39ee347a39
The gfx1151 llama.cpp prebuilt bundles a regressed TheRock ROCm runtime whose libhsa-runtime64 segfaults at HIP init on bare-metal Strix Halo (unslothai#6276), so every model / embedder load crashes (code -11) while a Vulkan build works. Extend the existing WSL-only _wsl_system_rocm_lib_dirs (which already shadows a bad bundled HIP with the system one) to also return the system ROCm lib dirs on bare-metal Strix Halo (detected via /proc/cpuinfo) when a system HIP runtime is present. The bundle still supplies libggml-hip / librocblas gfx1151 kernels; the driver-matched system libamdhip64 / libhsa-runtime64 drive the GPU. Applied identically to the serve-time and install-validation mirrors so a prebuilt that validates runs the same. No-op without system ROCm or on any non-Strix arch; UNSLOTH_LLAMA_NO_SYSTEM_ROCM=1 opts out. The complete fix is repinning the bundled runtime to a good TheRock nightly in the llama.cpp prebuilt CI; this is the in-repo mitigation. |
||
|---|---|---|
| .. | ||
| data_recipe | ||
| export | ||
| inference | ||
| rag | ||
| training | ||
| __init__.py | ||
| _torchao_stub.py | ||
| import_guards.py | ||
| tool_healing.py | ||