slurm_train.sh

#!/usr/bin/env bash

set -x

PARTITION=$1
JOB_NAME=$2
GPUS=$3
PY_ARGS=${@:4}

GPUS_PER_NODE=${GPUS_PER_NODE:-8}
CPUS_PER_TASK=${CPUS_PER_TASK:-5}
SRUN_ARGS=${SRUN_ARGS:-""}


srun -p ${PARTITION} \
    --job-name=${JOB_NAME} \
    --gres=gpu:${GPUS_PER_NODE} \
    --ntasks=${GPUS} \
    --ntasks-per-node=${GPUS_PER_NODE} \
    --cpus-per-task=${CPUS_PER_TASK} \
    --kill-on-bad-exit=1 \
    ${SRUN_ARGS} \
    python -u train.py --launcher slurm ${PY_ARGS}