These are the local AI runtimes I have tested on my Strix Halo system (Ryzen
AI Max+ / Radeon 8060S, gfx1151), each at the version it was tested at. Rows
I have not run here are labelled as such: an untested row records the absence
of a test on this machine, nothing about whether the runtime works.
Overall matrix:
| Runtime / framework | Status | Backend | Tested version | Notes |
|---|---|---|---|---|
| llama.cpp | Tested (2026-08-24) | ROCm/HIP | c060ca974 (build 10603) · ROCm 7.14 | benchmark · setup guide |
| llama.cpp | Tested (2026-08-24) | Vulkan (RADV) | c060ca974 (build 10603) · Mesa 26.1.7 | benchmark · setup guide |
| PyTorch | Unverified (checked 2026-08-24) | — | — | no runs made on this machine |
Scope: inference servers (vLLM, SGLang) are not evaluated here; they will be added to the matrix only once actually attempted on this hardware.
llama.cpp
ROCm / HIP backend
Status: Tested (2026-08-24) · llama.cpp c060ca974 (build 10603), -DGGML_HIP=ON, ROCm 7.14 · Setup: Install ROCm 7.14 · Results: ROCm vs Vulkan benchmark
Vulkan backend
Status: Tested (2026-08-24) · llama.cpp c060ca974 (build 10603), -DGGML_VULKAN=ON, Mesa 26.1.7 (RADV, kisak-mesa PPA) · Setup: Vulkan compute guide · Results: benchmark
PyTorch
ROCm wheel / source build for gfx1151
Status: Unverified (checked 2026-08-24) · No PyTorch runs are recorded on
this machine. To test it properly: pin the version, record exactly how it was
installed for gfx1151 (ROCm wheel index, env vars such as
HSA_OVERRIDE_GFX_VERSION if used, or a source build), and run a two-line GPU
tensor check (device available + one matmul on-device) before upgrading this
row to Tested.
What’s actually installed on this box (as captured)
- llama.cpp
c060ca974(build 10603) in two builds from the same commit — HIP and Vulkan - ROCm 7.14, Mesa 26.1.7 (kisak-mesa PPA, RADV)
- The box also carries
libnvidia-gl-595despite being all-AMD: its Vulkan ICD is present but non-functional (no NVIDIA GPU), and the loader skips it harmlessly — a headlessvulkaninforun showedERROR ... libGLX_nvidia.so.0: Could not get 'vkCreateInstance'followed bySkipping this driver. Nothing on this page depends on that package.
How to verify a combination (recipe)
- Install the runtime at a pinned version.
- Run the smallest real workload (not just an import check).
- Record: version string, exact commands, output proof, memory usage.
- Update the matrix row; keep “Tested” tied to that version — older entries stay dated rather than silently upgraded.
Sources and upstream documentation
- llama.cpp repository — backend build options behind the Tested rows.
- AMD ROCm documentation — the runtime the ROCm row runs on.