8f68efbf0f
Extend the prebuilt installer to detect AMD ROCm GPUs and select the appropriate llama.cpp binary: - Add has_rocm field (default False) to HostInfo dataclass - Extend detect_host() to probe for ROCm via hipcc, amd-smi, rocm-smi, /opt/rocm, and ROCM_PATH env var (skipped on macOS) - Linux x86_64 ROCm hosts (without NVIDIA) try the upstream ROCm 7.2 prebuilt first, with a log warning that it may fall back to source build on other ROCm versions - Windows x86_64 ROCm hosts try the HIP prebuilt; if not found, a log message is printed before falling through to CPU - Add "linux-rocm" and "windows-hip" install kinds to runtime_patterns_for_choice() with the correct shared library globs (libggml-hip.so* for Linux, *.dll for Windows) The ROCm path is only entered when has_usable_nvidia is False, so NVIDIA always takes precedence on dual-GPU systems. The source build fallback (via setup.sh with -DGGML_HIP=ON) compiles against the exact GPU target via rocminfo, which is more reliable for consumer GPUs (e.g. gfx1151) that may not be in the prebuilt. |
||
|---|---|---|
| .. | ||
| backend | ||
| frontend | ||
| __init__.py | ||
| install_llama_prebuilt.py | ||
| install_python_stack.py | ||
| LICENSE.AGPL-3.0 | ||
| setup.bat | ||
| setup.ps1 | ||
| setup.sh | ||
| Unsloth_Studio_Colab.ipynb | ||