Meeting growing demands for low latency and cost efficiency in production-grade large language model (LLM) serving systems requires integrating advanced optimization techniques. However, dynamic and unpredictable input-output lengths of LLM, compounded by these optimizations, exacerbate the issues of workload variability, making it difficult to maintain high efficiency on AI accelerators, especially DSAs with tile-based programming models. To address this challenge, we introduce XY-Serve, a versatile, Ascend NPU native, end-to-end production LLM-serving system. The core idea is an abstraction mechanism that smooths out the workload variability by decomposing computations into unified, hardware-friendly, fine-grained meta primitives. Then, kernels can efficiently execute without concerning the irregularity of workload. After this abstraction mechanism, for Attention, we propose a meta-kernel that computes the basic pattern of GEMM-Softmax-GEMM with architectural-aware tile sizes. For Linear, we introduce a virtual padding scheme that adapts to dynamic shape changes while using highly efficient GEMM primitives with assorted fixed tile sizes. XY-Serve sits harmoniously with vLLM. Experimental results show up to 95% end-to-end throughput improvement compared with current publicly available baselines on Ascend NPUs. We also set a new performance record for Linear (average 14.6% faster) and Attention (average 21.5% faster) kernels relative to existing libraries. Lastly, we demonstrate the generality of our technologies on GPU platform.

Key-value (KV) caching is an important technique to accelerate the inference of large language models (LLMs), but incurs significant memory overhead.To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment.In this paper, we introduce DecoQuant, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache.Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors.Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range.Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor.Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant.Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a ∼75% reduction in memory footprint while maintaining comparable generation quality.

Recently, deep learning-based image enhancement algorithms achieved state-of-the-art (SOTA) performance on several publicly available datasets. However, most existing methods fail to meet practical requirements either for visual perception or for computation efficiency, especially for high-resolution images. In this paper, we propose a novel real-time image enhancer via learnable spatial-aware 3dimentional lookup tables(3D LUTs), which well considers global scenario and local spatial information. Specifically, we introduce a light weight two-head weight predictor that has two outputs. One is a 1D weight vector used for image-level scenario adaptation, the other is a 3D weight map aimed for pixel-wise category fusion. We learn the spatial-aware 3D LUTs and fuse them according to the aforementioned weights in an end-to-end manner. The fused LUT is then used to transform the source image into the target tone in an efficient way. Extensive results show that our model outperforms SOTA image enhancement methods on public datasets both subjectively and objectively, and that our model only takes about 4ms to process a 4K resolution image on one NVIDIA V100 GPU.

Visual tracking is one of the most challenging computer vision problems. In order to achieve high performance visual tracking in various negative scenarios, a novel cascaded Siamese network is proposed and developed based on two different deep learning networks: a matching subnetwork and a classification subnetwork. The matching subnetwork is a fully convolutional Siamese network. According to the similarity score between the exemplar image and the candidate image, it aims to search possible object positions and crop scaled candidate patches. The classification subnet-work is designed to further evaluate the cropped candidate patches and determine the optimal tracking results based on the classification score. The matching subnetwork is trained offline and fixed online, while the classification subnetwork performs stochastic gradient descent online to learn more target-specific information. To improve the tracking performance further, an effective classification subnetwork update method based on both similarity and classification scores is utilized for updating the classification subnetwork. Extensive experimental results demonstrate that our proposed approach achieves state-of-the-art performance in recent benchmarks.

Compared with visible object tracking, thermal infrared (TIR) object tracking can track an arbitrary target in total darkness since it cannot be influenced by illumination variations. However, there are many unwanted attributes that constrain the potentials of TIR tracking, such as the absence of visual color patterns and low resolutions. Recently, structured output support vector machine (SOSVM) and discriminative correlation filter (DCF) have been successfully applied to visible object tracking, respectively. Motivated by these, in this paper, we propose a large margin structured convolution operator (LMSCO) to achieve efficient TIR object tracking. To improve the tracking performance, we employ the spatial regularization and implicit interpolation to obtain continuous deep feature maps, including deep appearance features and deep motion features, of the TIR targets. Finally, a collaborative optimization strategy is exploited to significantly update the operators. Our approach not only inherits the advantage of the strong discriminative capability of SOSVM but also achieves accurate and robust tracking with higher-dimensional features and more dense samples. To the best of our knowledge, we are the first to incorporate the advantages of DCF and SOSVM for TIR object tracking. Comprehensive evaluations on two thermal infrared tracking benchmarks, i.e. VOT-TIR2015 and VOT-TIR2016, clearly demonstrate that our LMSCO tracker achieves impressive results and outperforms most state-of-the-art trackers in terms of accuracy and robustness with sufficient frame rate.