snapbatch 0.4.0

`snapbatch` is a replacement of `sbatch` to create a snapshot of current working directory, and submit the command to `sbatch`.

SnapBATCH

Motivation

On slurm, if your task is queuing and you change the codes, the final launched code will be the modified version. Usually this behavior is not what we want.

snapbatch replaces sbatch to solve this problem.

Install

pip install snapbatch

Usage

snapbatch [-J your_job_name] [OPTIONS(1)...] [ : [OPTIONS(N)...]] script(0) [args(0)...]

snapbatch is a replacement of sbatch to create a snapshot of current working directory, and submit the command to sbatch.

This command simply:

1. commits the dirty changes of files monitored by git AND all untracked .py/.sh to a new branch.
2. mirros this branch to the path of environment SNAPBATCH_PATH, default to ~/snapbatches. (with git worktree, friendly to merge/commit/find/diff on these new workplaces than directly copying.)
3. runs sbatch --chdir /copied_path/relative/path {--arg xxx ...} (the following args to snapbatch)

Purge branches

Please first manually move or delete the ~/snapbatches dir. (too dangerous to automate), then run the following command under the git working directory,

snapbatch_purge [n]

It keeps the last n snapbatch branches, default 0.

Author: mingding.thu dot gmail.com

Other tools

snapbatch-dryrun [-J your_job_name] [OPTIONS(1)...] [ : [OPTIONS(N)...]] script(0) [args(0)...]

Only mirror the codes and print the sbatch command.

snapbatch-rsc [-J your_job_name] [OPTIONS(1)...] [ : [OPTIONS(N)...]] script(0) [args(0)...]

submit to the FAIR RSC cluster on dev server.

snapbatch-launch

Motivation

Sometimes, we develop codes on a SLURM cluster and want to run it on another cluster without management systems.

snapbatch-launch first mirrors the codes and launches a python or shell file on multiple machines with SLURM / torchrun environment variables, pretending that they are launched by srun / torchrun.

Usage

First speficify environment variable SNAPBATCH_PATH as a path on a shared filesystem.

snapbatch-launch [-h] [-H HOSTFILE] [-J JOB_NAME][--job-id JOB_ID] [--chdir CHDIR] [--env_style {slurm,torchrun}] [-i INCLUDE] [-e EXCLUDE] [--num_nodes NUM_NODES] [--num_gpus NUM_GPUS] [--master_port MASTER_PORT] [--master_addr MASTER_ADDR] [--launcher LAUNCHER] [--launcher_args LAUNCHER_ARGS] [--force_multi]
user_script ...(user_args)

Logs

snapbatch-launch will create a subfolder snapbatch_backup_logs under the mirrored working directory. It will capture and save outputs from different ranks (rank_i.log).

run tail -f .../snapbatch_backup_logs/rank_0.log to see the realtime output of the rank 0.

Stop

The codes are modified from deepspeed, based on pdsh. You need to manually kill the processes on different nodes due to the disadvantage of pdsh. An example is

pdsh -w ssh:node[0-1] "ps -ef | grep jobname | awk '{print \$2}' | xargs kill -9"