Mixed precision quantization has been adopted to accelerate large language models (LLMs) serving by leveraging high-throughput low-precision compute units in GPUs while preserving outliers in higher precision to maintain model accuracy. However, existing methods focus on mitigating single-dimensional channel-wise outliers, leading to model accuracy degradation when scaled to 4-bit precision.

Comprehensive defect simulation in the integrated circuit design stage is the basic guarantee to improve testability, which is also the basic requirement of current automotive-grade chip design. For example, there are 6 types of basic defect models for a single MOS transistor in a circuit, and the defect set of analog/hybrid integrated circuits containing thousands of devices is extremely large. It is very time-consuming to inject a large number of defects one by one and solve them in a large-scale matrix. In fact, the complete defect simulation often takes months or even years, and there is no good solution at present. In this paper, a circuit splitting method based on signal flow and graph analysis is proposed, which models the control relationship between the device current and voltage nodes of the circuit with a directed dependency graph through signal flow analysis. Then, more weak connections are obtained by identifying and breaking the feedback loops of larger strongly connected components, and the weak connectivity in strongly connected components are split to realize circuit splitting. The acceleration can then be achieved by parallel defect simulation. At the same time, in order to investigate the observability of the internal defects in the previous strongly connected components in the subsequent strongly connected components, the defects were simulated according to the one-way transfer relationship between each strongly connected components in directed acyclic graph. The equivalent compression method of the defects and the signal in the transfer process is proposed, which reduces the number of simulation times of the defects. Finally, this paper takes the benchmark circuit Bandgap, LDO and SARADC circuits as examples, and uses HSPICE simulation to verify the effectiveness of the graph splitting method. Compared with the latest transistor level circuit splitting scheme, the circuit splitting algorithm proposed in this paper can split more and smaller circuit blocks. Using the circuit blocks we have divided to implement the parallel defect simulation scheme and the defect transfer simulation scheme is more effective than the traditional full-circuit and full-defect simulation one by one.

ABSTRACT Learning path recommendation is crucial for guiding learners through a series of courses in a logical sequence based on their previous learning experiences. This is particularly important for improving learning outcomes in massive open online courses (MOOCs) for diverse learners. Because both the historical learning courses and recommended learning paths can be represented as sequential patterns (SPs); it is reasonable to approach this problem through SP mining (SPM). In addition to support, we incorporate three factors, that is, course learning days, grades and engagement, to model frequent high‐utility SPs (FHUSPs). When recommending a learning path, FHUSPs that align with the target user's learning history and are common among successful learners, while rare among less successful ones, are prioritised. If there are insufficient matching FHUSPs, we address this by recommending additional courses based on the joint competency and complementarity of learners similar to the target learner. Experimental results on a real‐world dataset demonstrate that our method provides highly accurate and relevant recommendations.

Graph Neural Network (GNN) has emerged as an important workload for learning on graphs. With the size of graph data and the complexity of GNN model architectures increasing, developing an efficient GNN system grows more important. As GNN has heavy neural computation workloads on a large graph, it is crucial to partition the entire workload into smaller parts for parallel execution and optimization. However, existing approaches separately partition graph data and GNN operations, resulting in inefficiency and large data movement overhead.

Recently, 3D medical image reconstruction (MIR) and segmentation (MIS) based on deep neural networks have been developed with promising results, and attention mechanism has been further designed for performance enhancement. However, the large size of 3D volume images poses a great computational challenge to traditional attention methods. In this paper, we propose a folded attention (FA) approach to improve the computational efficiency of traditional attention methods on 3D medical images. The main idea is that we apply tensor folding and unfolding operations to construct four small sub-affinity matrices to approximate the original affinity matrix. Through four consecutive sub-attention modules of FA, each element in the feature tensor can aggregate spatial-channel information from all other elements. Compared to traditional attention methods, with the moderate improvement of accuracy, FA can substantially reduce the computational complexity and GPU memory consumption. We demonstrate the superiority of our method on two challenging tasks for 3D MIR and MIS, which are quantitative susceptibility mapping and multiple sclerosis lesion segmentation.