#!/usr/bin/env bash
# Build the immutable upstream llama.cpp candidate used by GGUF fallbacks.
set -euo pipefail

readonly REPO="https://github.com/ggml-org/llama.cpp.git"
readonly TAG="b9637"
readonly COMMIT="aedb2a5e9ca3d4064148bbb919e0ddc0c1b70ab3"
SRC_DIR="${SRC_DIR:-$HOME/src/llama.cpp-qwen3-next-$TAG}"
BUILD_JOBS="${BUILD_JOBS:-$(getconf _NPROCESSORS_ONLN 2>/dev/null || echo 8)}"

command -v git >/dev/null 2>&1 || { echo "git is required" >&2; exit 2; }
command -v cmake >/dev/null 2>&1 || { echo "cmake is required" >&2; exit 2; }

if [[ ! -d "$SRC_DIR/.git" ]]; then
  git clone --filter=blob:none "$REPO" "$SRC_DIR"
fi

git -C "$SRC_DIR" fetch --depth 1 origin "refs/tags/$TAG:refs/tags/$TAG"
git -C "$SRC_DIR" checkout --detach "$COMMIT"
actual_commit="$(git -C "$SRC_DIR" rev-parse HEAD)"
[[ "$actual_commit" == "$COMMIT" ]] || {
  echo "runtime pin mismatch: expected $COMMIT, got $actual_commit" >&2
  exit 3
}

cmake -S "$SRC_DIR" -B "$SRC_DIR/build" \
  -DGGML_CUDA=ON \
  -DCMAKE_BUILD_TYPE=Release
cmake --build "$SRC_DIR/build" --config Release --parallel "$BUILD_JOBS"
"$SRC_DIR/build/bin/llama-server" --version