# 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

```text
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:

```text
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.