llama.cpp b10997：扩展 RDNA3.5 上 MoE tile 启发式，MoE 性能提升约 11%
原标题：b10997
AI 摘要
llama.cpp 发布 b10997 版本，主要改动是扩展 HIP 后端在 RDNA3.5 架构上的 MoE ncols_opt tile 启发式规则，将条件从 GGML_CUDA_CC_IS_RDNA3_0 改为 GGML_CUDA_CC_IS_RDNA3，代码位于 ggml/src/ggml-cuda/mmq.cu，dense 调度逻辑不变。在 AMD Radeon 8060S（gfx1151）上测试，MUL_MAT 与 MUL_MAT_ID 全部通过，MoE 模型性能提升约 11%（Q4_K 提升 16.2%，Q2_K 提升 6.2%），token 生成速度基本不变或小幅提升。
正文节选
<details open> HIP: broaden MoE ncols_opt tile heuristic on RDNA3.5 architecture (#28935) It's found the MoE ncols_opt tile heuristic needs to be broadened to include the RDNA3.5 architecture. The code change is implemented in ggml/src/ggml-cuda/mmq.cu and just change the GGML_CUDA_CC_IS_RDNA3_0 to GGML_CUDA_CC_IS_RDNA3 in the condition. The dense dispatch logic remains unchanged. The Test machine configuration we used is AMD Radeon 8060S, gfx1151 (RDNA3.5), 20 CU, wave32 + AMD Ryzen AI MAX+