval-Augmented Generation (RAG) paradigm is introduced to guide the LLMs toward correct repair. In addition, we use LLMs to create a static bit width optimization program to identify the optimized bit widths for variables. Moreover, LLM-driven HLS optimization strategies are introduced to add/tune pragmas in HLS-C programs for circuit optimization. Experimental results demonstrate that the proposed LLM-driven automated framework can achieve much higher repair pass rates in 24 real-world applications compared with the traditional scripts and the direct application of LLMs for program repair.

eess.SY↗

arXiv2024

Classification-Based Automatic HDL Code Generation Using LLMs

While large language models (LLMs) have demonstrated the ability to generate hardware description language (HDL) code for digital circuits, they still suffer from the hallucination problem, which leads to the generation of incorrect HDL code or misunderstanding of specifications. In this work, we introduce a human-expert-inspired method to mitigate the hallucination of LLMs and improve the performance in HDL code generation. We first let LLMs classify the type of the circuit based on the specifications. Then, according to the type of the circuit, we split the tasks into several sub-procedures, including information extraction and human-like design flow using Electronic Design Automation (EDA) tools. Besides, we also use a search method to mitigate the variation in code generation. Experimental results show that our method can significantly improve the functional correctness of the generated Verilog and reduce the hallucination of LLMs.

cs.AR↗

arXiv2024

Class-Aware Pruning for Efficient Neural Networks

Deep neural networks (DNNs) have demonstrated remarkable success in various fields. However, the large number of floating-point operations (FLOPs) in DNNs poses challenges for their deployment in resource-constrained applications, e.g., edge devices. To address the problem, pruning has been introduced to reduce the computational cost in executing DNNs. Previous pruning strategies are based on weight values, gradient values and activation outputs. Different from previous pruning solutions, in this paper, we propose a class-aware pruning technique to compress DNNs, which provides a novel perspective to reduce the computational cost of DNNs. In each iteration, the neural network training is modified to facilitate the class-aware pruning. Afterwards, the importance of filters with respect to the number of classes is evaluated. The filters that are only important for a few number of classes are removed. The neural network is then retrained to compensate for the incurred accuracy loss. The pruning iterations end until no filter can be removed anymore, indicating that the remaining filters are very important for many classes. This pruning technique outperforms previous pruning solutions in terms of accuracy, pruning ratio and the reduction of FLOPs. Experimental results confirm that this class-aware pruning technique can significantly reduce the number of weights and FLOPs, while maintaining a high inference accuracy.

cs.AI↗

arXiv2024

Early-Exit with Class Exclusion for Efficient Inference of Neural Networks

Deep neural networks (DNNs) have been successfully applied in various fields. In DNNs, a large number of multiply-accumulate (MAC) operations are required to be performed, posing critical challenges in applying them in resource-constrained platforms, e.g., edge devices. To address this challenge, in this paper, we propose a class-based early-exit for dynamic inference. Instead of pushing DNNs to make a dynamic decision at intermediate layers, we take advantage of the learned features in these layers to exclude as many irrelevant classes as possible, so that later layers only have to determine the target class among the remaining classes. When only one class remains at a layer, this class is the corresponding classification result. Experimental results demonstrate the computational cost of DNNs in inference can be reduced significantly with the proposed early-exit technique. The codes can be found at https://github.com/HWAI-TUDa/EarlyClassExclusion.

cs.LG↗

arXiv2023

OplixNet: Towards Area-Efficient Optical Split-Complex Networks with Real-to-Complex Data Assignment and Knowledge Distillation

Having the potential for high speed, high throughput, and low energy cost, optical neural networks (ONNs) have emerged as a promising candidate for accelerating deep learning tasks. In conventional ONNs, light amplitudes are modulated at the input and detected at the output. However, the light phases are still ignored in conventional structures, although they can also carry information for computing. To address this issue, in this paper, we propose a framework called OplixNet to compress the areas of ONNs by modulating input image data into the amplitudes and phase parts of light signals. The input and output parts of the ONNs are redesigned to make full use of both amplitude and phase information. Moreover, mutual learning across different ONN structures is introduced to maintain the accuracy. Experimental results demonstrate that the proposed framework significantly reduces the areas of ONNs with the accuracy within an acceptable range. For instance, 75.03% area is reduced with a 0.33% accuracy decrease on fully connected neural network (FCNN) and 74.88% area is reduced with a 2.38% accuracy decrease on ResNet-32.

eess.SY↗