You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi,
I've set up an environment using conda with deepmd-gpu-2.2.10 and dpgen. When training with DeePMD-kit using CUDA_VISIBLE_DEVICES=0,1,2,3 mpirun -l -launcher=fork -hosts=localhost -np 4 dp, all 4 GPUs are fully utilized.
However, when running LAMMPS with the default lmp command, only one GPU is used. Attempting to use CUDA_VISIBLE_DEVICES=0,1,2,3 mpirun -np 4 lmp fails to execute the LAMMPS task.
Questions:
Does the default LAMMPS in the conda-built deepmd-gpu-2.2.10 not support MPI parallelization?
But I noticed that lmp should be compiled with MPI support when I run
lmp -h
Compiler: GNU C++ 7.5.0 with OpenMP 4.5
C++ standard: C++11
MPI v4.0: MPICH Version: 4.1.1
MPICH Release date: Mon Mar 6 14:14:15 CST 2023
MPICH ABI: 15:0:3
Is GPU acceleration not supported in this LAMMPS build? (I noticed the GPU package is not listed in the Installed packages)
How can I enable multi-GPU support for LAMMPS in this environment? Do I have to rebuild and install LAMMPS from the source code with build-in or plugin from DeePMD-kit? My ultimate goal is to fully utilize all GPU resources by running lmp in parallel on GPU nodes.
Also, I've spent considerable time trying to compile LAMMPS from source, but have consistently failed due to various library or version incompatibility issues. This has been extremely frustrating. Any guidance on properly configuring LAMMPS for multi-GPU usage within the conda-built
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
Hi,
I've set up an environment using conda with deepmd-gpu-2.2.10 and dpgen. When training with DeePMD-kit using
CUDA_VISIBLE_DEVICES=0,1,2,3 mpirun -l -launcher=fork -hosts=localhost -np 4 dp
, all 4 GPUs are fully utilized.However, when running LAMMPS with the default
lmp
command, only one GPU is used. Attempting to useCUDA_VISIBLE_DEVICES=0,1,2,3 mpirun -np 4 lmp
fails to execute the LAMMPS task.Questions:
Does the default LAMMPS in the conda-built deepmd-gpu-2.2.10 not support MPI parallelization?
But I noticed that
lmp
should be compiled with MPI support when I runlmp -h
Is GPU acceleration not supported in this LAMMPS build? (I noticed the GPU package is not listed in the Installed packages)
How can I enable multi-GPU support for LAMMPS in this environment? Do I have to rebuild and install LAMMPS from the source code with build-in or plugin from DeePMD-kit? My ultimate goal is to fully utilize all GPU resources by running
lmp
in parallel on GPU nodes.Also, I've spent considerable time trying to compile LAMMPS from source, but have consistently failed due to various library or version incompatibility issues. This has been extremely frustrating. Any guidance on properly configuring LAMMPS for multi-GPU usage within the conda-built
environment would be greatly appreciated.
Beta Was this translation helpful? Give feedback.
All reactions