NVR: Vector Runahead on NPUs for Sparse Memory Access

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 Vector Runahead (NVR), a prefetching mechanism tailored for NPUs to address cache miss problems in sparse DNN workloads. Rather than optimising memory patterns with high overhead and poor portability, NVR adapts runahead execution to the unique architecture of NPUs. 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 SOTA 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.

GSQ-Tuning: Group-Shared Exponents Integer in Fully Quantized Training for LLMs On-Device Fine-tuning

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.

Is an Ultra Large Natural Image-Based Foundation Model Superior to a Retina-Specific Model for Detecting Ocular and Systemic Diseases?

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.

GSRender: Deduplicated Occupancy Prediction via Weakly Supervised 3D Gaussian Splatting

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.

Panoptic-FlashOcc: An Efficient Baseline to Marry Semantic Occupancy with Panoptic via Instance Center

Panoptic occupancy poses a novel challenge by aiming to integrate instance occupancy and semantic occupancy within a unified framework. However, there is still a lack of efficient solutions for panoptic occupancy. In this paper, we propose Panoptic-FlashOcc, a straightforward yet robust 2D feature framework that enables realtime panoptic occupancy. Building upon the lightweight design of FlashOcc, our approach simultaneously learns semantic occupancy and class-aware instance clustering in a single network, these outputs are jointly incorporated through panoptic occupancy procession for panoptic occupancy. This approach effectively addresses the drawbacks of high memory and computation requirements associated with three-dimensional voxel-level representations. With its straightforward and efficient design that facilitates easy deployment, Panoptic-FlashOcc demonstrates remarkable achievements in panoptic occupancy prediction. On the Occ3D-nuScenes benchmark, it achieves exceptional performance, with 38.5 RayIoU and 29.1 mIoU for semantic occupancy, operating at a rapid speed of 43.9 FPS. Furthermore, it attains a notable score of 16.0 RayPQ for panoptic occupancy, accompanied by a fast inference speed of 30.2 FPS. These results surpass the performance of existing methodologies in terms of both speed and accuracy. The source code and trained models can be found at the following github repository: https://github.com/Yzichen/FlashOCC.

M&M VTO: Multi-Garment Virtual Try-On and Editing

We present M&M VTO, a mix and match virtual try-on method that takes as input multiple garment images, text description for garment layout and an image of a person. An example input includes: an image of a shirt, an image of a pair of pants, "rolled sleeves, shirt tucked in", and an image of a person. The output is a visualization of how those garments (in the desired layout) would look like on the given person. Key contributions of our method are: 1) a single stage diffusion based model, with no super resolution cascading, that allows to mix and match multiple garments at 1024x512 resolution preserving and warping intricate garment details, 2) architecture design (VTO UNet Diffusion Transformer) to disentangle denoising from person specific features, allowing for a highly effective finetuning strategy for identity preservation (6MB model per individual vs 4GB achieved with, e.g., dreambooth finetuning); solving a common identity loss problem in current virtual try-on methods, 3) layout control for multiple garments via text inputs specifically finetuned over PaLI-3 for virtual try-on task. Experimental results indicate that M&M VTO achieves state-of-the-art performance both qualitatively and quantitatively, as well as opens up new opportunities for virtual try-on via language-guided and multi-garment try-on.

PillarHist: A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram

Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder named PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance.

I-LLM: Efficient Integer-Only Inference for Fully-Quantized Low-Bit Large Language Models

Post-training quantization (PTQ) serves as a potent technique to accelerate the inference of large language models (LLMs). Nonetheless, existing works still necessitate a considerable number of floating-point (FP) operations during inference, including additional quantization and de-quantization, as well as non-linear operators such as RMSNorm and Softmax. This limitation hinders the deployment of LLMs on the edge and cloud devices. In this paper, we identify the primary obstacle to integer-only quantization for LLMs lies in the large fluctuation of activations across channels and tokens in both linear and non-linear operations. To address this issue, we propose I-LLM, a novel integer-only fully-quantized PTQ framework tailored for LLMs. Specifically, (1) we develop Fully-Smooth Block-Reconstruction (FSBR) to aggressively smooth inter-channel variations of all activations and weights. (2) to alleviate degradation caused by inter-token variations, we introduce a novel approach called Dynamic Integer-only MatMul (DI-MatMul). This method enables dynamic quantization in full-integer matrix multiplication by dynamically quantizing the input and outputs with integer-only operations. (3) we design DI-ClippedSoftmax, DI-Exp, and DI-Normalization, which utilize bit shift to execute non-linear operators efficiently while maintaining accuracy. The experiment shows that our I-LLM achieves comparable accuracy to the FP baseline and outperforms non-integer quantization methods. For example, I-LLM can operate at W4A4 with negligible loss of accuracy. To our knowledge, we are the first to bridge the gap between integer-only quantization and LLMs. We've published our code on anonymous.4open.science, aiming to contribute to the advancement of this field.

Post-Training Quantization for Re-parameterization via Coarse & Fine Weight Splitting

Although neural networks have made remarkable advancements in various applications, they require substantial computational and memory resources. Network quantization is a powerful technique to compress neural networks, allowing for more efficient and scalable AI deployments. Recently, Re-parameterization has emerged as a promising technique to enhance model performance while simultaneously alleviating the computational burden in various computer vision tasks. However, the accuracy drops significantly when applying quantization on the re-parameterized networks. We identify that the primary challenge arises from the large variation in weight distribution across the original branches. To address this issue, we propose a coarse & fine weight splitting (CFWS) method to reduce quantization error of weight, and develop an improved KL metric to determine optimal quantization scales for activation. To the best of our knowledge, our approach is the first work that enables post-training quantization applicable on re-parameterized networks. For example, the quantized RepVGG-A1 model exhibits a mere 0.3% accuracy loss. The code is in https://github.com/NeonHo/Coarse-Fine-Weight-Split.git

FlashOcc: Fast and Memory-Efficient Occupancy Prediction via Channel-to-Height Plugin

Given the capability of mitigating the long-tail deficiencies and intricate-shaped absence prevalent in 3D object detection, occupancy prediction has become a pivotal component in autonomous driving systems. However, the procession of three-dimensional voxel-level representations inevitably introduces large overhead in both memory and computation, obstructing the deployment of to-date occupancy prediction approaches. In contrast to the trend of making the model larger and more complicated, we argue that a desirable framework should be deployment-friendly to diverse chips while maintaining high precision. To this end, we propose a plug-and-play paradigm, namely FlashOCC, to consolidate rapid and memory-efficient occupancy prediction while maintaining high precision. Particularly, our FlashOCC makes two improvements based on the contemporary voxel-level occupancy prediction approaches. Firstly, the features are kept in the BEV, enabling the employment of efficient 2D convolutional layers for feature extraction. Secondly, a channel-to-height transformation is introduced to lift the output logits from the BEV into the 3D space. We apply the FlashOCC to diverse occupancy prediction baselines on the challenging Occ3D-nuScenes benchmarks and conduct extensive experiments to validate the effectiveness. The results substantiate the superiority of our plug-and-play paradigm over previous state-of-the-art methods in terms of precision, runtime efficiency, and memory costs, demonstrating its potential for deployment. The code will be made available.