HiDeS — Diffusion LLM Acceleration

Algorithm–system co-design that accelerates block-diffusion LLM inference through hierarchical delta sparsity.

Advisor: Prof. Celine Lin, Georgia Tech  ·  Jan 2026 – present  ·  Under review at HPCA ‘26

HiDeS is an algorithm–system co-design approach that exploits hierarchical sparsity to accelerate block-diffusion model inference through selective recomputation across three granularities — layer, token, and column.

  • Built custom CUDA kernels for attention and MLP to support selective recomputation.
  • Developed on top of the Fast-dLLM v2 block-diffusion model.
  • Achieved 2.33× and 2.67× speedups on NVIDIA A100 and H100 GPUs respectively.