Skip to content

Latest commit

 

History

History
71 lines (55 loc) · 3.07 KB

README.md

File metadata and controls

71 lines (55 loc) · 3.07 KB

JZ-hydra-submitit-launcher

GitHub Workflow Build Status

A Jean Zay (JZ) tailored Hydra submitit launcher based on Hydra and its submitit-launcher plugin. Basically it extends the submitit-launcher plugin with defaults that make sense for JZ.

Install

This package can be installed from pypi (using --user is required on JZ):

pip install --user jz-hydra-submitit-launcher

You can also install it from source:

git clone https://github.com/zaccharieramzi/jz-hydra-submitit-launcher.git
cd jz-hydra-submitit-launcher
pip install --user .

The main command is then installed in your local binaries. You need to add them to your path in order for the command to be found (and avoid the error -bash: hydra-submitit-launch: command not found):

echo "export PATH=$PATH:$HOME/.local/bin" >> $HOME/.bashrc

Don't forget to source $HOME/.bashrc before using the command.

Use

The primary use is with the hydra-submitit-launch command with your script name and the config type:

hydra-submitit-launch my_app.py dev

Available configs

6 different configs are available:

  • dev: with 2 hours, 1 gpu, and qos_gpu-dev.
  • t3: with 20 hours, 1 gpu, and qos_gpu-t3.
  • t4: with 100 hours, 1 gpu, and qos_gpu-t4.
  • 4gpus_dev: with 2 hours, 4 gpus, and qos_gpu-dev.
  • 4gpus_t3: with 20 hours, 4 gpus, and qos_gpu-t3.
  • 4gpus_t4: with 100 hours, 4 gpus, and qos_gpu-t4.

By default, all the configs select 32Gb GPUs, use a single node and use the gpu_p1 partition.

Advanced configs

You can override the SLURM config, the same way you would with any hydra configuration. The parameters you can override are defined in the hydra-submitit-launcher plugin doc.

For example, if you want to use the gpu_p2 partition, you would need to do:

hydra-submitit-launch my_app.py dev hydra.launcher.setup=null hydra.launcher.partition=gpu_p2 hydra.launcher.cpus_per_task=3

In order to change the timeout on the SLURM job to for example 10 hours, you would need to do:

hydra-submitit-launch my_app.py base +hours=10

This will automatically select the right qos for you (and if you want to force a certain qos, you can always add hydra.launcher.qos=<your-qos>).

Logs

Your SLURM logs are stored in the following path: launch-dir/multirun/day-of-launch/time-of-launch/.submitit/job_array_task_id/ This is the default from hydra and submitit and this can be changed using both the submitit plugin parameters (see here) and the hydra job configurations (see here).

References