The rdna35-perf branch of strix-llama.cpp improves prompt processing on Strix Halo hardware, achieving 21–54% speed gains over Vulkan and 8–174% (up to 2.7×) over the stock ROCm kernel at 128k context length. These performance improvements come without altering model output quality. The branch contains ten ROCm/HIP‑only commits built atop the halo‑box/strix-llama.cpp fork.