Skip to content

build.md

Latest commit

 

History

History
47 lines (37 loc) · 1.7 KB

build.md

File metadata and controls

47 lines (37 loc) · 1.7 KB

Build from source code

Requirements

  • cudatoolkit-dev >= 10.1
  • protobuf >= 3.13
  • cmake >= 3.18

To install cudatoolkit-dev, you could run conda install -c conda-forge cudatoolkit-dev or follow the official guide, the runfile installation with --toolkit arg is recommended.

After installation, check the installation of nvcc and static libraries (*.a) in ${CUDA_PATH}/lib64.

To install cmake

$ curl -O -L -C - https://github.com/Kitware/CMake/releases/download/v3.18.2/cmake-3.18.2-Linux-x86_64.sh
$ sh cmake-3.18.2-Linux-x86_64.sh --skip-license
$ rm cmake-3.18.2-Linux-x86_64.sh && ln -s ${CMAKE_PATH}/bin/cmake /usr/bin/cmake

Protobuf need to be built and installed from source.

$ curl -O -L -C - https://github.com/protocolbuffers/protobuf/releases/download/v3.13.0/protobuf-cpp-3.13.0.tar.gz
$ tar xf protobuf-cpp-3.13.0.tar.gz
$ cd protobuf-3.13.0 && ./autogen.sh
$ ./configure "CFLAGS=-fPIC" "CXXFLAGS=-fPIC"
$ make -j && make install && ldconfig && cd .. && rm -rf protobuf-3.13.0

make install and ldconfig may need to run with sudo. If you are encountered with any problem, check this

Build

To build all targets.

$ mkdir build && cd build
$ cmake -DCMAKE_BUILD_TYPE=Release -DFP16_MODE=ON .. && make -j

You can also add -DDEBUG_MODE=ON to output intermediate result for debugging.

To build lightseq wheels.

$ pip wheel $PROJECT_DIR --no-deps -w $PROJECT_DIR/output/

To install python lightseq in development models

$ ENABLE_FP32=1 ENABLE_DEBUG=1 pip3 install -e $PROJECT_DIR