Efficient LLM inference, from algorithms to silicon.

• 🔬 Researcher in AI Infrastructure at Beihang University, focusing on efficient LLM inference and compute-in-memory (CiM) architectures
• ⚡ Creator of ASD — a bounded-regret Approximate Speculative Decoding acceptance policy for speculative decoding (arXiv:2608.03447)
• 📷 Photography enthusiast — find my works on Xiaohongshu and Douyin
• 📫 Reach me at yanruo.f@gmail.com (or ynfeng@buaa.edu.cn)

• Current: AI Infrastructure for Efficient Inference — speculative decoding (ASD), serving systems; currently working on efficient inference for Diffusion Language Models (DLMs)
• Past: Compute-in-Memory Co-optimization — noise-aware training, robust neural network & LLM deployment on analog CiM hardware (HPD, STE, ROMER, KV-cache protection)

ASD — Approximate Speculative Decoding acceptance policy (arXiv:2608.03447). Relaxes strict greedy verification by admitting draft tokens whose regret stays within a per-request bounded budget; exact verification is recovered at budget B = 0. Apache-2.0.

(auto-updated daily from my homepage via GitHub Actions)