Chiyue Wei

Excited to share our latest work on Diffusion LLM serving, Optimus: Elastic Decoding for Efficient Diffusion LLM Serving, delivering up to 6.1× throughput gain over autoregressive decoding. Check out the paper and code.

Started my research internship in the AI-System Co-Design team at Meta, where I work on MTIA, Meta’s in-house AI accelerator.

Our work EVA: Accelerating LLM Decoding via an Efficient Vector Quantization Architecture is accepted to ISCA 2026! Check out the paper and code.

Two papers accepted to HPCA 2026! One features our work Focus: A Streaming Concentration Architecture for Efficient Vision-Language Models, now a 🏆 Best Paper Nominee! Check out the paper and code.

Check out our work DPad, a training-free acceleration method for Diffusion LLMs, now available on arXiv!

Wrapped up my internship at NVIDIA, where I worked on the FlashInfer project — developing high-performance, customizable attention kernels with CuTe DSL, optimized for Blackwell GPUs.

Honored to be named a DAC 2025 Young Fellow.

Our works Phi, Transitive Array, and Ecco were presented at ISCA 2025 — check out the slides for Phi.

Started my summer internship at NVIDIA, focusing on LLM inference framework optimization within the Deep Learning Frameworks team.

Three papers accepted by ISCA 2025! Topics include acceleration for Spiking Neural Networks, General Matrix Multiplications, and Large Language Models.

Presented Prosperity at HPCA 2025 in Las Vegas! Check out the slides and video.

Our paper Prosperity: Accelerating Spiking Neural Networks via Product Sparsity is accepted by HPCA 2025.