RedKnot: Efficient Long-Context LLM Serving with Head-Aware KV Reuse and SegPagedAttention

Liu, Yang, Luo, Zhaokai, Jin, Huayi · arXiv · 2026