cx-git-agent
95 lines
2.9 KiB
Bash
95 lines
2.9 KiB
Bash
#!/usr/bin/env bash
|
|
# cxos/vendor/llama-cpp/build.sh — build the pinned llama.cpp for CxLLM-Arch.
|
|
# shellcheck disable=SC2012
|
|
#
|
|
# Prerequisites:
|
|
# * fetch.sh has run and verified the source under src/llama.cpp-<ver>/
|
|
# * Host has cmake, ninja-build, gcc (and optional Vulkan/CUDA/HIP/OpenCL
|
|
# SDKs for the corresponding backends).
|
|
#
|
|
# Usage:
|
|
# build.sh # CPU only
|
|
# build.sh --backend vulkan # +Vulkan
|
|
# build.sh --backend cuda # +CUDA
|
|
# build.sh --backend hip # +ROCm/HIP
|
|
# build.sh --backend opencl # +OpenCL
|
|
# build.sh --backend vulkan --backend cuda # multiple
|
|
# build.sh --dry # print plan; no compile
|
|
#
|
|
# Outputs (relative to repo root):
|
|
# dist/cxllm-arch/llama-cpp/include/...
|
|
# dist/cxllm-arch/llama-cpp/lib/libllama.a
|
|
# dist/cxllm-arch/llama-cpp/lib/libggml.a
|
|
set -euo pipefail
|
|
|
|
BACKENDS=()
|
|
DRY=0
|
|
JOBS="${JOBS:-$(nproc 2>/dev/null || echo 2)}"
|
|
|
|
while (( "$#" )); do
|
|
case "$1" in
|
|
--backend) BACKENDS+=("$2"); shift 2 ;;
|
|
--dry) DRY=1; shift ;;
|
|
--jobs) JOBS="$2"; shift 2 ;;
|
|
-h|--help)
|
|
sed -n '2,18p' "$0" | sed 's/^# \{0,1\}//'
|
|
exit 0 ;;
|
|
*) echo "build.sh: unknown arg: $1" >&2; exit 2 ;;
|
|
esac
|
|
done
|
|
|
|
HERE="$(cd "$(dirname "$0")" && pwd)"
|
|
REPO_ROOT="$(cd "$HERE/../../.." && pwd)"
|
|
PINNED="$HERE/PINNED.json"
|
|
|
|
VERSION="$(python3 -c 'import json,sys;print(json.load(open(sys.argv[1]))["version"])' "$PINNED")"
|
|
EXTRACTED_DIR="$(python3 -c 'import json,sys;print(json.load(open(sys.argv[1]))["extracted_dir"])' "$PINNED")"
|
|
SRC="$HERE/src/$EXTRACTED_DIR"
|
|
OUT="$REPO_ROOT/dist/cxllm-arch/llama-cpp"
|
|
|
|
CMAKE_FLAGS=(
|
|
-DCMAKE_BUILD_TYPE=Release
|
|
-DCMAKE_INSTALL_PREFIX="$OUT"
|
|
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
|
|
-DLLAMA_BUILD_TESTS=OFF
|
|
-DLLAMA_BUILD_EXAMPLES=OFF
|
|
-DLLAMA_BUILD_SERVER=OFF
|
|
)
|
|
|
|
for b in "${BACKENDS[@]:-}"; do
|
|
case "$b" in
|
|
cpu) ;;
|
|
vulkan) CMAKE_FLAGS+=(-DGGML_VULKAN=ON) ;;
|
|
cuda) CMAKE_FLAGS+=(-DGGML_CUDA=ON) ;;
|
|
hip|rocm) CMAKE_FLAGS+=(-DGGML_HIP=ON) ;;
|
|
opencl) CMAKE_FLAGS+=(-DGGML_OPENCL=ON) ;;
|
|
"") ;;
|
|
*) echo "build.sh: unknown backend: $b" >&2; exit 2 ;;
|
|
esac
|
|
done
|
|
|
|
echo "==> llama.cpp ${VERSION}"
|
|
echo " src : ${SRC}"
|
|
echo " out : ${OUT}"
|
|
echo " flags : ${CMAKE_FLAGS[*]}"
|
|
echo " jobs : ${JOBS}"
|
|
|
|
if [[ "$DRY" == "1" ]]; then
|
|
echo "==> dry-run; not compiling"
|
|
exit 0
|
|
fi
|
|
|
|
if [[ ! -d "$SRC" ]]; then
|
|
echo "build.sh: source not extracted; run fetch.sh first" >&2
|
|
exit 1
|
|
fi
|
|
|
|
BUILD="$SRC/build"
|
|
mkdir -p "$BUILD" "$OUT"
|
|
|
|
cmake -S "$SRC" -B "$BUILD" -G Ninja "${CMAKE_FLAGS[@]}"
|
|
cmake --build "$BUILD" --parallel "$JOBS"
|
|
cmake --install "$BUILD"
|
|
|
|
echo "==> built: $OUT/lib/$(ls "$OUT/lib" 2>/dev/null | head -1 || echo '<install>')"
|