llama.cpp HIP builds get faster prompt processing on AMD GPUs

A proposed llama.cpp change turns on for HIP builds. The benchmark used a gfx1151 AMD chip, RDNA3.5/ with 40 compute units, and compared the change with the latest main code.

On Qwen3.5-27B Q4_K_M, rose by about 4.0% to 7.0%, depending on prompt length. On Qwen3-0.6B BF16, the gain was around 3% for shorter prompts and only 0.6% at 32,768 tokens.

With FA=0, the two builds were almost the same. The practical result is a small speed gain for reading prompts on certain AMD GPU paths, not a broad change for every setup.

Key points

  • The change enables in llama.cpp HIP builds.
  • Qwen3.5-27B Q4_K_M showed about 4.0% to 7.0% faster .
  • Qwen3-0.6B BF16 showed around 3% gains on shorter prompts and 0.6% at 32,768 tokens.
  • FA=0 results were almost unchanged.
  • The impact is mainly for people running locally on AMD GPUs.
Read original