--Deep Neural Networks are increasingly leveraging sparsity to reduce the scaling up of model parameter size. However, reducing wall-clock time through sparsity and pruning remains challenging due to irregular memory access patterns, leading to frequent cache misses. In this paper, we present NPU V ector Runahead (NVR), a prefetching mechanism tailored for NPUs to address cache miss problems in sparse DNN workloads. NVR provides a general micro-architectural solution for sparse DNN workloads without requiring compiler or algorithmic support, operating as a decoupled, speculative, lightweight hardware sub-thread alongside the NPU, with minimal hardware overhead (under 5%). NVR achieves an average 90% reduction in cache misses compared to SOT A prefetching in general-purpose processors, delivering 4x average speedup on sparse workloads versus NPUs without prefetching. Moreover, we investigate the advantages of incorporating a small cache (16KB) into the NPU combined with NVR. Our evaluation shows that expanding this modest cache delivers 5x higher performance benefits than increasing the L2 cache size by the same amount. Fortunately, these workloads are typically over-parameterised [3], where up to 90% of parameters in prevalent models can be pruned while maintaining comparable performance [4]. This redundancy presents an opportunity to leverage sparsity to reduce such intensive resource demands. Theoretically, more fine-grained sparsity patterns yield higher acceleration by skipping more zero-valued elements.

Camera-based multi-view 3D detection has emerged as an attractive solution for autonomous driving due to its low cost and broad applicability. However, despite the strong performance of PETR-based methods in 3D perception benchmarks, their direct INT8 quantization for onboard deployment leads to drastic accuracy drops-up to 58.2% in mAP and 36.9% in NDS on the NuScenes dataset. In this work, we propose Q-PETR, a quantization-aware position embedding transformation that re-engineers key components of the PETR framework to reconcile the discrepancy between the dynamic ranges of positional encodings and image features, and to adapt the cross-attention mechanism for low-bit inference. By redesigning the positional encoding module and introducing an adaptive quantization strategy, Q-PETR maintains floating-point performance with a performance degradation of less than 1% under standard 8-bit per-tensor post-training quantization. Moreover, compared to its FP32 counterpart, Q-PETR achieves a two-fold speedup and reduces memory usage by three times, thereby offering a deployment-friendly solution for resource-constrained onboard devices. Extensive experiments across various PETR-series models validate the strong generalization and practical benefits of our approach.

Diabetic macular ischemia (DMI), marked by the loss of retinal capillaries in the macular area, contributes to vision impairment in patients with diabetes. Although color fundus photographs (CFPs), combined with artificial intelligence (AI), have been extensively applied in detecting various eye diseases, including diabetic retinopathy (DR), their applications in detecting DMI remain unexplored, partly due to skepticism among ophthalmologists regarding its feasibility. In this study, we propose a graph neural network-based multispectral view learning (GNN-MSVL) model designed to detect DMI from CFPs. The model leverages higher spectral resolution to capture subtle changes in fundus reflectance caused by ischemic tissue, enhancing sensitivity to DMI-related features. The proposed approach begins with computational multispectral imaging (CMI) to reconstruct 24-wavelength multispectral fundus images from CFPs. ResNeXt101 is employed as the backbone for multi-view learning to extract features from the reconstructed images. Additionally, a GNN with a customized jumper connection strategy is designed to enhance cross-spectral relationships, facilitating comprehensive and efficient multispectral view learning. The study included a total of 1,078 macula-centered CFPs from 1,078 eyes of 592 patients with diabetes, of which 530 CFPs from 530 eyes of 300 patients were diagnosed with DMI. The model achieved an accuracy of 84.7 percent and an area under the receiver operating characteristic curve (AUROC) of 0.900 (95 percent CI: 0.852-0.937) on eye-level, outperforming both the baseline model trained from CFPs and human experts (p-values less than 0.01). These findings suggest that AI-based CFP analysis holds promise for detecting DMI, contributing to its early and low-cost screening.

Large Language Models (LLMs) fine-tuning technologies have achieved remarkable results. However, traditional LLM fine-tuning approaches face significant challenges: they require large Floating Point (FP) computation, raising privacy concerns when handling sensitive data, and are impractical for resource-constrained edge devices. While Parameter-Efficient Fine-Tuning (PEFT) techniques reduce trainable parameters, their reliance on floating-point arithmetic creates fundamental incompatibilities with edge hardware. In this work, we introduce a novel framework for on-device LLM fine-tuning that eliminates the need for floating-point operations in both inference and training, named GSQ-Tuning. At its core is the Group-Shared Exponents Integer format, which efficiently represents model parameters in integer format using shared exponents among parameter groups. When combined with LoRA-like adapters, this enables fully integer-based fine-tuning that is both memory and compute efficient. We demonstrate that our approach achieves accuracy comparable to FP16-based fine-tuning while significantly reducing memory usage (50%). Moreover, compared to FP8, our method can reduce 5x power consumption and 11x chip area with same performance, making large-scale model adaptation feasible on edge devices.

The advent of foundation models (FMs) is transforming medical domain. In ophthalmology, RETFound, a retina-specific FM pre-trained sequentially on 1.4 million natural images and 1.6 million retinal images, has demonstrated high adaptability across clinical applications. Conversely, DINOv2, a general-purpose vision FM pre-trained on 142 million natural images, has shown promise in non-medical domains. However, its applicability to clinical tasks remains underexplored. To address this, we conducted head-to-head evaluations by fine-tuning RETFound and three DINOv2 models (large, base, small) for ocular disease detection and systemic disease prediction tasks, across eight standardized open-source ocular datasets, as well as the Moorfields AlzEye and the UK Biobank datasets. DINOv2-large model outperformed RETFound in detecting diabetic retinopathy (AUROC=0.850-0.952 vs 0.823-0.944, across three datasets, all P<=0.007) and multi-class eye diseases (AUROC=0.892 vs. 0.846, P<0.001). In glaucoma, DINOv2-base model outperformed RETFound (AUROC=0.958 vs 0.940, P<0.001). Conversely, RETFound achieved superior performance over all DINOv2 models in predicting heart failure, myocardial infarction, and ischaemic stroke (AUROC=0.732-0.796 vs 0.663-0.771, all P<0.001). These trends persisted even with 10% of the fine-tuning data. These findings showcase the distinct scenarios where general-purpose and domain-specific FMs excel, highlighting the importance of aligning FM selection with task-specific requirements to optimise clinical performance.

The substantial computational and memory demands of Large Language Models (LLMs) hinder their deployment. Block Floating Point (BFP) has proven effective in accelerating linear operations, a cornerstone of LLM workloads. However, as sequence lengths grow, nonlinear operations, such as Attention, increasingly become performance bottlenecks due to their quadratic computational complexity. These nonlinear operations are predominantly executed using inefficient floating-point formats, which renders the system challenging to optimize software efficiency and hardware overhead. In this paper, we delve into the limitations and potential of applying BFP to nonlinear operations. Given our findings, we introduce a hardware-software co-design framework (DB-Attn), including: (i) DBFP, an advanced BFP version, overcomes nonlinear operation challenges with a pivot-focus strategy for diverse data and an adaptive grouping strategy for flexible exponent sharing. (ii) DH-LUT, a novel lookup table algorithm dedicated to accelerating nonlinear operations with DBFP format. (iii) An RTL-level DBFP-based engine is implemented to support DB-Attn, applicable to FPGA and ASIC. Results show that DB-Attn provides significant performance improvements with negligible accuracy loss, achieving 74% GPU speedup on Softmax of LLaMA and 10x low overhead performance improvement over SOTA designs.

Mamba is an efficient sequence model that rivals Transformers and demonstrates significant potential as a foundational architecture for various tasks. Quantization is commonly used in neural networks to reduce model size and computational latency. However, applying quantization to Mamba remains underexplored, and existing quantization methods, which have been effective for CNN and Transformer models, appear inadequate for Mamba models (e.g., Quarot suffers a 21% accuracy drop on Vim-T We have pioneered the exploration of this issue and identified several key challenges. First, significant outliers are present in gate projections, output projections, and matrix multiplications. Second, Mamba's unique parallel scan further amplifies these outliers, leading to uneven and heavy-tailed data distributions. Third, even with the application of the Hadamard transform, the variance across channels in weights and activations still remains inconsistent. To these ends, we propose MambaQuant, a post-training quantization (PTQ) framework consisting of: 1) Karhunen-Loève Transformation (KLT) enhanced rotation, rendering the rotation matrix adaptable to diverse channel distributions. Experiments show that MambaQuant can quantize both weights and activations into 8-bit with less than 1% accuracy loss for Mamba-based vision and language tasks. To the best of our knowledge, MambaQuant is the first comprehensive PTQ design for the Mamba family, paving the way for further advancements in its application. Mamba (Gu & Dao, 2023) is a modern sequence model that competes with the Transformer (Vaswani et al., 2017), particularly noted for its ability to handle extremely long sequences. The model's design is inspired by the Structured State Space model (S4) (Gu et al., 2021) and integrates features from recurrent, convolutional, and continuous-time models to effectively capture long-term periodic dependencies.

Multimodal large language models (MLLMs) have garnered widespread attention due to their ability to understand multimodal input. However, their large parameter sizes and substantial computational demands severely hinder their practical deployment and application.While quantization is an effective way to reduce model size and inference latency, its application to MLLMs remains underexplored. In this paper, we propose MQuant, a post-training quantization (PTQ) framework designed to tackle the unique challenges of multimodal large language models (MLLMs). Conventional quantization often struggles with MLLMs because of (a) high inference latency from large visual token counts, (b) distributional disparities between visual and textual tokens, and (c) extreme outliers introduced by Hadamard-based transformations. To address these issues, MQuant introduces: Modality-Specific Static Quantization (MSQ), assigning distinct static scales for visual vs. textual tokens; Attention-Invariant Flexible Switching (AIFS), reordering tokens to preserve casual attention while eliminating expensive token-wise scale computations; Rotation Magnitude Suppression (RMS), mitigating weight outliers arising from online Hadamard rotations. On five mainstream MLLMs (including Qwen-VL, MiniCPM-V, CogVLM2), MQuant under W4A8 achieves near-floating-point accuracy (<1% degradation) while reducing inference latency by up to 30%, significantly outperforming existing PTQ baselines. Our MQuant effectively bridges the gap for efficient and accurate MLLMs inference in resource-constrained devices. Code will be released.

Post-training quantization (PTQ) has emerged as a widely adopted technique for compressing and accelerating Large Language Models (LLMs). The major challenge in LLM quantization is that uneven and heavy-tailed data distributions can expand the quantization range, thereby reducing bit precision for most values. Recent methods attempt to eliminate outliers and balance inter-channel differences by employing linear transformations; however, they remain heuristic and are often overlook optimizing the data distribution across the entire quantization space.In this paper, we introduce Quantization Space Utilization Rate (QSUR), a novel metric that effectively assesses the quantizability of transformed data by measuring the space utilization of the data in the quantization space. We complement QSUR with mathematical derivations that examine the effects and limitations of various transformations, guiding our development of Orthogonal and Scaling Transformation-based Quantization (OSTQuant). OSQuant employs a learnable equivalent transformation, consisting of an orthogonal transformation and a scaling transformation, to optimize the distributions of weights and activations across the entire quantization space. Futhermore, we propose the KL-Top loss function, designed to mitigate noise during optimization while retaining richer semantic information within the limited calibration data imposed by PTQ. OSTQuant outperforms existing work on various LLMs and benchmarks. In the W4-only setting, it retains 99.5\% of the floating-point accuracy. In the more challenging W4A4KV4 configuration, OSTQuant reduces the performance gap by 32\% on the LLaMA-3-8B model compared to state-of-the-art methods. \href{https://github.com/BrotherHappy/OSTQuant}{https://github.com/BrotherHappy/OSTQuant}.

3D occupancy perception is gaining increasing attention due to its capability to offer detailed and precise environment representations. Previous weakly-supervised NeRF methods balance efficiency and accuracy, with mIoU varying by 5-10 points due to sampling count along camera rays. Recently, real-time Gaussian splatting has gained widespread popularity in 3D reconstruction, and the occupancy prediction task can also be viewed as a reconstruction task. Consequently, we propose GSRender, which naturally employs 3D Gaussian Splatting for occupancy prediction, simplifying the sampling process. In addition, the limitations of 2D supervision result in duplicate predictions along the same camera ray. We implemented the Ray Compensation (RC) module, which mitigates this issue by compensating for features from adjacent frames. Finally, we redesigned the loss to eliminate the impact of dynamic objects from adjacent frames. Extensive experiments demonstrate that our approach achieves SOTA (state-of-the-art) results in RayIoU (+6.0), while narrowing the gap with 3D supervision methods. Our code will be released soon.