Available here main.pdf
Check out the excellent Python framework Ray Tune here.
A script for using Ray with Slurm is at ./ray_with_slurm. Take a look at the
slurm_launch.py command line utility.
$ python3 ray_with_slurm/slurm_launch.py --help
usage: slurm_launch.py [-h] [--exp-name EXP_NAME] [--num-nodes NUM_NODES]
[--num-gpus NUM_GPUS] [--num-cpus NUM_CPUS]
[--load-env LOAD_ENV] --script SCRIPT
[--script-args SCRIPT_ARGS] [--dry]
Launch a job on slurm using ray.
optional arguments:
-h, --help show this help message and exit
--exp-name EXP_NAME, --name EXP_NAME
The job name and path to logging file (exp_name.log).
--num-nodes NUM_NODES, -n NUM_NODES
Number of nodes to use.
--num-gpus NUM_GPUS Number of GPUs to use for each node. (Default: 0)
--num-cpus NUM_CPUS Number of CPUs to use for each node. (Default: 2)
--load-env LOAD_ENV, -e LOAD_ENV
The script to load your environment (e.g., 'module load
cuda/10.1')
--script SCRIPT, -s SCRIPT
The script you wish to execute. For example: --script
'test.py'. Note that the script must be a path/file.
--script-args SCRIPT_ARGS, -a SCRIPT_ARGS
The script arguments you wish to add. For example:
--script-args '--lr 1e-3 --logdir runs'. Note the quotes.
--dry Whether to only generate the job file (in `./jobs`),
without running it. Dry run.
Example usage: python3 slurm-launch.py --script job.py --script-args 'rllib train
--run PPO --env CartPole-v0'