Eval bug: Vulkan on Intel ARC – Gemma 4 models (especially with mtp) produce unusable, garbled output

用户在 Docker 中运行 llama.cpp server (vulkan 后端) 加载 Gemma 4 系列模型(Gemma4-12B qat+mtp 和 Gemma4-26B-A4B),使用 Intel ARC B580 显卡。输出包含乱码(符号、不同语种混杂)。同配置下 CPU 推理正常,

![[VPS] 出个台湾 seednet 家宽,适合用于 AI](https://www.chat-gpts.plus/wp-content/uploads/2026/06/ai_cover_5-650-768x403.jpg)
![[程序员] 两天从零开发了 DeepSeek 专属 Agent,我学到了什么?](https://www.chat-gpts.plus/wp-content/uploads/2026/06/ai_cover_4-656-768x403.jpg)
![[OpenAI] Codex Desktop app 现在几乎每日一更啊](https://www.chat-gpts.plus/wp-content/uploads/2026/06/ai_cover_3-659-768x403.jpg)



![[推广] 🚀 Claude360 GPT/Claude/Gemini 直连中转。留 ID 送体验额度。](https://www.chat-gpts.plus/wp-content/uploads/2026/06/ai_cover_4-655-768x403.jpg)
![[Claude Code] codex 与 Claude 相比,仍然存在差距](https://www.chat-gpts.plus/wp-content/uploads/2026/06/ai_cover_3-658-768x403.jpg)