- cudatoolkit-dev >= 10.1 < 11
- 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. LightSeq can be run on cuda11, but it is not supported to compile with cuda11, cuda11 has build-in cub, it will conflict with submodule cub.
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/cmakeProtobuf 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-3.13.0make install and ldconfig may need to run with sudo. If you are encountered with any problem, check this
HDF5 also need to be installed.
$ curl -O -L -C - https://github.com/HDFGroup/hdf5/archive/refs/tags/hdf5-1_12_0.tar.gz
$ tar -xzvf hdf5-1_12_0.tar.gz
$ cd hdf5-hdf5-1_12_0
$ ./configure --prefix=/usr/local/hdf5 "CFLAGS=-fPIC" "CXXFLAGS=-fPIC"
$ make
$ make install
$ cd ..If cmake fails with "Could NOT find HDF5", try update PATH via export PATH="$PATH:/usr/local/hdf5".
To build all targets.
$ mkdir build && cd build
$ cmake -DCMAKE_BUILD_TYPE=Release -DFP16_MODE=ON .. && make -jYou 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
$ PATH=/usr/local/hdf5/:$PATH ENABLE_FP32=1 ENABLE_DEBUG=1 pip3 install -e $PROJECT_DIR