TL;DR: A performance-tuned llama.cpp fork targets AMD Strix Halo (gfx1151) with flash-attention and MoE-prefill fixes plus a bundled Mesa driver.
Summary: Nathanw1014 released a llama.cpp fork optimized for AMD Strix Halo (gfx1151, Ryzen AI Max). It includes flash-attention and mixture-of-experts prefill fixes with a bundled current Mesa driver, using Vulkan and HIP backends. Distribution options include a portable directory, Docker, and distrobox.
Why it matters: For AI builders running local LLMs on AMD's Ryzen AI Max devices, this fork provides a ready-made llama.cpp path with better MoE prefill performance. Try the portable build or Docker image to benchmark speculative decoding and MoE-heavy models against upstream llama.cpp.
Source: github_topics