ynFeng/ynFeng

Yuannuo Feng

PhD student at Beihang University — efficient LLM inference, speculative decoding, diffusion language models, compute-in-memory.

• Homepage: kissmetothemoon.github.io
• GitHub: @Kissmetothemoon
• Google Scholar: profile

Models

• approximate-speculative-decoding — ASD as a custom_generate method for Transformers (5.19.x). Deterministic, bounded-regret acceptance policy for greedy assisted decoding; token-identical to strict verification at budget 0. Usage:

model.generate(..., assistant_model=draft,
    custom_generate="ynFeng/approximate-speculative-decoding", trust_remote_code=True,
    assistant_asd_budget=2.0, assistant_asd_local_ratio=0.25, assistant_asd_max_mismatches=2)

Papers

• Approximate Speculative Decoding — ASD: training-free bounded-regret verifier for speculative decoding
• ASSERT: Adaptive Stochastic Sampling for Robust Diffusion Models on Analog Compute-in-Memory Hardware — ASP-DAC 2027
• Selective KV Cache Protection for Noise-Resilient LLM Inference on Analog Compute-in-Memory Systems
• ROMER: Expert Replacement and Router Calibration for Robust MoE LLMs on Analog Compute-in-Memory Systems
• SetCommit: Coordinated Token Commitments for Few-Step Diffusion Language Model Decoding — code (arXiv preprint forthcoming)