arXiv subjects

Zecheng Liu

Publications and source records attributed to Zecheng Liu.

3 recordsLinked to original sources

arXiv2025

Dynamic Expert Quantization for Scalable Mixture-of-Experts Inference

Mixture-of-Experts (MoE) has become a practical architecture for scaling LLM capacity while keeping per-token compute modest, but deploying MoE models on a single, memory-limited GPU remains difficult because expert weights dominate the HBM footprint. Existing expert offloading and prefetching systems reduce the resident set, yet they often pay expert-loading costs on the critical path when activation becomes dense. Post-training quantization (PTQ) lowers the footprint without transfers, but prevailing pipelines fix expert bit-widths offline and assume routing remains stable, even though MoE expert utilization is heavy-tailed and the hot set can shift across workloads. We present DynaExq, a runtime-aware mixed-precision serving system that treats single-GPU MoE inference under a hard HBM envelope as an online, budget-constrained precision allocation problem. The key insight is to keep the experts that dominate runtime traffic resident at higher precision, while maintaining a low-precision fallback for the remaining experts, so the system can reduce transfer volume and avoid the waiting latency that limits offloading and prefetching under dense activation. DynaExq estimates long-horizon expert hotness from router traces, selects a per-layer high-precision resident set via a budget-feasible top-$n$ rule, and applies promotions and demotions asynchronously through stable expert handles so the forward pass always executes on a fully materialized expert version. Across Qwen3-MoE-30B/80B and six benchmarks, DynaExq improves accuracy over static PTQ on Qwen3-80B (73.09% to 77.57%) under comparable device-memory budgets and achieves up to 2.73x higher throughput than offloading/prefetch baselines at batch size 32.

cs.PF↗

arXiv2023

SFusion: Self-attention based N-to-One Multimodal Fusion Block

People perceive the world with different senses, such as sight, hearing, smell, and touch. Processing and fusing information from multiple modalities enables Artificial Intelligence to understand the world around us more easily. However, when there are missing modalities, the number of available modalities is different in diverse situations, which leads to an N-to-One fusion problem. To solve this problem, we propose a self-attention based fusion block called SFusion. Different from preset formulations or convolution based methods, the proposed block automatically learns to fuse available modalities without synthesizing or zero-padding missing ones. Specifically, the feature representations extracted from upstream processing model are projected as tokens and fed into self-attention module to generate latent multimodal correlations. Then, a modal attention mechanism is introduced to build a shared representation, which can be applied by the downstream decision model. The proposed SFusion can be easily integrated into existing multimodal analysis networks. In this work, we apply SFusion to different backbone networks for human activity recognition and brain tumor segmentation tasks. Extensive experimental results show that the SFusion block achieves better performance than the competing fusion strategies. Our code is available at https://github.com/scut-cszcl/SFusion.

cs.CV↗

arXiv2022

Learning Multi-Modal Brain Tumor Segmentation from Privileged Semi-Paired MRI Images with Curriculum Disentanglement Learning

Due to the difficulties of obtaining multimodal paired images in clinical practice, recent studies propose to train brain tumor segmentation models with unpaired images and capture complementary information through modality translation. However, these models cannot fully exploit the complementary information from different modalities. In this work, we thus present a novel two-step (intra-modality and inter-modality) curriculum disentanglement learning framework to effectively utilize privileged semi-paired images, i.e. limited paired images that are only available in training, for brain tumor segmentation. Specifically, in the first step, we propose to conduct reconstruction and segmentation with augmented intra-modality style-consistent images. In the second step, the model jointly performs reconstruction, unsupervised/supervised translation, and segmentation for both unpaired and paired inter-modality images. A content consistency loss and a supervised translation loss are proposed to leverage complementary information from different modalities in this step. Through these two steps, our method effectively extracts modality-specific style codes describing the attenuation of tissue features and image contrast, and modality-invariant content codes containing anatomical and functional information from the input images. Experiments on three brain tumor segmentation tasks show that our model outperforms competing segmentation models based on unpaired images.

cs.CV↗