build.sh

#!/usr/bin/env bash
set -ex

echo "Building MLC ${MLC_VERSION} (commit=${MLC_COMMIT})"

# install LLVM the upstream way instead of apt because of:
# https://discourse.llvm.org/t/llvm-15-0-7-missing-libpolly-a/67942 
wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
./llvm.sh ${LLVM_VERSION} all
ln -sf /usr/bin/llvm-config-* /usr/bin/llvm-config

# could NOT find zstd (missing: zstd_LIBRARY zstd_INCLUDE_DIR)
apt-get update
apt-get install -y --no-install-recommends libzstd-dev ccache
rm -rf /var/lib/apt/lists/*
apt-get clean
	
# clone the sources
git clone https://github.com/mlc-ai/mlc-llm /opt/mlc-llm
cd /opt/mlc-llm
git checkout ${MLC_COMMIT}
git submodule update --init --recursive
    
# apply patches to the source
if [ -s /tmp/mlc/patch.diff ]; then 
	git apply /tmp/mlc/patch.diff 
fi

git status
git diff --submodule=diff

# add extras to the source
cp /tmp/mlc/benchmark.py /opt/mlc-llm/

# flashinfer build references 'python'
ln -sf /usr/bin/python3 /usr/bin/python

# disable pytorch: https://github.com/apache/tvm/issues/9362
# -DUSE_LIBTORCH=$(pip3 show torch | grep Location: | cut -d' ' -f2)/torch
mkdir build
cd build

cmake -G Ninja \
	-DCMAKE_CXX_STANDARD=17 \
	-DCMAKE_CUDA_STANDARD=17 \
	-DCMAKE_CUDA_ARCHITECTURES=${CUDAARCHS} \
	-DUSE_CUDA=ON \
	-DUSE_CUDNN=ON \
	-DUSE_CUBLAS=ON \
	-DUSE_CURAND=ON \
	-DUSE_CUTLASS=ON \
	-DUSE_THRUST=ON \
	-DUSE_GRAPH_EXECUTOR_CUDA_GRAPH=ON \
	-DUSE_STACKVM_RUNTIME=ON \
	-DUSE_LLVM="/usr/bin/llvm-config --link-static" \
	-DHIDE_PRIVATE_SYMBOLS=ON \
	-DSUMMARIZE=ON \
	../
	
ninja

#rm -rf CMakeFiles tvm/CMakeFiles
#rm -rf tokenizers/CMakeFiles tokenizers/release

# build TVM python module
cd /opt/mlc-llm/3rdparty/tvm/python

TVM_LIBRARY_PATH=/opt/mlc-llm/build/tvm \
python3 setup.py --verbose bdist_wheel --dist-dir /opt

pip3 install --no-cache-dir --verbose /opt/tvm*.whl
#pip3 show tvm && python3 -c 'import tvm'
#rm -rf dist build

# build mlc-llm python module
cd /opt/mlc-llm

if [ -f setup.py ]; then
	python3 setup.py --verbose bdist_wheel --dist-dir /opt
fi

cd python
python3 setup.py --verbose bdist_wheel --dist-dir /opt

pip3 install --no-cache-dir --verbose /opt/mlc*.whl

# make sure it loads
#cd /
#pip3 show mlc_llm
#python3 -m mlc_llm.build --help
#python3 -c "from mlc_chat import ChatModule; print(ChatModule)"
    
# make the CUTLASS sources available for model builder
ln -s /opt/mlc-llm/3rdparty/tvm/3rdparty /usr/local/lib/python${PYTHON_VERSION}/dist-packages/tvm/3rdparty

# upload wheels
twine upload --verbose /opt/tvm*.whl || echo "failed to upload wheel to ${TWINE_REPOSITORY_URL}"
twine upload --verbose /opt/mlc_llm*.whl || echo "failed to upload wheel to ${TWINE_REPOSITORY_URL}"
twine upload --verbose /opt/mlc_chat*.whl || echo "failed to upload wheel to ${TWINE_REPOSITORY_URL}"