Abstract:Deploying large language models (LLMs) on smartphones poses significant engineering challenges due to stringent constraints on memory, latency, and runtime flexibility. In this work, we present a hardware-aware framework for efficient on-device inference of a LLaMA-based multilingual foundation model supporting multiple use cases on Samsung Galaxy S24 and S25 devices with SM8650 and SM8750 Qualcomm chipsets respectively. Our approach integrates application-specific LoRAs as runtime inputs to a single frozen inference graph, enabling dynamic task switching without recompilation or memory overhead. We further introduce a multi-stream decoding mechanism that concurrently generates stylistic variations - such as formal, polite, or jovial responses - within a single forward pass, reducing latency by up to 6x. To accelerate token generation, we apply Dynamic Self-Speculative Decoding (DS2D), a tree-based strategy that predicts future tokens without requiring a draft model, yielding up to 2.3x speedup in decode time. Combined with quantization to INT4 and architecture-level optimizations, our system achieves 4-6x overall improvements in memory and latency while maintaining accuracy across 9 languages and 8 tasks. These results demonstrate practical feasibility of deploying multi-use-case LLMs on edge devices, advancing the commercial viability of Generative AI in mobile platforms.
Abstract:Diffusion Transformers (DiT) have established a new state-of-the-art in high-fidelity image synthesis; however, their massive computational complexity and memory requirements hinder local deployment on resource-constrained edge devices. In this paper, we introduce EdgeDiT, a family of hardware-efficient generative transformers specifically engineered for mobile Neural Processing Units (NPUs), such as the Qualcomm Hexagon and Apple Neural Engine (ANE). By leveraging a hardware-aware optimization framework, we systematically identify and prune structural redundancies within the DiT backbone that are particularly taxing for mobile data-flows. Our approach yields a series of lightweight models that achieve a 20-30% reduction in parameters, a 36-46% decrease in FLOPs, and a 1.65-fold reduction in on-device latency without sacrificing the scaling advantages or the expressive capacity of the original transformer architecture. Extensive benchmarking demonstrates that EdgeDiT offers a superior Pareto-optimal trade-off between Frechet Inception Distance (FID) and inference latency compared to both optimized mobile U-Nets and vanilla DiT variants. By enabling responsive, private, and offline generative AI directly on-device, EdgeDiT provides a scalable blueprint for transitioning large-scale foundation models from high-end GPUs to the palm of the user.
Abstract:Image enhancement is a critical task in computer vision and photography that is often entangled with noise. This renders the traditional Image Signal Processing (ISP) ineffective compared to the advances in deep learning. However, the success of such methods is increasingly associated with the ease of their deployment on edge devices, such as smartphones. This work presents a novel mobile-friendly network for image de-noising obtained with Entropy-Regularized differentiable Neural Architecture Search (NAS) on a hardware-aware search space for a U-Net architecture, which is first-of-its-kind. The designed model has 12% less parameters, with ~2-fold improvement in ondevice latency and 1.5-fold improvement in the memory footprint for a 0.7% drop in PSNR, when deployed and profiled on Samsung Galaxy S24 Ultra. Compared to the SOTA Swin-Transformer for Image Restoration, the proposed network had competitive accuracy with ~18-fold reduction in GMACs. Further, the network was tested successfully for Gaussian de-noising with 3 intensities on 4 benchmarks and real-world de-noising on 1 benchmark demonstrating its generalization ability.
Abstract:Image deblurring is a critical stage in mobile image signal processing pipelines, where the ability to restore fine structures and textures must be balanced with real-time constraints on edge devices. While recent deep networks such as transformers and activation-free architectures achieve state-of-the-art (SOTA) accuracy, their efficiency is typically measured in FLOPs or parameters, which do not correlate with latency on embedded hardware. We propose a hardware-aware adaptation framework that restructures existing models through sensitivity-guided block substitution, surrogate distillation, and training-free multi-objective search driven by device profiling. Applied to the 36-block NAFNet baseline, the optimized variants achieve up to 55% reduction in GMACs compared to the recent transformer-based SOTA while maintaining competitive accuracy. Most importantly, on-device deployment yields a 1.25X latency improvement over the baseline. Experiments on motion deblurring (GoPro), defocus deblurring (DPDD), and auxiliary benchmarks (RealBlur-J/R, HIDE) demonstrate the generality of the approach, while comparisons with prior efficient baselines confirm its accuracy-efficiency trade-off. These results establish feedback-driven adaptation as a principled strategy for bridging the gap between algorithmic design and deployment-ready deblurring models.