Misc. bug: llama-server silently ignores -md (speculative decoding never engages) when the unified KV cache is active

用户在 llama-server 中运行推测解码时,通过 -md 参数加载草稿模型,并在启动日志中观察到 kv_unified = 'true' 。尽管草稿模型已加载到内存,但推测解码从未启用:生成阶段无草稿初始化日志、无接受统计、速度与无草稿基线相同。当前确认的最可能原因与 --spec-type
![Eval bug: [SYCL] Fence expiration time out - Sudden hang at exact same point in prompt processing](https://www.chat-gpts.plus/wp-content/uploads/2026/07/25350-c256f83d-768x403.jpg)
![[Claude] poe 上的 fable5,可以稳定的保持多久的记忆,有没有大佬知道啊](https://www.chat-gpts.plus/wp-content/uploads/2026/07/ai_cover_5-160-768x403.jpg)






