ggml-hip: enable -ffast-math for HIP builds by a-huk · Pull Request #23862 · ggml-org/llama.cpp
Llama 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
在 gfx1151 (RDNA3.5/Strix Halo, 40 CUs) 上进行的基准测试显示，为 HIP 构建启用 `-ffast-math` 带来了性能提升。对于 Qwen3.5-27B Q4_K_M 模型（head_dim=256，TILE 路径），在不同提示长度下，性能提升范围为 4.0% 到 7.0%。…
Benchmarked on gfx1151 (RDNA3.5/Strix Halo, 40 CUs) against latest master ( 2f6c815 ). These two models cover both the FA paths.
Qwen3.5-27B Q4_K_M — head_dim=256 (TILE path), FA=1, prompt t/s:
pp upstream + fast-math delta 512 307 321 +4.6% 2048 302 323 +7.0% 4096 300 312 +4.0% 8192 288 307 +6.6% 16384 263 280 +6.5% 32768 233 246 +5.6% Qwen3-0.6B BF16 — head_dim=64 (MMA path), FA=1, prompt t/s:
pp upstream + fast-math delta 512 9325 9642 +3.4% 2048 8609 8892 +3.3% 4096 7485 7661 +2.4% 8192 5770 5846 +1.3% 16384 3581 3636 +1.5% 32768 2037 2049 +0.6% FA=0 results were essentially the same across both builds.