Build and runtime details
Source identity
- Repository:
https://github.com/ggml-org/llama.cpp.git - Commit:
465e49b9cea78a68b9c244ffb48d0ee24a82873d - Tag at build:
b10830 - Local delta:
patches/gfx1151-disable-direct-host-buft.patch
The patch reverses cleanly from the built source and applies cleanly to the pinned
commit. It changes only ggml/src/ggml-cuda/ggml-cuda.cu.
Toolchain identity
- CMake 4.4.3
- Ninja 1.13.2
- Git 2.55.0.windows.3
- TheRock distribution
gfx1151-7.13.0 - HIP 7.13.99004-3309c6114a
- AMD Clang 23.0.0
- Windows SDK resource compiler 10.0.26100.0
CMake configuration
CMAKE_BUILD_TYPE=Release
BUILD_SHARED_LIBS=OFF
GGML_BACKEND_DL=OFF
GGML_HIP=ON
GGML_HIP_GRAPHS=ON
GGML_HIP_MMQ_MFMA=ON
GGML_HIP_NO_VMM=ON
GPU_TARGETS=gfx1151
LLAMA_BUILD_SERVER=ON
LLAMA_BUILD_TESTS=OFF
The script also pins hipcc, llvm-ar, llvm-ranlib, the ROCm CMake package
directories, Ninja, and the Windows resource compiler by explicit path.
Direct PE imports
The executable directly imports:
KERNEL32.dll
MSVCP140.dll
SHELL32.dll
WS2_32.dll
amdhip64_7.dll
hipblas.dll
ADVAPI32.dll
VCRUNTIME140.dll
VCRUNTIME140_1.dll
api-ms-win-crt-*.dll
ROCm and Microsoft runtime DLLs are not bundled. Windows can load a same-named HIP runtime from System32 before a runtime placed only on PATH. If behavior differs from this report, verify the DLLs actually loaded by the process rather than assuming the requested TheRock directory won the search order.
Reproducibility boundary
The script fixes source, patch, configuration, and tool paths. Bit-for-bit binary reproducibility has not been established across machines, because compiler/linker metadata and installed SDK components may differ. Functional reproducibility is the current claim; compare the reported version, PE imports, tests, and checksum of the distributed artifact.