Chuqi Zhang

•Exported the video generation model and optimized the convolution to increase MAC unit utilization. Achieved 16x speedup on temporal block, 2x speedup on camera blocks, and reduced one sync point in tensor parallelism. •Benchmarked video quality under fake quantization, testing MLP weights at INT8/INT7/INT6 precision to access compression trade-offs and mitigate memory-bound bottlenecks in large MLP layers. •Implemented and trained model with camera MLA, reducing tensor size by 16x to enable efficient tensor parallelism across SoCs, resolving inter-SoC communication bottlenecks. •Redesigned a hardware-friendly decoder by replacing expensive LayerNorm with Dynamic Tanh, enabling hardware compilation/execution and further pruned channels and reduced ResNet blocks to achieve a 4x faster decoder.

Implemented Token Merging on Large Language and Vision Assistant(LLaVA) model and LLaMA by modifying transformers library. Evaluated llava-1.5-7b model on MLLM Evaluation benchmark(MME), Science Question Answering(ScienceQA), and Text based Visual Question Answering(TextVQA) with different number of Token Merging. Evaluated Llama-3.2-11b-Vision-Instruct model on TextVQA and Massive Multi-discipline Multimodal Understanding (MMMU) benchmark with different number of Token Merging.

Proposed a semi-supervised 3D medical image segmentation framework by labeling only one slice of each 3D image, addressing the semi-supervised 3D medical image segmentation problem by modifying and applying an unsupervised object tracking method. Implemented an attention-enhanced convolutional VoxelMorph to learn a transformation network, capturing semantic correspondences between 2D slices in 3D images. The network generates pseudo-labels for slices in training sets lacking ground truth labels, and further refines them into Gaussian pseudo-labels. Employed forward search and backward tracking to train the network based on consistency loss, resulting in a model achieving a Dice coefficient of 0.731 while using annotations for less than 1% of the data.

Completed the final project “Breast Cancer Classification”, with the data set being the Breast Cancer Histopathological Image Classification (BreakHis). Pre-processed the data using color transfer and utilized VGG16 as the network. Achieved a lightweight compressed model with 91.52% accuracy and a 14.91% compression rate through weight pruning and weight clustering.