Skip to content

Latest commit

 

History

History
138 lines (125 loc) · 5.53 KB

install_dp.md

File metadata and controls

138 lines (125 loc) · 5.53 KB

Environment settings

The enviroment of rid-kit software is a bit complex, it uses dflow to manage the workflow in kubernetes enviroment refered to as management environment. The actual rid-kit code runs on local or remote (slurm, cloud) enviroment refered to as computation environment. The settings of the management enviroment has already been described in Tutorial, only settings of computation enviroment is given as follows.

Configuration of computation environment

Note:

If you want to set environments by hand, please follow settings of this section.

You can also set computation environment of rid through docker, just run "docker pull pkufjhdocker/rid-kit:latest" to get the docker image we built. Or build the image through the dockerfile in rid-kit/install directory.

Install tensorflow's C++ interface and other necessary packages

We recommend using conda to manage the python enviroment. Use the command

conda create -n rid-dp python=3.9 libtensorflow_cc=2.6.2=*cuda110* tensorflow=2.6.2=*cuda110* nccl mdtraj numpy scikit-learn cmake dpdata cython -c conda-forge

to get the compiled libtensorflow_cc and other necessary packages. Install the cudatoolkit which manage the nvcc and other commands.

conda install cudatoolkit-dev -c conda-forge

After installation, activate the enviroment and set library path

conda activate rid-dp
export LIBRARY_PATH=${CONDA_PREFIX}/lib:$LIBRARY_PATH
export LD_LIBRARY_PATH=${CONDA_PREFIX}/lib:$LD_LIBRARY_PATH

Install the DeepMD-kit's C++ interface

Now go to the source code directory of DeePMD-kit and make a build place.

cd $deepmd_source_dir/source
mkdir build 
cd build

gcc version need to be newer than 6.1.0. In slurm enviroment, if the software is managed by Enviroment Module, one can load gcc such as module load gcc/7.2.0 or other available versions.

set CC and CXX enviroment variables

export CC=/yourpath/to/gcc-7.2.0
export CXX=/yourpath/to/g++-7.2.0

install DeepMD into conda path

cmake -DTENSORFLOW_ROOT=$CONDA_PREFIX -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX -DUSE_CUDA_TOOLKIT=TRUE ..

You should use CUDA version>= 11.0, otherwise the following error will occur

/home/software/deepmd-kit/source/lib/src/cuda/neighbor_list.cu:4:10: 致命错误：cub/block/block_scan.cuh：没有那个文件或目录
 #include <cub/block/block_scan.cuh>
          ^~~~~~~~~~~~~~~~~~~~~~~~~~
编译中断。
CMake Error at deepmd_op_cuda_generated_neighbor_list.cu.o.release.cmake:222 (message):
  Error generating
  /home/software/deepmd-kit/source/build/lib/src/cuda/CMakeFiles/deepmd_op_cuda.dir//./deepmd_op_cuda_generated_neighbor_list.cu.o

In slurm enviroment, if the software is managed by Enviroment Module, one can load CUDA toolkits such as module load cuda/11.1. If the cmake has been executed successfully, then run the following make commands to build the package:

make -j4
make install

If everything works fine, you should find executable dp_gmx_patch in $CONDA_PREFIX.

Install plumed2.6.1

You need to copy compiled DeePFE.cpp to the plumed directory. This file locates at rid-kit/install/DeePFE.cpp.

tar -xvzf plumed-2.6.1.tgz
cp DeePFE.cpp plumed-2.6.1/src/bias
cd plumed-2.6.1
./configure --prefix=$CONDA_PREFIX \
                   CXXFLAGS="-O3 -I $CONDA_PREFIX/include/" \
                   LDFLAGS=" -L$CONDA_PREFIX/lib -ltensorflow_cc -ltensorflow_framework" \
make -j 6
make install

Set the enviroment variables

export PLUMED_KERNEL=${CONDA_PREFIX}/lib/libplumedKernel.@SOEXT@

Install gromacs 2020.2

Note that in order to compile gromacs with GPU support, you need to have CUDA toolkit installed. In slurm enviroment, if the software is managed by Enviroment Module, one can load CUDA toolkits such as module load cuda/11.1. Also note that in most slurm environment, there is no network connection at computing node, but if you want to compile gromacs without fftw installed, you will have to compile it on login-in node where network connection is on.

Also note that you should modify the cmake/gmxManageNvccConfig.cmake file, comment out thecompute 30 architecture otherwise you would get error like

nvcc fatal   : Unsupported gpu architecture 'compute_30'

tar -xzvf gromacs-2020.2.tar.gz
cd gromacs-2020.2
plumed patch -p  # make sure you use the correct gromacs version
dp_gmx_patch -d $gromacs_root -v $version -p #version=2020.2 is supported
mkdir build
cd build
export CMAKE_PREFIX_PATH="/path/to/fftw-3.3.9" # fftw libraries
cmake .. -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX \
                -DGMX_GPU=CUDA  \
                -DGMX_SIMD=avx_512
make -j 4
make install

Set the enviroment variables

source ${CONDA_PREFIX}/bin/GMXRC

If you want to install Lammps into your enviroment, you can execute the following part, otherwise neglect it

Install Lammps-stable_23Jun2022_update1

cd $deepmd_source_dir/source/build
make lammps

Then you will see USER-DEEPMD directory is generated in build dir.

tar -xzvf stable_23Jun2022_update1.tar.gz
cd lammps-stable_23Jun2022_update1/src
cp -r $deepmd_source_dir/source/build/USER-DEEPMD .
make yes-user-deepmd

make packages you will use

make yes-extra-fix
make yes-extra-dump
make yes-kspace

make plumed

make lib-plumed args="-p $CONDA_PREFIX -m runtime"
make yes-plumed
make serial -j4 #build serial version of lammps
ln -s /home/dongdong/software/lammps-stable_23Jun2022_update1/src/lmp_serial /home/dongdong/software/anaconda3/envs/rid_lmp/bin/lmp_serial