_

#!/bin/bash
set -e

# Pass --clean to force full rebuild from scratch
if [[ "$1" == "--clean" ]]; then
    rm -rf build
fi

mkdir -p build
cd build

export ROCM_PATH=/opt/rocm
export HIP_PATH=/opt/rocm

# ccache: needed for HIP files
export CCACHE_CPP2=yes
export CCACHE_NODIRECT=1

# Only run cmake configure if cache missing or --clean was used
if [[ ! -f CMakeCache.txt ]]; then
    cmake .. \
      -G Ninja \
      -DGGML_HIP=ON \
      -DAMDGPU_TARGETS=gfx1100 \
      -DGGML_HIP_ROCWMMA_FATTN=ON \
      -DCMAKE_BUILD_TYPE=Release \
      -DLLAMA_BUILD_SERVER=ON \
      -DLLAMA_CURL=ON \
      -DGGML_CCACHE=ON \
      -DCMAKE_C_COMPILER_LAUNCHER=ccache \
      -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
      -DCMAKE_HIP_COMPILER_LAUNCHER=ccache \
      -DCMAKE_PREFIX_PATH="/opt/rocm;/opt/rocm/hip;/opt/rocm/lib/cmake" \
      -DHIP_ROOT_DIR=/opt/rocm \
      -DROCM_PATH=/opt/rocm \
      -DHIP_PATH=/opt/rocm \
      -DCMAKE_HIP_ARCHITECTURES="gfx1100" \
      -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600"
fi

cmake --build . --config Release -j$(nproc)

# Archive binaries for quick rollback
ARCHIVE_DIR=~/src/llama.cpp/bin/GPU/ROCm-turboquant/$(date +%Y%m%d%H%M%S)
mkdir -p "$ARCHIVE_DIR"
cp bin/llama-server bin/lib*.so* "$ARCHIVE_DIR/"
echo "Archived to $ARCHIVE_DIR"