arXiv subjects

H. T. Kung

Publications and source records attributed to H. T. Kung.

At least 19 recordsLinked to original sources

arXiv2026

Towards Training Private LLMs: Exploring Fine-Tuning Language Models on Apple Silicon with RDMA over Thunderbolt

Private large language model (LLM) fine-tuning is increasingly important for organizations that need to adapt models using sensitive data, but it often exceeds the memory capacity of commodity datacenter accelerators. Apple Silicon offers a different design point through large unified memory and lower complete-system cost, while recent Apple software support enables distributed execution over RDMA-over-Thunderbolt (TB). This paper studies whether Apple Silicon can serve as a practical platform for private LLM fine-tuning. We characterize RDMA-over-TB communication on Mac Studio nodes, showing that the measured bandwidth is far below nominal TB specifications. Next, we extend Apple's implementation with multi-trunk communication, persistent worker threads, and CPU-side gradient overlap to better exploit multiple direct TB links for LLM fine-tuning workloads. Finally, on a four-node Mac Studio cluster that fine-tunes a Qwen3-9B, our optimizations improve weak-scaling throughput by up to 1.6X over the single-trunk, non-overlapped baseline and reach 936 tokens/s for sequence length 17408. We further compare Apple Silicon with an NVIDIA H100 platform to quantify the trade-off between memory capacity, throughput, and acquisition cost, showing that Apple Silicon can provide a cost-effective solution for private LLM fine-tuning.

cs.DC↗

arXiv2026

Decoding the Skew: Distribution-Aware MoE Inference with Adaptive Kernel Dispatch

Mixture-of-Experts (MoE) inference consists of sparse expert GEMMs whose shapes vary with the runtime routing distribution. Existing serving systems typically select fused-MoE kernels using static token-count buckets, ignoring the per-expert routing distribution that determines tile padding, memory reuse, and kernel efficiency. We introduce a distribution-aware framework for modeling and benchmarking MoE inference. The framework combines the compact Effective Experts metric with a Dirichlet-based reverse-modeling procedure that generates controllable routing distributions for systematic hardware studies. Using it, we show that the best fused-MoE kernel changes with routing skew and token count. We further present DA-MoE, a GPU-resident kernel-dispatch runtime for NVIDIA GPUs that matches the live routing histogram to offline-tuned distributions and selects a near-optimal fused-MoE kernel without CPU--GPU synchronization. On HumanEval-X serving traces, DA-MoE improves geomean fused-MoE latency by 1.16X on DeepSeek-V3 and 1.29X on Kimi K2, with peak speedups of 1.40X and 1.56X.

cs.AR↗

arXiv2026

ReRound: Reconstructive Rounding to Resolve Midpoint Ambiguity in Calibration-Free LLM Quantization

ReRound (Reconstructive Rounding) is a post-training quantization method that addresses the midpoint ambiguity inherent in standard round-to-nearest (RTN) schemes when quantizing weights near the centers of quantization intervals. Starting from a pretrained LLM, ReRound trains a conditional diffusion model to produce continuous reconstructions of low-bit weights for the LLM. These reconstructed weights act as a guidance signal to disambiguate the rounding direction of weights located close to interval midpoints. To integrate this reconstruction-guided rounding with conventional RTN, ReRound introduces a tolerance metric measuring how far the quantized weight (not the final quantized integer) is away from the midpoint: quantized weights within a tolerance region around midpoints are quantized using diffusion-based reconstructions, whereas weights closer to quantization boundaries are quantized with RTN. By sweeping the tolerance parameter, ReRound generates multiple candidate quantized integer weight matrices and selects the de-quantized weight matrix candidate whose leading singular values most closely match those of the original full-precision weights. This selected candidate determines the tolerance parameter ReRound uses. ReRound is particularly effective for smaller LLMs. Across a range of such models, it consistently outperforms standard RTN for 3-bit and 4-bit weight quantization. ReRound achieves superior accuracy compared to an extensive set of calibration-free methods, remains competitive with calibration-dependent approaches, and operates entirely offline, introducing no additional overhead during low-bit inference. The ReRound strategy represents a new approach for low-bit quantization. The method applies to AI models beyond LLMs. This paper focuses on its applications to small LLMs.

cs.LG↗

arXiv2026

Structured Testbench Generation for LLM-Driven HDL Design and Verification-Oriented Data Curation

Automated testbench generation has become a critical bottleneck in large language model (LLM)-driven Register Transfer Level (RTL) workflows, where large numbers of candidate designs must be verified rapidly and reliably. Existing prompt-based approaches treat testbench generation as unconstrained code synthesis, yielding stochastic outputs with high token cost, low reproducibility, and insufficient coverage. To address this gap, we present STG, a Structured Testbench Generation framework that exploits the inherent structure of hardware designs to generate deterministic testbenches. As a direct verification tool, STG runs 720x faster than an iterative LLM-based testbench generation flow and higher rate of successful compilation, achieves higher coverage, and reduces false-pass verdicts on incorrect DUTs. STG also helps identify errors in RTL generation benchmarks by exposing faulty benchmark testbenches. As a data curation engine, it is 11x faster than LLM-based filtering on a single CPU core with 127x less energy, and the resulting distilled models provide state-of-the-art performance in our multi-benchmark evaluation. As a test-time scaling oracle, it reduces node count by 14-47\%. Our models are available at https://huggingface.co/collections/AS-SiliconMind/siliconmind-v12.

cs.AI↗

arXiv2026

SiliconMind-V1: Multi-Agent Distillation and Debug-Reasoning Workflows for Verilog Code Generation

Large language models (LLMs) have recently emerged as a promising approach for automating Verilog code generation; however, existing methods primarily emphasize syntactic correctness and often rely on commercial models or external verification tools, which introduces concerns regarding cost, data privacy, and limited guarantees of functional correctness. This work proposes a unified multi-agent framework for reasoning-oriented training data generation with integrated testbench-driven verification, enabling locally fine-tuned LLMs, SiliconMind-V1, to iteratively generate, test, and debug Register-Transfer Level (RTL) designs through test-time scaling. Experimental results on representative benchmarks (VerilogEval-v2, RTLLM-v2, and CVDP) demonstrate that the proposed approach outperforms the state-of-the-art QiMeng-CodeV-R1 in functional correctness while using fewer training resources.

cs.AR↗

arXiv2026

Alfa: Attentive Low-Rank Filter Adaptation for Structure-Aware Cross-Domain Personalized Gaze Estimation

Pre-trained gaze models learn to identify useful patterns commonly found across users, but subtle user-specific variations (i.e., eyelid shape or facial structure) can degrade model performance. Test-time personalization (TTP) adapts pre-trained models to these user-specific domain shifts using only a few unlabeled samples. Efficient fine-tuning is critical in performing this domain adaptation: data and computation resources can be limited-especially for on-device customization. While popular parameter-efficient fine-tuning (PEFT) methods address adaptation costs by updating only a small set of weights, they may not be taking full advantage of structures encoded in pre-trained filters. To more effectively leverage existing structures learned during pre-training, we reframe personalization as a process to reweight existing features rather than learning entirely new ones. We present Attentive Low-Rank Filter Adaptation (Alfa) to adapt gaze models by reweighting semantic patterns in pre-trained filters. With Alfa, singular value decomposition (SVD) extracts dominant spatial components that capture eye and facial characteristics across users. Via an attention mechanism, we need only a few unlabeled samples to adjust and reweight pre-trained structures, selectively amplifying those relevant to a target user. Alfa achieves the lowest average gaze errors across four cross-dataset gaze benchmarks, outperforming existing TTP methods and low-rank adaptation (LoRA)-based variants. We also show that Alfa's attentive low-rank methods can be applied to applications beyond vision, such as diffusion-based language models.

cs.CV↗

arXiv2025

EDIT: Early Diffusion Inference Termination for dLLMs Based on Dynamics of Training Gradients

Diffusion-based large language models (dLLMs) refine token generations through iterative denoising, but answers often stabilize before all steps complete. We propose EDIT (Early Diffusion Inference Termination), an inference-time criterion that adaptively stops denoising once sufficient reasoning stability relative to training-time reasoning is detected. EDIT monitors the alignment between token activations and a reasoning map derived from AdamW-aggregated LoRA updates captured during supervised fine-tuning (SFT). During training, optimization dynamics generate rich metadata about parameter importance that in prior methods is typically discarded upon model release. We preserve this information as a compact representation of learned reasoning pathways. During inference, alignment scores are converted to a distribution over the tokens already unmasked at the current denoising step, and convergence is detected when KL divergence between consecutive steps falls below a threshold on the matched unmasked (visible) tokens. Across reasoning benchmarks, EDIT reduces diffusion steps by 11.8% to 68.3% while preserving or improving accuracy in most settings, with approximately 0.02% storage overhead (about 1.5-2 MB for all QKV modules across 32 blocks in an 8 GB model). By utilizing training-gradient dynamics, our work opens a new research direction for reducing dLLM inference time and cost.

cs.AI↗

arXiv2025

MDM: Manhattan Distance Mapping of DNN Weights for Parasitic-Resistance-Resilient Memristive Crossbars

Manhattan Distance Mapping (MDM) is a post-training deep neural network (DNN) weight mapping technique for memristive bit-sliced compute-in-memory (CIM) crossbars that reduces parasitic resistance (PR) nonidealities. PR limits crossbar efficiency by mapping DNN matrices into small crossbar tiles, reducing CIM-based speedup. Each crossbar executes one tile, requiring digital synchronization before the next layer. At this granularity, designers either deploy many small crossbars in parallel or reuse a few sequentially-both increasing analog-to-digital conversions, latency, I/O pressure, and chip area. MDM alleviates PR effects by optimizing active-memristor placement. Exploiting bit-level structured sparsity, it feeds activations from the denser low-order side and reorders rows according to the Manhattan distance, relocating active cells toward regions less affected by PR and thus lowering the nonideality factor (NF). Applied to DNN models on ImageNet-1k, MDM reduces NF by up to 46% and improves accuracy under analog distortion by an average of 3.6% in ResNets. Overall, it provides a lightweight, spatially informed method for scaling CIM DNN accelerators.

cs.AR↗

arXiv2025

PQS (Prune, Quantize, and Sort): Low-Bitwidth Accumulation of Dot Products in Neural Network Computations

We present PQS, which uses three techniques together - Prune, Quantize, and Sort - to achieve low-bitwidth accumulation of dot products in neural network computations. In conventional quantized (e.g., 8-bit) dot products, partial results are accumulated into wide (e.g., 32-bit) accumulators to avoid overflows when accumulating intermediate partial sums. However, such wide accumulators increase memory bandwidth usage and reduce energy efficiency. We show that iterative N:M pruning in floating point followed by quantization to 8 (or fewer) bits, and accumulation of partial products in a sorted order ("small to large") allows for accurate, compressed models with short dot product lengths that do not require wide accumulators. We design, analyze, and implement the PQS algorithm to eliminate accumulation overflows at inference time for several neural networks. Our method offers a 2.5x reduction in accumulator bitwidth while achieving model accuracy on par with floating-point baselines for multiple image classification tasks.

cs.LG↗

arXiv2025

MGS: Markov Greedy Sums for Accurate Low-Bitwidth Floating-Point Accumulation

We offer a novel approach, MGS (Markov Greedy Sums), to improve the accuracy of low-bitwidth floating-point dot products in neural network computations. In conventional 32-bit floating-point summation, adding values with different exponents may lead to loss of precision in the mantissa of the smaller term, which is right-shifted to align with the larger term's exponent. Such shifting (a.k.a. 'swamping') is a significant source of numerical errors in accumulation when implementing low-bitwidth dot products (e.g., 8-bit floating point) as the mantissa has a small number of bits. We avoid most swamping errors by arranging the terms in dot product summation based on their exponents and summing the mantissas without overflowing the low-bitwidth accumulator. We design, analyze, and implement the algorithm to minimize 8-bit floating point error at inference time for several neural networks. In contrast to traditional sequential summation, our method has significantly lowered numerical errors, achieving classification accuracy on par with high-precision floating-point baselines for multiple image classification tasks. Our dMAC hardware units can reduce power consumption by up to 34.1\% relative to conventional MAC units.

cs.AR↗

arXiv2025

ElaLoRA: Elastic & Learnable Low-Rank Adaptation for Efficient Model Fine-Tuning

Low-Rank Adaptation (LoRA) has become a widely adopted technique for fine-tuning large-scale pre-trained models with minimal parameter updates. However, existing methods rely on fixed ranks or focus solely on either rank pruning or expansion, failing to adapt ranks dynamically to match the importance of different layers during training. In this work, we propose ElaLoRA, an adaptive low-rank adaptation framework that dynamically prunes and expands ranks based on gradient-derived importance scores. To the best of our knowledge, ElaLoRA is the first method that enables both rank pruning and expansion during fine-tuning. Experiments across multiple benchmarks demonstrate that ElaLoRA consistently outperforms existing PEFT methods across different parameter budgets. Furthermore, our studies validate that layers receiving higher rank allocations contribute more significantly to model performance, providing theoretical justification for our adaptive strategy. By introducing a principled and adaptive rank allocation mechanism, ElaLoRA offers a scalable and efficient fine-tuning solution, particularly suited for resource-constrained environments.

cs.LG↗

arXiv2024

GazeGen: Gaze-Driven User Interaction for Visual Content Generation

We present GazeGen, a user interaction system that generates visual content (images and videos) for locations indicated by the user's eye gaze. GazeGen allows intuitive manipulation of visual content by targeting regions of interest with gaze. Using advanced techniques in object detection and generative AI, GazeGen performs gaze-controlled image adding/deleting, repositioning, and surface style changes of image objects, and converts static images into videos. Central to GazeGen is the DFT Gaze (Distilled and Fine-Tuned Gaze) agent, an ultra-lightweight model with only 281K parameters, performing accurate real-time gaze predictions tailored to individual users' eyes on small edge devices. GazeGen is the first system to combine visual content generation with real-time gaze estimation, made possible exclusively by DFT Gaze. This real-time gaze estimation enables various visual content generation tasks, all controlled by the user's gaze. The input for DFT Gaze is the user's eye images, while the inputs for visual content generation are the user's view and the predicted gaze point from DFT Gaze. To achieve efficient gaze predictions, we derive the small model from a large model (10x larger) via novel knowledge distillation and personal adaptation techniques. We integrate knowledge distillation with a masked autoencoder, developing a compact yet powerful gaze estimation model. This model is further fine-tuned with Adapters, enabling highly accurate and personalized gaze predictions with minimal user input. DFT Gaze ensures low-latency and precise gaze tracking, supporting a wide range of gaze-driven tasks. We validate the performance of DFT Gaze on AEA and OpenEDS2020 benchmarks, demonstrating low angular gaze error and low latency on the edge device (Raspberry Pi 4). Furthermore, we describe applications of GazeGen, illustrating its versatility and effectiveness in various usage scenarios.

cs.CV↗

arXiv2024

Efficient Reprogramming of Memristive Crossbars for DNNs: Weight Sorting and Bit Stucking

We introduce a novel approach to reduce the number of times required for reprogramming memristors on bit-sliced compute-in-memory crossbars for deep neural networks (DNNs). Our idea addresses the limited non-volatile memory endurance, which restrict the number of times they can be reprogrammed. To reduce reprogramming demands, we employ two techniques: (1) we organize weights into sorted sections to schedule reprogramming of similar crossbars, maximizing memristor state reuse, and (2) we reprogram only a fraction of randomly selected memristors in low-order columns, leveraging their bit-level distribution and recognizing their relatively small impact on model accuracy. We evaluate our approach for state-of-the-art models on the ImageNet-1K dataset. We demonstrate a substantial reduction in crossbar reprogramming by 3.7x for ResNet-50 and 21x for ViT-Base, while maintaining model accuracy within a 1% margin.

cs.AR↗

arXiv2024

Sorted Weight Sectioning for Energy-Efficient Unstructured Sparse DNNs on Compute-in-Memory Crossbars

We introduce $\textit{sorted weight sectioning}$ (SWS): a weight allocation algorithm that places sorted deep neural network (DNN) weight sections on bit-sliced compute-in-memory (CIM) crossbars to reduce analog-to-digital converter (ADC) energy consumption. Data conversions are the most energy-intensive process in crossbar operation. SWS effectively reduces this cost leveraging (1) small weights and (2) zero weights (weight sparsity). DNN weights follow bell-shaped distributions, with most weights near zero. Using SWS, we only need low-order crossbar columns for sections with low-magnitude weights. This reduces the quantity and resolution of ADCs used, exponentially decreasing ADC energy costs without significantly degrading DNN accuracy. Unstructured sparsification further sharpens the weight distribution with small accuracy loss. However, it presents challenges in hardware tracking of zeros: we cannot switch zero rows to other layer weights in unsorted crossbars without index matching. SWS efficiently addresses unstructured sparse models using offline remapping of zeros into earlier sections, which reveals full sparsity potential and maximizes energy efficiency. Our method reduces ADC energy use by 89.5% on unstructured sparse BERT models. Overall, this paper introduces a novel algorithm to promote energy-efficient CIM crossbars for unstructured sparse DNN workloads.

cs.AR↗

arXiv2024

DeepMachining: Online Prediction of Machining Errors of Lathe Machines

We describe DeepMachining, a deep learning-based AI system for online prediction of machining errors of lathe machine operations. We have built and evaluated DeepMachining based on manufacturing data from factories. Specifically, we first pretrain a deep learning model for a given lathe machine's operations to learn the salient features of machining states. Then, we fine-tune the pretrained model to adapt to specific machining tasks. We demonstrate that DeepMachining achieves high prediction accuracy for multiple tasks that involve different workpieces and cutting tools. To the best of our knowledge, this work is one of the first factory experiments using pre-trained deep-learning models to predict machining errors of lathe machines.

cs.LG↗

arXiv2024

Gen4Gen: Generative Data Pipeline for Generative Multi-Concept Composition

Recent text-to-image diffusion models are able to learn and synthesize images containing novel, personalized concepts (e.g., their own pets or specific items) with just a few examples for training. This paper tackles two interconnected issues within this realm of personalizing text-to-image diffusion models. First, current personalization techniques fail to reliably extend to multiple concepts -- we hypothesize this to be due to the mismatch between complex scenes and simple text descriptions in the pre-training dataset (e.g., LAION). Second, given an image containing multiple personalized concepts, there lacks a holistic metric that evaluates performance on not just the degree of resemblance of personalized concepts, but also whether all concepts are present in the image and whether the image accurately reflects the overall text description. To address these issues, we introduce Gen4Gen, a semi-automated dataset creation pipeline utilizing generative models to combine personalized concepts into complex compositions along with text-descriptions. Using this, we create a dataset called MyCanvas, that can be used to benchmark the task of multi-concept personalization. In addition, we design a comprehensive metric comprising two scores (CP-CLIP and TI-CLIP) for better quantifying the performance of multi-concept, personalized text-to-image diffusion methods. We provide a simple baseline built on top of Custom Diffusion with empirical prompting strategies for future researchers to evaluate on MyCanvas. We show that by improving data quality and prompting strategies, we can significantly increase multi-concept personalized image generation quality, without requiring any modifications to model architecture or training algorithms.

cs.CV↗

arXiv2023

SpeedLimit: Neural Architecture Search for Quantized Transformer Models

While research in the field of transformer models has primarily focused on enhancing performance metrics such as accuracy and perplexity, practical applications in industry often necessitate a rigorous consideration of inference latency constraints. Addressing this challenge, we introduce SpeedLimit, a novel Neural Architecture Search (NAS) technique that optimizes accuracy whilst adhering to an upper-bound latency constraint. Our method incorporates 8-bit integer quantization in the search process to outperform the current state-of-the-art technique. Our results underline the feasibility and efficacy of seeking an optimal balance between performance and latency, providing new avenues for deploying state-of-the-art transformer models in latency-sensitive environments.

cs.LG↗

arXiv2023

StitchNet: Composing Neural Networks from Pre-Trained Fragments

We propose StitchNet, a novel neural network creation paradigm that stitches together fragments (one or more consecutive network layers) from multiple pre-trained neural networks. StitchNet allows the creation of high-performing neural networks without the large compute and data requirements needed under traditional model creation processes via backpropagation training. We leverage Centered Kernel Alignment (CKA) as a compatibility measure to efficiently guide the selection of these fragments in composing a network for a given task tailored to specific accuracy needs and computing resource constraints. We then show that these fragments can be stitched together to create neural networks with accuracy comparable to that of traditionally trained networks at a fraction of computing resource and data requirements. Finally, we explore a novel on-the-fly personalized model creation and inference application enabled by this new paradigm. The code is available at https://github.com/steerapi/stitchnet.

cs.LG↗