Started working with Prof. Celine Lin on HiDeS — accelerating block-diffusion LLM inference through hierarchical delta sparsity and custom CUDA kernels.