#!/bin/bash #SBATCH -A YOUR_ACCOUNT # EDIT: your SLURM account #SBATCH -p gpu # EDIT: your GPU partition name #SBATCH --nodes 1 #SBATCH --gpus-per-node 1 #SBATCH --ntasks-per-node 1 #SBATCH -c 16 #SBATCH -t 00:30:00 #SBATCH --job-name=gcb-meter-eval #SBATCH -o logs/eval-%j.out #SBATCH -e logs/eval-%j.out # SLURM copies the submitted script into a per-job spool dir before running it on this # cluster, so locating the repo via ${BASH_SOURCE[0]} resolves to that spool path, not # the real one ("/var/lib/slurm/..." errors downstream) -- a real failure mode hit # repeatedly in practice. $SLURM_SUBMIT_DIR is set by sbatch to the directory it was # invoked from, immune to that copy, and matches this repo's own submit-from-root # convention (see scripts/slurm/README.md point 6). STOICHEIA_ROOT="${SLURM_SUBMIT_DIR:-$PWD}" export STOICHEIA_ROOT # Usage: sbatch scripts/slurm/meter_eval.sbatch [predict args...] set -euo pipefail MODEL="${1:?usage: sbatch scripts/eval.sbatch [args]}" shift || true EXTRA="${*:---scan-split --norma}" METER_ROOT=$STOICHEIA_ROOT source $METER_ROOT/env.sh echo "model=$MODEL extra='$EXTRA' $(date)" apptainer exec --nv $APPTAINER_BINDS $SIF bash -lc " set -e export PYTHONPATH=$METER_ROOT:$STOICHEIA_ROOT export METER_DATA=$METER_DATA STOICHEIA_DATA=$STOICHEIA_DATA MACRONIZER_SRC=$MACRONIZER_SRC cd $METER_ROOT python -m meter.predict --model $MODEL --micro 32 $EXTRA " echo "EVAL FINISHED $(date)"