Vulkan build for AMD Strix Halo (Radeon 8060S / gfx1151), Ubuntu x86-64.
A fork of llama.cpp with adaptive speculative decoding and a tuned Vulkan backend.
See the README for the measured
numbers and the ready-to-run llama-server commands.
Install
tar xzf llama-b10641-bin-ubuntu-vulkan-x64.tar.gz
cd llama-b10641/
./llama --versionSelf-contained: the shared libraries sit next to the executables and resolve through an
$ORIGIN rpath, so there is nothing to install and no LD_LIBRARY_PATH to set. Run the
binaries from the folder you extracted.
Requirements
- Linux, x86-64
- A Vulkan 1.3 driver. Measured on Mesa RADV 26.0.8.
- Check your device is visible:
vulkaninfo --summary
Built portable: GGML_NATIVE=OFF with every CPU variant compiled in and selected at run
time, GGML_BACKEND_DL=ON, GGML_VULKAN=ON. Exact flags in
.github/workflows/release-vulkan.yml. The benchmark figures in the README came from a
GGML_NATIVE=ON build; the work is on the GPU, so the two measure the same.
Highlights
- Prefill on Ornith-1.5-35B-A3B at ubatch 2048: 1.9x mainline (1648 t/s against 870)
- Generation on Qwen3.8-27B with an FP4 sidecar and adaptive drafting: 65.6 t/s, 4.7x bare decode
- Adaptive draft sizing (
--spec-draft-adaptive) works on MTP and DFlash2, and is hardware independent
Known issue
-ub 2048 at a context depth of 65536 or beyond can time out the GPU compute ring on this
hardware. It reproduces on stock upstream llama.cpp, so it is not specific to this fork.
Use -ub 512 if you serve long context.
Verified on this build: $ORIGIN rpath resolves, the RPC, Vulkan and CPU backends load
dynamically, the Radeon 8060S is detected, and the CPU variant resolves at run time to a
generic build rather than the host's native one.