Submitit Launcher plugin

  

The Submitit Launcher plugin provides a SLURM Launcher based on Submitit.

Installation

pip install hydra-submitit-launcher --upgrade

Usage

Once installed, add hydra/launcher=submitit_slurm to your command line. Alternatively, override hydra/launcher in your config:

defaults:
  - override hydra/launcher: submitit_slurm

There are several standard approaches for configuring plugins. Check this page for more information. Note that this plugin expects a valid environment in the target host. Usually this means a shared file system between the launching host and the target host.

The Submitit Plugin implements 2 different launchers: submitit_slurm to run on a SLURM cluster, and submitit_local for basic local tests.

Discover the SLURM Launcher parameters (Expand)

$ python your_app.py hydra/launcher=submitit_slurm --cfg hydra -p hydra.launcher

# @package hydra.launcher
submitit_folder: ${hydra.sweep.dir/.submitit/%j
timeout_min: 60
cpus_per_task: null
gpus_per_node: null
tasks_per_node: 1
mem_gb: null
nodes: 1
name: ${hydra.job.name}
_target_: hydra_plugins.hydra_submitit_launcher.submitit_launcher.SlurmLauncher
partition: null
comment: null
constraint: null
exclude: null
cpus_per_gpu: null
gpus_per_task: null
mem_per_gpu: null
mem_per_cpu: null
signal_delay_s: 120
max_num_timeout: 0
additional_parameters: {}
array_parallelism: 256
setup: null

Discover the Local Launcher parameters (Expand)

$ python example/my_app.py hydra/launcher=submitit_local --cfg hydra -p hydra.launcher

# @package hydra.launcher
_target_: hydra_plugins.hydra_submitit_launcher.submitit_launcher.LocalLauncher
submitit_folder: ${hydra.sweep.dir}/.submitit/%j
timeout_min: 60
cpus_per_task: 1
gpus_per_node: 0
tasks_per_node: 1
mem_gb: 4
nodes: 1
name: ${hydra.job.name}

You can set all these parameters in your configuration file and/or override them in the command-line:

python foo.py --multirun hydra/launcher=submitit_slurm hydra.launcher.timeout_min=3

For more details, including descriptions for each parameter, check out the config file. You can also check the Submitit documentation.

Caution: use of --multirun is required for the launcher to be picked up.

Example

An example application using this launcher is provided in the plugin repository.

Starting the app with python my_app.py task=1,2,3 --multirun (see Multi-run for details) will launch 3 executions (you can override the launcher to run locally for testing by adding hydra/launcher=submitit_local):

$ python my_app.py task=1,2,3 --multirun
[HYDRA] Sweep output dir : multirun/2020-05-28/15-05-22
[HYDRA]        #0 : task=1
[HYDRA]        #1 : task=2
[HYDRA]        #2 : task=3

You will be able to see the output of the app in the output dir:

$ tree
.
├── 0
│   └── my_app.log
├── 1
│   └── my_app.log
├── 2
│   └── my_app.log
└── multirun.yaml

$ cat 0/my_app.log 
[2020-05-28 15:05:23,511][__main__][INFO] - Process ID 15887 executing task 1 ...
[2020-05-28 15:05:24,514][submitit][INFO] - Job completed successfully