Lbx binary pkg

vllm.whl

cd vllm/
CUDA_VISIBLE_DEVICES=11 python setup.py bdist_wheel

sglang.whl

cd sglang/python
python -m build

flash_attn.whl

cd flash-attention
CUDA_VISIBLE_DEVICES=11 python setup.py bdist_wheel

flashinfer.whl

cd flashinfer/python/ 
FLASHINFER_BUILD_VERSION="999.0.4+cu124torch2.3" TORCH_CUDA_ARCH_LIST="8.0 8.9" python -m build --no-isolation
 
The "-Wno-switch-bool" parameter needs to be removed in setup.py. nvcc 12.4 does not have this option.

onnxruntime_openvino.whl

download openvino

https://storage.openvinotoolkit.org/repositories/openvino/packages/2024.4/linux https://docs.openvino.ai/2024/get-started/install-openvino/install-openvino-archive-linux.html

mkdir /opt/intel
mv openvino_xxx /opt/intel
cd /opt/intel/openvino_xxx
./install_dependencies/install_openvino_dependencies.sh

build onnxruntime

https://onnxruntime.ai/docs/build/eps.html#openvino

source /opt/intel/openvino_xxx/setupvars.sh
./build.sh --config RelWithDebInfo --use_openvino CPU --build_shared_lib --build_wheel --allow_running_as_root --skip_tests --compile_no_warning_as_error --skip_submodule_sync

if build error with libprotobuf or googletest, use onnxruntime/cmake/deps.txt manual build protobuf and googletest

wget https://github.com/protocolbuffers/protobuf/archive/refs/tags/v21.12.zip
unzip v21.12.zip
cd protobuf-21.12
mkdir build && cd build
cmake .. -DONNX_USE_PROTOBUF_SHARED=ON -Dprotobuf_BUILD_TESTS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON
make
make install
rm -rf build/ and rebuild onnxruntime_openvino