Eval bug: Significant performance difference between llama-prism and main llama.cpp builds (same base version)

该报错表现为同一模型、同一参数下,llama-prism 分支与 main llama.cpp 构建出现 2-3 倍吞吐差异(Eval bug: Significant performance difference between llama-prism and main llama.cpp bui
![[Bug]: Abnormally bad performance on AMD ROCM gfx1030 (W6800, V620, 6900XT 6800XT)](https://www.chat-gpts.plus/wp-content/uploads/2026/10/38107-c0716f30-768x403.jpg)
![[Bug]: persistent_topk silently drops top-k candidates when many values share a coarse histogram bin](https://www.chat-gpts.plus/wp-content/uploads/2026/10/51782-b37011fe-768x403.jpg)




![[Bug]: [Usage]: RTX PRO 5000 (SM120) deadlocks at 3-4W during vLLM 0.32.0 startup with 70B NVFP4 model, even with --enforce-eager](https://www.chat-gpts.plus/wp-content/uploads/2026/10/63536-a00b257b-768x403.jpg)

