Due to the sparse nature of Mixture-of-Experts (MoE) models, they are particularly suitable for hybrid CPU/GPU inference, especially in low-concurrency scenarios. This hybrid approach leverages both the large, cost-effective memory capacity of CPU/DRAM and the high bandwidth of GPU/VRAM. However, existing hybrid solutions remain bottlenecked by CPU computation limits and CPU-GPU synchronization overheads, severely restricting their ability to efficiently run state-of-the-art large MoE models, such as the 671B DeepSeek-V3/R1.

Radiomics, dosiomics and contouromics have been combined to predict adaptive radiotherapy eligibility in nasopharyngeal carcinoma. However, the commonly-used feature concatenation ignores the complementary or consistency relationship across different omics feature spaces. Also, the number of features increases with the concatenation of omics, leading to the curse of dimensionality and potential overfitting. To address the issues, in this study, multi-omics collaborative learning MOCL is developed. In MOCL, a priori knowledge-driven consistency regularization associated with Shannon entropy is designed to automatically explore the weighting consistency across different omics feature spaces. In addition, a label soften strategy is adopted to enlarge the margins between different classes, rendering more freedom for the models to fit the soft label matrix. To avoid overfitting, we design a regularized term deduced from manifold learning to keep samples in the label space as close as possible if they are in the same manifold in the feature space. Experimental results on 311 nasopharyngeal carcinoma patients collected from the Hong Kong Queen Elizabeth Hospital demonstrate the promising performance of MOCL.

With the development of sensors, more and more multimodal data are accumulated, especially in biomedical and bioinformatics fields. Therefore, multimodal data analysis becomes very important and urgent. In this study, we combine multi-kernel learning and transfer learning, and propose a feature-level multi-modality fusion model with insufficient training samples. To be specific, we firstly extend kernel Ridge regression to its multi-kernel version under the lp-norm constraint to explore complementary patterns contained in multimodal data. Then we use marginal probability distribution adaption to minimize the distribution differences between the source domain and the target domain to solve the problem of insufficient training samples. Based on epilepsy EEG data provided by the University of Bonn, we construct 12 multi-modality & transfer scenarios to evaluate our model. Experimental results show that compared with baselines, our model performs better on most scenarios.