Performance-tuned llama.cpp for AMD Strix Halo (gfx1151): FA + MoE-prefill fixes with a bundled current Mesa driver. Vulkan and HIP; portable dir, Docker, and distrobox.
// readme
Strix Halo llama.cpp toolbox (FA + MoE-prefill fixes)
A ready-to-run llama.cpp for AMD Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S / gfx1151), tuned for long-context, quantized-KV workloads. It bundles a set of Flash-Attention and MoE-prefill fixes plus a current GPU driver, so quantized KV cache is fast instead of a penalty.
The measurements behind these fixes (matrices, methodology, raw data) live in the companion evidence pack.
Speedups at a glance
v0.6.4 payload vs stock master, same box, same session (amd_iommu=off, pp512 / tg32 t/s, -b 512 -ub 512,
-r 3, captured 2026-08-18 between 05:28 and 06:04 UTC). Baseline is stock upstream 9f0d017 at f16 KV, the
commit this release merged; fixed is the released v0.6.4 payload (baf6360b on strix-halo-vulkan). Both arms
were built with the same pinned glslc and run against the same pinned Mesa, so only llama.cpp differs. Every cell
traces to a raw run under benchmarks/results/glance-20260818/.
| Model (arch) | KV |…