[speculative/mtp] Flash-Attention crash in gemma4-assistant: Query head dimension mismatch (neq0 != HSK / D=2048/4096)
![[speculative/mtp] Flash-Attention crash in gemma4-assistant: Query head dimension mismatch (neq0 != HSK / D=2048/4096)](https://www.chat-gpts.plus/wp-content/uploads/2026/11/29419-eee099a4-768x403.jpg)
该报错通常发生在 llama-server 启用 MTP 草稿模型( --spec-type draft-mtp )做投机解码、且开启 Flash-Attention( -fa on )时:draft 图传入 FA 的查询头维度是 2048/4096,而内核期望 256,于是触发维度断言崩溃。优先排

![[Bug]: Scale-out `/render` endpoints and `/inference/v1/generate` ignore the `X-Vllm-Priority` header](https://www.chat-gpts.plus/wp-content/uploads/2026/11/67007-e3d022c9-768x403.jpg)






