8k_v3/eval/submit.slurm

#!/bin/bash





#SBATCH --job-name=8kv3
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1          # crucial - only 1 task per dist per node!
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=1
#SBATCH --time=8000
#SBATCH --partition=hopper-prod
#SBATCH --mem=64G
#SBATCH --qos=normal

#SBATCH --output=/fsx/craffel/lingua_logs/8k_v3/eval/logs/%j.stdout
#SBATCH --error=/fsx/craffel/lingua_logs/8k_v3/eval/logs/%j.stderr

#SBATCH --begin=now+0minutes
#SBATCH --mail-type=ALL
#SBATCH --mail-user=craffel@huggingface.co
#SBATCH --requeue

#SBATCH --open-mode=append

# Mimic the effect of "conda init", which doesn't work for scripts
eval "$(/fsx/craffel/miniconda3/bin/conda shell.bash hook)"
source activate /fsx/craffel/miniconda3/envs/lingua_250401

cd /fsx/craffel/lingua_logs/8k_v3/eval/code/

export OMP_NUM_THREADS=1
export LAUNCH_WITH="SBATCH"
export DUMP_DIR=/fsx/craffel/lingua_logs/8k_v3/eval
export TMPDIR=/scratch



srun -o $DUMP_DIR/logs/%j_%t.out -e $DUMP_DIR/logs/%j_%t.err -n 1 -N 1 python -u -m apps.main.eval config=$DUMP_DIR/base_config.yaml