resource Configuration
resource Configuration
resource.json File
This file configures cluster resources for training, molecular dynamics, and DFT calculations (SCF, relaxation, and AIMD), including compute nodes, CPUs, GPUs, and software such as LAMMPS, VASP, PWmat, and MatPL.
Parameters are grouped into train, explore, DFT, and direct modules. The resource fields have the same meaning in every module.
For initial training-set preparation (init_bulk), configure explore, DFT, and direct. In the Mcloud example below, explore defines the foundation-model MD environment, direct defines the DIRECT sampling environment, and DFT defines the environment for SCF or AIMD calculations.
{
"DFT": {
"command":"mpirun -np 4 PWmat",
"task_run_num":1,
"number_node": 1,
"cpu_per_node": 4,
"gpu_per_node": 4,
"group_size": 1,
"queue_name": "3080ti,3090",
"custom_flags": [
],
"_custom_flags": [
],
"module_list": [
"compiler/2022.0.2",
"mkl/2022.0.2",
"mpi/2021.5.1",
"cuda/11.6",
"pwmat"
]
},
"explore": {
"command": "python sevennet_md.py",
"group_size": 1,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti",
"custom_flags": [],
"source_list": [
"/share/app/anaconda3/envs/SevenNet/env.sh"
],
"module_list": [
],
"env_list": [
]
},
"direct": {
"command": "python direct.py",
"group_size": 1,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti",
"custom_flags": [],
"source_list": [
"/share/app/anaconda3/envs/m3gnet/env.sh"
],
"module_list": [
],
"env_list": [
]
}
}
For active learning (run), configure all four modules. In the Mcloud example below, train defines the training environment, explore defines the LAMMPS MD environment, direct defines the DIRECT sampling environment, and DFT defines the environment for SCF calculations or foundation-model labeling.
{
"train": {
"command": "MatPL",
"group_size": 1,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti,3090",
"custom_flags": [
],
"source_list": [
"/share/app/MATPL/MatPL-2025.3/env.sh"
],
"module_list": [
],
"env_list":[
]
},
"explore": {
"command": "mpirun -np 1 lmp_mpi -in in.lammps",
"group_size": 1,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti,3090",
"custom_flags": [],
"source_list": [],
"module_list": [
"lammps4matpl/2025.3"
],
"env_list":[]
},
"DFT": {
"command": "mpirun -np 4 PWmat",
"number_node": 1,
"cpu_per_node": 4,
"gpu_per_node": 4,
"group_size": 1,
"queue_name": "3080ti,3090",
"custom_flags": [],
"source_list": [],
"module_list": [],
"env_list":[
"compiler/2022.0.2",
"mkl/2022.0.2",
"mpi/2021.5.1",
"cuda/11.6",
"pwmat"
]
}
}
Parameter Details
The parameters fall into three categories:
command, which defines the executable command;
resource fields for each computational task:
number_node、cpu_per_node、gpu_per_node、group_size、queue_name、custom_flags；
and custom_flags, source_list, module_list, and env_list, which load software and environment variables.
The fields are described below.
command
Required. Sets the command for the module.
Examples for different tasks:
For DFT calculations:
PWmat:
"command":"mpirun -np 4 PWmat"
VASP:
"command":"vasp_std"
CP2K:
"command":"mpirun -np $SLURM_NTASKS cp2k.popt"
For LAMMPS:
"command":"mpirun -np 10 lmp_mpi"
The value after -np is the CPU process count and must match cpu_per_node. When gpu_per_node is set, these processes are distributed evenly across the configured GPUs.
For model training with MatPL:
"command":"MatPL"
group_size
Groups multiple submitted calculations. Tasks within a group run sequentially, while groups run in parallel.
For example, with 34 SCF calculations and "group_size": 5, PWact creates 7 groups (six groups of five and one group of four), represented by seven Slurm jobs. The Slurm jobs are submitted in parallel, while the calculations inside each job run sequentially.
- For training or exploration, the default is
1, so each task forms its own group and all tasks are submitted together. - For AIMD or SCF tasks under
DFT, the default is-1, placing all tasks in one group for sequential execution. Set this value according to the workload to avoid submitting enough DFT jobs to consume all cluster resources.
number_node
Sets the number of compute nodes per Slurm job. The default is 1.
In the train module, this value is automatically set to 1.
gpu_per_node
Sets the number of GPUs per node. The default is 0. For PWmat DFT calculations (SCF, relaxation, or AIMD), it must match the GPU count specified by command.
In the train module, this value is automatically set to 1.
cpu_per_node
Sets the number of CPUs per node. The default is 1, and it must be greater than or equal to gpu_per_node.
In the train module, this value is automatically set to 1.
queue_name
Required. Sets the cluster partitions as a comma-separated string, for example "queue_name":"cpu, 3080ti, 3090".
custom_flags
Optional list of additional #SBATCH options for the Slurm script. For example:
"custom_flags": [
"#SBATCH -x gn43,gn66"
]
PWact adds "#SBATCH -x gn43,gn66" to the Slurm script. The #SBATCH prefix may be omitted, leaving only -x gn43,gn66.
For example, with the following settings:
"number_node": 1,
"cpu_per_node": 4,
"gpu_per_node": 4,
"queue_name": "3080ti,3090",
"custom_flags": [
"#SBATCH -x gn43,gn66"
]
the generated Slurm header is:
#SBATCH --nodes=1
#SBATCH --gres=gpu:4
#SBATCH --ntasks-per-node=4
#SBATCH --partition=3080ti,3090
#SBATCH -x gn43,gn66
Because some systems do not support --gpus-per-task, PWact does not generate it automatically. Add it through custom_flags when needed.
source_list
Optional list of environment scripts to source at runtime. For example:
"source_list": [
"~/anaconda3/etc/profile.d/conda.sh"
]
PWact prefixes each string with source and writes it to the Slurm script.
module_list
Optional list of software modules to load in the Slurm script.
For example:
"module_list": [
"cuda/11.8",
"intel/2020"
]
PWact writes module load cuda/11.8 and module load intel/2020 to the Slurm script.
env_list
Optional list of arbitrary environment commands written verbatim to the Slurm script.
For example:
"env_list": [
"source the/path/MatPL-2025.3/env.sh"
]
The string is copied directly into the Slurm script.
With the queue_name, custom_flags, source_list, module_list, and env_list settings above, the generated Slurm script contains:
#SBATCH --partition=3080ti,3090
#SBATCH -x gn43,gn66
source /opt/rh/devtoolset-8/enable
module load cuda/11.8
module load intel/2020
source the/path/MatPL-2025.3/env.sh
Configuration Example: train Module
The train module requires the MatPL Python environment. To use the MatPL installation on Mcloud, configure:
"train": {
"command": "MatPL",
"group_size": 1,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti,3090",
"custom_flags": [
],
"source_list": [
"/share/app/MATPL/MatPL-2025.3/env.sh"
],
"env_list": [
],
"module_list": [
]
}
Each training task uses one node, one GPU, and one CPU from either the 3080ti or 3090 partition.
For a MatPL installation built from source, use:
"train":{
"command": "MatPL",
"group_size": 1,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti,3090",
"custom_flags": [
],
"source_list": [
"~/anaconda3/etc/profile.d/conda.sh",
"the/path/MatPL-2025.3/env.sh"
],
"env_list": [
"conda activate matpl-2025.3"
],
"module_list": [
"cuda/11.8-share",
"intel/2020"
]
}
Here, "~/anaconda3/etc/profile.d/conda.sh" is the Conda initialization script, matpl-2025.3 is the MatPL Python environment, and the/path/MatPL-2025.3 is the source path.
Configuration Example: explore Module
For the explore module with the LAMMPS installation on Mcloud, load lammps4matpl as follows:
"explore": {
"command": "mpirun -np 1 lmp_mpi",
"group_size": 2,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti,3090",
"custom_flags": [
],
"source_list": [
],
"module_list": [
"lammps4matpl/2025.3"
],
"env_list":[
]
}
Each LAMMPS task uses one node, one GPU, and one CPU, with two LAMMPS tasks per group.
For LAMMPS built from source, use:
"explore": {
"command": "mpirun -np 1 lmp_mpi",
"group_size": 2,
"number_node": 1,
"gpu_per_node": 1,
"cpu_per_node": 1,
"queue_name": "3080ti,3090",
"custom_flags": [
],
"source_list": [
"the/path/of/lammps/env.sh"
],
"module_list": [
"cuda/11.8-share",
"intel/2020"
],
"env_list":[
]
}
Here, the/path/of/lammps/env.sh is the environment file in the LAMMPS source tree.
Configuration Example: DFT Module
The following example configures PWmat in the DFT module:
"DFT": {
"command":"PWmat",
"number_node": 1,
"cpu_per_node": 4,
"gpu_per_node": 4,
"group_size": 5,
"queue_name": "3080ti,1080ti,3090",
"custom_flags": [
"#SBATCH -x gn18,gn17"
],
"module_list": [
"compiler/2022.0.2",
"mkl/2022.0.2",
"mpi/2021.5.1",
"cuda/11.6"
],
"env_list":[
]
}
Each PWmat task uses one node, four GPUs, and four CPUs, with five DFT tasks per group.