redknot

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