August 6, 2026 Bole: Efficient Tree Speculation for Hybrid-Attention Language Models 内核-运行时协同设计：通过闭式树验证、值域分块GPU核和因子化推测状态，将混合注意力LLM的树投机解码吞吐量提升至自回归解码的4.72× tree-speculative-decoding hybrid-attention linear-attention gdn llm-inference sglang gpu-kernel-optimization closed-form-parallelization